1 //===-- X86ISelLowering.cpp - X86 DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that X86 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #include "X86ISelLowering.h"
16 #include "Utils/X86ShuffleDecode.h"
17 #include "X86CallingConv.h"
18 #include "X86FrameLowering.h"
19 #include "X86InstrBuilder.h"
20 #include "X86MachineFunctionInfo.h"
21 #include "X86TargetMachine.h"
22 #include "X86TargetObjectFile.h"
23 #include "llvm/ADT/SmallBitVector.h"
24 #include "llvm/ADT/SmallSet.h"
25 #include "llvm/ADT/Statistic.h"
26 #include "llvm/ADT/StringExtras.h"
27 #include "llvm/ADT/StringSwitch.h"
28 #include "llvm/Analysis/EHPersonalities.h"
29 #include "llvm/CodeGen/IntrinsicLowering.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineFunction.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/MachineJumpTableInfo.h"
34 #include "llvm/CodeGen/MachineModuleInfo.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/WinEHFuncInfo.h"
37 #include "llvm/IR/CallSite.h"
38 #include "llvm/IR/CallingConv.h"
39 #include "llvm/IR/Constants.h"
40 #include "llvm/IR/DerivedTypes.h"
41 #include "llvm/IR/Function.h"
42 #include "llvm/IR/GlobalAlias.h"
43 #include "llvm/IR/GlobalVariable.h"
44 #include "llvm/IR/Instructions.h"
45 #include "llvm/IR/Intrinsics.h"
46 #include "llvm/MC/MCAsmInfo.h"
47 #include "llvm/MC/MCContext.h"
48 #include "llvm/MC/MCExpr.h"
49 #include "llvm/MC/MCSymbol.h"
50 #include "llvm/Support/CommandLine.h"
51 #include "llvm/Support/Debug.h"
52 #include "llvm/Support/ErrorHandling.h"
53 #include "llvm/Support/MathExtras.h"
54 #include "llvm/Target/TargetOptions.h"
55 #include "X86IntrinsicsInfo.h"
61 #define DEBUG_TYPE "x86-isel"
63 STATISTIC(NumTailCalls, "Number of tail calls");
65 static cl::opt<bool> ExperimentalVectorWideningLegalization(
66 "x86-experimental-vector-widening-legalization", cl::init(false),
67 cl::desc("Enable an experimental vector type legalization through widening "
68 "rather than promotion."),
71 X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
72 const X86Subtarget &STI)
73 : TargetLowering(TM), Subtarget(&STI) {
74 X86ScalarSSEf64 = Subtarget->hasSSE2();
75 X86ScalarSSEf32 = Subtarget->hasSSE1();
76 MVT PtrVT = MVT::getIntegerVT(8 * TM.getPointerSize());
78 // Set up the TargetLowering object.
80 // X86 is weird. It always uses i8 for shift amounts and setcc results.
81 setBooleanContents(ZeroOrOneBooleanContent);
82 // X86-SSE is even stranger. It uses -1 or 0 for vector masks.
83 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
85 // For 64-bit, since we have so many registers, use the ILP scheduler.
86 // For 32-bit, use the register pressure specific scheduling.
87 // For Atom, always use ILP scheduling.
88 if (Subtarget->isAtom())
89 setSchedulingPreference(Sched::ILP);
90 else if (Subtarget->is64Bit())
91 setSchedulingPreference(Sched::ILP);
93 setSchedulingPreference(Sched::RegPressure);
94 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
95 setStackPointerRegisterToSaveRestore(RegInfo->getStackRegister());
97 // Bypass expensive divides on Atom when compiling with O2.
98 if (TM.getOptLevel() >= CodeGenOpt::Default) {
99 if (Subtarget->hasSlowDivide32())
100 addBypassSlowDiv(32, 8);
101 if (Subtarget->hasSlowDivide64() && Subtarget->is64Bit())
102 addBypassSlowDiv(64, 16);
105 if (Subtarget->isTargetKnownWindowsMSVC()) {
106 // Setup Windows compiler runtime calls.
107 setLibcallName(RTLIB::SDIV_I64, "_alldiv");
108 setLibcallName(RTLIB::UDIV_I64, "_aulldiv");
109 setLibcallName(RTLIB::SREM_I64, "_allrem");
110 setLibcallName(RTLIB::UREM_I64, "_aullrem");
111 setLibcallName(RTLIB::MUL_I64, "_allmul");
112 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::X86_StdCall);
113 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::X86_StdCall);
114 setLibcallCallingConv(RTLIB::SREM_I64, CallingConv::X86_StdCall);
115 setLibcallCallingConv(RTLIB::UREM_I64, CallingConv::X86_StdCall);
116 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::X86_StdCall);
119 if (Subtarget->isTargetDarwin()) {
120 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
121 setUseUnderscoreSetJmp(false);
122 setUseUnderscoreLongJmp(false);
123 } else if (Subtarget->isTargetWindowsGNU()) {
124 // MS runtime is weird: it exports _setjmp, but longjmp!
125 setUseUnderscoreSetJmp(true);
126 setUseUnderscoreLongJmp(false);
128 setUseUnderscoreSetJmp(true);
129 setUseUnderscoreLongJmp(true);
132 // Set up the register classes.
133 addRegisterClass(MVT::i8, &X86::GR8RegClass);
134 addRegisterClass(MVT::i16, &X86::GR16RegClass);
135 addRegisterClass(MVT::i32, &X86::GR32RegClass);
136 if (Subtarget->is64Bit())
137 addRegisterClass(MVT::i64, &X86::GR64RegClass);
139 for (MVT VT : MVT::integer_valuetypes())
140 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
142 // We don't accept any truncstore of integer registers.
143 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
144 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
145 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
146 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
147 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
148 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
150 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
152 // SETOEQ and SETUNE require checking two conditions.
153 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
154 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
155 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
156 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
157 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
158 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
160 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
162 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
163 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
164 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
166 if (Subtarget->is64Bit()) {
167 if (!Subtarget->useSoftFloat() && Subtarget->hasAVX512())
168 // f32/f64 are legal, f80 is custom.
169 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
171 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
172 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
173 } else if (!Subtarget->useSoftFloat()) {
174 // We have an algorithm for SSE2->double, and we turn this into a
175 // 64-bit FILD followed by conditional FADD for other targets.
176 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
177 // We have an algorithm for SSE2, and we turn this into a 64-bit
178 // FILD or VCVTUSI2SS/SD for other targets.
179 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
182 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
184 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
185 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
187 if (!Subtarget->useSoftFloat()) {
188 // SSE has no i16 to fp conversion, only i32
189 if (X86ScalarSSEf32) {
190 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
191 // f32 and f64 cases are Legal, f80 case is not
192 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
194 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
195 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
198 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
199 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
202 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
204 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
205 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
207 if (!Subtarget->useSoftFloat()) {
208 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
209 // are Legal, f80 is custom lowered.
210 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
211 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
213 if (X86ScalarSSEf32) {
214 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
215 // f32 and f64 cases are Legal, f80 case is not
216 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
218 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
219 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
222 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
223 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Expand);
224 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Expand);
227 // Handle FP_TO_UINT by promoting the destination to a larger signed
229 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
230 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
231 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
233 if (Subtarget->is64Bit()) {
234 if (!Subtarget->useSoftFloat() && Subtarget->hasAVX512()) {
235 // FP_TO_UINT-i32/i64 is legal for f32/f64, but custom for f80.
236 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
237 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Custom);
239 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
240 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
242 } else if (!Subtarget->useSoftFloat()) {
243 // Since AVX is a superset of SSE3, only check for SSE here.
244 if (Subtarget->hasSSE1() && !Subtarget->hasSSE3())
245 // Expand FP_TO_UINT into a select.
246 // FIXME: We would like to use a Custom expander here eventually to do
247 // the optimal thing for SSE vs. the default expansion in the legalizer.
248 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
250 // With AVX512 we can use vcvts[ds]2usi for f32/f64->i32, f80 is custom.
251 // With SSE3 we can use fisttpll to convert to a signed i64; without
252 // SSE, we're stuck with a fistpll.
253 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
255 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Custom);
258 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
259 if (!X86ScalarSSEf64) {
260 setOperationAction(ISD::BITCAST , MVT::f32 , Expand);
261 setOperationAction(ISD::BITCAST , MVT::i32 , Expand);
262 if (Subtarget->is64Bit()) {
263 setOperationAction(ISD::BITCAST , MVT::f64 , Expand);
264 // Without SSE, i64->f64 goes through memory.
265 setOperationAction(ISD::BITCAST , MVT::i64 , Expand);
269 // Scalar integer divide and remainder are lowered to use operations that
270 // produce two results, to match the available instructions. This exposes
271 // the two-result form to trivial CSE, which is able to combine x/y and x%y
272 // into a single instruction.
274 // Scalar integer multiply-high is also lowered to use two-result
275 // operations, to match the available instructions. However, plain multiply
276 // (low) operations are left as Legal, as there are single-result
277 // instructions for this in x86. Using the two-result multiply instructions
278 // when both high and low results are needed must be arranged by dagcombine.
279 for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) {
280 setOperationAction(ISD::MULHS, VT, Expand);
281 setOperationAction(ISD::MULHU, VT, Expand);
282 setOperationAction(ISD::SDIV, VT, Expand);
283 setOperationAction(ISD::UDIV, VT, Expand);
284 setOperationAction(ISD::SREM, VT, Expand);
285 setOperationAction(ISD::UREM, VT, Expand);
287 // Add/Sub overflow ops with MVT::Glues are lowered to EFLAGS dependences.
288 setOperationAction(ISD::ADDC, VT, Custom);
289 setOperationAction(ISD::ADDE, VT, Custom);
290 setOperationAction(ISD::SUBC, VT, Custom);
291 setOperationAction(ISD::SUBE, VT, Custom);
294 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
295 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
296 setOperationAction(ISD::BR_CC , MVT::f32, Expand);
297 setOperationAction(ISD::BR_CC , MVT::f64, Expand);
298 setOperationAction(ISD::BR_CC , MVT::f80, Expand);
299 setOperationAction(ISD::BR_CC , MVT::f128, Expand);
300 setOperationAction(ISD::BR_CC , MVT::i8, Expand);
301 setOperationAction(ISD::BR_CC , MVT::i16, Expand);
302 setOperationAction(ISD::BR_CC , MVT::i32, Expand);
303 setOperationAction(ISD::BR_CC , MVT::i64, Expand);
304 setOperationAction(ISD::SELECT_CC , MVT::f32, Expand);
305 setOperationAction(ISD::SELECT_CC , MVT::f64, Expand);
306 setOperationAction(ISD::SELECT_CC , MVT::f80, Expand);
307 setOperationAction(ISD::SELECT_CC , MVT::f128, Expand);
308 setOperationAction(ISD::SELECT_CC , MVT::i8, Expand);
309 setOperationAction(ISD::SELECT_CC , MVT::i16, Expand);
310 setOperationAction(ISD::SELECT_CC , MVT::i32, Expand);
311 setOperationAction(ISD::SELECT_CC , MVT::i64, Expand);
312 if (Subtarget->is64Bit())
313 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
314 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
315 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
316 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
317 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
319 if (Subtarget->is32Bit() && Subtarget->isTargetKnownWindowsMSVC()) {
320 // On 32 bit MSVC, `fmodf(f32)` is not defined - only `fmod(f64)`
321 // is. We should promote the value to 64-bits to solve this.
322 // This is what the CRT headers do - `fmodf` is an inline header
323 // function casting to f64 and calling `fmod`.
324 setOperationAction(ISD::FREM , MVT::f32 , Promote);
326 setOperationAction(ISD::FREM , MVT::f32 , Expand);
329 setOperationAction(ISD::FREM , MVT::f64 , Expand);
330 setOperationAction(ISD::FREM , MVT::f80 , Expand);
331 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
333 // Promote the i8 variants and force them on up to i32 which has a shorter
335 setOperationAction(ISD::CTTZ , MVT::i8 , Promote);
336 AddPromotedToType (ISD::CTTZ , MVT::i8 , MVT::i32);
337 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i8 , Promote);
338 AddPromotedToType (ISD::CTTZ_ZERO_UNDEF , MVT::i8 , MVT::i32);
339 if (Subtarget->hasBMI()) {
340 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i16 , Expand);
341 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32 , Expand);
342 if (Subtarget->is64Bit())
343 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
345 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
346 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
347 if (Subtarget->is64Bit())
348 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
351 if (Subtarget->hasLZCNT()) {
352 // When promoting the i8 variants, force them to i32 for a shorter
354 setOperationAction(ISD::CTLZ , MVT::i8 , Promote);
355 AddPromotedToType (ISD::CTLZ , MVT::i8 , MVT::i32);
356 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Promote);
357 AddPromotedToType (ISD::CTLZ_ZERO_UNDEF, MVT::i8 , MVT::i32);
358 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Expand);
359 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Expand);
360 if (Subtarget->is64Bit())
361 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
363 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
364 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
365 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
366 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Custom);
367 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Custom);
368 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Custom);
369 if (Subtarget->is64Bit()) {
370 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
371 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Custom);
375 // Special handling for half-precision floating point conversions.
376 // If we don't have F16C support, then lower half float conversions
377 // into library calls.
378 if (Subtarget->useSoftFloat() || !Subtarget->hasF16C()) {
379 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
380 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
383 // There's never any support for operations beyond MVT::f32.
384 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
385 setOperationAction(ISD::FP16_TO_FP, MVT::f80, Expand);
386 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
387 setOperationAction(ISD::FP_TO_FP16, MVT::f80, Expand);
389 setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
390 setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
391 setLoadExtAction(ISD::EXTLOAD, MVT::f80, MVT::f16, Expand);
392 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
393 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
394 setTruncStoreAction(MVT::f80, MVT::f16, Expand);
396 if (Subtarget->hasPOPCNT()) {
397 setOperationAction(ISD::CTPOP , MVT::i8 , Promote);
399 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
400 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
401 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
402 if (Subtarget->is64Bit())
403 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
406 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
408 if (!Subtarget->hasMOVBE())
409 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
411 // These should be promoted to a larger select which is supported.
412 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
413 // X86 wants to expand cmov itself.
414 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
415 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
416 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
417 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
418 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
419 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
420 setOperationAction(ISD::SELECT , MVT::f128 , Custom);
421 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
422 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
423 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
424 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
425 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
426 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
427 setOperationAction(ISD::SETCC , MVT::f128 , Custom);
428 setOperationAction(ISD::SETCCE , MVT::i8 , Custom);
429 setOperationAction(ISD::SETCCE , MVT::i16 , Custom);
430 setOperationAction(ISD::SETCCE , MVT::i32 , Custom);
431 if (Subtarget->is64Bit()) {
432 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
433 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
434 setOperationAction(ISD::SETCCE , MVT::i64 , Custom);
436 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
437 // NOTE: EH_SJLJ_SETJMP/_LONGJMP supported here is NOT intended to support
438 // SjLj exception handling but a light-weight setjmp/longjmp replacement to
439 // support continuation, user-level threading, and etc.. As a result, no
440 // other SjLj exception interfaces are implemented and please don't build
441 // your own exception handling based on them.
442 // LLVM/Clang supports zero-cost DWARF exception handling.
443 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
444 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
447 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
448 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
449 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
450 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
451 if (Subtarget->is64Bit())
452 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
453 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
454 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
455 if (Subtarget->is64Bit()) {
456 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
457 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
458 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
459 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
460 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
462 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
463 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
464 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
465 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
466 if (Subtarget->is64Bit()) {
467 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
468 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
469 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
472 if (Subtarget->hasSSE1())
473 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
475 setOperationAction(ISD::ATOMIC_FENCE , MVT::Other, Custom);
477 // Expand certain atomics
478 for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) {
479 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Custom);
480 setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
481 setOperationAction(ISD::ATOMIC_STORE, VT, Custom);
484 if (Subtarget->hasCmpxchg16b()) {
485 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i128, Custom);
488 // FIXME - use subtarget debug flags
489 if (!Subtarget->isTargetDarwin() && !Subtarget->isTargetELF() &&
490 !Subtarget->isTargetCygMing() && !Subtarget->isTargetWin64()) {
491 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
494 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
495 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
497 setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
498 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
500 setOperationAction(ISD::TRAP, MVT::Other, Legal);
501 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Legal);
503 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
504 setOperationAction(ISD::VASTART , MVT::Other, Custom);
505 setOperationAction(ISD::VAEND , MVT::Other, Expand);
506 if (Subtarget->is64Bit()) {
507 setOperationAction(ISD::VAARG , MVT::Other, Custom);
508 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
510 // TargetInfo::CharPtrBuiltinVaList
511 setOperationAction(ISD::VAARG , MVT::Other, Expand);
512 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
515 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
516 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
518 setOperationAction(ISD::DYNAMIC_STACKALLOC, PtrVT, Custom);
520 // GC_TRANSITION_START and GC_TRANSITION_END need custom lowering.
521 setOperationAction(ISD::GC_TRANSITION_START, MVT::Other, Custom);
522 setOperationAction(ISD::GC_TRANSITION_END, MVT::Other, Custom);
524 if (!Subtarget->useSoftFloat() && X86ScalarSSEf64) {
525 // f32 and f64 use SSE.
526 // Set up the FP register classes.
527 addRegisterClass(MVT::f32, &X86::FR32RegClass);
528 addRegisterClass(MVT::f64, &X86::FR64RegClass);
530 // Use ANDPD to simulate FABS.
531 setOperationAction(ISD::FABS , MVT::f64, Custom);
532 setOperationAction(ISD::FABS , MVT::f32, Custom);
534 // Use XORP to simulate FNEG.
535 setOperationAction(ISD::FNEG , MVT::f64, Custom);
536 setOperationAction(ISD::FNEG , MVT::f32, Custom);
538 // Use ANDPD and ORPD to simulate FCOPYSIGN.
539 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
540 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
542 // Lower this to FGETSIGNx86 plus an AND.
543 setOperationAction(ISD::FGETSIGN, MVT::i64, Custom);
544 setOperationAction(ISD::FGETSIGN, MVT::i32, Custom);
546 // We don't support sin/cos/fmod
547 setOperationAction(ISD::FSIN , MVT::f64, Expand);
548 setOperationAction(ISD::FCOS , MVT::f64, Expand);
549 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
550 setOperationAction(ISD::FSIN , MVT::f32, Expand);
551 setOperationAction(ISD::FCOS , MVT::f32, Expand);
552 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
554 // Expand FP immediates into loads from the stack, except for the special
556 addLegalFPImmediate(APFloat(+0.0)); // xorpd
557 addLegalFPImmediate(APFloat(+0.0f)); // xorps
558 } else if (!Subtarget->useSoftFloat() && X86ScalarSSEf32) {
559 // Use SSE for f32, x87 for f64.
560 // Set up the FP register classes.
561 addRegisterClass(MVT::f32, &X86::FR32RegClass);
562 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
564 // Use ANDPS to simulate FABS.
565 setOperationAction(ISD::FABS , MVT::f32, Custom);
567 // Use XORP to simulate FNEG.
568 setOperationAction(ISD::FNEG , MVT::f32, Custom);
570 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
572 // Use ANDPS and ORPS to simulate FCOPYSIGN.
573 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
574 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
576 // We don't support sin/cos/fmod
577 setOperationAction(ISD::FSIN , MVT::f32, Expand);
578 setOperationAction(ISD::FCOS , MVT::f32, Expand);
579 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
581 // Special cases we handle for FP constants.
582 addLegalFPImmediate(APFloat(+0.0f)); // xorps
583 addLegalFPImmediate(APFloat(+0.0)); // FLD0
584 addLegalFPImmediate(APFloat(+1.0)); // FLD1
585 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
586 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
588 if (!TM.Options.UnsafeFPMath) {
589 setOperationAction(ISD::FSIN , MVT::f64, Expand);
590 setOperationAction(ISD::FCOS , MVT::f64, Expand);
591 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
593 } else if (!Subtarget->useSoftFloat()) {
594 // f32 and f64 in x87.
595 // Set up the FP register classes.
596 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
597 addRegisterClass(MVT::f32, &X86::RFP32RegClass);
599 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
600 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
601 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
602 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
604 if (!TM.Options.UnsafeFPMath) {
605 setOperationAction(ISD::FSIN , MVT::f64, Expand);
606 setOperationAction(ISD::FSIN , MVT::f32, Expand);
607 setOperationAction(ISD::FCOS , MVT::f64, Expand);
608 setOperationAction(ISD::FCOS , MVT::f32, Expand);
609 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
610 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
612 addLegalFPImmediate(APFloat(+0.0)); // FLD0
613 addLegalFPImmediate(APFloat(+1.0)); // FLD1
614 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
615 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
616 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
617 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
618 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
619 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
622 // We don't support FMA.
623 setOperationAction(ISD::FMA, MVT::f64, Expand);
624 setOperationAction(ISD::FMA, MVT::f32, Expand);
626 // Long double always uses X87, except f128 in MMX.
627 if (!Subtarget->useSoftFloat()) {
628 if (Subtarget->is64Bit() && Subtarget->hasMMX()) {
629 addRegisterClass(MVT::f128, &X86::FR128RegClass);
630 ValueTypeActions.setTypeAction(MVT::f128, TypeSoftenFloat);
631 setOperationAction(ISD::FABS , MVT::f128, Custom);
632 setOperationAction(ISD::FNEG , MVT::f128, Custom);
633 setOperationAction(ISD::FCOPYSIGN, MVT::f128, Custom);
636 addRegisterClass(MVT::f80, &X86::RFP80RegClass);
637 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
638 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
640 APFloat TmpFlt = APFloat::getZero(APFloat::x87DoubleExtended);
641 addLegalFPImmediate(TmpFlt); // FLD0
643 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
646 APFloat TmpFlt2(+1.0);
647 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
649 addLegalFPImmediate(TmpFlt2); // FLD1
650 TmpFlt2.changeSign();
651 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
654 if (!TM.Options.UnsafeFPMath) {
655 setOperationAction(ISD::FSIN , MVT::f80, Expand);
656 setOperationAction(ISD::FCOS , MVT::f80, Expand);
657 setOperationAction(ISD::FSINCOS, MVT::f80, Expand);
660 setOperationAction(ISD::FFLOOR, MVT::f80, Expand);
661 setOperationAction(ISD::FCEIL, MVT::f80, Expand);
662 setOperationAction(ISD::FTRUNC, MVT::f80, Expand);
663 setOperationAction(ISD::FRINT, MVT::f80, Expand);
664 setOperationAction(ISD::FNEARBYINT, MVT::f80, Expand);
665 setOperationAction(ISD::FMA, MVT::f80, Expand);
668 // Always use a library call for pow.
669 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
670 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
671 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
673 setOperationAction(ISD::FLOG, MVT::f80, Expand);
674 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
675 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
676 setOperationAction(ISD::FEXP, MVT::f80, Expand);
677 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
678 setOperationAction(ISD::FMINNUM, MVT::f80, Expand);
679 setOperationAction(ISD::FMAXNUM, MVT::f80, Expand);
681 // First set operation action for all vector types to either promote
682 // (for widening) or expand (for scalarization). Then we will selectively
683 // turn on ones that can be effectively codegen'd.
684 for (MVT VT : MVT::vector_valuetypes()) {
685 setOperationAction(ISD::ADD , VT, Expand);
686 setOperationAction(ISD::SUB , VT, Expand);
687 setOperationAction(ISD::FADD, VT, Expand);
688 setOperationAction(ISD::FNEG, VT, Expand);
689 setOperationAction(ISD::FSUB, VT, Expand);
690 setOperationAction(ISD::MUL , VT, Expand);
691 setOperationAction(ISD::FMUL, VT, Expand);
692 setOperationAction(ISD::SDIV, VT, Expand);
693 setOperationAction(ISD::UDIV, VT, Expand);
694 setOperationAction(ISD::FDIV, VT, Expand);
695 setOperationAction(ISD::SREM, VT, Expand);
696 setOperationAction(ISD::UREM, VT, Expand);
697 setOperationAction(ISD::LOAD, VT, Expand);
698 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
699 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT,Expand);
700 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
701 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT,Expand);
702 setOperationAction(ISD::INSERT_SUBVECTOR, VT,Expand);
703 setOperationAction(ISD::FABS, VT, Expand);
704 setOperationAction(ISD::FSIN, VT, Expand);
705 setOperationAction(ISD::FSINCOS, VT, Expand);
706 setOperationAction(ISD::FCOS, VT, Expand);
707 setOperationAction(ISD::FSINCOS, VT, Expand);
708 setOperationAction(ISD::FREM, VT, Expand);
709 setOperationAction(ISD::FMA, VT, Expand);
710 setOperationAction(ISD::FPOWI, VT, Expand);
711 setOperationAction(ISD::FSQRT, VT, Expand);
712 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
713 setOperationAction(ISD::FFLOOR, VT, Expand);
714 setOperationAction(ISD::FCEIL, VT, Expand);
715 setOperationAction(ISD::FTRUNC, VT, Expand);
716 setOperationAction(ISD::FRINT, VT, Expand);
717 setOperationAction(ISD::FNEARBYINT, VT, Expand);
718 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
719 setOperationAction(ISD::MULHS, VT, Expand);
720 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
721 setOperationAction(ISD::MULHU, VT, Expand);
722 setOperationAction(ISD::SDIVREM, VT, Expand);
723 setOperationAction(ISD::UDIVREM, VT, Expand);
724 setOperationAction(ISD::FPOW, VT, Expand);
725 setOperationAction(ISD::CTPOP, VT, Expand);
726 setOperationAction(ISD::CTTZ, VT, Expand);
727 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
728 setOperationAction(ISD::CTLZ, VT, Expand);
729 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
730 setOperationAction(ISD::SHL, VT, Expand);
731 setOperationAction(ISD::SRA, VT, Expand);
732 setOperationAction(ISD::SRL, VT, Expand);
733 setOperationAction(ISD::ROTL, VT, Expand);
734 setOperationAction(ISD::ROTR, VT, Expand);
735 setOperationAction(ISD::BSWAP, VT, Expand);
736 setOperationAction(ISD::SETCC, VT, Expand);
737 setOperationAction(ISD::FLOG, VT, Expand);
738 setOperationAction(ISD::FLOG2, VT, Expand);
739 setOperationAction(ISD::FLOG10, VT, Expand);
740 setOperationAction(ISD::FEXP, VT, Expand);
741 setOperationAction(ISD::FEXP2, VT, Expand);
742 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
743 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
744 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
745 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
746 setOperationAction(ISD::SIGN_EXTEND_INREG, VT,Expand);
747 setOperationAction(ISD::TRUNCATE, VT, Expand);
748 setOperationAction(ISD::SIGN_EXTEND, VT, Expand);
749 setOperationAction(ISD::ZERO_EXTEND, VT, Expand);
750 setOperationAction(ISD::ANY_EXTEND, VT, Expand);
751 setOperationAction(ISD::VSELECT, VT, Expand);
752 setOperationAction(ISD::SELECT_CC, VT, Expand);
753 for (MVT InnerVT : MVT::vector_valuetypes()) {
754 setTruncStoreAction(InnerVT, VT, Expand);
756 setLoadExtAction(ISD::SEXTLOAD, InnerVT, VT, Expand);
757 setLoadExtAction(ISD::ZEXTLOAD, InnerVT, VT, Expand);
759 // N.b. ISD::EXTLOAD legality is basically ignored except for i1-like
760 // types, we have to deal with them whether we ask for Expansion or not.
761 // Setting Expand causes its own optimisation problems though, so leave
763 if (VT.getVectorElementType() == MVT::i1)
764 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
766 // EXTLOAD for MVT::f16 vectors is not legal because f16 vectors are
767 // split/scalarized right now.
768 if (VT.getVectorElementType() == MVT::f16)
769 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
773 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
774 // with -msoft-float, disable use of MMX as well.
775 if (!Subtarget->useSoftFloat() && Subtarget->hasMMX()) {
776 addRegisterClass(MVT::x86mmx, &X86::VR64RegClass);
777 // No operations on x86mmx supported, everything uses intrinsics.
780 // MMX-sized vectors (other than x86mmx) are expected to be expanded
781 // into smaller operations.
782 for (MVT MMXTy : {MVT::v8i8, MVT::v4i16, MVT::v2i32, MVT::v1i64}) {
783 setOperationAction(ISD::MULHS, MMXTy, Expand);
784 setOperationAction(ISD::AND, MMXTy, Expand);
785 setOperationAction(ISD::OR, MMXTy, Expand);
786 setOperationAction(ISD::XOR, MMXTy, Expand);
787 setOperationAction(ISD::SCALAR_TO_VECTOR, MMXTy, Expand);
788 setOperationAction(ISD::SELECT, MMXTy, Expand);
789 setOperationAction(ISD::BITCAST, MMXTy, Expand);
791 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v1i64, Expand);
793 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE1()) {
794 addRegisterClass(MVT::v4f32, &X86::VR128RegClass);
796 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
797 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
798 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
799 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
800 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
801 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
802 setOperationAction(ISD::FABS, MVT::v4f32, Custom);
803 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
804 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
805 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
806 setOperationAction(ISD::VSELECT, MVT::v4f32, Custom);
807 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
808 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
809 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Custom);
812 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE2()) {
813 addRegisterClass(MVT::v2f64, &X86::VR128RegClass);
815 // FIXME: Unfortunately, -soft-float and -no-implicit-float mean XMM
816 // registers cannot be used even for integer operations.
817 addRegisterClass(MVT::v16i8, &X86::VR128RegClass);
818 addRegisterClass(MVT::v8i16, &X86::VR128RegClass);
819 addRegisterClass(MVT::v4i32, &X86::VR128RegClass);
820 addRegisterClass(MVT::v2i64, &X86::VR128RegClass);
822 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
823 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
824 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
825 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
826 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
827 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
828 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
829 setOperationAction(ISD::UMUL_LOHI, MVT::v4i32, Custom);
830 setOperationAction(ISD::SMUL_LOHI, MVT::v4i32, Custom);
831 setOperationAction(ISD::MULHU, MVT::v8i16, Legal);
832 setOperationAction(ISD::MULHS, MVT::v8i16, Legal);
833 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
834 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
835 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
836 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
837 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
838 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
839 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
840 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
841 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
842 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
843 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
844 setOperationAction(ISD::FABS, MVT::v2f64, Custom);
846 setOperationAction(ISD::SMAX, MVT::v8i16, Legal);
847 setOperationAction(ISD::UMAX, MVT::v16i8, Legal);
848 setOperationAction(ISD::SMIN, MVT::v8i16, Legal);
849 setOperationAction(ISD::UMIN, MVT::v16i8, Legal);
851 setOperationAction(ISD::SETCC, MVT::v2i64, Custom);
852 setOperationAction(ISD::SETCC, MVT::v16i8, Custom);
853 setOperationAction(ISD::SETCC, MVT::v8i16, Custom);
854 setOperationAction(ISD::SETCC, MVT::v4i32, Custom);
856 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
857 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
858 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
859 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
860 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
862 setOperationAction(ISD::CTPOP, MVT::v16i8, Custom);
863 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom);
864 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
865 setOperationAction(ISD::CTPOP, MVT::v2i64, Custom);
867 setOperationAction(ISD::CTTZ, MVT::v16i8, Custom);
868 setOperationAction(ISD::CTTZ, MVT::v8i16, Custom);
869 setOperationAction(ISD::CTTZ, MVT::v4i32, Custom);
870 // ISD::CTTZ v2i64 - scalarization is faster.
871 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i8, Custom);
872 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i16, Custom);
873 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i32, Custom);
874 // ISD::CTTZ_ZERO_UNDEF v2i64 - scalarization is faster.
876 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
877 for (auto VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32 }) {
878 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
879 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
880 setOperationAction(ISD::VSELECT, VT, Custom);
881 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
884 // We support custom legalizing of sext and anyext loads for specific
885 // memory vector types which we can load as a scalar (or sequence of
886 // scalars) and extend in-register to a legal 128-bit vector type. For sext
887 // loads these must work with a single scalar load.
888 for (MVT VT : MVT::integer_vector_valuetypes()) {
889 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i8, Custom);
890 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i16, Custom);
891 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v8i8, Custom);
892 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i8, Custom);
893 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i16, Custom);
894 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i32, Custom);
895 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i8, Custom);
896 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i16, Custom);
897 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8i8, Custom);
900 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
901 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
902 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
903 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
904 setOperationAction(ISD::VSELECT, MVT::v2f64, Custom);
905 setOperationAction(ISD::VSELECT, MVT::v2i64, Custom);
906 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
907 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
909 if (Subtarget->is64Bit()) {
910 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
911 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
914 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
915 for (auto VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32 }) {
916 setOperationAction(ISD::AND, VT, Promote);
917 AddPromotedToType (ISD::AND, VT, MVT::v2i64);
918 setOperationAction(ISD::OR, VT, Promote);
919 AddPromotedToType (ISD::OR, VT, MVT::v2i64);
920 setOperationAction(ISD::XOR, VT, Promote);
921 AddPromotedToType (ISD::XOR, VT, MVT::v2i64);
922 setOperationAction(ISD::LOAD, VT, Promote);
923 AddPromotedToType (ISD::LOAD, VT, MVT::v2i64);
924 setOperationAction(ISD::SELECT, VT, Promote);
925 AddPromotedToType (ISD::SELECT, VT, MVT::v2i64);
928 // Custom lower v2i64 and v2f64 selects.
929 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
930 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
931 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
932 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
934 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
935 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
937 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
939 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Custom);
940 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
941 // As there is no 64-bit GPR available, we need build a special custom
942 // sequence to convert from v2i32 to v2f32.
943 if (!Subtarget->is64Bit())
944 setOperationAction(ISD::UINT_TO_FP, MVT::v2f32, Custom);
946 setOperationAction(ISD::FP_EXTEND, MVT::v2f32, Custom);
947 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Custom);
949 for (MVT VT : MVT::fp_vector_valuetypes())
950 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2f32, Legal);
952 setOperationAction(ISD::BITCAST, MVT::v2i32, Custom);
953 setOperationAction(ISD::BITCAST, MVT::v4i16, Custom);
954 setOperationAction(ISD::BITCAST, MVT::v8i8, Custom);
957 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE41()) {
958 for (MVT RoundedTy : {MVT::f32, MVT::f64, MVT::v4f32, MVT::v2f64}) {
959 setOperationAction(ISD::FFLOOR, RoundedTy, Legal);
960 setOperationAction(ISD::FCEIL, RoundedTy, Legal);
961 setOperationAction(ISD::FTRUNC, RoundedTy, Legal);
962 setOperationAction(ISD::FRINT, RoundedTy, Legal);
963 setOperationAction(ISD::FNEARBYINT, RoundedTy, Legal);
966 setOperationAction(ISD::SMAX, MVT::v16i8, Legal);
967 setOperationAction(ISD::SMAX, MVT::v4i32, Legal);
968 setOperationAction(ISD::UMAX, MVT::v8i16, Legal);
969 setOperationAction(ISD::UMAX, MVT::v4i32, Legal);
970 setOperationAction(ISD::SMIN, MVT::v16i8, Legal);
971 setOperationAction(ISD::SMIN, MVT::v4i32, Legal);
972 setOperationAction(ISD::UMIN, MVT::v8i16, Legal);
973 setOperationAction(ISD::UMIN, MVT::v4i32, Legal);
975 // FIXME: Do we need to handle scalar-to-vector here?
976 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
978 // We directly match byte blends in the backend as they match the VSELECT
980 setOperationAction(ISD::VSELECT, MVT::v16i8, Legal);
982 // SSE41 brings specific instructions for doing vector sign extend even in
983 // cases where we don't have SRA.
984 for (MVT VT : MVT::integer_vector_valuetypes()) {
985 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i8, Custom);
986 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i16, Custom);
987 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i32, Custom);
990 // SSE41 also has vector sign/zero extending loads, PMOV[SZ]X
991 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
992 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
993 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
994 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
995 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
996 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
998 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
999 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
1000 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
1001 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
1002 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
1003 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
1005 // i8 and i16 vectors are custom because the source register and source
1006 // source memory operand types are not the same width. f32 vectors are
1007 // custom since the immediate controlling the insert encodes additional
1009 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
1010 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
1011 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
1012 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
1014 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
1015 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
1016 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
1017 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
1019 // FIXME: these should be Legal, but that's only for the case where
1020 // the index is constant. For now custom expand to deal with that.
1021 if (Subtarget->is64Bit()) {
1022 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
1023 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
1027 if (Subtarget->hasSSE2()) {
1028 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v2i64, Custom);
1029 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v4i32, Custom);
1030 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v8i16, Custom);
1032 setOperationAction(ISD::SRL, MVT::v8i16, Custom);
1033 setOperationAction(ISD::SRL, MVT::v16i8, Custom);
1035 setOperationAction(ISD::SHL, MVT::v8i16, Custom);
1036 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
1038 setOperationAction(ISD::SRA, MVT::v8i16, Custom);
1039 setOperationAction(ISD::SRA, MVT::v16i8, Custom);
1041 // In the customized shift lowering, the legal cases in AVX2 will be
1043 setOperationAction(ISD::SRL, MVT::v2i64, Custom);
1044 setOperationAction(ISD::SRL, MVT::v4i32, Custom);
1046 setOperationAction(ISD::SHL, MVT::v2i64, Custom);
1047 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
1049 setOperationAction(ISD::SRA, MVT::v2i64, Custom);
1050 setOperationAction(ISD::SRA, MVT::v4i32, Custom);
1053 if (Subtarget->hasXOP()) {
1054 setOperationAction(ISD::ROTL, MVT::v16i8, Custom);
1055 setOperationAction(ISD::ROTL, MVT::v8i16, Custom);
1056 setOperationAction(ISD::ROTL, MVT::v4i32, Custom);
1057 setOperationAction(ISD::ROTL, MVT::v2i64, Custom);
1058 setOperationAction(ISD::ROTL, MVT::v32i8, Custom);
1059 setOperationAction(ISD::ROTL, MVT::v16i16, Custom);
1060 setOperationAction(ISD::ROTL, MVT::v8i32, Custom);
1061 setOperationAction(ISD::ROTL, MVT::v4i64, Custom);
1064 if (!Subtarget->useSoftFloat() && Subtarget->hasFp256()) {
1065 addRegisterClass(MVT::v32i8, &X86::VR256RegClass);
1066 addRegisterClass(MVT::v16i16, &X86::VR256RegClass);
1067 addRegisterClass(MVT::v8i32, &X86::VR256RegClass);
1068 addRegisterClass(MVT::v8f32, &X86::VR256RegClass);
1069 addRegisterClass(MVT::v4i64, &X86::VR256RegClass);
1070 addRegisterClass(MVT::v4f64, &X86::VR256RegClass);
1072 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
1073 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
1074 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
1076 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
1077 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
1078 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
1079 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
1080 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
1081 setOperationAction(ISD::FFLOOR, MVT::v8f32, Legal);
1082 setOperationAction(ISD::FCEIL, MVT::v8f32, Legal);
1083 setOperationAction(ISD::FTRUNC, MVT::v8f32, Legal);
1084 setOperationAction(ISD::FRINT, MVT::v8f32, Legal);
1085 setOperationAction(ISD::FNEARBYINT, MVT::v8f32, Legal);
1086 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
1087 setOperationAction(ISD::FABS, MVT::v8f32, Custom);
1089 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
1090 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
1091 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
1092 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
1093 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
1094 setOperationAction(ISD::FFLOOR, MVT::v4f64, Legal);
1095 setOperationAction(ISD::FCEIL, MVT::v4f64, Legal);
1096 setOperationAction(ISD::FTRUNC, MVT::v4f64, Legal);
1097 setOperationAction(ISD::FRINT, MVT::v4f64, Legal);
1098 setOperationAction(ISD::FNEARBYINT, MVT::v4f64, Legal);
1099 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
1100 setOperationAction(ISD::FABS, MVT::v4f64, Custom);
1102 // (fp_to_int:v8i16 (v8f32 ..)) requires the result type to be promoted
1103 // even though v8i16 is a legal type.
1104 setOperationAction(ISD::FP_TO_SINT, MVT::v8i16, Promote);
1105 setOperationAction(ISD::FP_TO_UINT, MVT::v8i16, Promote);
1106 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1108 setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Promote);
1109 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1110 setOperationAction(ISD::FP_ROUND, MVT::v4f32, Legal);
1112 setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Custom);
1113 setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Custom);
1115 for (MVT VT : MVT::fp_vector_valuetypes())
1116 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4f32, Legal);
1118 setOperationAction(ISD::SRL, MVT::v16i16, Custom);
1119 setOperationAction(ISD::SRL, MVT::v32i8, Custom);
1121 setOperationAction(ISD::SHL, MVT::v16i16, Custom);
1122 setOperationAction(ISD::SHL, MVT::v32i8, Custom);
1124 setOperationAction(ISD::SRA, MVT::v16i16, Custom);
1125 setOperationAction(ISD::SRA, MVT::v32i8, Custom);
1127 setOperationAction(ISD::SETCC, MVT::v32i8, Custom);
1128 setOperationAction(ISD::SETCC, MVT::v16i16, Custom);
1129 setOperationAction(ISD::SETCC, MVT::v8i32, Custom);
1130 setOperationAction(ISD::SETCC, MVT::v4i64, Custom);
1132 setOperationAction(ISD::SELECT, MVT::v4f64, Custom);
1133 setOperationAction(ISD::SELECT, MVT::v4i64, Custom);
1134 setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
1136 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i64, Custom);
1137 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i32, Custom);
1138 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1139 setOperationAction(ISD::ZERO_EXTEND, MVT::v4i64, Custom);
1140 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i32, Custom);
1141 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i16, Custom);
1142 setOperationAction(ISD::ANY_EXTEND, MVT::v4i64, Custom);
1143 setOperationAction(ISD::ANY_EXTEND, MVT::v8i32, Custom);
1144 setOperationAction(ISD::ANY_EXTEND, MVT::v16i16, Custom);
1145 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1146 setOperationAction(ISD::TRUNCATE, MVT::v8i16, Custom);
1147 setOperationAction(ISD::TRUNCATE, MVT::v4i32, Custom);
1149 setOperationAction(ISD::CTPOP, MVT::v32i8, Custom);
1150 setOperationAction(ISD::CTPOP, MVT::v16i16, Custom);
1151 setOperationAction(ISD::CTPOP, MVT::v8i32, Custom);
1152 setOperationAction(ISD::CTPOP, MVT::v4i64, Custom);
1154 setOperationAction(ISD::CTTZ, MVT::v32i8, Custom);
1155 setOperationAction(ISD::CTTZ, MVT::v16i16, Custom);
1156 setOperationAction(ISD::CTTZ, MVT::v8i32, Custom);
1157 setOperationAction(ISD::CTTZ, MVT::v4i64, Custom);
1158 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v32i8, Custom);
1159 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i16, Custom);
1160 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i32, Custom);
1161 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i64, Custom);
1163 if (Subtarget->hasAnyFMA()) {
1164 setOperationAction(ISD::FMA, MVT::v8f32, Legal);
1165 setOperationAction(ISD::FMA, MVT::v4f64, Legal);
1166 setOperationAction(ISD::FMA, MVT::v4f32, Legal);
1167 setOperationAction(ISD::FMA, MVT::v2f64, Legal);
1168 setOperationAction(ISD::FMA, MVT::f32, Legal);
1169 setOperationAction(ISD::FMA, MVT::f64, Legal);
1172 if (Subtarget->hasInt256()) {
1173 setOperationAction(ISD::ADD, MVT::v4i64, Legal);
1174 setOperationAction(ISD::ADD, MVT::v8i32, Legal);
1175 setOperationAction(ISD::ADD, MVT::v16i16, Legal);
1176 setOperationAction(ISD::ADD, MVT::v32i8, Legal);
1178 setOperationAction(ISD::SUB, MVT::v4i64, Legal);
1179 setOperationAction(ISD::SUB, MVT::v8i32, Legal);
1180 setOperationAction(ISD::SUB, MVT::v16i16, Legal);
1181 setOperationAction(ISD::SUB, MVT::v32i8, Legal);
1183 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1184 setOperationAction(ISD::MUL, MVT::v8i32, Legal);
1185 setOperationAction(ISD::MUL, MVT::v16i16, Legal);
1186 setOperationAction(ISD::MUL, MVT::v32i8, Custom);
1188 setOperationAction(ISD::UMUL_LOHI, MVT::v8i32, Custom);
1189 setOperationAction(ISD::SMUL_LOHI, MVT::v8i32, Custom);
1190 setOperationAction(ISD::MULHU, MVT::v16i16, Legal);
1191 setOperationAction(ISD::MULHS, MVT::v16i16, Legal);
1193 setOperationAction(ISD::SMAX, MVT::v32i8, Legal);
1194 setOperationAction(ISD::SMAX, MVT::v16i16, Legal);
1195 setOperationAction(ISD::SMAX, MVT::v8i32, Legal);
1196 setOperationAction(ISD::UMAX, MVT::v32i8, Legal);
1197 setOperationAction(ISD::UMAX, MVT::v16i16, Legal);
1198 setOperationAction(ISD::UMAX, MVT::v8i32, Legal);
1199 setOperationAction(ISD::SMIN, MVT::v32i8, Legal);
1200 setOperationAction(ISD::SMIN, MVT::v16i16, Legal);
1201 setOperationAction(ISD::SMIN, MVT::v8i32, Legal);
1202 setOperationAction(ISD::UMIN, MVT::v32i8, Legal);
1203 setOperationAction(ISD::UMIN, MVT::v16i16, Legal);
1204 setOperationAction(ISD::UMIN, MVT::v8i32, Legal);
1206 // The custom lowering for UINT_TO_FP for v8i32 becomes interesting
1207 // when we have a 256bit-wide blend with immediate.
1208 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Custom);
1210 // AVX2 also has wider vector sign/zero extending loads, VPMOV[SZ]X
1211 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1212 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1213 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1214 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1215 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1216 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1218 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1219 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1220 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1221 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1222 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1223 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1225 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
1226 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
1227 setOperationAction(ISD::ADD, MVT::v16i16, Custom);
1228 setOperationAction(ISD::ADD, MVT::v32i8, Custom);
1230 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
1231 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
1232 setOperationAction(ISD::SUB, MVT::v16i16, Custom);
1233 setOperationAction(ISD::SUB, MVT::v32i8, Custom);
1235 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1236 setOperationAction(ISD::MUL, MVT::v8i32, Custom);
1237 setOperationAction(ISD::MUL, MVT::v16i16, Custom);
1238 setOperationAction(ISD::MUL, MVT::v32i8, Custom);
1240 setOperationAction(ISD::SMAX, MVT::v32i8, Custom);
1241 setOperationAction(ISD::SMAX, MVT::v16i16, Custom);
1242 setOperationAction(ISD::SMAX, MVT::v8i32, Custom);
1243 setOperationAction(ISD::UMAX, MVT::v32i8, Custom);
1244 setOperationAction(ISD::UMAX, MVT::v16i16, Custom);
1245 setOperationAction(ISD::UMAX, MVT::v8i32, Custom);
1246 setOperationAction(ISD::SMIN, MVT::v32i8, Custom);
1247 setOperationAction(ISD::SMIN, MVT::v16i16, Custom);
1248 setOperationAction(ISD::SMIN, MVT::v8i32, Custom);
1249 setOperationAction(ISD::UMIN, MVT::v32i8, Custom);
1250 setOperationAction(ISD::UMIN, MVT::v16i16, Custom);
1251 setOperationAction(ISD::UMIN, MVT::v8i32, Custom);
1254 // In the customized shift lowering, the legal cases in AVX2 will be
1256 setOperationAction(ISD::SRL, MVT::v4i64, Custom);
1257 setOperationAction(ISD::SRL, MVT::v8i32, Custom);
1259 setOperationAction(ISD::SHL, MVT::v4i64, Custom);
1260 setOperationAction(ISD::SHL, MVT::v8i32, Custom);
1262 setOperationAction(ISD::SRA, MVT::v4i64, Custom);
1263 setOperationAction(ISD::SRA, MVT::v8i32, Custom);
1265 // Custom lower several nodes for 256-bit types.
1266 for (MVT VT : MVT::vector_valuetypes()) {
1267 if (VT.getScalarSizeInBits() >= 32) {
1268 setOperationAction(ISD::MLOAD, VT, Legal);
1269 setOperationAction(ISD::MSTORE, VT, Legal);
1271 // Extract subvector is special because the value type
1272 // (result) is 128-bit but the source is 256-bit wide.
1273 if (VT.is128BitVector()) {
1274 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1276 // Do not attempt to custom lower other non-256-bit vectors
1277 if (!VT.is256BitVector())
1280 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1281 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1282 setOperationAction(ISD::VSELECT, VT, Custom);
1283 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1284 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1285 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1286 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1287 setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
1290 if (Subtarget->hasInt256())
1291 setOperationAction(ISD::VSELECT, MVT::v32i8, Legal);
1293 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
1294 for (auto VT : { MVT::v32i8, MVT::v16i16, MVT::v8i32 }) {
1295 setOperationAction(ISD::AND, VT, Promote);
1296 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
1297 setOperationAction(ISD::OR, VT, Promote);
1298 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
1299 setOperationAction(ISD::XOR, VT, Promote);
1300 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
1301 setOperationAction(ISD::LOAD, VT, Promote);
1302 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
1303 setOperationAction(ISD::SELECT, VT, Promote);
1304 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
1308 if (!Subtarget->useSoftFloat() && Subtarget->hasAVX512()) {
1309 addRegisterClass(MVT::v16i32, &X86::VR512RegClass);
1310 addRegisterClass(MVT::v16f32, &X86::VR512RegClass);
1311 addRegisterClass(MVT::v8i64, &X86::VR512RegClass);
1312 addRegisterClass(MVT::v8f64, &X86::VR512RegClass);
1314 addRegisterClass(MVT::i1, &X86::VK1RegClass);
1315 addRegisterClass(MVT::v8i1, &X86::VK8RegClass);
1316 addRegisterClass(MVT::v16i1, &X86::VK16RegClass);
1318 for (MVT VT : MVT::fp_vector_valuetypes())
1319 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8f32, Legal);
1321 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i32, MVT::v16i8, Legal);
1322 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i32, MVT::v16i8, Legal);
1323 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i32, MVT::v16i16, Legal);
1324 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i32, MVT::v16i16, Legal);
1325 setLoadExtAction(ISD::ZEXTLOAD, MVT::v32i16, MVT::v32i8, Legal);
1326 setLoadExtAction(ISD::SEXTLOAD, MVT::v32i16, MVT::v32i8, Legal);
1327 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i8, Legal);
1328 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i8, Legal);
1329 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i16, Legal);
1330 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i16, Legal);
1331 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i32, Legal);
1332 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i32, Legal);
1334 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
1335 setOperationAction(ISD::SETCC, MVT::i1, Custom);
1336 setOperationAction(ISD::SELECT_CC, MVT::i1, Expand);
1337 setOperationAction(ISD::XOR, MVT::i1, Legal);
1338 setOperationAction(ISD::OR, MVT::i1, Legal);
1339 setOperationAction(ISD::AND, MVT::i1, Legal);
1340 setOperationAction(ISD::SUB, MVT::i1, Custom);
1341 setOperationAction(ISD::ADD, MVT::i1, Custom);
1342 setOperationAction(ISD::MUL, MVT::i1, Custom);
1343 setOperationAction(ISD::LOAD, MVT::v16f32, Legal);
1344 setOperationAction(ISD::LOAD, MVT::v8f64, Legal);
1345 setOperationAction(ISD::LOAD, MVT::v8i64, Legal);
1346 setOperationAction(ISD::LOAD, MVT::v16i32, Legal);
1347 setOperationAction(ISD::LOAD, MVT::v16i1, Legal);
1349 setOperationAction(ISD::FADD, MVT::v16f32, Legal);
1350 setOperationAction(ISD::FSUB, MVT::v16f32, Legal);
1351 setOperationAction(ISD::FMUL, MVT::v16f32, Legal);
1352 setOperationAction(ISD::FDIV, MVT::v16f32, Legal);
1353 setOperationAction(ISD::FSQRT, MVT::v16f32, Legal);
1354 setOperationAction(ISD::FNEG, MVT::v16f32, Custom);
1355 setOperationAction(ISD::FABS, MVT::v16f32, Custom);
1357 setOperationAction(ISD::FADD, MVT::v8f64, Legal);
1358 setOperationAction(ISD::FSUB, MVT::v8f64, Legal);
1359 setOperationAction(ISD::FMUL, MVT::v8f64, Legal);
1360 setOperationAction(ISD::FDIV, MVT::v8f64, Legal);
1361 setOperationAction(ISD::FSQRT, MVT::v8f64, Legal);
1362 setOperationAction(ISD::FNEG, MVT::v8f64, Custom);
1363 setOperationAction(ISD::FABS, MVT::v8f64, Custom);
1364 setOperationAction(ISD::FMA, MVT::v8f64, Legal);
1365 setOperationAction(ISD::FMA, MVT::v16f32, Legal);
1367 setOperationAction(ISD::FP_TO_SINT, MVT::v16i32, Legal);
1368 setOperationAction(ISD::FP_TO_UINT, MVT::v16i32, Legal);
1369 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1370 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1371 setOperationAction(ISD::SINT_TO_FP, MVT::v16i32, Legal);
1372 setOperationAction(ISD::SINT_TO_FP, MVT::v8i1, Custom);
1373 setOperationAction(ISD::SINT_TO_FP, MVT::v16i1, Custom);
1374 setOperationAction(ISD::SINT_TO_FP, MVT::v16i8, Promote);
1375 setOperationAction(ISD::SINT_TO_FP, MVT::v16i16, Promote);
1376 setOperationAction(ISD::UINT_TO_FP, MVT::v16i32, Legal);
1377 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1378 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1379 setOperationAction(ISD::UINT_TO_FP, MVT::v16i8, Custom);
1380 setOperationAction(ISD::UINT_TO_FP, MVT::v16i16, Custom);
1381 setOperationAction(ISD::FP_ROUND, MVT::v8f32, Legal);
1382 setOperationAction(ISD::FP_EXTEND, MVT::v8f32, Legal);
1384 setTruncStoreAction(MVT::v8i64, MVT::v8i8, Legal);
1385 setTruncStoreAction(MVT::v8i64, MVT::v8i16, Legal);
1386 setTruncStoreAction(MVT::v8i64, MVT::v8i32, Legal);
1387 setTruncStoreAction(MVT::v16i32, MVT::v16i8, Legal);
1388 setTruncStoreAction(MVT::v16i32, MVT::v16i16, Legal);
1389 if (Subtarget->hasVLX()){
1390 setTruncStoreAction(MVT::v4i64, MVT::v4i8, Legal);
1391 setTruncStoreAction(MVT::v4i64, MVT::v4i16, Legal);
1392 setTruncStoreAction(MVT::v4i64, MVT::v4i32, Legal);
1393 setTruncStoreAction(MVT::v8i32, MVT::v8i8, Legal);
1394 setTruncStoreAction(MVT::v8i32, MVT::v8i16, Legal);
1396 setTruncStoreAction(MVT::v2i64, MVT::v2i8, Legal);
1397 setTruncStoreAction(MVT::v2i64, MVT::v2i16, Legal);
1398 setTruncStoreAction(MVT::v2i64, MVT::v2i32, Legal);
1399 setTruncStoreAction(MVT::v4i32, MVT::v4i8, Legal);
1400 setTruncStoreAction(MVT::v4i32, MVT::v4i16, Legal);
1402 setOperationAction(ISD::MLOAD, MVT::v8i32, Custom);
1403 setOperationAction(ISD::MLOAD, MVT::v8f32, Custom);
1404 setOperationAction(ISD::MSTORE, MVT::v8i32, Custom);
1405 setOperationAction(ISD::MSTORE, MVT::v8f32, Custom);
1407 setOperationAction(ISD::TRUNCATE, MVT::i1, Custom);
1408 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1409 setOperationAction(ISD::TRUNCATE, MVT::v8i32, Custom);
1410 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i1, Custom);
1411 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i1, Custom);
1412 if (Subtarget->hasDQI()) {
1413 setOperationAction(ISD::TRUNCATE, MVT::v2i1, Custom);
1414 setOperationAction(ISD::TRUNCATE, MVT::v4i1, Custom);
1416 setOperationAction(ISD::SINT_TO_FP, MVT::v8i64, Legal);
1417 setOperationAction(ISD::UINT_TO_FP, MVT::v8i64, Legal);
1418 setOperationAction(ISD::FP_TO_SINT, MVT::v8i64, Legal);
1419 setOperationAction(ISD::FP_TO_UINT, MVT::v8i64, Legal);
1420 if (Subtarget->hasVLX()) {
1421 setOperationAction(ISD::SINT_TO_FP, MVT::v4i64, Legal);
1422 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Legal);
1423 setOperationAction(ISD::UINT_TO_FP, MVT::v4i64, Legal);
1424 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Legal);
1425 setOperationAction(ISD::FP_TO_SINT, MVT::v4i64, Legal);
1426 setOperationAction(ISD::FP_TO_SINT, MVT::v2i64, Legal);
1427 setOperationAction(ISD::FP_TO_UINT, MVT::v4i64, Legal);
1428 setOperationAction(ISD::FP_TO_UINT, MVT::v2i64, Legal);
1431 if (Subtarget->hasVLX()) {
1432 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1433 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1434 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1435 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1436 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
1437 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1438 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
1439 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1441 setOperationAction(ISD::TRUNCATE, MVT::v8i1, Custom);
1442 setOperationAction(ISD::TRUNCATE, MVT::v16i1, Custom);
1443 setOperationAction(ISD::TRUNCATE, MVT::v16i16, Custom);
1444 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i32, Custom);
1445 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i64, Custom);
1446 setOperationAction(ISD::ANY_EXTEND, MVT::v16i32, Custom);
1447 setOperationAction(ISD::ANY_EXTEND, MVT::v8i64, Custom);
1448 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i32, Custom);
1449 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i64, Custom);
1450 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i8, Custom);
1451 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i16, Custom);
1452 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1453 if (Subtarget->hasDQI()) {
1454 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i32, Custom);
1455 setOperationAction(ISD::SIGN_EXTEND, MVT::v2i64, Custom);
1457 setOperationAction(ISD::FFLOOR, MVT::v16f32, Legal);
1458 setOperationAction(ISD::FFLOOR, MVT::v8f64, Legal);
1459 setOperationAction(ISD::FCEIL, MVT::v16f32, Legal);
1460 setOperationAction(ISD::FCEIL, MVT::v8f64, Legal);
1461 setOperationAction(ISD::FTRUNC, MVT::v16f32, Legal);
1462 setOperationAction(ISD::FTRUNC, MVT::v8f64, Legal);
1463 setOperationAction(ISD::FRINT, MVT::v16f32, Legal);
1464 setOperationAction(ISD::FRINT, MVT::v8f64, Legal);
1465 setOperationAction(ISD::FNEARBYINT, MVT::v16f32, Legal);
1466 setOperationAction(ISD::FNEARBYINT, MVT::v8f64, Legal);
1468 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f64, Custom);
1469 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i64, Custom);
1470 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16f32, Custom);
1471 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i32, Custom);
1472 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i1, Custom);
1474 setOperationAction(ISD::SETCC, MVT::v16i1, Custom);
1475 setOperationAction(ISD::SETCC, MVT::v8i1, Custom);
1477 setOperationAction(ISD::MUL, MVT::v8i64, Custom);
1479 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i1, Custom);
1480 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i1, Custom);
1481 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v16i1, Custom);
1482 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i1, Custom);
1483 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i1, Custom);
1484 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i1, Custom);
1485 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i1, Custom);
1486 setOperationAction(ISD::SELECT, MVT::v8f64, Custom);
1487 setOperationAction(ISD::SELECT, MVT::v8i64, Custom);
1488 setOperationAction(ISD::SELECT, MVT::v16f32, Custom);
1489 setOperationAction(ISD::SELECT, MVT::v16i1, Custom);
1490 setOperationAction(ISD::SELECT, MVT::v8i1, Custom);
1492 setOperationAction(ISD::SMAX, MVT::v16i32, Legal);
1493 setOperationAction(ISD::SMAX, MVT::v8i64, Legal);
1494 setOperationAction(ISD::UMAX, MVT::v16i32, Legal);
1495 setOperationAction(ISD::UMAX, MVT::v8i64, Legal);
1496 setOperationAction(ISD::SMIN, MVT::v16i32, Legal);
1497 setOperationAction(ISD::SMIN, MVT::v8i64, Legal);
1498 setOperationAction(ISD::UMIN, MVT::v16i32, Legal);
1499 setOperationAction(ISD::UMIN, MVT::v8i64, Legal);
1501 setOperationAction(ISD::ADD, MVT::v8i64, Legal);
1502 setOperationAction(ISD::ADD, MVT::v16i32, Legal);
1504 setOperationAction(ISD::SUB, MVT::v8i64, Legal);
1505 setOperationAction(ISD::SUB, MVT::v16i32, Legal);
1507 setOperationAction(ISD::MUL, MVT::v16i32, Legal);
1509 setOperationAction(ISD::SRL, MVT::v8i64, Custom);
1510 setOperationAction(ISD::SRL, MVT::v16i32, Custom);
1512 setOperationAction(ISD::SHL, MVT::v8i64, Custom);
1513 setOperationAction(ISD::SHL, MVT::v16i32, Custom);
1515 setOperationAction(ISD::SRA, MVT::v8i64, Custom);
1516 setOperationAction(ISD::SRA, MVT::v16i32, Custom);
1518 setOperationAction(ISD::AND, MVT::v8i64, Legal);
1519 setOperationAction(ISD::OR, MVT::v8i64, Legal);
1520 setOperationAction(ISD::XOR, MVT::v8i64, Legal);
1521 setOperationAction(ISD::AND, MVT::v16i32, Legal);
1522 setOperationAction(ISD::OR, MVT::v16i32, Legal);
1523 setOperationAction(ISD::XOR, MVT::v16i32, Legal);
1525 if (Subtarget->hasCDI()) {
1526 setOperationAction(ISD::CTLZ, MVT::v8i64, Legal);
1527 setOperationAction(ISD::CTLZ, MVT::v16i32, Legal);
1528 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i64, Legal);
1529 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v16i32, Legal);
1531 setOperationAction(ISD::CTLZ, MVT::v8i16, Custom);
1532 setOperationAction(ISD::CTLZ, MVT::v16i8, Custom);
1533 setOperationAction(ISD::CTLZ, MVT::v16i16, Custom);
1534 setOperationAction(ISD::CTLZ, MVT::v32i8, Custom);
1535 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i16, Custom);
1536 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v16i8, Custom);
1537 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v16i16, Custom);
1538 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v32i8, Custom);
1540 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i64, Custom);
1541 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i32, Custom);
1543 if (Subtarget->hasVLX()) {
1544 setOperationAction(ISD::CTLZ, MVT::v4i64, Legal);
1545 setOperationAction(ISD::CTLZ, MVT::v8i32, Legal);
1546 setOperationAction(ISD::CTLZ, MVT::v2i64, Legal);
1547 setOperationAction(ISD::CTLZ, MVT::v4i32, Legal);
1548 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i64, Legal);
1549 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i32, Legal);
1550 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v2i64, Legal);
1551 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i32, Legal);
1553 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i64, Custom);
1554 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i32, Custom);
1555 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v2i64, Custom);
1556 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i32, Custom);
1558 setOperationAction(ISD::CTLZ, MVT::v4i64, Custom);
1559 setOperationAction(ISD::CTLZ, MVT::v8i32, Custom);
1560 setOperationAction(ISD::CTLZ, MVT::v2i64, Custom);
1561 setOperationAction(ISD::CTLZ, MVT::v4i32, Custom);
1562 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i64, Custom);
1563 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i32, Custom);
1564 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v2i64, Custom);
1565 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i32, Custom);
1567 } // Subtarget->hasCDI()
1569 if (Subtarget->hasDQI()) {
1570 setOperationAction(ISD::MUL, MVT::v2i64, Legal);
1571 setOperationAction(ISD::MUL, MVT::v4i64, Legal);
1572 setOperationAction(ISD::MUL, MVT::v8i64, Legal);
1574 // Custom lower several nodes.
1575 for (MVT VT : MVT::vector_valuetypes()) {
1576 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1578 setOperationAction(ISD::AND, VT, Legal);
1579 setOperationAction(ISD::OR, VT, Legal);
1580 setOperationAction(ISD::XOR, VT, Legal);
1582 if ((VT.is128BitVector() || VT.is256BitVector()) && EltSize >= 32) {
1583 setOperationAction(ISD::MGATHER, VT, Custom);
1584 setOperationAction(ISD::MSCATTER, VT, Custom);
1586 // Extract subvector is special because the value type
1587 // (result) is 256/128-bit but the source is 512-bit wide.
1588 if (VT.is128BitVector() || VT.is256BitVector()) {
1589 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1591 if (VT.getVectorElementType() == MVT::i1)
1592 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
1594 // Do not attempt to custom lower other non-512-bit vectors
1595 if (!VT.is512BitVector())
1598 if (EltSize >= 32) {
1599 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1600 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1601 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1602 setOperationAction(ISD::VSELECT, VT, Legal);
1603 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1604 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1605 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1606 setOperationAction(ISD::MLOAD, VT, Legal);
1607 setOperationAction(ISD::MSTORE, VT, Legal);
1608 setOperationAction(ISD::MGATHER, VT, Legal);
1609 setOperationAction(ISD::MSCATTER, VT, Custom);
1612 for (auto VT : { MVT::v64i8, MVT::v32i16, MVT::v16i32 }) {
1613 setOperationAction(ISD::SELECT, VT, Promote);
1614 AddPromotedToType (ISD::SELECT, VT, MVT::v8i64);
1618 if (!Subtarget->useSoftFloat() && Subtarget->hasBWI()) {
1619 addRegisterClass(MVT::v32i16, &X86::VR512RegClass);
1620 addRegisterClass(MVT::v64i8, &X86::VR512RegClass);
1622 addRegisterClass(MVT::v32i1, &X86::VK32RegClass);
1623 addRegisterClass(MVT::v64i1, &X86::VK64RegClass);
1625 setOperationAction(ISD::LOAD, MVT::v32i16, Legal);
1626 setOperationAction(ISD::LOAD, MVT::v64i8, Legal);
1627 setOperationAction(ISD::SETCC, MVT::v32i1, Custom);
1628 setOperationAction(ISD::SETCC, MVT::v64i1, Custom);
1629 setOperationAction(ISD::ADD, MVT::v32i16, Legal);
1630 setOperationAction(ISD::ADD, MVT::v64i8, Legal);
1631 setOperationAction(ISD::SUB, MVT::v32i16, Legal);
1632 setOperationAction(ISD::SUB, MVT::v64i8, Legal);
1633 setOperationAction(ISD::MUL, MVT::v32i16, Legal);
1634 setOperationAction(ISD::MULHS, MVT::v32i16, Legal);
1635 setOperationAction(ISD::MULHU, MVT::v32i16, Legal);
1636 setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i1, Custom);
1637 setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i1, Custom);
1638 setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i16, Custom);
1639 setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i8, Custom);
1640 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v32i1, Custom);
1641 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v64i1, Custom);
1642 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v32i16, Custom);
1643 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v64i8, Custom);
1644 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v32i16, Custom);
1645 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v64i8, Custom);
1646 setOperationAction(ISD::SELECT, MVT::v32i1, Custom);
1647 setOperationAction(ISD::SELECT, MVT::v64i1, Custom);
1648 setOperationAction(ISD::SIGN_EXTEND, MVT::v32i8, Custom);
1649 setOperationAction(ISD::ZERO_EXTEND, MVT::v32i8, Custom);
1650 setOperationAction(ISD::SIGN_EXTEND, MVT::v32i16, Custom);
1651 setOperationAction(ISD::ZERO_EXTEND, MVT::v32i16, Custom);
1652 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v32i16, Custom);
1653 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v64i8, Custom);
1654 setOperationAction(ISD::SIGN_EXTEND, MVT::v64i8, Custom);
1655 setOperationAction(ISD::ZERO_EXTEND, MVT::v64i8, Custom);
1656 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v32i1, Custom);
1657 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v64i1, Custom);
1658 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v32i16, Custom);
1659 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v64i8, Custom);
1660 setOperationAction(ISD::VSELECT, MVT::v32i16, Legal);
1661 setOperationAction(ISD::VSELECT, MVT::v64i8, Legal);
1662 setOperationAction(ISD::TRUNCATE, MVT::v32i1, Custom);
1663 setOperationAction(ISD::TRUNCATE, MVT::v64i1, Custom);
1664 setOperationAction(ISD::TRUNCATE, MVT::v32i8, Custom);
1665 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v32i1, Custom);
1666 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v64i1, Custom);
1668 setOperationAction(ISD::SMAX, MVT::v64i8, Legal);
1669 setOperationAction(ISD::SMAX, MVT::v32i16, Legal);
1670 setOperationAction(ISD::UMAX, MVT::v64i8, Legal);
1671 setOperationAction(ISD::UMAX, MVT::v32i16, Legal);
1672 setOperationAction(ISD::SMIN, MVT::v64i8, Legal);
1673 setOperationAction(ISD::SMIN, MVT::v32i16, Legal);
1674 setOperationAction(ISD::UMIN, MVT::v64i8, Legal);
1675 setOperationAction(ISD::UMIN, MVT::v32i16, Legal);
1677 setTruncStoreAction(MVT::v32i16, MVT::v32i8, Legal);
1678 setTruncStoreAction(MVT::v16i16, MVT::v16i8, Legal);
1679 if (Subtarget->hasVLX())
1680 setTruncStoreAction(MVT::v8i16, MVT::v8i8, Legal);
1682 if (Subtarget->hasCDI()) {
1683 setOperationAction(ISD::CTLZ, MVT::v32i16, Custom);
1684 setOperationAction(ISD::CTLZ, MVT::v64i8, Custom);
1685 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v32i16, Custom);
1686 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v64i8, Custom);
1689 for (auto VT : { MVT::v64i8, MVT::v32i16 }) {
1690 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1691 setOperationAction(ISD::VSELECT, VT, Legal);
1695 if (!Subtarget->useSoftFloat() && Subtarget->hasVLX()) {
1696 addRegisterClass(MVT::v4i1, &X86::VK4RegClass);
1697 addRegisterClass(MVT::v2i1, &X86::VK2RegClass);
1699 setOperationAction(ISD::SETCC, MVT::v4i1, Custom);
1700 setOperationAction(ISD::SETCC, MVT::v2i1, Custom);
1701 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i1, Custom);
1702 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i1, Custom);
1703 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v8i1, Custom);
1704 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v4i1, Custom);
1705 setOperationAction(ISD::SELECT, MVT::v4i1, Custom);
1706 setOperationAction(ISD::SELECT, MVT::v2i1, Custom);
1707 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i1, Custom);
1708 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i1, Custom);
1709 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i1, Custom);
1710 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i1, Custom);
1712 setOperationAction(ISD::AND, MVT::v8i32, Legal);
1713 setOperationAction(ISD::OR, MVT::v8i32, Legal);
1714 setOperationAction(ISD::XOR, MVT::v8i32, Legal);
1715 setOperationAction(ISD::AND, MVT::v4i32, Legal);
1716 setOperationAction(ISD::OR, MVT::v4i32, Legal);
1717 setOperationAction(ISD::XOR, MVT::v4i32, Legal);
1718 setOperationAction(ISD::SRA, MVT::v2i64, Custom);
1719 setOperationAction(ISD::SRA, MVT::v4i64, Custom);
1721 setOperationAction(ISD::SMAX, MVT::v2i64, Legal);
1722 setOperationAction(ISD::SMAX, MVT::v4i64, Legal);
1723 setOperationAction(ISD::UMAX, MVT::v2i64, Legal);
1724 setOperationAction(ISD::UMAX, MVT::v4i64, Legal);
1725 setOperationAction(ISD::SMIN, MVT::v2i64, Legal);
1726 setOperationAction(ISD::SMIN, MVT::v4i64, Legal);
1727 setOperationAction(ISD::UMIN, MVT::v2i64, Legal);
1728 setOperationAction(ISD::UMIN, MVT::v4i64, Legal);
1731 // We want to custom lower some of our intrinsics.
1732 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1733 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
1734 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
1735 if (!Subtarget->is64Bit()) {
1736 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
1737 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i64, Custom);
1740 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1741 // handle type legalization for these operations here.
1743 // FIXME: We really should do custom legalization for addition and
1744 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1745 // than generic legalization for 64-bit multiplication-with-overflow, though.
1746 for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) {
1747 if (VT == MVT::i64 && !Subtarget->is64Bit())
1749 // Add/Sub/Mul with overflow operations are custom lowered.
1750 setOperationAction(ISD::SADDO, VT, Custom);
1751 setOperationAction(ISD::UADDO, VT, Custom);
1752 setOperationAction(ISD::SSUBO, VT, Custom);
1753 setOperationAction(ISD::USUBO, VT, Custom);
1754 setOperationAction(ISD::SMULO, VT, Custom);
1755 setOperationAction(ISD::UMULO, VT, Custom);
1758 if (!Subtarget->is64Bit()) {
1759 // These libcalls are not available in 32-bit.
1760 setLibcallName(RTLIB::SHL_I128, nullptr);
1761 setLibcallName(RTLIB::SRL_I128, nullptr);
1762 setLibcallName(RTLIB::SRA_I128, nullptr);
1765 // Combine sin / cos into one node or libcall if possible.
1766 if (Subtarget->hasSinCos()) {
1767 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
1768 setLibcallName(RTLIB::SINCOS_F64, "sincos");
1769 if (Subtarget->isTargetDarwin()) {
1770 // For MacOSX, we don't want the normal expansion of a libcall to sincos.
1771 // We want to issue a libcall to __sincos_stret to avoid memory traffic.
1772 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
1773 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
1777 if (Subtarget->isTargetWin64()) {
1778 setOperationAction(ISD::SDIV, MVT::i128, Custom);
1779 setOperationAction(ISD::UDIV, MVT::i128, Custom);
1780 setOperationAction(ISD::SREM, MVT::i128, Custom);
1781 setOperationAction(ISD::UREM, MVT::i128, Custom);
1782 setOperationAction(ISD::SDIVREM, MVT::i128, Custom);
1783 setOperationAction(ISD::UDIVREM, MVT::i128, Custom);
1786 // We have target-specific dag combine patterns for the following nodes:
1787 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1788 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1789 setTargetDAGCombine(ISD::BITCAST);
1790 setTargetDAGCombine(ISD::VSELECT);
1791 setTargetDAGCombine(ISD::SELECT);
1792 setTargetDAGCombine(ISD::SHL);
1793 setTargetDAGCombine(ISD::SRA);
1794 setTargetDAGCombine(ISD::SRL);
1795 setTargetDAGCombine(ISD::OR);
1796 setTargetDAGCombine(ISD::AND);
1797 setTargetDAGCombine(ISD::ADD);
1798 setTargetDAGCombine(ISD::FADD);
1799 setTargetDAGCombine(ISD::FSUB);
1800 setTargetDAGCombine(ISD::FNEG);
1801 setTargetDAGCombine(ISD::FMA);
1802 setTargetDAGCombine(ISD::SUB);
1803 setTargetDAGCombine(ISD::LOAD);
1804 setTargetDAGCombine(ISD::MLOAD);
1805 setTargetDAGCombine(ISD::STORE);
1806 setTargetDAGCombine(ISD::MSTORE);
1807 setTargetDAGCombine(ISD::TRUNCATE);
1808 setTargetDAGCombine(ISD::ZERO_EXTEND);
1809 setTargetDAGCombine(ISD::ANY_EXTEND);
1810 setTargetDAGCombine(ISD::SIGN_EXTEND);
1811 setTargetDAGCombine(ISD::SIGN_EXTEND_INREG);
1812 setTargetDAGCombine(ISD::SINT_TO_FP);
1813 setTargetDAGCombine(ISD::UINT_TO_FP);
1814 setTargetDAGCombine(ISD::SETCC);
1815 setTargetDAGCombine(ISD::BUILD_VECTOR);
1816 setTargetDAGCombine(ISD::MUL);
1817 setTargetDAGCombine(ISD::XOR);
1818 setTargetDAGCombine(ISD::MSCATTER);
1819 setTargetDAGCombine(ISD::MGATHER);
1821 computeRegisterProperties(Subtarget->getRegisterInfo());
1823 MaxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1824 MaxStoresPerMemsetOptSize = 8;
1825 MaxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1826 MaxStoresPerMemcpyOptSize = 4;
1827 MaxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1828 MaxStoresPerMemmoveOptSize = 4;
1829 setPrefLoopAlignment(4); // 2^4 bytes.
1831 // A predictable cmov does not hurt on an in-order CPU.
1832 // FIXME: Use a CPU attribute to trigger this, not a CPU model.
1833 PredictableSelectIsExpensive = !Subtarget->isAtom();
1834 EnableExtLdPromotion = true;
1835 setPrefFunctionAlignment(4); // 2^4 bytes.
1837 verifyIntrinsicTables();
1840 // This has so far only been implemented for 64-bit MachO.
1841 bool X86TargetLowering::useLoadStackGuardNode() const {
1842 return Subtarget->isTargetMachO() && Subtarget->is64Bit();
1845 TargetLoweringBase::LegalizeTypeAction
1846 X86TargetLowering::getPreferredVectorAction(EVT VT) const {
1847 if (ExperimentalVectorWideningLegalization &&
1848 VT.getVectorNumElements() != 1 &&
1849 VT.getVectorElementType().getSimpleVT() != MVT::i1)
1850 return TypeWidenVector;
1852 return TargetLoweringBase::getPreferredVectorAction(VT);
1855 EVT X86TargetLowering::getSetCCResultType(const DataLayout &DL, LLVMContext &,
1858 return Subtarget->hasAVX512() ? MVT::i1: MVT::i8;
1860 if (VT.isSimple()) {
1861 MVT VVT = VT.getSimpleVT();
1862 const unsigned NumElts = VVT.getVectorNumElements();
1863 const MVT EltVT = VVT.getVectorElementType();
1864 if (VVT.is512BitVector()) {
1865 if (Subtarget->hasAVX512())
1866 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1867 EltVT == MVT::f32 || EltVT == MVT::f64)
1869 case 8: return MVT::v8i1;
1870 case 16: return MVT::v16i1;
1872 if (Subtarget->hasBWI())
1873 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1875 case 32: return MVT::v32i1;
1876 case 64: return MVT::v64i1;
1880 if (VVT.is256BitVector() || VVT.is128BitVector()) {
1881 if (Subtarget->hasVLX())
1882 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1883 EltVT == MVT::f32 || EltVT == MVT::f64)
1885 case 2: return MVT::v2i1;
1886 case 4: return MVT::v4i1;
1887 case 8: return MVT::v8i1;
1889 if (Subtarget->hasBWI() && Subtarget->hasVLX())
1890 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1892 case 8: return MVT::v8i1;
1893 case 16: return MVT::v16i1;
1894 case 32: return MVT::v32i1;
1899 return VT.changeVectorElementTypeToInteger();
1902 /// Helper for getByValTypeAlignment to determine
1903 /// the desired ByVal argument alignment.
1904 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1907 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1908 if (VTy->getBitWidth() == 128)
1910 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1911 unsigned EltAlign = 0;
1912 getMaxByValAlign(ATy->getElementType(), EltAlign);
1913 if (EltAlign > MaxAlign)
1914 MaxAlign = EltAlign;
1915 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1916 for (auto *EltTy : STy->elements()) {
1917 unsigned EltAlign = 0;
1918 getMaxByValAlign(EltTy, EltAlign);
1919 if (EltAlign > MaxAlign)
1920 MaxAlign = EltAlign;
1927 /// Return the desired alignment for ByVal aggregate
1928 /// function arguments in the caller parameter area. For X86, aggregates
1929 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1930 /// are at 4-byte boundaries.
1931 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty,
1932 const DataLayout &DL) const {
1933 if (Subtarget->is64Bit()) {
1934 // Max of 8 and alignment of type.
1935 unsigned TyAlign = DL.getABITypeAlignment(Ty);
1942 if (Subtarget->hasSSE1())
1943 getMaxByValAlign(Ty, Align);
1947 /// Returns the target specific optimal type for load
1948 /// and store operations as a result of memset, memcpy, and memmove
1949 /// lowering. If DstAlign is zero that means it's safe to destination
1950 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1951 /// means there isn't a need to check it against alignment requirement,
1952 /// probably because the source does not need to be loaded. If 'IsMemset' is
1953 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
1954 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
1955 /// source is constant so it does not need to be loaded.
1956 /// It returns EVT::Other if the type should be determined using generic
1957 /// target-independent logic.
1959 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1960 unsigned DstAlign, unsigned SrcAlign,
1961 bool IsMemset, bool ZeroMemset,
1963 MachineFunction &MF) const {
1964 const Function *F = MF.getFunction();
1965 if ((!IsMemset || ZeroMemset) &&
1966 !F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1968 (!Subtarget->isUnalignedMem16Slow() ||
1969 ((DstAlign == 0 || DstAlign >= 16) &&
1970 (SrcAlign == 0 || SrcAlign >= 16)))) {
1972 // FIXME: Check if unaligned 32-byte accesses are slow.
1973 if (Subtarget->hasInt256())
1975 if (Subtarget->hasFp256())
1978 if (Subtarget->hasSSE2())
1980 if (Subtarget->hasSSE1())
1982 } else if (!MemcpyStrSrc && Size >= 8 &&
1983 !Subtarget->is64Bit() &&
1984 Subtarget->hasSSE2()) {
1985 // Do not use f64 to lower memcpy if source is string constant. It's
1986 // better to use i32 to avoid the loads.
1990 // This is a compromise. If we reach here, unaligned accesses may be slow on
1991 // this target. However, creating smaller, aligned accesses could be even
1992 // slower and would certainly be a lot more code.
1993 if (Subtarget->is64Bit() && Size >= 8)
1998 bool X86TargetLowering::isSafeMemOpType(MVT VT) const {
2000 return X86ScalarSSEf32;
2001 else if (VT == MVT::f64)
2002 return X86ScalarSSEf64;
2007 X86TargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
2012 switch (VT.getSizeInBits()) {
2014 // 8-byte and under are always assumed to be fast.
2018 *Fast = !Subtarget->isUnalignedMem16Slow();
2021 *Fast = !Subtarget->isUnalignedMem32Slow();
2023 // TODO: What about AVX-512 (512-bit) accesses?
2026 // Misaligned accesses of any size are always allowed.
2030 /// Return the entry encoding for a jump table in the
2031 /// current function. The returned value is a member of the
2032 /// MachineJumpTableInfo::JTEntryKind enum.
2033 unsigned X86TargetLowering::getJumpTableEncoding() const {
2034 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
2036 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2037 Subtarget->isPICStyleGOT())
2038 return MachineJumpTableInfo::EK_Custom32;
2040 // Otherwise, use the normal jump table encoding heuristics.
2041 return TargetLowering::getJumpTableEncoding();
2044 bool X86TargetLowering::useSoftFloat() const {
2045 return Subtarget->useSoftFloat();
2049 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
2050 const MachineBasicBlock *MBB,
2051 unsigned uid,MCContext &Ctx) const{
2052 assert(MBB->getParent()->getTarget().getRelocationModel() == Reloc::PIC_ &&
2053 Subtarget->isPICStyleGOT());
2054 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
2056 return MCSymbolRefExpr::create(MBB->getSymbol(),
2057 MCSymbolRefExpr::VK_GOTOFF, Ctx);
2060 /// Returns relocation base for the given PIC jumptable.
2061 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
2062 SelectionDAG &DAG) const {
2063 if (!Subtarget->is64Bit())
2064 // This doesn't have SDLoc associated with it, but is not really the
2065 // same as a Register.
2066 return DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(),
2067 getPointerTy(DAG.getDataLayout()));
2071 /// This returns the relocation base for the given PIC jumptable,
2072 /// the same as getPICJumpTableRelocBase, but as an MCExpr.
2073 const MCExpr *X86TargetLowering::
2074 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
2075 MCContext &Ctx) const {
2076 // X86-64 uses RIP relative addressing based on the jump table label.
2077 if (Subtarget->isPICStyleRIPRel())
2078 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
2080 // Otherwise, the reference is relative to the PIC base.
2081 return MCSymbolRefExpr::create(MF->getPICBaseSymbol(), Ctx);
2084 std::pair<const TargetRegisterClass *, uint8_t>
2085 X86TargetLowering::findRepresentativeClass(const TargetRegisterInfo *TRI,
2087 const TargetRegisterClass *RRC = nullptr;
2089 switch (VT.SimpleTy) {
2091 return TargetLowering::findRepresentativeClass(TRI, VT);
2092 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
2093 RRC = Subtarget->is64Bit() ? &X86::GR64RegClass : &X86::GR32RegClass;
2096 RRC = &X86::VR64RegClass;
2098 case MVT::f32: case MVT::f64:
2099 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
2100 case MVT::v4f32: case MVT::v2f64:
2101 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
2103 RRC = &X86::VR128RegClass;
2106 return std::make_pair(RRC, Cost);
2109 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
2110 unsigned &Offset) const {
2111 if (!Subtarget->isTargetLinux())
2114 if (Subtarget->is64Bit()) {
2115 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
2117 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
2129 Value *X86TargetLowering::getSafeStackPointerLocation(IRBuilder<> &IRB) const {
2130 if (!Subtarget->isTargetAndroid())
2131 return TargetLowering::getSafeStackPointerLocation(IRB);
2133 // Android provides a fixed TLS slot for the SafeStack pointer. See the
2134 // definition of TLS_SLOT_SAFESTACK in
2135 // https://android.googlesource.com/platform/bionic/+/master/libc/private/bionic_tls.h
2136 unsigned AddressSpace, Offset;
2137 if (Subtarget->is64Bit()) {
2138 // %fs:0x48, unless we're using a Kernel code model, in which case it's %gs:
2140 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
2150 return ConstantExpr::getIntToPtr(
2151 ConstantInt::get(Type::getInt32Ty(IRB.getContext()), Offset),
2152 Type::getInt8PtrTy(IRB.getContext())->getPointerTo(AddressSpace));
2155 bool X86TargetLowering::isNoopAddrSpaceCast(unsigned SrcAS,
2156 unsigned DestAS) const {
2157 assert(SrcAS != DestAS && "Expected different address spaces!");
2159 return SrcAS < 256 && DestAS < 256;
2162 //===----------------------------------------------------------------------===//
2163 // Return Value Calling Convention Implementation
2164 //===----------------------------------------------------------------------===//
2166 #include "X86GenCallingConv.inc"
2168 bool X86TargetLowering::CanLowerReturn(
2169 CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg,
2170 const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const {
2171 SmallVector<CCValAssign, 16> RVLocs;
2172 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2173 return CCInfo.CheckReturn(Outs, RetCC_X86);
2176 const MCPhysReg *X86TargetLowering::getScratchRegisters(CallingConv::ID) const {
2177 static const MCPhysReg ScratchRegs[] = { X86::R11, 0 };
2182 X86TargetLowering::LowerReturn(SDValue Chain,
2183 CallingConv::ID CallConv, bool isVarArg,
2184 const SmallVectorImpl<ISD::OutputArg> &Outs,
2185 const SmallVectorImpl<SDValue> &OutVals,
2186 SDLoc dl, SelectionDAG &DAG) const {
2187 MachineFunction &MF = DAG.getMachineFunction();
2188 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2190 SmallVector<CCValAssign, 16> RVLocs;
2191 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, *DAG.getContext());
2192 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
2195 SmallVector<SDValue, 6> RetOps;
2196 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2197 // Operand #1 = Bytes To Pop
2198 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(), dl,
2201 // Copy the result values into the output registers.
2202 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2203 CCValAssign &VA = RVLocs[i];
2204 assert(VA.isRegLoc() && "Can only return in registers!");
2205 SDValue ValToCopy = OutVals[i];
2206 EVT ValVT = ValToCopy.getValueType();
2208 // Promote values to the appropriate types.
2209 if (VA.getLocInfo() == CCValAssign::SExt)
2210 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
2211 else if (VA.getLocInfo() == CCValAssign::ZExt)
2212 ValToCopy = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), ValToCopy);
2213 else if (VA.getLocInfo() == CCValAssign::AExt) {
2214 if (ValVT.isVector() && ValVT.getVectorElementType() == MVT::i1)
2215 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
2217 ValToCopy = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), ValToCopy);
2219 else if (VA.getLocInfo() == CCValAssign::BCvt)
2220 ValToCopy = DAG.getBitcast(VA.getLocVT(), ValToCopy);
2222 assert(VA.getLocInfo() != CCValAssign::FPExt &&
2223 "Unexpected FP-extend for return value.");
2225 // If this is x86-64, and we disabled SSE, we can't return FP values,
2226 // or SSE or MMX vectors.
2227 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
2228 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
2229 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
2230 report_fatal_error("SSE register return with SSE disabled");
2232 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
2233 // llvm-gcc has never done it right and no one has noticed, so this
2234 // should be OK for now.
2235 if (ValVT == MVT::f64 &&
2236 (Subtarget->is64Bit() && !Subtarget->hasSSE2()))
2237 report_fatal_error("SSE2 register return with SSE2 disabled");
2239 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
2240 // the RET instruction and handled by the FP Stackifier.
2241 if (VA.getLocReg() == X86::FP0 ||
2242 VA.getLocReg() == X86::FP1) {
2243 // If this is a copy from an xmm register to ST(0), use an FPExtend to
2244 // change the value to the FP stack register class.
2245 if (isScalarFPTypeInSSEReg(VA.getValVT()))
2246 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
2247 RetOps.push_back(ValToCopy);
2248 // Don't emit a copytoreg.
2252 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
2253 // which is returned in RAX / RDX.
2254 if (Subtarget->is64Bit()) {
2255 if (ValVT == MVT::x86mmx) {
2256 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
2257 ValToCopy = DAG.getBitcast(MVT::i64, ValToCopy);
2258 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
2260 // If we don't have SSE2 available, convert to v4f32 so the generated
2261 // register is legal.
2262 if (!Subtarget->hasSSE2())
2263 ValToCopy = DAG.getBitcast(MVT::v4f32, ValToCopy);
2268 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
2269 Flag = Chain.getValue(1);
2270 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2273 // All x86 ABIs require that for returning structs by value we copy
2274 // the sret argument into %rax/%eax (depending on ABI) for the return.
2275 // We saved the argument into a virtual register in the entry block,
2276 // so now we copy the value out and into %rax/%eax.
2278 // Checking Function.hasStructRetAttr() here is insufficient because the IR
2279 // may not have an explicit sret argument. If FuncInfo.CanLowerReturn is
2280 // false, then an sret argument may be implicitly inserted in the SelDAG. In
2281 // either case FuncInfo->setSRetReturnReg() will have been called.
2282 if (unsigned SRetReg = FuncInfo->getSRetReturnReg()) {
2283 SDValue Val = DAG.getCopyFromReg(Chain, dl, SRetReg,
2284 getPointerTy(MF.getDataLayout()));
2287 = (Subtarget->is64Bit() && !Subtarget->isTarget64BitILP32()) ?
2288 X86::RAX : X86::EAX;
2289 Chain = DAG.getCopyToReg(Chain, dl, RetValReg, Val, Flag);
2290 Flag = Chain.getValue(1);
2292 // RAX/EAX now acts like a return value.
2294 DAG.getRegister(RetValReg, getPointerTy(DAG.getDataLayout())));
2297 RetOps[0] = Chain; // Update chain.
2299 // Add the flag if we have it.
2301 RetOps.push_back(Flag);
2303 return DAG.getNode(X86ISD::RET_FLAG, dl, MVT::Other, RetOps);
2306 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2307 if (N->getNumValues() != 1)
2309 if (!N->hasNUsesOfValue(1, 0))
2312 SDValue TCChain = Chain;
2313 SDNode *Copy = *N->use_begin();
2314 if (Copy->getOpcode() == ISD::CopyToReg) {
2315 // If the copy has a glue operand, we conservatively assume it isn't safe to
2316 // perform a tail call.
2317 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2319 TCChain = Copy->getOperand(0);
2320 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
2323 bool HasRet = false;
2324 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2326 if (UI->getOpcode() != X86ISD::RET_FLAG)
2328 // If we are returning more than one value, we can definitely
2329 // not make a tail call see PR19530
2330 if (UI->getNumOperands() > 4)
2332 if (UI->getNumOperands() == 4 &&
2333 UI->getOperand(UI->getNumOperands()-1).getValueType() != MVT::Glue)
2346 X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
2347 ISD::NodeType ExtendKind) const {
2349 // TODO: Is this also valid on 32-bit?
2350 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
2351 ReturnMVT = MVT::i8;
2353 ReturnMVT = MVT::i32;
2355 EVT MinVT = getRegisterType(Context, ReturnMVT);
2356 return VT.bitsLT(MinVT) ? MinVT : VT;
2359 /// Lower the result values of a call into the
2360 /// appropriate copies out of appropriate physical registers.
2363 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2364 CallingConv::ID CallConv, bool isVarArg,
2365 const SmallVectorImpl<ISD::InputArg> &Ins,
2366 SDLoc dl, SelectionDAG &DAG,
2367 SmallVectorImpl<SDValue> &InVals) const {
2369 // Assign locations to each value returned by this call.
2370 SmallVector<CCValAssign, 16> RVLocs;
2371 bool Is64Bit = Subtarget->is64Bit();
2372 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2374 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2376 // Copy all of the result registers out of their specified physreg.
2377 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2378 CCValAssign &VA = RVLocs[i];
2379 EVT CopyVT = VA.getLocVT();
2381 // If this is x86-64, and we disabled SSE, we can't return FP values
2382 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64 || CopyVT == MVT::f128) &&
2383 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
2384 report_fatal_error("SSE register return with SSE disabled");
2387 // If we prefer to use the value in xmm registers, copy it out as f80 and
2388 // use a truncate to move it from fp stack reg to xmm reg.
2389 bool RoundAfterCopy = false;
2390 if ((VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) &&
2391 isScalarFPTypeInSSEReg(VA.getValVT())) {
2393 RoundAfterCopy = (CopyVT != VA.getLocVT());
2396 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
2397 CopyVT, InFlag).getValue(1);
2398 SDValue Val = Chain.getValue(0);
2401 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
2402 // This truncation won't change the value.
2403 DAG.getIntPtrConstant(1, dl));
2405 if (VA.isExtInLoc() && VA.getValVT().getScalarType() == MVT::i1)
2406 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
2408 InFlag = Chain.getValue(2);
2409 InVals.push_back(Val);
2415 //===----------------------------------------------------------------------===//
2416 // C & StdCall & Fast Calling Convention implementation
2417 //===----------------------------------------------------------------------===//
2418 // StdCall calling convention seems to be standard for many Windows' API
2419 // routines and around. It differs from C calling convention just a little:
2420 // callee should clean up the stack, not caller. Symbols should be also
2421 // decorated in some fancy way :) It doesn't support any vector arguments.
2422 // For info on fast calling convention see Fast Calling Convention (tail call)
2423 // implementation LowerX86_32FastCCCallTo.
2425 /// CallIsStructReturn - Determines whether a call uses struct return
2427 enum StructReturnType {
2432 static StructReturnType
2433 callIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
2435 return NotStructReturn;
2437 const ISD::ArgFlagsTy &Flags = Outs[0].Flags;
2438 if (!Flags.isSRet())
2439 return NotStructReturn;
2440 if (Flags.isInReg())
2441 return RegStructReturn;
2442 return StackStructReturn;
2445 /// Determines whether a function uses struct return semantics.
2446 static StructReturnType
2447 argsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
2449 return NotStructReturn;
2451 const ISD::ArgFlagsTy &Flags = Ins[0].Flags;
2452 if (!Flags.isSRet())
2453 return NotStructReturn;
2454 if (Flags.isInReg())
2455 return RegStructReturn;
2456 return StackStructReturn;
2459 /// Make a copy of an aggregate at address specified by "Src" to address
2460 /// "Dst" with size and alignment information specified by the specific
2461 /// parameter attribute. The copy will be passed as a byval function parameter.
2463 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2464 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2466 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), dl, MVT::i32);
2468 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2469 /*isVolatile*/false, /*AlwaysInline=*/true,
2470 /*isTailCall*/false,
2471 MachinePointerInfo(), MachinePointerInfo());
2474 /// Return true if the calling convention is one that we can guarantee TCO for.
2475 static bool canGuaranteeTCO(CallingConv::ID CC) {
2476 return (CC == CallingConv::Fast || CC == CallingConv::GHC ||
2477 CC == CallingConv::HiPE || CC == CallingConv::HHVM);
2480 /// Return true if we might ever do TCO for calls with this calling convention.
2481 static bool mayTailCallThisCC(CallingConv::ID CC) {
2483 // C calling conventions:
2484 case CallingConv::C:
2485 case CallingConv::X86_64_Win64:
2486 case CallingConv::X86_64_SysV:
2487 // Callee pop conventions:
2488 case CallingConv::X86_ThisCall:
2489 case CallingConv::X86_StdCall:
2490 case CallingConv::X86_VectorCall:
2491 case CallingConv::X86_FastCall:
2494 return canGuaranteeTCO(CC);
2498 /// Return true if the function is being made into a tailcall target by
2499 /// changing its ABI.
2500 static bool shouldGuaranteeTCO(CallingConv::ID CC, bool GuaranteedTailCallOpt) {
2501 return GuaranteedTailCallOpt && canGuaranteeTCO(CC);
2504 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2506 CI->getParent()->getParent()->getFnAttribute("disable-tail-calls");
2507 if (!CI->isTailCall() || Attr.getValueAsString() == "true")
2511 CallingConv::ID CalleeCC = CS.getCallingConv();
2512 if (!mayTailCallThisCC(CalleeCC))
2519 X86TargetLowering::LowerMemArgument(SDValue Chain,
2520 CallingConv::ID CallConv,
2521 const SmallVectorImpl<ISD::InputArg> &Ins,
2522 SDLoc dl, SelectionDAG &DAG,
2523 const CCValAssign &VA,
2524 MachineFrameInfo *MFI,
2526 // Create the nodes corresponding to a load from this parameter slot.
2527 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2528 bool AlwaysUseMutable = shouldGuaranteeTCO(
2529 CallConv, DAG.getTarget().Options.GuaranteedTailCallOpt);
2530 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
2533 // If value is passed by pointer we have address passed instead of the value
2535 bool ExtendedInMem = VA.isExtInLoc() &&
2536 VA.getValVT().getScalarType() == MVT::i1;
2538 if (VA.getLocInfo() == CCValAssign::Indirect || ExtendedInMem)
2539 ValVT = VA.getLocVT();
2541 ValVT = VA.getValVT();
2543 // FIXME: For now, all byval parameter objects are marked mutable. This can be
2544 // changed with more analysis.
2545 // In case of tail call optimization mark all arguments mutable. Since they
2546 // could be overwritten by lowering of arguments in case of a tail call.
2547 if (Flags.isByVal()) {
2548 unsigned Bytes = Flags.getByValSize();
2549 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2550 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
2551 return DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2553 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2554 VA.getLocMemOffset(), isImmutable);
2555 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2556 SDValue Val = DAG.getLoad(
2557 ValVT, dl, Chain, FIN,
2558 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI), false,
2560 return ExtendedInMem ?
2561 DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val) : Val;
2565 // FIXME: Get this from tablegen.
2566 static ArrayRef<MCPhysReg> get64BitArgumentGPRs(CallingConv::ID CallConv,
2567 const X86Subtarget *Subtarget) {
2568 assert(Subtarget->is64Bit());
2570 if (Subtarget->isCallingConvWin64(CallConv)) {
2571 static const MCPhysReg GPR64ArgRegsWin64[] = {
2572 X86::RCX, X86::RDX, X86::R8, X86::R9
2574 return makeArrayRef(std::begin(GPR64ArgRegsWin64), std::end(GPR64ArgRegsWin64));
2577 static const MCPhysReg GPR64ArgRegs64Bit[] = {
2578 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
2580 return makeArrayRef(std::begin(GPR64ArgRegs64Bit), std::end(GPR64ArgRegs64Bit));
2583 // FIXME: Get this from tablegen.
2584 static ArrayRef<MCPhysReg> get64BitArgumentXMMs(MachineFunction &MF,
2585 CallingConv::ID CallConv,
2586 const X86Subtarget *Subtarget) {
2587 assert(Subtarget->is64Bit());
2588 if (Subtarget->isCallingConvWin64(CallConv)) {
2589 // The XMM registers which might contain var arg parameters are shadowed
2590 // in their paired GPR. So we only need to save the GPR to their home
2592 // TODO: __vectorcall will change this.
2596 const Function *Fn = MF.getFunction();
2597 bool NoImplicitFloatOps = Fn->hasFnAttribute(Attribute::NoImplicitFloat);
2598 bool isSoftFloat = Subtarget->useSoftFloat();
2599 assert(!(isSoftFloat && NoImplicitFloatOps) &&
2600 "SSE register cannot be used when SSE is disabled!");
2601 if (isSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
2602 // Kernel mode asks for SSE to be disabled, so there are no XMM argument
2606 static const MCPhysReg XMMArgRegs64Bit[] = {
2607 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2608 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2610 return makeArrayRef(std::begin(XMMArgRegs64Bit), std::end(XMMArgRegs64Bit));
2613 SDValue X86TargetLowering::LowerFormalArguments(
2614 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
2615 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc dl, SelectionDAG &DAG,
2616 SmallVectorImpl<SDValue> &InVals) const {
2617 MachineFunction &MF = DAG.getMachineFunction();
2618 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2619 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
2621 const Function* Fn = MF.getFunction();
2622 if (Fn->hasExternalLinkage() &&
2623 Subtarget->isTargetCygMing() &&
2624 Fn->getName() == "main")
2625 FuncInfo->setForceFramePointer(true);
2627 MachineFrameInfo *MFI = MF.getFrameInfo();
2628 bool Is64Bit = Subtarget->is64Bit();
2629 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2631 assert(!(isVarArg && canGuaranteeTCO(CallConv)) &&
2632 "Var args not supported with calling convention fastcc, ghc or hipe");
2634 // Assign locations to all of the incoming arguments.
2635 SmallVector<CCValAssign, 16> ArgLocs;
2636 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2638 // Allocate shadow area for Win64
2640 CCInfo.AllocateStack(32, 8);
2642 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
2644 unsigned LastVal = ~0U;
2646 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2647 CCValAssign &VA = ArgLocs[i];
2648 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
2650 assert(VA.getValNo() != LastVal &&
2651 "Don't support value assigned to multiple locs yet");
2653 LastVal = VA.getValNo();
2655 if (VA.isRegLoc()) {
2656 EVT RegVT = VA.getLocVT();
2657 const TargetRegisterClass *RC;
2658 if (RegVT == MVT::i32)
2659 RC = &X86::GR32RegClass;
2660 else if (Is64Bit && RegVT == MVT::i64)
2661 RC = &X86::GR64RegClass;
2662 else if (RegVT == MVT::f32)
2663 RC = &X86::FR32RegClass;
2664 else if (RegVT == MVT::f64)
2665 RC = &X86::FR64RegClass;
2666 else if (RegVT == MVT::f128)
2667 RC = &X86::FR128RegClass;
2668 else if (RegVT.is512BitVector())
2669 RC = &X86::VR512RegClass;
2670 else if (RegVT.is256BitVector())
2671 RC = &X86::VR256RegClass;
2672 else if (RegVT.is128BitVector())
2673 RC = &X86::VR128RegClass;
2674 else if (RegVT == MVT::x86mmx)
2675 RC = &X86::VR64RegClass;
2676 else if (RegVT == MVT::i1)
2677 RC = &X86::VK1RegClass;
2678 else if (RegVT == MVT::v8i1)
2679 RC = &X86::VK8RegClass;
2680 else if (RegVT == MVT::v16i1)
2681 RC = &X86::VK16RegClass;
2682 else if (RegVT == MVT::v32i1)
2683 RC = &X86::VK32RegClass;
2684 else if (RegVT == MVT::v64i1)
2685 RC = &X86::VK64RegClass;
2687 llvm_unreachable("Unknown argument type!");
2689 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2690 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2692 // If this is an 8 or 16-bit value, it is really passed promoted to 32
2693 // bits. Insert an assert[sz]ext to capture this, then truncate to the
2695 if (VA.getLocInfo() == CCValAssign::SExt)
2696 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2697 DAG.getValueType(VA.getValVT()));
2698 else if (VA.getLocInfo() == CCValAssign::ZExt)
2699 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2700 DAG.getValueType(VA.getValVT()));
2701 else if (VA.getLocInfo() == CCValAssign::BCvt)
2702 ArgValue = DAG.getBitcast(VA.getValVT(), ArgValue);
2704 if (VA.isExtInLoc()) {
2705 // Handle MMX values passed in XMM regs.
2706 if (RegVT.isVector() && VA.getValVT().getScalarType() != MVT::i1)
2707 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(), ArgValue);
2709 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2712 assert(VA.isMemLoc());
2713 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
2716 // If value is passed via pointer - do a load.
2717 if (VA.getLocInfo() == CCValAssign::Indirect)
2718 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
2719 MachinePointerInfo(), false, false, false, 0);
2721 InVals.push_back(ArgValue);
2724 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2725 // All x86 ABIs require that for returning structs by value we copy the
2726 // sret argument into %rax/%eax (depending on ABI) for the return. Save
2727 // the argument into a virtual register so that we can access it from the
2729 if (Ins[i].Flags.isSRet()) {
2730 unsigned Reg = FuncInfo->getSRetReturnReg();
2732 MVT PtrTy = getPointerTy(DAG.getDataLayout());
2733 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrTy));
2734 FuncInfo->setSRetReturnReg(Reg);
2736 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[i]);
2737 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
2742 unsigned StackSize = CCInfo.getNextStackOffset();
2743 // Align stack specially for tail calls.
2744 if (shouldGuaranteeTCO(CallConv,
2745 MF.getTarget().Options.GuaranteedTailCallOpt))
2746 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
2748 // If the function takes variable number of arguments, make a frame index for
2749 // the start of the first vararg value... for expansion of llvm.va_start. We
2750 // can skip this if there are no va_start calls.
2751 if (MFI->hasVAStart() &&
2752 (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
2753 CallConv != CallingConv::X86_ThisCall))) {
2754 FuncInfo->setVarArgsFrameIndex(
2755 MFI->CreateFixedObject(1, StackSize, true));
2758 // Figure out if XMM registers are in use.
2759 assert(!(Subtarget->useSoftFloat() &&
2760 Fn->hasFnAttribute(Attribute::NoImplicitFloat)) &&
2761 "SSE register cannot be used when SSE is disabled!");
2763 // 64-bit calling conventions support varargs and register parameters, so we
2764 // have to do extra work to spill them in the prologue.
2765 if (Is64Bit && isVarArg && MFI->hasVAStart()) {
2766 // Find the first unallocated argument registers.
2767 ArrayRef<MCPhysReg> ArgGPRs = get64BitArgumentGPRs(CallConv, Subtarget);
2768 ArrayRef<MCPhysReg> ArgXMMs = get64BitArgumentXMMs(MF, CallConv, Subtarget);
2769 unsigned NumIntRegs = CCInfo.getFirstUnallocated(ArgGPRs);
2770 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(ArgXMMs);
2771 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
2772 "SSE register cannot be used when SSE is disabled!");
2774 // Gather all the live in physical registers.
2775 SmallVector<SDValue, 6> LiveGPRs;
2776 SmallVector<SDValue, 8> LiveXMMRegs;
2778 for (MCPhysReg Reg : ArgGPRs.slice(NumIntRegs)) {
2779 unsigned GPR = MF.addLiveIn(Reg, &X86::GR64RegClass);
2781 DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64));
2783 if (!ArgXMMs.empty()) {
2784 unsigned AL = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2785 ALVal = DAG.getCopyFromReg(Chain, dl, AL, MVT::i8);
2786 for (MCPhysReg Reg : ArgXMMs.slice(NumXMMRegs)) {
2787 unsigned XMMReg = MF.addLiveIn(Reg, &X86::VR128RegClass);
2788 LiveXMMRegs.push_back(
2789 DAG.getCopyFromReg(Chain, dl, XMMReg, MVT::v4f32));
2794 // Get to the caller-allocated home save location. Add 8 to account
2795 // for the return address.
2796 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2797 FuncInfo->setRegSaveFrameIndex(
2798 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
2799 // Fixup to set vararg frame on shadow area (4 x i64).
2801 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
2803 // For X86-64, if there are vararg parameters that are passed via
2804 // registers, then we must store them to their spots on the stack so
2805 // they may be loaded by deferencing the result of va_next.
2806 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
2807 FuncInfo->setVarArgsFPOffset(ArgGPRs.size() * 8 + NumXMMRegs * 16);
2808 FuncInfo->setRegSaveFrameIndex(MFI->CreateStackObject(
2809 ArgGPRs.size() * 8 + ArgXMMs.size() * 16, 16, false));
2812 // Store the integer parameter registers.
2813 SmallVector<SDValue, 8> MemOps;
2814 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
2815 getPointerTy(DAG.getDataLayout()));
2816 unsigned Offset = FuncInfo->getVarArgsGPOffset();
2817 for (SDValue Val : LiveGPRs) {
2818 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
2819 RSFIN, DAG.getIntPtrConstant(Offset, dl));
2821 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2822 MachinePointerInfo::getFixedStack(
2823 DAG.getMachineFunction(),
2824 FuncInfo->getRegSaveFrameIndex(), Offset),
2826 MemOps.push_back(Store);
2830 if (!ArgXMMs.empty() && NumXMMRegs != ArgXMMs.size()) {
2831 // Now store the XMM (fp + vector) parameter registers.
2832 SmallVector<SDValue, 12> SaveXMMOps;
2833 SaveXMMOps.push_back(Chain);
2834 SaveXMMOps.push_back(ALVal);
2835 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2836 FuncInfo->getRegSaveFrameIndex(), dl));
2837 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2838 FuncInfo->getVarArgsFPOffset(), dl));
2839 SaveXMMOps.insert(SaveXMMOps.end(), LiveXMMRegs.begin(),
2841 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
2842 MVT::Other, SaveXMMOps));
2845 if (!MemOps.empty())
2846 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2849 if (isVarArg && MFI->hasMustTailInVarArgFunc()) {
2850 // Find the largest legal vector type.
2851 MVT VecVT = MVT::Other;
2852 // FIXME: Only some x86_32 calling conventions support AVX512.
2853 if (Subtarget->hasAVX512() &&
2854 (Is64Bit || (CallConv == CallingConv::X86_VectorCall ||
2855 CallConv == CallingConv::Intel_OCL_BI)))
2856 VecVT = MVT::v16f32;
2857 else if (Subtarget->hasAVX())
2859 else if (Subtarget->hasSSE2())
2862 // We forward some GPRs and some vector types.
2863 SmallVector<MVT, 2> RegParmTypes;
2864 MVT IntVT = Is64Bit ? MVT::i64 : MVT::i32;
2865 RegParmTypes.push_back(IntVT);
2866 if (VecVT != MVT::Other)
2867 RegParmTypes.push_back(VecVT);
2869 // Compute the set of forwarded registers. The rest are scratch.
2870 SmallVectorImpl<ForwardedRegister> &Forwards =
2871 FuncInfo->getForwardedMustTailRegParms();
2872 CCInfo.analyzeMustTailForwardedRegisters(Forwards, RegParmTypes, CC_X86);
2874 // Conservatively forward AL on x86_64, since it might be used for varargs.
2875 if (Is64Bit && !CCInfo.isAllocated(X86::AL)) {
2876 unsigned ALVReg = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2877 Forwards.push_back(ForwardedRegister(ALVReg, X86::AL, MVT::i8));
2880 // Copy all forwards from physical to virtual registers.
2881 for (ForwardedRegister &F : Forwards) {
2882 // FIXME: Can we use a less constrained schedule?
2883 SDValue RegVal = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
2884 F.VReg = MF.getRegInfo().createVirtualRegister(getRegClassFor(F.VT));
2885 Chain = DAG.getCopyToReg(Chain, dl, F.VReg, RegVal);
2889 // Some CCs need callee pop.
2890 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
2891 MF.getTarget().Options.GuaranteedTailCallOpt)) {
2892 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
2894 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
2895 // If this is an sret function, the return should pop the hidden pointer.
2896 if (!Is64Bit && !canGuaranteeTCO(CallConv) &&
2897 !Subtarget->getTargetTriple().isOSMSVCRT() &&
2898 argsAreStructReturn(Ins) == StackStructReturn)
2899 FuncInfo->setBytesToPopOnReturn(4);
2903 // RegSaveFrameIndex is X86-64 only.
2904 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
2905 if (CallConv == CallingConv::X86_FastCall ||
2906 CallConv == CallingConv::X86_ThisCall)
2907 // fastcc functions can't have varargs.
2908 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
2911 FuncInfo->setArgumentStackSize(StackSize);
2913 if (WinEHFuncInfo *EHInfo = MF.getWinEHFuncInfo()) {
2914 EHPersonality Personality = classifyEHPersonality(Fn->getPersonalityFn());
2915 if (Personality == EHPersonality::CoreCLR) {
2917 // TODO: Add a mechanism to frame lowering that will allow us to indicate
2918 // that we'd prefer this slot be allocated towards the bottom of the frame
2919 // (i.e. near the stack pointer after allocating the frame). Every
2920 // funclet needs a copy of this slot in its (mostly empty) frame, and the
2921 // offset from the bottom of this and each funclet's frame must be the
2922 // same, so the size of funclets' (mostly empty) frames is dictated by
2923 // how far this slot is from the bottom (since they allocate just enough
2924 // space to accomodate holding this slot at the correct offset).
2925 int PSPSymFI = MFI->CreateStackObject(8, 8, /*isSS=*/false);
2926 EHInfo->PSPSymFrameIdx = PSPSymFI;
2934 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
2935 SDValue StackPtr, SDValue Arg,
2936 SDLoc dl, SelectionDAG &DAG,
2937 const CCValAssign &VA,
2938 ISD::ArgFlagsTy Flags) const {
2939 unsigned LocMemOffset = VA.getLocMemOffset();
2940 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
2941 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
2943 if (Flags.isByVal())
2944 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
2946 return DAG.getStore(
2947 Chain, dl, Arg, PtrOff,
2948 MachinePointerInfo::getStack(DAG.getMachineFunction(), LocMemOffset),
2952 /// Emit a load of return address if tail call
2953 /// optimization is performed and it is required.
2955 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
2956 SDValue &OutRetAddr, SDValue Chain,
2957 bool IsTailCall, bool Is64Bit,
2958 int FPDiff, SDLoc dl) const {
2959 // Adjust the Return address stack slot.
2960 EVT VT = getPointerTy(DAG.getDataLayout());
2961 OutRetAddr = getReturnAddressFrameIndex(DAG);
2963 // Load the "old" Return address.
2964 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
2965 false, false, false, 0);
2966 return SDValue(OutRetAddr.getNode(), 1);
2969 /// Emit a store of the return address if tail call
2970 /// optimization is performed and it is required (FPDiff!=0).
2971 static SDValue EmitTailCallStoreRetAddr(SelectionDAG &DAG, MachineFunction &MF,
2972 SDValue Chain, SDValue RetAddrFrIdx,
2973 EVT PtrVT, unsigned SlotSize,
2974 int FPDiff, SDLoc dl) {
2975 // Store the return address to the appropriate stack slot.
2976 if (!FPDiff) return Chain;
2977 // Calculate the new stack slot for the return address.
2978 int NewReturnAddrFI =
2979 MF.getFrameInfo()->CreateFixedObject(SlotSize, (int64_t)FPDiff - SlotSize,
2981 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, PtrVT);
2982 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
2983 MachinePointerInfo::getFixedStack(
2984 DAG.getMachineFunction(), NewReturnAddrFI),
2989 /// Returns a vector_shuffle mask for an movs{s|d}, movd
2990 /// operation of specified width.
2991 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
2993 unsigned NumElems = VT.getVectorNumElements();
2994 SmallVector<int, 8> Mask;
2995 Mask.push_back(NumElems);
2996 for (unsigned i = 1; i != NumElems; ++i)
2998 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3002 X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
3003 SmallVectorImpl<SDValue> &InVals) const {
3004 SelectionDAG &DAG = CLI.DAG;
3006 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
3007 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
3008 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
3009 SDValue Chain = CLI.Chain;
3010 SDValue Callee = CLI.Callee;
3011 CallingConv::ID CallConv = CLI.CallConv;
3012 bool &isTailCall = CLI.IsTailCall;
3013 bool isVarArg = CLI.IsVarArg;
3015 MachineFunction &MF = DAG.getMachineFunction();
3016 bool Is64Bit = Subtarget->is64Bit();
3017 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
3018 StructReturnType SR = callIsStructReturn(Outs);
3019 bool IsSibcall = false;
3020 X86MachineFunctionInfo *X86Info = MF.getInfo<X86MachineFunctionInfo>();
3021 auto Attr = MF.getFunction()->getFnAttribute("disable-tail-calls");
3023 if (Attr.getValueAsString() == "true")
3026 if (Subtarget->isPICStyleGOT() &&
3027 !MF.getTarget().Options.GuaranteedTailCallOpt) {
3028 // If we are using a GOT, disable tail calls to external symbols with
3029 // default visibility. Tail calling such a symbol requires using a GOT
3030 // relocation, which forces early binding of the symbol. This breaks code
3031 // that require lazy function symbol resolution. Using musttail or
3032 // GuaranteedTailCallOpt will override this.
3033 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
3034 if (!G || (!G->getGlobal()->hasLocalLinkage() &&
3035 G->getGlobal()->hasDefaultVisibility()))
3039 bool IsMustTail = CLI.CS && CLI.CS->isMustTailCall();
3041 // Force this to be a tail call. The verifier rules are enough to ensure
3042 // that we can lower this successfully without moving the return address
3045 } else if (isTailCall) {
3046 // Check if it's really possible to do a tail call.
3047 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
3048 isVarArg, SR != NotStructReturn,
3049 MF.getFunction()->hasStructRetAttr(), CLI.RetTy,
3050 Outs, OutVals, Ins, DAG);
3052 // Sibcalls are automatically detected tailcalls which do not require
3054 if (!MF.getTarget().Options.GuaranteedTailCallOpt && isTailCall)
3061 assert(!(isVarArg && canGuaranteeTCO(CallConv)) &&
3062 "Var args not supported with calling convention fastcc, ghc or hipe");
3064 // Analyze operands of the call, assigning locations to each operand.
3065 SmallVector<CCValAssign, 16> ArgLocs;
3066 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
3068 // Allocate shadow area for Win64
3070 CCInfo.AllocateStack(32, 8);
3072 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3074 // Get a count of how many bytes are to be pushed on the stack.
3075 unsigned NumBytes = CCInfo.getAlignedCallFrameSize();
3077 // This is a sibcall. The memory operands are available in caller's
3078 // own caller's stack.
3080 else if (MF.getTarget().Options.GuaranteedTailCallOpt &&
3081 canGuaranteeTCO(CallConv))
3082 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
3085 if (isTailCall && !IsSibcall && !IsMustTail) {
3086 // Lower arguments at fp - stackoffset + fpdiff.
3087 unsigned NumBytesCallerPushed = X86Info->getBytesToPopOnReturn();
3089 FPDiff = NumBytesCallerPushed - NumBytes;
3091 // Set the delta of movement of the returnaddr stackslot.
3092 // But only set if delta is greater than previous delta.
3093 if (FPDiff < X86Info->getTCReturnAddrDelta())
3094 X86Info->setTCReturnAddrDelta(FPDiff);
3097 unsigned NumBytesToPush = NumBytes;
3098 unsigned NumBytesToPop = NumBytes;
3100 // If we have an inalloca argument, all stack space has already been allocated
3101 // for us and be right at the top of the stack. We don't support multiple
3102 // arguments passed in memory when using inalloca.
3103 if (!Outs.empty() && Outs.back().Flags.isInAlloca()) {
3105 if (!ArgLocs.back().isMemLoc())
3106 report_fatal_error("cannot use inalloca attribute on a register "
3108 if (ArgLocs.back().getLocMemOffset() != 0)
3109 report_fatal_error("any parameter with the inalloca attribute must be "
3110 "the only memory argument");
3114 Chain = DAG.getCALLSEQ_START(
3115 Chain, DAG.getIntPtrConstant(NumBytesToPush, dl, true), dl);
3117 SDValue RetAddrFrIdx;
3118 // Load return address for tail calls.
3119 if (isTailCall && FPDiff)
3120 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
3121 Is64Bit, FPDiff, dl);
3123 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3124 SmallVector<SDValue, 8> MemOpChains;
3127 // Walk the register/memloc assignments, inserting copies/loads. In the case
3128 // of tail call optimization arguments are handle later.
3129 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3130 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3131 // Skip inalloca arguments, they have already been written.
3132 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3133 if (Flags.isInAlloca())
3136 CCValAssign &VA = ArgLocs[i];
3137 EVT RegVT = VA.getLocVT();
3138 SDValue Arg = OutVals[i];
3139 bool isByVal = Flags.isByVal();
3141 // Promote the value if needed.
3142 switch (VA.getLocInfo()) {
3143 default: llvm_unreachable("Unknown loc info!");
3144 case CCValAssign::Full: break;
3145 case CCValAssign::SExt:
3146 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
3148 case CCValAssign::ZExt:
3149 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
3151 case CCValAssign::AExt:
3152 if (Arg.getValueType().isVector() &&
3153 Arg.getValueType().getVectorElementType() == MVT::i1)
3154 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
3155 else if (RegVT.is128BitVector()) {
3156 // Special case: passing MMX values in XMM registers.
3157 Arg = DAG.getBitcast(MVT::i64, Arg);
3158 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
3159 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
3161 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
3163 case CCValAssign::BCvt:
3164 Arg = DAG.getBitcast(RegVT, Arg);
3166 case CCValAssign::Indirect: {
3167 // Store the argument.
3168 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
3169 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
3170 Chain = DAG.getStore(
3171 Chain, dl, Arg, SpillSlot,
3172 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3179 if (VA.isRegLoc()) {
3180 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3181 if (isVarArg && IsWin64) {
3182 // Win64 ABI requires argument XMM reg to be copied to the corresponding
3183 // shadow reg if callee is a varargs function.
3184 unsigned ShadowReg = 0;
3185 switch (VA.getLocReg()) {
3186 case X86::XMM0: ShadowReg = X86::RCX; break;
3187 case X86::XMM1: ShadowReg = X86::RDX; break;
3188 case X86::XMM2: ShadowReg = X86::R8; break;
3189 case X86::XMM3: ShadowReg = X86::R9; break;
3192 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
3194 } else if (!IsSibcall && (!isTailCall || isByVal)) {
3195 assert(VA.isMemLoc());
3196 if (!StackPtr.getNode())
3197 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
3198 getPointerTy(DAG.getDataLayout()));
3199 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
3200 dl, DAG, VA, Flags));
3204 if (!MemOpChains.empty())
3205 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
3207 if (Subtarget->isPICStyleGOT()) {
3208 // ELF / PIC requires GOT in the EBX register before function calls via PLT
3211 RegsToPass.push_back(std::make_pair(
3212 unsigned(X86::EBX), DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(),
3213 getPointerTy(DAG.getDataLayout()))));
3215 // If we are tail calling and generating PIC/GOT style code load the
3216 // address of the callee into ECX. The value in ecx is used as target of
3217 // the tail jump. This is done to circumvent the ebx/callee-saved problem
3218 // for tail calls on PIC/GOT architectures. Normally we would just put the
3219 // address of GOT into ebx and then call target@PLT. But for tail calls
3220 // ebx would be restored (since ebx is callee saved) before jumping to the
3223 // Note: The actual moving to ECX is done further down.
3224 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
3225 if (G && !G->getGlobal()->hasLocalLinkage() &&
3226 G->getGlobal()->hasDefaultVisibility())
3227 Callee = LowerGlobalAddress(Callee, DAG);
3228 else if (isa<ExternalSymbolSDNode>(Callee))
3229 Callee = LowerExternalSymbol(Callee, DAG);
3233 if (Is64Bit && isVarArg && !IsWin64 && !IsMustTail) {
3234 // From AMD64 ABI document:
3235 // For calls that may call functions that use varargs or stdargs
3236 // (prototype-less calls or calls to functions containing ellipsis (...) in
3237 // the declaration) %al is used as hidden argument to specify the number
3238 // of SSE registers used. The contents of %al do not need to match exactly
3239 // the number of registers, but must be an ubound on the number of SSE
3240 // registers used and is in the range 0 - 8 inclusive.
3242 // Count the number of XMM registers allocated.
3243 static const MCPhysReg XMMArgRegs[] = {
3244 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
3245 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
3247 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs);
3248 assert((Subtarget->hasSSE1() || !NumXMMRegs)
3249 && "SSE registers cannot be used when SSE is disabled");
3251 RegsToPass.push_back(std::make_pair(unsigned(X86::AL),
3252 DAG.getConstant(NumXMMRegs, dl,
3256 if (isVarArg && IsMustTail) {
3257 const auto &Forwards = X86Info->getForwardedMustTailRegParms();
3258 for (const auto &F : Forwards) {
3259 SDValue Val = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
3260 RegsToPass.push_back(std::make_pair(unsigned(F.PReg), Val));
3264 // For tail calls lower the arguments to the 'real' stack slots. Sibcalls
3265 // don't need this because the eligibility check rejects calls that require
3266 // shuffling arguments passed in memory.
3267 if (!IsSibcall && isTailCall) {
3268 // Force all the incoming stack arguments to be loaded from the stack
3269 // before any new outgoing arguments are stored to the stack, because the
3270 // outgoing stack slots may alias the incoming argument stack slots, and
3271 // the alias isn't otherwise explicit. This is slightly more conservative
3272 // than necessary, because it means that each store effectively depends
3273 // on every argument instead of just those arguments it would clobber.
3274 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
3276 SmallVector<SDValue, 8> MemOpChains2;
3279 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3280 CCValAssign &VA = ArgLocs[i];
3283 assert(VA.isMemLoc());
3284 SDValue Arg = OutVals[i];
3285 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3286 // Skip inalloca arguments. They don't require any work.
3287 if (Flags.isInAlloca())
3289 // Create frame index.
3290 int32_t Offset = VA.getLocMemOffset()+FPDiff;
3291 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
3292 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
3293 FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
3295 if (Flags.isByVal()) {
3296 // Copy relative to framepointer.
3297 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset(), dl);
3298 if (!StackPtr.getNode())
3299 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
3300 getPointerTy(DAG.getDataLayout()));
3301 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
3304 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
3308 // Store relative to framepointer.
3309 MemOpChains2.push_back(DAG.getStore(
3310 ArgChain, dl, Arg, FIN,
3311 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3316 if (!MemOpChains2.empty())
3317 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains2);
3319 // Store the return address to the appropriate stack slot.
3320 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx,
3321 getPointerTy(DAG.getDataLayout()),
3322 RegInfo->getSlotSize(), FPDiff, dl);
3325 // Build a sequence of copy-to-reg nodes chained together with token chain
3326 // and flag operands which copy the outgoing args into registers.
3328 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3329 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3330 RegsToPass[i].second, InFlag);
3331 InFlag = Chain.getValue(1);
3334 if (DAG.getTarget().getCodeModel() == CodeModel::Large) {
3335 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
3336 // In the 64-bit large code model, we have to make all calls
3337 // through a register, since the call instruction's 32-bit
3338 // pc-relative offset may not be large enough to hold the whole
3340 } else if (Callee->getOpcode() == ISD::GlobalAddress) {
3341 // If the callee is a GlobalAddress node (quite common, every direct call
3342 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
3344 GlobalAddressSDNode* G = cast<GlobalAddressSDNode>(Callee);
3346 // We should use extra load for direct calls to dllimported functions in
3348 const GlobalValue *GV = G->getGlobal();
3349 if (!GV->hasDLLImportStorageClass()) {
3350 unsigned char OpFlags = 0;
3351 bool ExtraLoad = false;
3352 unsigned WrapperKind = ISD::DELETED_NODE;
3354 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
3355 // external symbols most go through the PLT in PIC mode. If the symbol
3356 // has hidden or protected visibility, or if it is static or local, then
3357 // we don't need to use the PLT - we can directly call it.
3358 if (Subtarget->isTargetELF() &&
3359 DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
3360 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
3361 OpFlags = X86II::MO_PLT;
3362 } else if (Subtarget->isPICStyleStubAny() &&
3363 !GV->isStrongDefinitionForLinker() &&
3364 (!Subtarget->getTargetTriple().isMacOSX() ||
3365 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3366 // PC-relative references to external symbols should go through $stub,
3367 // unless we're building with the leopard linker or later, which
3368 // automatically synthesizes these stubs.
3369 OpFlags = X86II::MO_DARWIN_STUB;
3370 } else if (Subtarget->isPICStyleRIPRel() && isa<Function>(GV) &&
3371 cast<Function>(GV)->hasFnAttribute(Attribute::NonLazyBind)) {
3372 // If the function is marked as non-lazy, generate an indirect call
3373 // which loads from the GOT directly. This avoids runtime overhead
3374 // at the cost of eager binding (and one extra byte of encoding).
3375 OpFlags = X86II::MO_GOTPCREL;
3376 WrapperKind = X86ISD::WrapperRIP;
3380 Callee = DAG.getTargetGlobalAddress(
3381 GV, dl, getPointerTy(DAG.getDataLayout()), G->getOffset(), OpFlags);
3383 // Add a wrapper if needed.
3384 if (WrapperKind != ISD::DELETED_NODE)
3385 Callee = DAG.getNode(X86ISD::WrapperRIP, dl,
3386 getPointerTy(DAG.getDataLayout()), Callee);
3387 // Add extra indirection if needed.
3389 Callee = DAG.getLoad(
3390 getPointerTy(DAG.getDataLayout()), dl, DAG.getEntryNode(), Callee,
3391 MachinePointerInfo::getGOT(DAG.getMachineFunction()), false, false,
3394 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3395 unsigned char OpFlags = 0;
3397 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
3398 // external symbols should go through the PLT.
3399 if (Subtarget->isTargetELF() &&
3400 DAG.getTarget().getRelocationModel() == Reloc::PIC_) {
3401 OpFlags = X86II::MO_PLT;
3402 } else if (Subtarget->isPICStyleStubAny() &&
3403 (!Subtarget->getTargetTriple().isMacOSX() ||
3404 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3405 // PC-relative references to external symbols should go through $stub,
3406 // unless we're building with the leopard linker or later, which
3407 // automatically synthesizes these stubs.
3408 OpFlags = X86II::MO_DARWIN_STUB;
3411 Callee = DAG.getTargetExternalSymbol(
3412 S->getSymbol(), getPointerTy(DAG.getDataLayout()), OpFlags);
3413 } else if (Subtarget->isTarget64BitILP32() &&
3414 Callee->getValueType(0) == MVT::i32) {
3415 // Zero-extend the 32-bit Callee address into a 64-bit according to x32 ABI
3416 Callee = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, Callee);
3419 // Returns a chain & a flag for retval copy to use.
3420 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3421 SmallVector<SDValue, 8> Ops;
3423 if (!IsSibcall && isTailCall) {
3424 Chain = DAG.getCALLSEQ_END(Chain,
3425 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3426 DAG.getIntPtrConstant(0, dl, true), InFlag, dl);
3427 InFlag = Chain.getValue(1);
3430 Ops.push_back(Chain);
3431 Ops.push_back(Callee);
3434 Ops.push_back(DAG.getConstant(FPDiff, dl, MVT::i32));
3436 // Add argument registers to the end of the list so that they are known live
3438 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3439 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3440 RegsToPass[i].second.getValueType()));
3442 // Add a register mask operand representing the call-preserved registers.
3443 const uint32_t *Mask = RegInfo->getCallPreservedMask(MF, CallConv);
3444 assert(Mask && "Missing call preserved mask for calling convention");
3446 // If this is an invoke in a 32-bit function using a funclet-based
3447 // personality, assume the function clobbers all registers. If an exception
3448 // is thrown, the runtime will not restore CSRs.
3449 // FIXME: Model this more precisely so that we can register allocate across
3450 // the normal edge and spill and fill across the exceptional edge.
3451 if (!Is64Bit && CLI.CS && CLI.CS->isInvoke()) {
3452 const Function *CallerFn = MF.getFunction();
3453 EHPersonality Pers =
3454 CallerFn->hasPersonalityFn()
3455 ? classifyEHPersonality(CallerFn->getPersonalityFn())
3456 : EHPersonality::Unknown;
3457 if (isFuncletEHPersonality(Pers))
3458 Mask = RegInfo->getNoPreservedMask();
3461 Ops.push_back(DAG.getRegisterMask(Mask));
3463 if (InFlag.getNode())
3464 Ops.push_back(InFlag);
3468 //// If this is the first return lowered for this function, add the regs
3469 //// to the liveout set for the function.
3470 // This isn't right, although it's probably harmless on x86; liveouts
3471 // should be computed from returns not tail calls. Consider a void
3472 // function making a tail call to a function returning int.
3473 MF.getFrameInfo()->setHasTailCall();
3474 return DAG.getNode(X86ISD::TC_RETURN, dl, NodeTys, Ops);
3477 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops);
3478 InFlag = Chain.getValue(1);
3480 // Create the CALLSEQ_END node.
3481 unsigned NumBytesForCalleeToPop;
3482 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
3483 DAG.getTarget().Options.GuaranteedTailCallOpt))
3484 NumBytesForCalleeToPop = NumBytes; // Callee pops everything
3485 else if (!Is64Bit && !canGuaranteeTCO(CallConv) &&
3486 !Subtarget->getTargetTriple().isOSMSVCRT() &&
3487 SR == StackStructReturn)
3488 // If this is a call to a struct-return function, the callee
3489 // pops the hidden struct pointer, so we have to push it back.
3490 // This is common for Darwin/X86, Linux & Mingw32 targets.
3491 // For MSVC Win32 targets, the caller pops the hidden struct pointer.
3492 NumBytesForCalleeToPop = 4;
3494 NumBytesForCalleeToPop = 0; // Callee pops nothing.
3496 // Returns a flag for retval copy to use.
3498 Chain = DAG.getCALLSEQ_END(Chain,
3499 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3500 DAG.getIntPtrConstant(NumBytesForCalleeToPop, dl,
3503 InFlag = Chain.getValue(1);
3506 // Handle result values, copying them out of physregs into vregs that we
3508 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3509 Ins, dl, DAG, InVals);
3512 //===----------------------------------------------------------------------===//
3513 // Fast Calling Convention (tail call) implementation
3514 //===----------------------------------------------------------------------===//
3516 // Like std call, callee cleans arguments, convention except that ECX is
3517 // reserved for storing the tail called function address. Only 2 registers are
3518 // free for argument passing (inreg). Tail call optimization is performed
3520 // * tailcallopt is enabled
3521 // * caller/callee are fastcc
3522 // On X86_64 architecture with GOT-style position independent code only local
3523 // (within module) calls are supported at the moment.
3524 // To keep the stack aligned according to platform abi the function
3525 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
3526 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
3527 // If a tail called function callee has more arguments than the caller the
3528 // caller needs to make sure that there is room to move the RETADDR to. This is
3529 // achieved by reserving an area the size of the argument delta right after the
3530 // original RETADDR, but before the saved framepointer or the spilled registers
3531 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
3543 /// Make the stack size align e.g 16n + 12 aligned for a 16-byte align
3546 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
3547 SelectionDAG& DAG) const {
3548 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3549 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
3550 unsigned StackAlignment = TFI.getStackAlignment();
3551 uint64_t AlignMask = StackAlignment - 1;
3552 int64_t Offset = StackSize;
3553 unsigned SlotSize = RegInfo->getSlotSize();
3554 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
3555 // Number smaller than 12 so just add the difference.
3556 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
3558 // Mask out lower bits, add stackalignment once plus the 12 bytes.
3559 Offset = ((~AlignMask) & Offset) + StackAlignment +
3560 (StackAlignment-SlotSize);
3565 /// Return true if the given stack call argument is already available in the
3566 /// same position (relatively) of the caller's incoming argument stack.
3568 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
3569 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
3570 const X86InstrInfo *TII) {
3571 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
3573 if (Arg.getOpcode() == ISD::CopyFromReg) {
3574 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
3575 if (!TargetRegisterInfo::isVirtualRegister(VR))
3577 MachineInstr *Def = MRI->getVRegDef(VR);
3580 if (!Flags.isByVal()) {
3581 if (!TII->isLoadFromStackSlot(Def, FI))
3584 unsigned Opcode = Def->getOpcode();
3585 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r ||
3586 Opcode == X86::LEA64_32r) &&
3587 Def->getOperand(1).isFI()) {
3588 FI = Def->getOperand(1).getIndex();
3589 Bytes = Flags.getByValSize();
3593 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
3594 if (Flags.isByVal())
3595 // ByVal argument is passed in as a pointer but it's now being
3596 // dereferenced. e.g.
3597 // define @foo(%struct.X* %A) {
3598 // tail call @bar(%struct.X* byval %A)
3601 SDValue Ptr = Ld->getBasePtr();
3602 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
3605 FI = FINode->getIndex();
3606 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
3607 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
3608 FI = FINode->getIndex();
3609 Bytes = Flags.getByValSize();
3613 assert(FI != INT_MAX);
3614 if (!MFI->isFixedObjectIndex(FI))
3616 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
3619 /// Check whether the call is eligible for tail call optimization. Targets
3620 /// that want to do tail call optimization should implement this function.
3621 bool X86TargetLowering::IsEligibleForTailCallOptimization(
3622 SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg,
3623 bool isCalleeStructRet, bool isCallerStructRet, Type *RetTy,
3624 const SmallVectorImpl<ISD::OutputArg> &Outs,
3625 const SmallVectorImpl<SDValue> &OutVals,
3626 const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const {
3627 if (!mayTailCallThisCC(CalleeCC))
3630 // If -tailcallopt is specified, make fastcc functions tail-callable.
3631 MachineFunction &MF = DAG.getMachineFunction();
3632 const Function *CallerF = MF.getFunction();
3634 // If the function return type is x86_fp80 and the callee return type is not,
3635 // then the FP_EXTEND of the call result is not a nop. It's not safe to
3636 // perform a tailcall optimization here.
3637 if (CallerF->getReturnType()->isX86_FP80Ty() && !RetTy->isX86_FP80Ty())
3640 CallingConv::ID CallerCC = CallerF->getCallingConv();
3641 bool CCMatch = CallerCC == CalleeCC;
3642 bool IsCalleeWin64 = Subtarget->isCallingConvWin64(CalleeCC);
3643 bool IsCallerWin64 = Subtarget->isCallingConvWin64(CallerCC);
3645 // Win64 functions have extra shadow space for argument homing. Don't do the
3646 // sibcall if the caller and callee have mismatched expectations for this
3648 if (IsCalleeWin64 != IsCallerWin64)
3651 if (DAG.getTarget().Options.GuaranteedTailCallOpt) {
3652 if (canGuaranteeTCO(CalleeCC) && CCMatch)
3657 // Look for obvious safe cases to perform tail call optimization that do not
3658 // require ABI changes. This is what gcc calls sibcall.
3660 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
3661 // emit a special epilogue.
3662 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3663 if (RegInfo->needsStackRealignment(MF))
3666 // Also avoid sibcall optimization if either caller or callee uses struct
3667 // return semantics.
3668 if (isCalleeStructRet || isCallerStructRet)
3671 // Do not sibcall optimize vararg calls unless all arguments are passed via
3673 if (isVarArg && !Outs.empty()) {
3674 // Optimizing for varargs on Win64 is unlikely to be safe without
3675 // additional testing.
3676 if (IsCalleeWin64 || IsCallerWin64)
3679 SmallVector<CCValAssign, 16> ArgLocs;
3680 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3683 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3684 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
3685 if (!ArgLocs[i].isRegLoc())
3689 // If the call result is in ST0 / ST1, it needs to be popped off the x87
3690 // stack. Therefore, if it's not used by the call it is not safe to optimize
3691 // this into a sibcall.
3692 bool Unused = false;
3693 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
3700 SmallVector<CCValAssign, 16> RVLocs;
3701 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(), RVLocs,
3703 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
3704 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3705 CCValAssign &VA = RVLocs[i];
3706 if (VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1)
3711 // If the calling conventions do not match, then we'd better make sure the
3712 // results are returned in the same way as what the caller expects.
3714 SmallVector<CCValAssign, 16> RVLocs1;
3715 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
3717 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
3719 SmallVector<CCValAssign, 16> RVLocs2;
3720 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
3722 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
3724 if (RVLocs1.size() != RVLocs2.size())
3726 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
3727 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
3729 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
3731 if (RVLocs1[i].isRegLoc()) {
3732 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
3735 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
3741 unsigned StackArgsSize = 0;
3743 // If the callee takes no arguments then go on to check the results of the
3745 if (!Outs.empty()) {
3746 // Check if stack adjustment is needed. For now, do not do this if any
3747 // argument is passed on the stack.
3748 SmallVector<CCValAssign, 16> ArgLocs;
3749 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3752 // Allocate shadow area for Win64
3754 CCInfo.AllocateStack(32, 8);
3756 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3757 StackArgsSize = CCInfo.getNextStackOffset();
3759 if (CCInfo.getNextStackOffset()) {
3760 // Check if the arguments are already laid out in the right way as
3761 // the caller's fixed stack objects.
3762 MachineFrameInfo *MFI = MF.getFrameInfo();
3763 const MachineRegisterInfo *MRI = &MF.getRegInfo();
3764 const X86InstrInfo *TII = Subtarget->getInstrInfo();
3765 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3766 CCValAssign &VA = ArgLocs[i];
3767 SDValue Arg = OutVals[i];
3768 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3769 if (VA.getLocInfo() == CCValAssign::Indirect)
3771 if (!VA.isRegLoc()) {
3772 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
3779 // If the tailcall address may be in a register, then make sure it's
3780 // possible to register allocate for it. In 32-bit, the call address can
3781 // only target EAX, EDX, or ECX since the tail call must be scheduled after
3782 // callee-saved registers are restored. These happen to be the same
3783 // registers used to pass 'inreg' arguments so watch out for those.
3784 if (!Subtarget->is64Bit() &&
3785 ((!isa<GlobalAddressSDNode>(Callee) &&
3786 !isa<ExternalSymbolSDNode>(Callee)) ||
3787 DAG.getTarget().getRelocationModel() == Reloc::PIC_)) {
3788 unsigned NumInRegs = 0;
3789 // In PIC we need an extra register to formulate the address computation
3791 unsigned MaxInRegs =
3792 (DAG.getTarget().getRelocationModel() == Reloc::PIC_) ? 2 : 3;
3794 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3795 CCValAssign &VA = ArgLocs[i];
3798 unsigned Reg = VA.getLocReg();
3801 case X86::EAX: case X86::EDX: case X86::ECX:
3802 if (++NumInRegs == MaxInRegs)
3810 bool CalleeWillPop =
3811 X86::isCalleePop(CalleeCC, Subtarget->is64Bit(), isVarArg,
3812 MF.getTarget().Options.GuaranteedTailCallOpt);
3814 if (unsigned BytesToPop =
3815 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn()) {
3816 // If we have bytes to pop, the callee must pop them.
3817 bool CalleePopMatches = CalleeWillPop && BytesToPop == StackArgsSize;
3818 if (!CalleePopMatches)
3820 } else if (CalleeWillPop && StackArgsSize > 0) {
3821 // If we don't have bytes to pop, make sure the callee doesn't pop any.
3829 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
3830 const TargetLibraryInfo *libInfo) const {
3831 return X86::createFastISel(funcInfo, libInfo);
3834 //===----------------------------------------------------------------------===//
3835 // Other Lowering Hooks
3836 //===----------------------------------------------------------------------===//
3838 static bool MayFoldLoad(SDValue Op) {
3839 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
3842 static bool MayFoldIntoStore(SDValue Op) {
3843 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
3846 static bool isTargetShuffle(unsigned Opcode) {
3848 default: return false;
3849 case X86ISD::BLENDI:
3850 case X86ISD::PSHUFB:
3851 case X86ISD::PSHUFD:
3852 case X86ISD::PSHUFHW:
3853 case X86ISD::PSHUFLW:
3855 case X86ISD::PALIGNR:
3856 case X86ISD::MOVLHPS:
3857 case X86ISD::MOVLHPD:
3858 case X86ISD::MOVHLPS:
3859 case X86ISD::MOVLPS:
3860 case X86ISD::MOVLPD:
3861 case X86ISD::MOVSHDUP:
3862 case X86ISD::MOVSLDUP:
3863 case X86ISD::MOVDDUP:
3866 case X86ISD::UNPCKL:
3867 case X86ISD::UNPCKH:
3868 case X86ISD::VPERMILPI:
3869 case X86ISD::VPERM2X128:
3870 case X86ISD::VPERMI:
3871 case X86ISD::VPERMV:
3872 case X86ISD::VPERMV3:
3877 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, MVT VT,
3878 SDValue V1, unsigned TargetMask,
3879 SelectionDAG &DAG) {
3881 default: llvm_unreachable("Unknown x86 shuffle node");
3882 case X86ISD::PSHUFD:
3883 case X86ISD::PSHUFHW:
3884 case X86ISD::PSHUFLW:
3885 case X86ISD::VPERMILPI:
3886 case X86ISD::VPERMI:
3887 return DAG.getNode(Opc, dl, VT, V1,
3888 DAG.getConstant(TargetMask, dl, MVT::i8));
3892 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, MVT VT,
3893 SDValue V1, SDValue V2, SelectionDAG &DAG) {
3895 default: llvm_unreachable("Unknown x86 shuffle node");
3896 case X86ISD::MOVLHPS:
3897 case X86ISD::MOVLHPD:
3898 case X86ISD::MOVHLPS:
3899 case X86ISD::MOVLPS:
3900 case X86ISD::MOVLPD:
3903 case X86ISD::UNPCKL:
3904 case X86ISD::UNPCKH:
3905 return DAG.getNode(Opc, dl, VT, V1, V2);
3909 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
3910 MachineFunction &MF = DAG.getMachineFunction();
3911 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3912 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
3913 int ReturnAddrIndex = FuncInfo->getRAIndex();
3915 if (ReturnAddrIndex == 0) {
3916 // Set up a frame object for the return address.
3917 unsigned SlotSize = RegInfo->getSlotSize();
3918 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize,
3921 FuncInfo->setRAIndex(ReturnAddrIndex);
3924 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy(DAG.getDataLayout()));
3927 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
3928 bool hasSymbolicDisplacement) {
3929 // Offset should fit into 32 bit immediate field.
3930 if (!isInt<32>(Offset))
3933 // If we don't have a symbolic displacement - we don't have any extra
3935 if (!hasSymbolicDisplacement)
3938 // FIXME: Some tweaks might be needed for medium code model.
3939 if (M != CodeModel::Small && M != CodeModel::Kernel)
3942 // For small code model we assume that latest object is 16MB before end of 31
3943 // bits boundary. We may also accept pretty large negative constants knowing
3944 // that all objects are in the positive half of address space.
3945 if (M == CodeModel::Small && Offset < 16*1024*1024)
3948 // For kernel code model we know that all object resist in the negative half
3949 // of 32bits address space. We may not accept negative offsets, since they may
3950 // be just off and we may accept pretty large positive ones.
3951 if (M == CodeModel::Kernel && Offset >= 0)
3957 /// Determines whether the callee is required to pop its own arguments.
3958 /// Callee pop is necessary to support tail calls.
3959 bool X86::isCalleePop(CallingConv::ID CallingConv,
3960 bool is64Bit, bool IsVarArg, bool GuaranteeTCO) {
3961 // If GuaranteeTCO is true, we force some calls to be callee pop so that we
3962 // can guarantee TCO.
3963 if (!IsVarArg && shouldGuaranteeTCO(CallingConv, GuaranteeTCO))
3966 switch (CallingConv) {
3969 case CallingConv::X86_StdCall:
3970 case CallingConv::X86_FastCall:
3971 case CallingConv::X86_ThisCall:
3972 case CallingConv::X86_VectorCall:
3977 /// \brief Return true if the condition is an unsigned comparison operation.
3978 static bool isX86CCUnsigned(unsigned X86CC) {
3980 default: llvm_unreachable("Invalid integer condition!");
3981 case X86::COND_E: return true;
3982 case X86::COND_G: return false;
3983 case X86::COND_GE: return false;
3984 case X86::COND_L: return false;
3985 case X86::COND_LE: return false;
3986 case X86::COND_NE: return true;
3987 case X86::COND_B: return true;
3988 case X86::COND_A: return true;
3989 case X86::COND_BE: return true;
3990 case X86::COND_AE: return true;
3994 static X86::CondCode TranslateIntegerX86CC(ISD::CondCode SetCCOpcode) {
3995 switch (SetCCOpcode) {
3996 default: llvm_unreachable("Invalid integer condition!");
3997 case ISD::SETEQ: return X86::COND_E;
3998 case ISD::SETGT: return X86::COND_G;
3999 case ISD::SETGE: return X86::COND_GE;
4000 case ISD::SETLT: return X86::COND_L;
4001 case ISD::SETLE: return X86::COND_LE;
4002 case ISD::SETNE: return X86::COND_NE;
4003 case ISD::SETULT: return X86::COND_B;
4004 case ISD::SETUGT: return X86::COND_A;
4005 case ISD::SETULE: return X86::COND_BE;
4006 case ISD::SETUGE: return X86::COND_AE;
4010 /// Do a one-to-one translation of a ISD::CondCode to the X86-specific
4011 /// condition code, returning the condition code and the LHS/RHS of the
4012 /// comparison to make.
4013 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, SDLoc DL, bool isFP,
4014 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
4016 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
4017 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
4018 // X > -1 -> X == 0, jump !sign.
4019 RHS = DAG.getConstant(0, DL, RHS.getValueType());
4020 return X86::COND_NS;
4022 if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
4023 // X < 0 -> X == 0, jump on sign.
4026 if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
4028 RHS = DAG.getConstant(0, DL, RHS.getValueType());
4029 return X86::COND_LE;
4033 return TranslateIntegerX86CC(SetCCOpcode);
4036 // First determine if it is required or is profitable to flip the operands.
4038 // If LHS is a foldable load, but RHS is not, flip the condition.
4039 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
4040 !ISD::isNON_EXTLoad(RHS.getNode())) {
4041 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
4042 std::swap(LHS, RHS);
4045 switch (SetCCOpcode) {
4051 std::swap(LHS, RHS);
4055 // On a floating point condition, the flags are set as follows:
4057 // 0 | 0 | 0 | X > Y
4058 // 0 | 0 | 1 | X < Y
4059 // 1 | 0 | 0 | X == Y
4060 // 1 | 1 | 1 | unordered
4061 switch (SetCCOpcode) {
4062 default: llvm_unreachable("Condcode should be pre-legalized away");
4064 case ISD::SETEQ: return X86::COND_E;
4065 case ISD::SETOLT: // flipped
4067 case ISD::SETGT: return X86::COND_A;
4068 case ISD::SETOLE: // flipped
4070 case ISD::SETGE: return X86::COND_AE;
4071 case ISD::SETUGT: // flipped
4073 case ISD::SETLT: return X86::COND_B;
4074 case ISD::SETUGE: // flipped
4076 case ISD::SETLE: return X86::COND_BE;
4078 case ISD::SETNE: return X86::COND_NE;
4079 case ISD::SETUO: return X86::COND_P;
4080 case ISD::SETO: return X86::COND_NP;
4082 case ISD::SETUNE: return X86::COND_INVALID;
4086 /// Is there a floating point cmov for the specific X86 condition code?
4087 /// Current x86 isa includes the following FP cmov instructions:
4088 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
4089 static bool hasFPCMov(unsigned X86CC) {
4105 /// Returns true if the target can instruction select the
4106 /// specified FP immediate natively. If false, the legalizer will
4107 /// materialize the FP immediate as a load from a constant pool.
4108 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4109 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
4110 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
4116 bool X86TargetLowering::shouldReduceLoadWidth(SDNode *Load,
4117 ISD::LoadExtType ExtTy,
4119 // "ELF Handling for Thread-Local Storage" specifies that R_X86_64_GOTTPOFF
4120 // relocation target a movq or addq instruction: don't let the load shrink.
4121 SDValue BasePtr = cast<LoadSDNode>(Load)->getBasePtr();
4122 if (BasePtr.getOpcode() == X86ISD::WrapperRIP)
4123 if (const auto *GA = dyn_cast<GlobalAddressSDNode>(BasePtr.getOperand(0)))
4124 return GA->getTargetFlags() != X86II::MO_GOTTPOFF;
4128 /// \brief Returns true if it is beneficial to convert a load of a constant
4129 /// to just the constant itself.
4130 bool X86TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
4132 assert(Ty->isIntegerTy());
4134 unsigned BitSize = Ty->getPrimitiveSizeInBits();
4135 if (BitSize == 0 || BitSize > 64)
4140 bool X86TargetLowering::isExtractSubvectorCheap(EVT ResVT,
4141 unsigned Index) const {
4142 if (!isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, ResVT))
4145 return (Index == 0 || Index == ResVT.getVectorNumElements());
4148 bool X86TargetLowering::isCheapToSpeculateCttz() const {
4149 // Speculate cttz only if we can directly use TZCNT.
4150 return Subtarget->hasBMI();
4153 bool X86TargetLowering::isCheapToSpeculateCtlz() const {
4154 // Speculate ctlz only if we can directly use LZCNT.
4155 return Subtarget->hasLZCNT();
4158 /// Return true if every element in Mask, beginning
4159 /// from position Pos and ending in Pos+Size is undef.
4160 static bool isUndefInRange(ArrayRef<int> Mask, unsigned Pos, unsigned Size) {
4161 for (unsigned i = Pos, e = Pos + Size; i != e; ++i)
4167 /// Return true if Val is undef or if its value falls within the
4168 /// specified range (L, H].
4169 static bool isUndefOrInRange(int Val, int Low, int Hi) {
4170 return (Val < 0) || (Val >= Low && Val < Hi);
4173 /// Val is either less than zero (undef) or equal to the specified value.
4174 static bool isUndefOrEqual(int Val, int CmpVal) {
4175 return (Val < 0 || Val == CmpVal);
4178 /// Return true if every element in Mask, beginning
4179 /// from position Pos and ending in Pos+Size, falls within the specified
4180 /// sequential range (Low, Low+Size]. or is undef.
4181 static bool isSequentialOrUndefInRange(ArrayRef<int> Mask,
4182 unsigned Pos, unsigned Size, int Low) {
4183 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
4184 if (!isUndefOrEqual(Mask[i], Low))
4189 /// Return true if the specified EXTRACT_SUBVECTOR operand specifies a vector
4190 /// extract that is suitable for instruction that extract 128 or 256 bit vectors
4191 static bool isVEXTRACTIndex(SDNode *N, unsigned vecWidth) {
4192 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4193 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4196 // The index should be aligned on a vecWidth-bit boundary.
4198 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4200 MVT VT = N->getSimpleValueType(0);
4201 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4202 bool Result = (Index * ElSize) % vecWidth == 0;
4207 /// Return true if the specified INSERT_SUBVECTOR
4208 /// operand specifies a subvector insert that is suitable for input to
4209 /// insertion of 128 or 256-bit subvectors
4210 static bool isVINSERTIndex(SDNode *N, unsigned vecWidth) {
4211 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4212 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4214 // The index should be aligned on a vecWidth-bit boundary.
4216 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4218 MVT VT = N->getSimpleValueType(0);
4219 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4220 bool Result = (Index * ElSize) % vecWidth == 0;
4225 bool X86::isVINSERT128Index(SDNode *N) {
4226 return isVINSERTIndex(N, 128);
4229 bool X86::isVINSERT256Index(SDNode *N) {
4230 return isVINSERTIndex(N, 256);
4233 bool X86::isVEXTRACT128Index(SDNode *N) {
4234 return isVEXTRACTIndex(N, 128);
4237 bool X86::isVEXTRACT256Index(SDNode *N) {
4238 return isVEXTRACTIndex(N, 256);
4241 static unsigned getExtractVEXTRACTImmediate(SDNode *N, unsigned vecWidth) {
4242 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4243 assert(isa<ConstantSDNode>(N->getOperand(1).getNode()) &&
4244 "Illegal extract subvector for VEXTRACT");
4247 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4249 MVT VecVT = N->getOperand(0).getSimpleValueType();
4250 MVT ElVT = VecVT.getVectorElementType();
4252 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4253 return Index / NumElemsPerChunk;
4256 static unsigned getInsertVINSERTImmediate(SDNode *N, unsigned vecWidth) {
4257 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4258 assert(isa<ConstantSDNode>(N->getOperand(2).getNode()) &&
4259 "Illegal insert subvector for VINSERT");
4262 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4264 MVT VecVT = N->getSimpleValueType(0);
4265 MVT ElVT = VecVT.getVectorElementType();
4267 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4268 return Index / NumElemsPerChunk;
4271 /// Return the appropriate immediate to extract the specified
4272 /// EXTRACT_SUBVECTOR index with VEXTRACTF128 and VINSERTI128 instructions.
4273 unsigned X86::getExtractVEXTRACT128Immediate(SDNode *N) {
4274 return getExtractVEXTRACTImmediate(N, 128);
4277 /// Return the appropriate immediate to extract the specified
4278 /// EXTRACT_SUBVECTOR index with VEXTRACTF64x4 and VINSERTI64x4 instructions.
4279 unsigned X86::getExtractVEXTRACT256Immediate(SDNode *N) {
4280 return getExtractVEXTRACTImmediate(N, 256);
4283 /// Return the appropriate immediate to insert at the specified
4284 /// INSERT_SUBVECTOR index with VINSERTF128 and VINSERTI128 instructions.
4285 unsigned X86::getInsertVINSERT128Immediate(SDNode *N) {
4286 return getInsertVINSERTImmediate(N, 128);
4289 /// Return the appropriate immediate to insert at the specified
4290 /// INSERT_SUBVECTOR index with VINSERTF46x4 and VINSERTI64x4 instructions.
4291 unsigned X86::getInsertVINSERT256Immediate(SDNode *N) {
4292 return getInsertVINSERTImmediate(N, 256);
4295 /// Returns true if Elt is a constant zero or a floating point constant +0.0.
4296 bool X86::isZeroNode(SDValue Elt) {
4297 return isNullConstant(Elt) || isNullFPConstant(Elt);
4300 // Build a vector of constants
4301 // Use an UNDEF node if MaskElt == -1.
4302 // Spilt 64-bit constants in the 32-bit mode.
4303 static SDValue getConstVector(ArrayRef<int> Values, MVT VT,
4305 SDLoc dl, bool IsMask = false) {
4307 SmallVector<SDValue, 32> Ops;
4310 MVT ConstVecVT = VT;
4311 unsigned NumElts = VT.getVectorNumElements();
4312 bool In64BitMode = DAG.getTargetLoweringInfo().isTypeLegal(MVT::i64);
4313 if (!In64BitMode && VT.getVectorElementType() == MVT::i64) {
4314 ConstVecVT = MVT::getVectorVT(MVT::i32, NumElts * 2);
4318 MVT EltVT = ConstVecVT.getVectorElementType();
4319 for (unsigned i = 0; i < NumElts; ++i) {
4320 bool IsUndef = Values[i] < 0 && IsMask;
4321 SDValue OpNode = IsUndef ? DAG.getUNDEF(EltVT) :
4322 DAG.getConstant(Values[i], dl, EltVT);
4323 Ops.push_back(OpNode);
4325 Ops.push_back(IsUndef ? DAG.getUNDEF(EltVT) :
4326 DAG.getConstant(0, dl, EltVT));
4328 SDValue ConstsNode = DAG.getNode(ISD::BUILD_VECTOR, dl, ConstVecVT, Ops);
4330 ConstsNode = DAG.getBitcast(VT, ConstsNode);
4334 /// Returns a vector of specified type with all zero elements.
4335 static SDValue getZeroVector(MVT VT, const X86Subtarget *Subtarget,
4336 SelectionDAG &DAG, SDLoc dl) {
4337 assert(VT.isVector() && "Expected a vector type");
4339 // Always build SSE zero vectors as <4 x i32> bitcasted
4340 // to their dest type. This ensures they get CSE'd.
4342 if (VT.is128BitVector()) { // SSE
4343 if (Subtarget->hasSSE2()) { // SSE2
4344 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4345 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4347 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
4348 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
4350 } else if (VT.is256BitVector()) { // AVX
4351 if (Subtarget->hasInt256()) { // AVX2
4352 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4353 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4354 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4356 // 256-bit logic and arithmetic instructions in AVX are all
4357 // floating-point, no support for integer ops. Emit fp zeroed vectors.
4358 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
4359 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4360 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops);
4362 } else if (VT.is512BitVector()) { // AVX-512
4363 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4364 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
4365 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4366 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
4367 } else if (VT.getVectorElementType() == MVT::i1) {
4369 assert((Subtarget->hasBWI() || VT.getVectorNumElements() <= 16)
4370 && "Unexpected vector type");
4371 assert((Subtarget->hasVLX() || VT.getVectorNumElements() >= 8)
4372 && "Unexpected vector type");
4373 SDValue Cst = DAG.getConstant(0, dl, MVT::i1);
4374 SmallVector<SDValue, 64> Ops(VT.getVectorNumElements(), Cst);
4375 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
4377 llvm_unreachable("Unexpected vector type");
4379 return DAG.getBitcast(VT, Vec);
4382 static SDValue ExtractSubVector(SDValue Vec, unsigned IdxVal,
4383 SelectionDAG &DAG, SDLoc dl,
4384 unsigned vectorWidth) {
4385 assert((vectorWidth == 128 || vectorWidth == 256) &&
4386 "Unsupported vector width");
4387 EVT VT = Vec.getValueType();
4388 EVT ElVT = VT.getVectorElementType();
4389 unsigned Factor = VT.getSizeInBits()/vectorWidth;
4390 EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT,
4391 VT.getVectorNumElements()/Factor);
4393 // Extract from UNDEF is UNDEF.
4394 if (Vec.getOpcode() == ISD::UNDEF)
4395 return DAG.getUNDEF(ResultVT);
4397 // Extract the relevant vectorWidth bits. Generate an EXTRACT_SUBVECTOR
4398 unsigned ElemsPerChunk = vectorWidth / ElVT.getSizeInBits();
4399 assert(isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2");
4401 // This is the index of the first element of the vectorWidth-bit chunk
4402 // we want. Since ElemsPerChunk is a power of 2 just need to clear bits.
4403 IdxVal &= ~(ElemsPerChunk - 1);
4405 // If the input is a buildvector just emit a smaller one.
4406 if (Vec.getOpcode() == ISD::BUILD_VECTOR)
4407 return DAG.getNode(ISD::BUILD_VECTOR, dl, ResultVT,
4408 makeArrayRef(Vec->op_begin() + IdxVal, ElemsPerChunk));
4410 SDValue VecIdx = DAG.getIntPtrConstant(IdxVal, dl);
4411 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec, VecIdx);
4414 /// Generate a DAG to grab 128-bits from a vector > 128 bits. This
4415 /// sets things up to match to an AVX VEXTRACTF128 / VEXTRACTI128
4416 /// or AVX-512 VEXTRACTF32x4 / VEXTRACTI32x4
4417 /// instructions or a simple subregister reference. Idx is an index in the
4418 /// 128 bits we want. It need not be aligned to a 128-bit boundary. That makes
4419 /// lowering EXTRACT_VECTOR_ELT operations easier.
4420 static SDValue Extract128BitVector(SDValue Vec, unsigned IdxVal,
4421 SelectionDAG &DAG, SDLoc dl) {
4422 assert((Vec.getValueType().is256BitVector() ||
4423 Vec.getValueType().is512BitVector()) && "Unexpected vector size!");
4424 return ExtractSubVector(Vec, IdxVal, DAG, dl, 128);
4427 /// Generate a DAG to grab 256-bits from a 512-bit vector.
4428 static SDValue Extract256BitVector(SDValue Vec, unsigned IdxVal,
4429 SelectionDAG &DAG, SDLoc dl) {
4430 assert(Vec.getValueType().is512BitVector() && "Unexpected vector size!");
4431 return ExtractSubVector(Vec, IdxVal, DAG, dl, 256);
4434 static SDValue InsertSubVector(SDValue Result, SDValue Vec,
4435 unsigned IdxVal, SelectionDAG &DAG,
4436 SDLoc dl, unsigned vectorWidth) {
4437 assert((vectorWidth == 128 || vectorWidth == 256) &&
4438 "Unsupported vector width");
4439 // Inserting UNDEF is Result
4440 if (Vec.getOpcode() == ISD::UNDEF)
4442 EVT VT = Vec.getValueType();
4443 EVT ElVT = VT.getVectorElementType();
4444 EVT ResultVT = Result.getValueType();
4446 // Insert the relevant vectorWidth bits.
4447 unsigned ElemsPerChunk = vectorWidth/ElVT.getSizeInBits();
4448 assert(isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2");
4450 // This is the index of the first element of the vectorWidth-bit chunk
4451 // we want. Since ElemsPerChunk is a power of 2 just need to clear bits.
4452 IdxVal &= ~(ElemsPerChunk - 1);
4454 SDValue VecIdx = DAG.getIntPtrConstant(IdxVal, dl);
4455 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec, VecIdx);
4458 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
4459 /// sets things up to match to an AVX VINSERTF128/VINSERTI128 or
4460 /// AVX-512 VINSERTF32x4/VINSERTI32x4 instructions or a
4461 /// simple superregister reference. Idx is an index in the 128 bits
4462 /// we want. It need not be aligned to a 128-bit boundary. That makes
4463 /// lowering INSERT_VECTOR_ELT operations easier.
4464 static SDValue Insert128BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4465 SelectionDAG &DAG, SDLoc dl) {
4466 assert(Vec.getValueType().is128BitVector() && "Unexpected vector size!");
4468 // For insertion into the zero index (low half) of a 256-bit vector, it is
4469 // more efficient to generate a blend with immediate instead of an insert*128.
4470 // We are still creating an INSERT_SUBVECTOR below with an undef node to
4471 // extend the subvector to the size of the result vector. Make sure that
4472 // we are not recursing on that node by checking for undef here.
4473 if (IdxVal == 0 && Result.getValueType().is256BitVector() &&
4474 Result.getOpcode() != ISD::UNDEF) {
4475 EVT ResultVT = Result.getValueType();
4476 SDValue ZeroIndex = DAG.getIntPtrConstant(0, dl);
4477 SDValue Undef = DAG.getUNDEF(ResultVT);
4478 SDValue Vec256 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Undef,
4481 // The blend instruction, and therefore its mask, depend on the data type.
4482 MVT ScalarType = ResultVT.getVectorElementType().getSimpleVT();
4483 if (ScalarType.isFloatingPoint()) {
4484 // Choose either vblendps (float) or vblendpd (double).
4485 unsigned ScalarSize = ScalarType.getSizeInBits();
4486 assert((ScalarSize == 64 || ScalarSize == 32) && "Unknown float type");
4487 unsigned MaskVal = (ScalarSize == 64) ? 0x03 : 0x0f;
4488 SDValue Mask = DAG.getConstant(MaskVal, dl, MVT::i8);
4489 return DAG.getNode(X86ISD::BLENDI, dl, ResultVT, Result, Vec256, Mask);
4492 const X86Subtarget &Subtarget =
4493 static_cast<const X86Subtarget &>(DAG.getSubtarget());
4495 // AVX2 is needed for 256-bit integer blend support.
4496 // Integers must be cast to 32-bit because there is only vpblendd;
4497 // vpblendw can't be used for this because it has a handicapped mask.
4499 // If we don't have AVX2, then cast to float. Using a wrong domain blend
4500 // is still more efficient than using the wrong domain vinsertf128 that
4501 // will be created by InsertSubVector().
4502 MVT CastVT = Subtarget.hasAVX2() ? MVT::v8i32 : MVT::v8f32;
4504 SDValue Mask = DAG.getConstant(0x0f, dl, MVT::i8);
4505 Vec256 = DAG.getBitcast(CastVT, Vec256);
4506 Vec256 = DAG.getNode(X86ISD::BLENDI, dl, CastVT, Result, Vec256, Mask);
4507 return DAG.getBitcast(ResultVT, Vec256);
4510 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 128);
4513 static SDValue Insert256BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4514 SelectionDAG &DAG, SDLoc dl) {
4515 assert(Vec.getValueType().is256BitVector() && "Unexpected vector size!");
4516 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 256);
4519 /// Insert i1-subvector to i1-vector.
4520 static SDValue Insert1BitVector(SDValue Op, SelectionDAG &DAG) {
4523 SDValue Vec = Op.getOperand(0);
4524 SDValue SubVec = Op.getOperand(1);
4525 SDValue Idx = Op.getOperand(2);
4527 if (!isa<ConstantSDNode>(Idx))
4530 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
4531 if (IdxVal == 0 && Vec.isUndef()) // the operation is legal
4534 MVT OpVT = Op.getSimpleValueType();
4535 MVT SubVecVT = SubVec.getSimpleValueType();
4536 unsigned NumElems = OpVT.getVectorNumElements();
4537 unsigned SubVecNumElems = SubVecVT.getVectorNumElements();
4539 assert(IdxVal + SubVecNumElems <= NumElems &&
4540 IdxVal % SubVecVT.getSizeInBits() == 0 &&
4541 "Unexpected index value in INSERT_SUBVECTOR");
4543 // There are 3 possible cases:
4544 // 1. Subvector should be inserted in the lower part (IdxVal == 0)
4545 // 2. Subvector should be inserted in the upper part
4546 // (IdxVal + SubVecNumElems == NumElems)
4547 // 3. Subvector should be inserted in the middle (for example v2i1
4548 // to v16i1, index 2)
4550 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
4551 SDValue Undef = DAG.getUNDEF(OpVT);
4552 SDValue WideSubVec =
4553 DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef, SubVec, ZeroIdx);
4555 return DAG.getNode(X86ISD::VSHLI, dl, OpVT, WideSubVec,
4556 DAG.getConstant(IdxVal, dl, MVT::i8));
4558 if (ISD::isBuildVectorAllZeros(Vec.getNode())) {
4559 unsigned ShiftLeft = NumElems - SubVecNumElems;
4560 unsigned ShiftRight = NumElems - SubVecNumElems - IdxVal;
4561 WideSubVec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, WideSubVec,
4562 DAG.getConstant(ShiftLeft, dl, MVT::i8));
4563 return ShiftRight ? DAG.getNode(X86ISD::VSRLI, dl, OpVT, WideSubVec,
4564 DAG.getConstant(ShiftRight, dl, MVT::i8)) : WideSubVec;
4568 // Zero lower bits of the Vec
4569 SDValue ShiftBits = DAG.getConstant(SubVecNumElems, dl, MVT::i8);
4570 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
4571 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
4572 // Merge them together
4573 return DAG.getNode(ISD::OR, dl, OpVT, Vec, WideSubVec);
4576 // Simple case when we put subvector in the upper part
4577 if (IdxVal + SubVecNumElems == NumElems) {
4578 // Zero upper bits of the Vec
4579 WideSubVec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec,
4580 DAG.getConstant(IdxVal, dl, MVT::i8));
4581 SDValue ShiftBits = DAG.getConstant(SubVecNumElems, dl, MVT::i8);
4582 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
4583 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
4584 return DAG.getNode(ISD::OR, dl, OpVT, Vec, WideSubVec);
4586 // Subvector should be inserted in the middle - use shuffle
4587 SmallVector<int, 64> Mask;
4588 for (unsigned i = 0; i < NumElems; ++i)
4589 Mask.push_back(i >= IdxVal && i < IdxVal + SubVecNumElems ?
4591 return DAG.getVectorShuffle(OpVT, dl, WideSubVec, Vec, Mask);
4594 /// Concat two 128-bit vectors into a 256 bit vector using VINSERTF128
4595 /// instructions. This is used because creating CONCAT_VECTOR nodes of
4596 /// BUILD_VECTORS returns a larger BUILD_VECTOR while we're trying to lower
4597 /// large BUILD_VECTORS.
4598 static SDValue Concat128BitVectors(SDValue V1, SDValue V2, EVT VT,
4599 unsigned NumElems, SelectionDAG &DAG,
4601 SDValue V = Insert128BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4602 return Insert128BitVector(V, V2, NumElems/2, DAG, dl);
4605 static SDValue Concat256BitVectors(SDValue V1, SDValue V2, EVT VT,
4606 unsigned NumElems, SelectionDAG &DAG,
4608 SDValue V = Insert256BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4609 return Insert256BitVector(V, V2, NumElems/2, DAG, dl);
4612 /// Returns a vector of specified type with all bits set.
4613 /// Always build ones vectors as <4 x i32> or <8 x i32>. For 256-bit types with
4614 /// no AVX2 supprt, use two <4 x i32> inserted in a <8 x i32> appropriately.
4615 /// Then bitcast to their original type, ensuring they get CSE'd.
4616 static SDValue getOnesVector(EVT VT, const X86Subtarget *Subtarget,
4617 SelectionDAG &DAG, SDLoc dl) {
4618 assert(VT.isVector() && "Expected a vector type");
4620 SDValue Cst = DAG.getConstant(~0U, dl, MVT::i32);
4622 if (VT.is512BitVector()) {
4623 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
4624 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4625 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
4626 } else if (VT.is256BitVector()) {
4627 if (Subtarget->hasInt256()) { // AVX2
4628 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4629 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4631 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4632 Vec = Concat128BitVectors(Vec, Vec, MVT::v8i32, 8, DAG, dl);
4634 } else if (VT.is128BitVector()) {
4635 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4637 llvm_unreachable("Unexpected vector type");
4639 return DAG.getBitcast(VT, Vec);
4642 /// Returns a vector_shuffle node for an unpackl operation.
4643 static SDValue getUnpackl(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4645 unsigned NumElems = VT.getVectorNumElements();
4646 SmallVector<int, 8> Mask;
4647 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
4649 Mask.push_back(i + NumElems);
4651 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4654 /// Returns a vector_shuffle node for an unpackh operation.
4655 static SDValue getUnpackh(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4657 unsigned NumElems = VT.getVectorNumElements();
4658 SmallVector<int, 8> Mask;
4659 for (unsigned i = 0, Half = NumElems/2; i != Half; ++i) {
4660 Mask.push_back(i + Half);
4661 Mask.push_back(i + NumElems + Half);
4663 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4666 /// Return a vector_shuffle of the specified vector of zero or undef vector.
4667 /// This produces a shuffle where the low element of V2 is swizzled into the
4668 /// zero/undef vector, landing at element Idx.
4669 /// This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
4670 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
4672 const X86Subtarget *Subtarget,
4673 SelectionDAG &DAG) {
4674 MVT VT = V2.getSimpleValueType();
4676 ? getZeroVector(VT, Subtarget, DAG, SDLoc(V2)) : DAG.getUNDEF(VT);
4677 unsigned NumElems = VT.getVectorNumElements();
4678 SmallVector<int, 16> MaskVec;
4679 for (unsigned i = 0; i != NumElems; ++i)
4680 // If this is the insertion idx, put the low elt of V2 here.
4681 MaskVec.push_back(i == Idx ? NumElems : i);
4682 return DAG.getVectorShuffle(VT, SDLoc(V2), V1, V2, &MaskVec[0]);
4685 /// Calculates the shuffle mask corresponding to the target-specific opcode.
4686 /// Returns true if the Mask could be calculated. Sets IsUnary to true if only
4687 /// uses one source. Note that this will set IsUnary for shuffles which use a
4688 /// single input multiple times, and in those cases it will
4689 /// adjust the mask to only have indices within that single input.
4690 /// FIXME: Add support for Decode*Mask functions that return SM_SentinelZero.
4691 static bool getTargetShuffleMask(SDNode *N, MVT VT,
4692 SmallVectorImpl<int> &Mask, bool &IsUnary) {
4693 unsigned NumElems = VT.getVectorNumElements();
4697 bool IsFakeUnary = false;
4698 switch(N->getOpcode()) {
4699 case X86ISD::BLENDI:
4700 ImmN = N->getOperand(N->getNumOperands()-1);
4701 DecodeBLENDMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4704 ImmN = N->getOperand(N->getNumOperands()-1);
4705 DecodeSHUFPMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4706 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4708 case X86ISD::UNPCKH:
4709 DecodeUNPCKHMask(VT, Mask);
4710 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4712 case X86ISD::UNPCKL:
4713 DecodeUNPCKLMask(VT, Mask);
4714 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4716 case X86ISD::MOVHLPS:
4717 DecodeMOVHLPSMask(NumElems, Mask);
4718 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4720 case X86ISD::MOVLHPS:
4721 DecodeMOVLHPSMask(NumElems, Mask);
4722 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4724 case X86ISD::PALIGNR:
4725 ImmN = N->getOperand(N->getNumOperands()-1);
4726 DecodePALIGNRMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4728 case X86ISD::PSHUFD:
4729 case X86ISD::VPERMILPI:
4730 ImmN = N->getOperand(N->getNumOperands()-1);
4731 DecodePSHUFMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4734 case X86ISD::PSHUFHW:
4735 ImmN = N->getOperand(N->getNumOperands()-1);
4736 DecodePSHUFHWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4739 case X86ISD::PSHUFLW:
4740 ImmN = N->getOperand(N->getNumOperands()-1);
4741 DecodePSHUFLWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4744 case X86ISD::PSHUFB: {
4746 SDValue MaskNode = N->getOperand(1);
4747 while (MaskNode->getOpcode() == ISD::BITCAST)
4748 MaskNode = MaskNode->getOperand(0);
4750 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4751 // If we have a build-vector, then things are easy.
4752 MVT VT = MaskNode.getSimpleValueType();
4753 assert(VT.isVector() &&
4754 "Can't produce a non-vector with a build_vector!");
4755 if (!VT.isInteger())
4758 int NumBytesPerElement = VT.getVectorElementType().getSizeInBits() / 8;
4760 SmallVector<uint64_t, 32> RawMask;
4761 for (int i = 0, e = MaskNode->getNumOperands(); i < e; ++i) {
4762 SDValue Op = MaskNode->getOperand(i);
4763 if (Op->getOpcode() == ISD::UNDEF) {
4764 RawMask.push_back((uint64_t)SM_SentinelUndef);
4767 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
4770 APInt MaskElement = CN->getAPIntValue();
4772 // We now have to decode the element which could be any integer size and
4773 // extract each byte of it.
4774 for (int j = 0; j < NumBytesPerElement; ++j) {
4775 // Note that this is x86 and so always little endian: the low byte is
4776 // the first byte of the mask.
4777 RawMask.push_back(MaskElement.getLoBits(8).getZExtValue());
4778 MaskElement = MaskElement.lshr(8);
4781 DecodePSHUFBMask(RawMask, Mask);
4785 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4789 SDValue Ptr = MaskLoad->getBasePtr();
4790 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4791 Ptr->getOpcode() == X86ISD::WrapperRIP)
4792 Ptr = Ptr->getOperand(0);
4794 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4795 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4798 if (auto *C = dyn_cast<Constant>(MaskCP->getConstVal())) {
4799 DecodePSHUFBMask(C, Mask);
4807 case X86ISD::VPERMI:
4808 ImmN = N->getOperand(N->getNumOperands()-1);
4809 DecodeVPERMMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4814 DecodeScalarMoveMask(VT, /* IsLoad */ false, Mask);
4816 case X86ISD::VPERM2X128:
4817 ImmN = N->getOperand(N->getNumOperands()-1);
4818 DecodeVPERM2X128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4819 if (Mask.empty()) return false;
4820 // Mask only contains negative index if an element is zero.
4821 if (std::any_of(Mask.begin(), Mask.end(),
4822 [](int M){ return M == SM_SentinelZero; }))
4825 case X86ISD::MOVSLDUP:
4826 DecodeMOVSLDUPMask(VT, Mask);
4829 case X86ISD::MOVSHDUP:
4830 DecodeMOVSHDUPMask(VT, Mask);
4833 case X86ISD::MOVDDUP:
4834 DecodeMOVDDUPMask(VT, Mask);
4837 case X86ISD::MOVLHPD:
4838 case X86ISD::MOVLPD:
4839 case X86ISD::MOVLPS:
4840 // Not yet implemented
4842 case X86ISD::VPERMV: {
4844 SDValue MaskNode = N->getOperand(0);
4845 while (MaskNode->getOpcode() == ISD::BITCAST)
4846 MaskNode = MaskNode->getOperand(0);
4848 unsigned MaskLoBits = Log2_64(VT.getVectorNumElements());
4849 SmallVector<uint64_t, 32> RawMask;
4850 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4851 // If we have a build-vector, then things are easy.
4852 assert(MaskNode.getSimpleValueType().isInteger() &&
4853 MaskNode.getSimpleValueType().getVectorNumElements() ==
4854 VT.getVectorNumElements());
4856 for (unsigned i = 0; i < MaskNode->getNumOperands(); ++i) {
4857 SDValue Op = MaskNode->getOperand(i);
4858 if (Op->getOpcode() == ISD::UNDEF)
4859 RawMask.push_back((uint64_t)SM_SentinelUndef);
4860 else if (isa<ConstantSDNode>(Op)) {
4861 APInt MaskElement = cast<ConstantSDNode>(Op)->getAPIntValue();
4862 RawMask.push_back(MaskElement.getLoBits(MaskLoBits).getZExtValue());
4866 DecodeVPERMVMask(RawMask, Mask);
4869 if (MaskNode->getOpcode() == X86ISD::VBROADCAST) {
4870 unsigned NumEltsInMask = MaskNode->getNumOperands();
4871 MaskNode = MaskNode->getOperand(0);
4872 auto *CN = dyn_cast<ConstantSDNode>(MaskNode);
4874 APInt MaskEltValue = CN->getAPIntValue();
4875 for (unsigned i = 0; i < NumEltsInMask; ++i)
4876 RawMask.push_back(MaskEltValue.getLoBits(MaskLoBits).getZExtValue());
4877 DecodeVPERMVMask(RawMask, Mask);
4880 // It may be a scalar load
4883 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4887 SDValue Ptr = MaskLoad->getBasePtr();
4888 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4889 Ptr->getOpcode() == X86ISD::WrapperRIP)
4890 Ptr = Ptr->getOperand(0);
4892 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4893 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4896 auto *C = dyn_cast<Constant>(MaskCP->getConstVal());
4898 DecodeVPERMVMask(C, VT, Mask);
4905 case X86ISD::VPERMV3: {
4907 SDValue MaskNode = N->getOperand(1);
4908 while (MaskNode->getOpcode() == ISD::BITCAST)
4909 MaskNode = MaskNode->getOperand(1);
4911 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4912 // If we have a build-vector, then things are easy.
4913 assert(MaskNode.getSimpleValueType().isInteger() &&
4914 MaskNode.getSimpleValueType().getVectorNumElements() ==
4915 VT.getVectorNumElements());
4917 SmallVector<uint64_t, 32> RawMask;
4918 unsigned MaskLoBits = Log2_64(VT.getVectorNumElements()*2);
4920 for (unsigned i = 0; i < MaskNode->getNumOperands(); ++i) {
4921 SDValue Op = MaskNode->getOperand(i);
4922 if (Op->getOpcode() == ISD::UNDEF)
4923 RawMask.push_back((uint64_t)SM_SentinelUndef);
4925 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
4928 APInt MaskElement = CN->getAPIntValue();
4929 RawMask.push_back(MaskElement.getLoBits(MaskLoBits).getZExtValue());
4932 DecodeVPERMV3Mask(RawMask, Mask);
4936 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4940 SDValue Ptr = MaskLoad->getBasePtr();
4941 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4942 Ptr->getOpcode() == X86ISD::WrapperRIP)
4943 Ptr = Ptr->getOperand(0);
4945 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4946 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4949 auto *C = dyn_cast<Constant>(MaskCP->getConstVal());
4951 DecodeVPERMV3Mask(C, VT, Mask);
4958 default: llvm_unreachable("unknown target shuffle node");
4961 // If we have a fake unary shuffle, the shuffle mask is spread across two
4962 // inputs that are actually the same node. Re-map the mask to always point
4963 // into the first input.
4966 if (M >= (int)Mask.size())
4972 /// Returns the scalar element that will make up the ith
4973 /// element of the result of the vector shuffle.
4974 static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG,
4977 return SDValue(); // Limit search depth.
4979 SDValue V = SDValue(N, 0);
4980 EVT VT = V.getValueType();
4981 unsigned Opcode = V.getOpcode();
4983 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
4984 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
4985 int Elt = SV->getMaskElt(Index);
4988 return DAG.getUNDEF(VT.getVectorElementType());
4990 unsigned NumElems = VT.getVectorNumElements();
4991 SDValue NewV = (Elt < (int)NumElems) ? SV->getOperand(0)
4992 : SV->getOperand(1);
4993 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG, Depth+1);
4996 // Recurse into target specific vector shuffles to find scalars.
4997 if (isTargetShuffle(Opcode)) {
4998 MVT ShufVT = V.getSimpleValueType();
4999 unsigned NumElems = ShufVT.getVectorNumElements();
5000 SmallVector<int, 16> ShuffleMask;
5003 if (!getTargetShuffleMask(N, ShufVT, ShuffleMask, IsUnary))
5006 int Elt = ShuffleMask[Index];
5008 return DAG.getUNDEF(ShufVT.getVectorElementType());
5010 SDValue NewV = (Elt < (int)NumElems) ? N->getOperand(0)
5012 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG,
5016 // Actual nodes that may contain scalar elements
5017 if (Opcode == ISD::BITCAST) {
5018 V = V.getOperand(0);
5019 EVT SrcVT = V.getValueType();
5020 unsigned NumElems = VT.getVectorNumElements();
5022 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
5026 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
5027 return (Index == 0) ? V.getOperand(0)
5028 : DAG.getUNDEF(VT.getVectorElementType());
5030 if (V.getOpcode() == ISD::BUILD_VECTOR)
5031 return V.getOperand(Index);
5036 /// Custom lower build_vector of v16i8.
5037 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
5038 unsigned NumNonZero, unsigned NumZero,
5040 const X86Subtarget* Subtarget,
5041 const TargetLowering &TLI) {
5049 // SSE4.1 - use PINSRB to insert each byte directly.
5050 if (Subtarget->hasSSE41()) {
5051 for (unsigned i = 0; i < 16; ++i) {
5052 bool isNonZero = (NonZeros & (1 << i)) != 0;
5056 V = getZeroVector(MVT::v16i8, Subtarget, DAG, dl);
5058 V = DAG.getUNDEF(MVT::v16i8);
5061 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
5062 MVT::v16i8, V, Op.getOperand(i),
5063 DAG.getIntPtrConstant(i, dl));
5070 // Pre-SSE4.1 - merge byte pairs and insert with PINSRW.
5071 for (unsigned i = 0; i < 16; ++i) {
5072 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
5073 if (ThisIsNonZero && First) {
5075 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
5077 V = DAG.getUNDEF(MVT::v8i16);
5082 SDValue ThisElt, LastElt;
5083 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
5084 if (LastIsNonZero) {
5085 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
5086 MVT::i16, Op.getOperand(i-1));
5088 if (ThisIsNonZero) {
5089 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
5090 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
5091 ThisElt, DAG.getConstant(8, dl, MVT::i8));
5093 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
5097 if (ThisElt.getNode())
5098 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
5099 DAG.getIntPtrConstant(i/2, dl));
5103 return DAG.getBitcast(MVT::v16i8, V);
5106 /// Custom lower build_vector of v8i16.
5107 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
5108 unsigned NumNonZero, unsigned NumZero,
5110 const X86Subtarget* Subtarget,
5111 const TargetLowering &TLI) {
5118 for (unsigned i = 0; i < 8; ++i) {
5119 bool isNonZero = (NonZeros & (1 << i)) != 0;
5123 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
5125 V = DAG.getUNDEF(MVT::v8i16);
5128 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
5129 MVT::v8i16, V, Op.getOperand(i),
5130 DAG.getIntPtrConstant(i, dl));
5137 /// Custom lower build_vector of v4i32 or v4f32.
5138 static SDValue LowerBuildVectorv4x32(SDValue Op, SelectionDAG &DAG,
5139 const X86Subtarget *Subtarget,
5140 const TargetLowering &TLI) {
5141 // Find all zeroable elements.
5142 std::bitset<4> Zeroable;
5143 for (int i=0; i < 4; ++i) {
5144 SDValue Elt = Op->getOperand(i);
5145 Zeroable[i] = (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt));
5147 assert(Zeroable.size() - Zeroable.count() > 1 &&
5148 "We expect at least two non-zero elements!");
5150 // We only know how to deal with build_vector nodes where elements are either
5151 // zeroable or extract_vector_elt with constant index.
5152 SDValue FirstNonZero;
5153 unsigned FirstNonZeroIdx;
5154 for (unsigned i=0; i < 4; ++i) {
5157 SDValue Elt = Op->getOperand(i);
5158 if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5159 !isa<ConstantSDNode>(Elt.getOperand(1)))
5161 // Make sure that this node is extracting from a 128-bit vector.
5162 MVT VT = Elt.getOperand(0).getSimpleValueType();
5163 if (!VT.is128BitVector())
5165 if (!FirstNonZero.getNode()) {
5167 FirstNonZeroIdx = i;
5171 assert(FirstNonZero.getNode() && "Unexpected build vector of all zeros!");
5172 SDValue V1 = FirstNonZero.getOperand(0);
5173 MVT VT = V1.getSimpleValueType();
5175 // See if this build_vector can be lowered as a blend with zero.
5177 unsigned EltMaskIdx, EltIdx;
5179 for (EltIdx = 0; EltIdx < 4; ++EltIdx) {
5180 if (Zeroable[EltIdx]) {
5181 // The zero vector will be on the right hand side.
5182 Mask[EltIdx] = EltIdx+4;
5186 Elt = Op->getOperand(EltIdx);
5187 // By construction, Elt is a EXTRACT_VECTOR_ELT with constant index.
5188 EltMaskIdx = cast<ConstantSDNode>(Elt.getOperand(1))->getZExtValue();
5189 if (Elt.getOperand(0) != V1 || EltMaskIdx != EltIdx)
5191 Mask[EltIdx] = EltIdx;
5195 // Let the shuffle legalizer deal with blend operations.
5196 SDValue VZero = getZeroVector(VT, Subtarget, DAG, SDLoc(Op));
5197 if (V1.getSimpleValueType() != VT)
5198 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), VT, V1);
5199 return DAG.getVectorShuffle(VT, SDLoc(V1), V1, VZero, &Mask[0]);
5202 // See if we can lower this build_vector to a INSERTPS.
5203 if (!Subtarget->hasSSE41())
5206 SDValue V2 = Elt.getOperand(0);
5207 if (Elt == FirstNonZero && EltIdx == FirstNonZeroIdx)
5210 bool CanFold = true;
5211 for (unsigned i = EltIdx + 1; i < 4 && CanFold; ++i) {
5215 SDValue Current = Op->getOperand(i);
5216 SDValue SrcVector = Current->getOperand(0);
5219 CanFold = SrcVector == V1 &&
5220 cast<ConstantSDNode>(Current.getOperand(1))->getZExtValue() == i;
5226 assert(V1.getNode() && "Expected at least two non-zero elements!");
5227 if (V1.getSimpleValueType() != MVT::v4f32)
5228 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), MVT::v4f32, V1);
5229 if (V2.getSimpleValueType() != MVT::v4f32)
5230 V2 = DAG.getNode(ISD::BITCAST, SDLoc(V2), MVT::v4f32, V2);
5232 // Ok, we can emit an INSERTPS instruction.
5233 unsigned ZMask = Zeroable.to_ulong();
5235 unsigned InsertPSMask = EltMaskIdx << 6 | EltIdx << 4 | ZMask;
5236 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
5238 SDValue Result = DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
5239 DAG.getIntPtrConstant(InsertPSMask, DL));
5240 return DAG.getBitcast(VT, Result);
5243 /// Return a vector logical shift node.
5244 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
5245 unsigned NumBits, SelectionDAG &DAG,
5246 const TargetLowering &TLI, SDLoc dl) {
5247 assert(VT.is128BitVector() && "Unknown type for VShift");
5248 MVT ShVT = MVT::v2i64;
5249 unsigned Opc = isLeft ? X86ISD::VSHLDQ : X86ISD::VSRLDQ;
5250 SrcOp = DAG.getBitcast(ShVT, SrcOp);
5251 MVT ScalarShiftTy = TLI.getScalarShiftAmountTy(DAG.getDataLayout(), VT);
5252 assert(NumBits % 8 == 0 && "Only support byte sized shifts");
5253 SDValue ShiftVal = DAG.getConstant(NumBits/8, dl, ScalarShiftTy);
5254 return DAG.getBitcast(VT, DAG.getNode(Opc, dl, ShVT, SrcOp, ShiftVal));
5258 LowerAsSplatVectorLoad(SDValue SrcOp, MVT VT, SDLoc dl, SelectionDAG &DAG) {
5260 // Check if the scalar load can be widened into a vector load. And if
5261 // the address is "base + cst" see if the cst can be "absorbed" into
5262 // the shuffle mask.
5263 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
5264 SDValue Ptr = LD->getBasePtr();
5265 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
5267 EVT PVT = LD->getValueType(0);
5268 if (PVT != MVT::i32 && PVT != MVT::f32)
5273 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
5274 FI = FINode->getIndex();
5276 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
5277 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
5278 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
5279 Offset = Ptr.getConstantOperandVal(1);
5280 Ptr = Ptr.getOperand(0);
5285 // FIXME: 256-bit vector instructions don't require a strict alignment,
5286 // improve this code to support it better.
5287 unsigned RequiredAlign = VT.getSizeInBits()/8;
5288 SDValue Chain = LD->getChain();
5289 // Make sure the stack object alignment is at least 16 or 32.
5290 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5291 if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
5292 if (MFI->isFixedObjectIndex(FI)) {
5293 // Can't change the alignment. FIXME: It's possible to compute
5294 // the exact stack offset and reference FI + adjust offset instead.
5295 // If someone *really* cares about this. That's the way to implement it.
5298 MFI->setObjectAlignment(FI, RequiredAlign);
5302 // (Offset % 16 or 32) must be multiple of 4. Then address is then
5303 // Ptr + (Offset & ~15).
5306 if ((Offset % RequiredAlign) & 3)
5308 int64_t StartOffset = Offset & ~int64_t(RequiredAlign - 1);
5311 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
5312 DAG.getConstant(StartOffset, DL, Ptr.getValueType()));
5315 int EltNo = (Offset - StartOffset) >> 2;
5316 unsigned NumElems = VT.getVectorNumElements();
5318 EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
5319 SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
5320 LD->getPointerInfo().getWithOffset(StartOffset),
5321 false, false, false, 0);
5323 SmallVector<int, 8> Mask(NumElems, EltNo);
5325 return DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), &Mask[0]);
5331 /// Given the initializing elements 'Elts' of a vector of type 'VT', see if the
5332 /// elements can be replaced by a single large load which has the same value as
5333 /// a build_vector or insert_subvector whose loaded operands are 'Elts'.
5335 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
5337 /// FIXME: we'd also like to handle the case where the last elements are zero
5338 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
5339 /// There's even a handy isZeroNode for that purpose.
5340 static SDValue EltsFromConsecutiveLoads(EVT VT, ArrayRef<SDValue> Elts,
5341 SDLoc &DL, SelectionDAG &DAG,
5342 bool isAfterLegalize) {
5343 unsigned NumElems = Elts.size();
5345 LoadSDNode *LDBase = nullptr;
5346 unsigned LastLoadedElt = -1U;
5348 // For each element in the initializer, see if we've found a load or an undef.
5349 // If we don't find an initial load element, or later load elements are
5350 // non-consecutive, bail out.
5351 for (unsigned i = 0; i < NumElems; ++i) {
5352 SDValue Elt = Elts[i];
5353 // Look through a bitcast.
5354 if (Elt.getNode() && Elt.getOpcode() == ISD::BITCAST)
5355 Elt = Elt.getOperand(0);
5356 if (!Elt.getNode() ||
5357 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
5360 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
5362 LDBase = cast<LoadSDNode>(Elt.getNode());
5366 if (Elt.getOpcode() == ISD::UNDEF)
5369 LoadSDNode *LD = cast<LoadSDNode>(Elt);
5370 EVT LdVT = Elt.getValueType();
5371 // Each loaded element must be the correct fractional portion of the
5372 // requested vector load.
5373 if (LdVT.getSizeInBits() != VT.getSizeInBits() / NumElems)
5375 if (!DAG.isConsecutiveLoad(LD, LDBase, LdVT.getSizeInBits() / 8, i))
5380 // If we have found an entire vector of loads and undefs, then return a large
5381 // load of the entire vector width starting at the base pointer. If we found
5382 // consecutive loads for the low half, generate a vzext_load node.
5383 if (LastLoadedElt == NumElems - 1) {
5384 assert(LDBase && "Did not find base load for merging consecutive loads");
5385 EVT EltVT = LDBase->getValueType(0);
5386 // Ensure that the input vector size for the merged loads matches the
5387 // cumulative size of the input elements.
5388 if (VT.getSizeInBits() != EltVT.getSizeInBits() * NumElems)
5391 if (isAfterLegalize &&
5392 !DAG.getTargetLoweringInfo().isOperationLegal(ISD::LOAD, VT))
5395 SDValue NewLd = SDValue();
5397 NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
5398 LDBase->getPointerInfo(), LDBase->isVolatile(),
5399 LDBase->isNonTemporal(), LDBase->isInvariant(),
5400 LDBase->getAlignment());
5402 if (LDBase->hasAnyUseOfValue(1)) {
5403 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5405 SDValue(NewLd.getNode(), 1));
5406 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5407 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5408 SDValue(NewLd.getNode(), 1));
5414 //TODO: The code below fires only for for loading the low v2i32 / v2f32
5415 //of a v4i32 / v4f32. It's probably worth generalizing.
5416 EVT EltVT = VT.getVectorElementType();
5417 if (NumElems == 4 && LastLoadedElt == 1 && (EltVT.getSizeInBits() == 32) &&
5418 DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
5419 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
5420 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
5422 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys, Ops, MVT::i64,
5423 LDBase->getPointerInfo(),
5424 LDBase->getAlignment(),
5425 false/*isVolatile*/, true/*ReadMem*/,
5428 // Make sure the newly-created LOAD is in the same position as LDBase in
5429 // terms of dependency. We create a TokenFactor for LDBase and ResNode, and
5430 // update uses of LDBase's output chain to use the TokenFactor.
5431 if (LDBase->hasAnyUseOfValue(1)) {
5432 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5433 SDValue(LDBase, 1), SDValue(ResNode.getNode(), 1));
5434 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5435 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5436 SDValue(ResNode.getNode(), 1));
5439 return DAG.getBitcast(VT, ResNode);
5444 /// LowerVectorBroadcast - Attempt to use the vbroadcast instruction
5445 /// to generate a splat value for the following cases:
5446 /// 1. A splat BUILD_VECTOR which uses a single scalar load, or a constant.
5447 /// 2. A splat shuffle which uses a scalar_to_vector node which comes from
5448 /// a scalar load, or a constant.
5449 /// The VBROADCAST node is returned when a pattern is found,
5450 /// or SDValue() otherwise.
5451 static SDValue LowerVectorBroadcast(SDValue Op, const X86Subtarget* Subtarget,
5452 SelectionDAG &DAG) {
5453 // VBROADCAST requires AVX.
5454 // TODO: Splats could be generated for non-AVX CPUs using SSE
5455 // instructions, but there's less potential gain for only 128-bit vectors.
5456 if (!Subtarget->hasAVX())
5459 MVT VT = Op.getSimpleValueType();
5462 assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&
5463 "Unsupported vector type for broadcast.");
5468 switch (Op.getOpcode()) {
5470 // Unknown pattern found.
5473 case ISD::BUILD_VECTOR: {
5474 auto *BVOp = cast<BuildVectorSDNode>(Op.getNode());
5475 BitVector UndefElements;
5476 SDValue Splat = BVOp->getSplatValue(&UndefElements);
5478 // We need a splat of a single value to use broadcast, and it doesn't
5479 // make any sense if the value is only in one element of the vector.
5480 if (!Splat || (VT.getVectorNumElements() - UndefElements.count()) <= 1)
5484 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5485 Ld.getOpcode() == ISD::ConstantFP);
5487 // Make sure that all of the users of a non-constant load are from the
5488 // BUILD_VECTOR node.
5489 if (!ConstSplatVal && !BVOp->isOnlyUserOf(Ld.getNode()))
5494 case ISD::VECTOR_SHUFFLE: {
5495 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5497 // Shuffles must have a splat mask where the first element is
5499 if ((!SVOp->isSplat()) || SVOp->getMaskElt(0) != 0)
5502 SDValue Sc = Op.getOperand(0);
5503 if (Sc.getOpcode() != ISD::SCALAR_TO_VECTOR &&
5504 Sc.getOpcode() != ISD::BUILD_VECTOR) {
5506 if (!Subtarget->hasInt256())
5509 // Use the register form of the broadcast instruction available on AVX2.
5510 if (VT.getSizeInBits() >= 256)
5511 Sc = Extract128BitVector(Sc, 0, DAG, dl);
5512 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Sc);
5515 Ld = Sc.getOperand(0);
5516 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5517 Ld.getOpcode() == ISD::ConstantFP);
5519 // The scalar_to_vector node and the suspected
5520 // load node must have exactly one user.
5521 // Constants may have multiple users.
5523 // AVX-512 has register version of the broadcast
5524 bool hasRegVer = Subtarget->hasAVX512() && VT.is512BitVector() &&
5525 Ld.getValueType().getSizeInBits() >= 32;
5526 if (!ConstSplatVal && ((!Sc.hasOneUse() || !Ld.hasOneUse()) &&
5533 unsigned ScalarSize = Ld.getValueType().getSizeInBits();
5534 bool IsGE256 = (VT.getSizeInBits() >= 256);
5536 // When optimizing for size, generate up to 5 extra bytes for a broadcast
5537 // instruction to save 8 or more bytes of constant pool data.
5538 // TODO: If multiple splats are generated to load the same constant,
5539 // it may be detrimental to overall size. There needs to be a way to detect
5540 // that condition to know if this is truly a size win.
5541 bool OptForSize = DAG.getMachineFunction().getFunction()->optForSize();
5543 // Handle broadcasting a single constant scalar from the constant pool
5545 // On Sandybridge (no AVX2), it is still better to load a constant vector
5546 // from the constant pool and not to broadcast it from a scalar.
5547 // But override that restriction when optimizing for size.
5548 // TODO: Check if splatting is recommended for other AVX-capable CPUs.
5549 if (ConstSplatVal && (Subtarget->hasAVX2() || OptForSize)) {
5550 EVT CVT = Ld.getValueType();
5551 assert(!CVT.isVector() && "Must not broadcast a vector type");
5553 // Splat f32, i32, v4f64, v4i64 in all cases with AVX2.
5554 // For size optimization, also splat v2f64 and v2i64, and for size opt
5555 // with AVX2, also splat i8 and i16.
5556 // With pattern matching, the VBROADCAST node may become a VMOVDDUP.
5557 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5558 (OptForSize && (ScalarSize == 64 || Subtarget->hasAVX2()))) {
5559 const Constant *C = nullptr;
5560 if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld))
5561 C = CI->getConstantIntValue();
5562 else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld))
5563 C = CF->getConstantFPValue();
5565 assert(C && "Invalid constant type");
5567 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5569 DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout()));
5570 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
5572 CVT, dl, DAG.getEntryNode(), CP,
5573 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
5574 false, false, Alignment);
5576 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5580 bool IsLoad = ISD::isNormalLoad(Ld.getNode());
5582 // Handle AVX2 in-register broadcasts.
5583 if (!IsLoad && Subtarget->hasInt256() &&
5584 (ScalarSize == 32 || (IsGE256 && ScalarSize == 64)))
5585 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5587 // The scalar source must be a normal load.
5591 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5592 (Subtarget->hasVLX() && ScalarSize == 64))
5593 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5595 // The integer check is needed for the 64-bit into 128-bit so it doesn't match
5596 // double since there is no vbroadcastsd xmm
5597 if (Subtarget->hasInt256() && Ld.getValueType().isInteger()) {
5598 if (ScalarSize == 8 || ScalarSize == 16 || ScalarSize == 64)
5599 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5602 // Unsupported broadcast.
5606 /// \brief For an EXTRACT_VECTOR_ELT with a constant index return the real
5607 /// underlying vector and index.
5609 /// Modifies \p ExtractedFromVec to the real vector and returns the real
5611 static int getUnderlyingExtractedFromVec(SDValue &ExtractedFromVec,
5613 int Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue();
5614 if (!isa<ShuffleVectorSDNode>(ExtractedFromVec))
5617 // For 256-bit vectors, LowerEXTRACT_VECTOR_ELT_SSE4 may have already
5619 // (extract_vector_elt (v8f32 %vreg1), Constant<6>)
5621 // (extract_vector_elt (vector_shuffle<2,u,u,u>
5622 // (extract_subvector (v8f32 %vreg0), Constant<4>),
5625 // In this case the vector is the extract_subvector expression and the index
5626 // is 2, as specified by the shuffle.
5627 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(ExtractedFromVec);
5628 SDValue ShuffleVec = SVOp->getOperand(0);
5629 MVT ShuffleVecVT = ShuffleVec.getSimpleValueType();
5630 assert(ShuffleVecVT.getVectorElementType() ==
5631 ExtractedFromVec.getSimpleValueType().getVectorElementType());
5633 int ShuffleIdx = SVOp->getMaskElt(Idx);
5634 if (isUndefOrInRange(ShuffleIdx, 0, ShuffleVecVT.getVectorNumElements())) {
5635 ExtractedFromVec = ShuffleVec;
5641 static SDValue buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) {
5642 MVT VT = Op.getSimpleValueType();
5644 // Skip if insert_vec_elt is not supported.
5645 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5646 if (!TLI.isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT))
5650 unsigned NumElems = Op.getNumOperands();
5654 SmallVector<unsigned, 4> InsertIndices;
5655 SmallVector<int, 8> Mask(NumElems, -1);
5657 for (unsigned i = 0; i != NumElems; ++i) {
5658 unsigned Opc = Op.getOperand(i).getOpcode();
5660 if (Opc == ISD::UNDEF)
5663 if (Opc != ISD::EXTRACT_VECTOR_ELT) {
5664 // Quit if more than 1 elements need inserting.
5665 if (InsertIndices.size() > 1)
5668 InsertIndices.push_back(i);
5672 SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0);
5673 SDValue ExtIdx = Op.getOperand(i).getOperand(1);
5674 // Quit if non-constant index.
5675 if (!isa<ConstantSDNode>(ExtIdx))
5677 int Idx = getUnderlyingExtractedFromVec(ExtractedFromVec, ExtIdx);
5679 // Quit if extracted from vector of different type.
5680 if (ExtractedFromVec.getValueType() != VT)
5683 if (!VecIn1.getNode())
5684 VecIn1 = ExtractedFromVec;
5685 else if (VecIn1 != ExtractedFromVec) {
5686 if (!VecIn2.getNode())
5687 VecIn2 = ExtractedFromVec;
5688 else if (VecIn2 != ExtractedFromVec)
5689 // Quit if more than 2 vectors to shuffle
5693 if (ExtractedFromVec == VecIn1)
5695 else if (ExtractedFromVec == VecIn2)
5696 Mask[i] = Idx + NumElems;
5699 if (!VecIn1.getNode())
5702 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
5703 SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, &Mask[0]);
5704 for (unsigned i = 0, e = InsertIndices.size(); i != e; ++i) {
5705 unsigned Idx = InsertIndices[i];
5706 NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx),
5707 DAG.getIntPtrConstant(Idx, DL));
5713 static SDValue ConvertI1VectorToInteger(SDValue Op, SelectionDAG &DAG) {
5714 assert(ISD::isBuildVectorOfConstantSDNodes(Op.getNode()) &&
5715 Op.getScalarValueSizeInBits() == 1 &&
5716 "Can not convert non-constant vector");
5717 uint64_t Immediate = 0;
5718 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5719 SDValue In = Op.getOperand(idx);
5720 if (In.getOpcode() != ISD::UNDEF)
5721 Immediate |= cast<ConstantSDNode>(In)->getZExtValue() << idx;
5725 MVT::getIntegerVT(std::max((int)Op.getValueType().getSizeInBits(), 8));
5726 return DAG.getConstant(Immediate, dl, VT);
5728 // Lower BUILD_VECTOR operation for v8i1 and v16i1 types.
5730 X86TargetLowering::LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const {
5732 MVT VT = Op.getSimpleValueType();
5733 assert((VT.getVectorElementType() == MVT::i1) &&
5734 "Unexpected type in LowerBUILD_VECTORvXi1!");
5737 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5738 SDValue Cst = DAG.getTargetConstant(0, dl, MVT::i1);
5739 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5740 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5743 if (ISD::isBuildVectorAllOnes(Op.getNode())) {
5744 SDValue Cst = DAG.getTargetConstant(1, dl, MVT::i1);
5745 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5746 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5749 if (ISD::isBuildVectorOfConstantSDNodes(Op.getNode())) {
5750 SDValue Imm = ConvertI1VectorToInteger(Op, DAG);
5751 if (Imm.getValueSizeInBits() == VT.getSizeInBits())
5752 return DAG.getBitcast(VT, Imm);
5753 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, Imm);
5754 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec,
5755 DAG.getIntPtrConstant(0, dl));
5758 // Vector has one or more non-const elements
5759 uint64_t Immediate = 0;
5760 SmallVector<unsigned, 16> NonConstIdx;
5761 bool IsSplat = true;
5762 bool HasConstElts = false;
5764 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5765 SDValue In = Op.getOperand(idx);
5766 if (In.getOpcode() == ISD::UNDEF)
5768 if (!isa<ConstantSDNode>(In))
5769 NonConstIdx.push_back(idx);
5771 Immediate |= cast<ConstantSDNode>(In)->getZExtValue() << idx;
5772 HasConstElts = true;
5776 else if (In != Op.getOperand(SplatIdx))
5780 // for splat use " (select i1 splat_elt, all-ones, all-zeroes)"
5782 return DAG.getNode(ISD::SELECT, dl, VT, Op.getOperand(SplatIdx),
5783 DAG.getConstant(1, dl, VT),
5784 DAG.getConstant(0, dl, VT));
5786 // insert elements one by one
5790 MVT ImmVT = MVT::getIntegerVT(std::max((int)VT.getSizeInBits(), 8));
5791 Imm = DAG.getConstant(Immediate, dl, ImmVT);
5793 else if (HasConstElts)
5794 Imm = DAG.getConstant(0, dl, VT);
5796 Imm = DAG.getUNDEF(VT);
5797 if (Imm.getValueSizeInBits() == VT.getSizeInBits())
5798 DstVec = DAG.getBitcast(VT, Imm);
5800 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, Imm);
5801 DstVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec,
5802 DAG.getIntPtrConstant(0, dl));
5805 for (unsigned i = 0; i < NonConstIdx.size(); ++i) {
5806 unsigned InsertIdx = NonConstIdx[i];
5807 DstVec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
5808 Op.getOperand(InsertIdx),
5809 DAG.getIntPtrConstant(InsertIdx, dl));
5814 /// \brief Return true if \p N implements a horizontal binop and return the
5815 /// operands for the horizontal binop into V0 and V1.
5817 /// This is a helper function of LowerToHorizontalOp().
5818 /// This function checks that the build_vector \p N in input implements a
5819 /// horizontal operation. Parameter \p Opcode defines the kind of horizontal
5820 /// operation to match.
5821 /// For example, if \p Opcode is equal to ISD::ADD, then this function
5822 /// checks if \p N implements a horizontal arithmetic add; if instead \p Opcode
5823 /// is equal to ISD::SUB, then this function checks if this is a horizontal
5826 /// This function only analyzes elements of \p N whose indices are
5827 /// in range [BaseIdx, LastIdx).
5828 static bool isHorizontalBinOp(const BuildVectorSDNode *N, unsigned Opcode,
5830 unsigned BaseIdx, unsigned LastIdx,
5831 SDValue &V0, SDValue &V1) {
5832 EVT VT = N->getValueType(0);
5834 assert(BaseIdx * 2 <= LastIdx && "Invalid Indices in input!");
5835 assert(VT.isVector() && VT.getVectorNumElements() >= LastIdx &&
5836 "Invalid Vector in input!");
5838 bool IsCommutable = (Opcode == ISD::ADD || Opcode == ISD::FADD);
5839 bool CanFold = true;
5840 unsigned ExpectedVExtractIdx = BaseIdx;
5841 unsigned NumElts = LastIdx - BaseIdx;
5842 V0 = DAG.getUNDEF(VT);
5843 V1 = DAG.getUNDEF(VT);
5845 // Check if N implements a horizontal binop.
5846 for (unsigned i = 0, e = NumElts; i != e && CanFold; ++i) {
5847 SDValue Op = N->getOperand(i + BaseIdx);
5850 if (Op->getOpcode() == ISD::UNDEF) {
5851 // Update the expected vector extract index.
5852 if (i * 2 == NumElts)
5853 ExpectedVExtractIdx = BaseIdx;
5854 ExpectedVExtractIdx += 2;
5858 CanFold = Op->getOpcode() == Opcode && Op->hasOneUse();
5863 SDValue Op0 = Op.getOperand(0);
5864 SDValue Op1 = Op.getOperand(1);
5866 // Try to match the following pattern:
5867 // (BINOP (extract_vector_elt A, I), (extract_vector_elt A, I+1))
5868 CanFold = (Op0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5869 Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5870 Op0.getOperand(0) == Op1.getOperand(0) &&
5871 isa<ConstantSDNode>(Op0.getOperand(1)) &&
5872 isa<ConstantSDNode>(Op1.getOperand(1)));
5876 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5877 unsigned I1 = cast<ConstantSDNode>(Op1.getOperand(1))->getZExtValue();
5879 if (i * 2 < NumElts) {
5880 if (V0.getOpcode() == ISD::UNDEF) {
5881 V0 = Op0.getOperand(0);
5882 if (V0.getValueType() != VT)
5886 if (V1.getOpcode() == ISD::UNDEF) {
5887 V1 = Op0.getOperand(0);
5888 if (V1.getValueType() != VT)
5891 if (i * 2 == NumElts)
5892 ExpectedVExtractIdx = BaseIdx;
5895 SDValue Expected = (i * 2 < NumElts) ? V0 : V1;
5896 if (I0 == ExpectedVExtractIdx)
5897 CanFold = I1 == I0 + 1 && Op0.getOperand(0) == Expected;
5898 else if (IsCommutable && I1 == ExpectedVExtractIdx) {
5899 // Try to match the following dag sequence:
5900 // (BINOP (extract_vector_elt A, I+1), (extract_vector_elt A, I))
5901 CanFold = I0 == I1 + 1 && Op1.getOperand(0) == Expected;
5905 ExpectedVExtractIdx += 2;
5911 /// \brief Emit a sequence of two 128-bit horizontal add/sub followed by
5912 /// a concat_vector.
5914 /// This is a helper function of LowerToHorizontalOp().
5915 /// This function expects two 256-bit vectors called V0 and V1.
5916 /// At first, each vector is split into two separate 128-bit vectors.
5917 /// Then, the resulting 128-bit vectors are used to implement two
5918 /// horizontal binary operations.
5920 /// The kind of horizontal binary operation is defined by \p X86Opcode.
5922 /// \p Mode specifies how the 128-bit parts of V0 and V1 are passed in input to
5923 /// the two new horizontal binop.
5924 /// When Mode is set, the first horizontal binop dag node would take as input
5925 /// the lower 128-bit of V0 and the upper 128-bit of V0. The second
5926 /// horizontal binop dag node would take as input the lower 128-bit of V1
5927 /// and the upper 128-bit of V1.
5929 /// HADD V0_LO, V0_HI
5930 /// HADD V1_LO, V1_HI
5932 /// Otherwise, the first horizontal binop dag node takes as input the lower
5933 /// 128-bit of V0 and the lower 128-bit of V1, and the second horizontal binop
5934 /// dag node takes the upper 128-bit of V0 and the upper 128-bit of V1.
5936 /// HADD V0_LO, V1_LO
5937 /// HADD V0_HI, V1_HI
5939 /// If \p isUndefLO is set, then the algorithm propagates UNDEF to the lower
5940 /// 128-bits of the result. If \p isUndefHI is set, then UNDEF is propagated to
5941 /// the upper 128-bits of the result.
5942 static SDValue ExpandHorizontalBinOp(const SDValue &V0, const SDValue &V1,
5943 SDLoc DL, SelectionDAG &DAG,
5944 unsigned X86Opcode, bool Mode,
5945 bool isUndefLO, bool isUndefHI) {
5946 EVT VT = V0.getValueType();
5947 assert(VT.is256BitVector() && VT == V1.getValueType() &&
5948 "Invalid nodes in input!");
5950 unsigned NumElts = VT.getVectorNumElements();
5951 SDValue V0_LO = Extract128BitVector(V0, 0, DAG, DL);
5952 SDValue V0_HI = Extract128BitVector(V0, NumElts/2, DAG, DL);
5953 SDValue V1_LO = Extract128BitVector(V1, 0, DAG, DL);
5954 SDValue V1_HI = Extract128BitVector(V1, NumElts/2, DAG, DL);
5955 EVT NewVT = V0_LO.getValueType();
5957 SDValue LO = DAG.getUNDEF(NewVT);
5958 SDValue HI = DAG.getUNDEF(NewVT);
5961 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5962 if (!isUndefLO && V0->getOpcode() != ISD::UNDEF)
5963 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V0_HI);
5964 if (!isUndefHI && V1->getOpcode() != ISD::UNDEF)
5965 HI = DAG.getNode(X86Opcode, DL, NewVT, V1_LO, V1_HI);
5967 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5968 if (!isUndefLO && (V0_LO->getOpcode() != ISD::UNDEF ||
5969 V1_LO->getOpcode() != ISD::UNDEF))
5970 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V1_LO);
5972 if (!isUndefHI && (V0_HI->getOpcode() != ISD::UNDEF ||
5973 V1_HI->getOpcode() != ISD::UNDEF))
5974 HI = DAG.getNode(X86Opcode, DL, NewVT, V0_HI, V1_HI);
5977 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LO, HI);
5980 /// Try to fold a build_vector that performs an 'addsub' to an X86ISD::ADDSUB
5982 static SDValue LowerToAddSub(const BuildVectorSDNode *BV,
5983 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
5984 MVT VT = BV->getSimpleValueType(0);
5985 if ((!Subtarget->hasSSE3() || (VT != MVT::v4f32 && VT != MVT::v2f64)) &&
5986 (!Subtarget->hasAVX() || (VT != MVT::v8f32 && VT != MVT::v4f64)))
5990 unsigned NumElts = VT.getVectorNumElements();
5991 SDValue InVec0 = DAG.getUNDEF(VT);
5992 SDValue InVec1 = DAG.getUNDEF(VT);
5994 assert((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v4f32 ||
5995 VT == MVT::v2f64) && "build_vector with an invalid type found!");
5997 // Odd-numbered elements in the input build vector are obtained from
5998 // adding two integer/float elements.
5999 // Even-numbered elements in the input build vector are obtained from
6000 // subtracting two integer/float elements.
6001 unsigned ExpectedOpcode = ISD::FSUB;
6002 unsigned NextExpectedOpcode = ISD::FADD;
6003 bool AddFound = false;
6004 bool SubFound = false;
6006 for (unsigned i = 0, e = NumElts; i != e; ++i) {
6007 SDValue Op = BV->getOperand(i);
6009 // Skip 'undef' values.
6010 unsigned Opcode = Op.getOpcode();
6011 if (Opcode == ISD::UNDEF) {
6012 std::swap(ExpectedOpcode, NextExpectedOpcode);
6016 // Early exit if we found an unexpected opcode.
6017 if (Opcode != ExpectedOpcode)
6020 SDValue Op0 = Op.getOperand(0);
6021 SDValue Op1 = Op.getOperand(1);
6023 // Try to match the following pattern:
6024 // (BINOP (extract_vector_elt A, i), (extract_vector_elt B, i))
6025 // Early exit if we cannot match that sequence.
6026 if (Op0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
6027 Op1.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
6028 !isa<ConstantSDNode>(Op0.getOperand(1)) ||
6029 !isa<ConstantSDNode>(Op1.getOperand(1)) ||
6030 Op0.getOperand(1) != Op1.getOperand(1))
6033 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
6037 // We found a valid add/sub node. Update the information accordingly.
6043 // Update InVec0 and InVec1.
6044 if (InVec0.getOpcode() == ISD::UNDEF) {
6045 InVec0 = Op0.getOperand(0);
6046 if (InVec0.getSimpleValueType() != VT)
6049 if (InVec1.getOpcode() == ISD::UNDEF) {
6050 InVec1 = Op1.getOperand(0);
6051 if (InVec1.getSimpleValueType() != VT)
6055 // Make sure that operands in input to each add/sub node always
6056 // come from a same pair of vectors.
6057 if (InVec0 != Op0.getOperand(0)) {
6058 if (ExpectedOpcode == ISD::FSUB)
6061 // FADD is commutable. Try to commute the operands
6062 // and then test again.
6063 std::swap(Op0, Op1);
6064 if (InVec0 != Op0.getOperand(0))
6068 if (InVec1 != Op1.getOperand(0))
6071 // Update the pair of expected opcodes.
6072 std::swap(ExpectedOpcode, NextExpectedOpcode);
6075 // Don't try to fold this build_vector into an ADDSUB if the inputs are undef.
6076 if (AddFound && SubFound && InVec0.getOpcode() != ISD::UNDEF &&
6077 InVec1.getOpcode() != ISD::UNDEF)
6078 return DAG.getNode(X86ISD::ADDSUB, DL, VT, InVec0, InVec1);
6083 /// Lower BUILD_VECTOR to a horizontal add/sub operation if possible.
6084 static SDValue LowerToHorizontalOp(const BuildVectorSDNode *BV,
6085 const X86Subtarget *Subtarget,
6086 SelectionDAG &DAG) {
6087 MVT VT = BV->getSimpleValueType(0);
6088 unsigned NumElts = VT.getVectorNumElements();
6089 unsigned NumUndefsLO = 0;
6090 unsigned NumUndefsHI = 0;
6091 unsigned Half = NumElts/2;
6093 // Count the number of UNDEF operands in the build_vector in input.
6094 for (unsigned i = 0, e = Half; i != e; ++i)
6095 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
6098 for (unsigned i = Half, e = NumElts; i != e; ++i)
6099 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
6102 // Early exit if this is either a build_vector of all UNDEFs or all the
6103 // operands but one are UNDEF.
6104 if (NumUndefsLO + NumUndefsHI + 1 >= NumElts)
6108 SDValue InVec0, InVec1;
6109 if ((VT == MVT::v4f32 || VT == MVT::v2f64) && Subtarget->hasSSE3()) {
6110 // Try to match an SSE3 float HADD/HSUB.
6111 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
6112 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
6114 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
6115 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
6116 } else if ((VT == MVT::v4i32 || VT == MVT::v8i16) && Subtarget->hasSSSE3()) {
6117 // Try to match an SSSE3 integer HADD/HSUB.
6118 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
6119 return DAG.getNode(X86ISD::HADD, DL, VT, InVec0, InVec1);
6121 if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
6122 return DAG.getNode(X86ISD::HSUB, DL, VT, InVec0, InVec1);
6125 if (!Subtarget->hasAVX())
6128 if ((VT == MVT::v8f32 || VT == MVT::v4f64)) {
6129 // Try to match an AVX horizontal add/sub of packed single/double
6130 // precision floating point values from 256-bit vectors.
6131 SDValue InVec2, InVec3;
6132 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, Half, InVec0, InVec1) &&
6133 isHorizontalBinOp(BV, ISD::FADD, DAG, Half, NumElts, InVec2, InVec3) &&
6134 ((InVec0.getOpcode() == ISD::UNDEF ||
6135 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6136 ((InVec1.getOpcode() == ISD::UNDEF ||
6137 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6138 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
6140 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, Half, InVec0, InVec1) &&
6141 isHorizontalBinOp(BV, ISD::FSUB, DAG, Half, NumElts, InVec2, InVec3) &&
6142 ((InVec0.getOpcode() == ISD::UNDEF ||
6143 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6144 ((InVec1.getOpcode() == ISD::UNDEF ||
6145 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6146 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
6147 } else if (VT == MVT::v8i32 || VT == MVT::v16i16) {
6148 // Try to match an AVX2 horizontal add/sub of signed integers.
6149 SDValue InVec2, InVec3;
6151 bool CanFold = true;
6153 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, Half, InVec0, InVec1) &&
6154 isHorizontalBinOp(BV, ISD::ADD, DAG, Half, NumElts, InVec2, InVec3) &&
6155 ((InVec0.getOpcode() == ISD::UNDEF ||
6156 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6157 ((InVec1.getOpcode() == ISD::UNDEF ||
6158 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6159 X86Opcode = X86ISD::HADD;
6160 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, Half, InVec0, InVec1) &&
6161 isHorizontalBinOp(BV, ISD::SUB, DAG, Half, NumElts, InVec2, InVec3) &&
6162 ((InVec0.getOpcode() == ISD::UNDEF ||
6163 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6164 ((InVec1.getOpcode() == ISD::UNDEF ||
6165 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6166 X86Opcode = X86ISD::HSUB;
6171 // Fold this build_vector into a single horizontal add/sub.
6172 // Do this only if the target has AVX2.
6173 if (Subtarget->hasAVX2())
6174 return DAG.getNode(X86Opcode, DL, VT, InVec0, InVec1);
6176 // Do not try to expand this build_vector into a pair of horizontal
6177 // add/sub if we can emit a pair of scalar add/sub.
6178 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
6181 // Convert this build_vector into a pair of horizontal binop followed by
6183 bool isUndefLO = NumUndefsLO == Half;
6184 bool isUndefHI = NumUndefsHI == Half;
6185 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, false,
6186 isUndefLO, isUndefHI);
6190 if ((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v8i32 ||
6191 VT == MVT::v16i16) && Subtarget->hasAVX()) {
6193 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
6194 X86Opcode = X86ISD::HADD;
6195 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
6196 X86Opcode = X86ISD::HSUB;
6197 else if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
6198 X86Opcode = X86ISD::FHADD;
6199 else if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
6200 X86Opcode = X86ISD::FHSUB;
6204 // Don't try to expand this build_vector into a pair of horizontal add/sub
6205 // if we can simply emit a pair of scalar add/sub.
6206 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
6209 // Convert this build_vector into two horizontal add/sub followed by
6211 bool isUndefLO = NumUndefsLO == Half;
6212 bool isUndefHI = NumUndefsHI == Half;
6213 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, true,
6214 isUndefLO, isUndefHI);
6221 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
6224 MVT VT = Op.getSimpleValueType();
6225 MVT ExtVT = VT.getVectorElementType();
6226 unsigned NumElems = Op.getNumOperands();
6228 // Generate vectors for predicate vectors.
6229 if (VT.getVectorElementType() == MVT::i1 && Subtarget->hasAVX512())
6230 return LowerBUILD_VECTORvXi1(Op, DAG);
6232 // Vectors containing all zeros can be matched by pxor and xorps later
6233 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
6234 // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
6235 // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
6236 if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32)
6239 return getZeroVector(VT, Subtarget, DAG, dl);
6242 // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
6243 // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use
6244 // vpcmpeqd on 256-bit vectors.
6245 if (Subtarget->hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) {
6246 if (VT == MVT::v4i32 || (VT == MVT::v8i32 && Subtarget->hasInt256()))
6249 if (!VT.is512BitVector())
6250 return getOnesVector(VT, Subtarget, DAG, dl);
6253 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(Op.getNode());
6254 if (SDValue AddSub = LowerToAddSub(BV, Subtarget, DAG))
6256 if (SDValue HorizontalOp = LowerToHorizontalOp(BV, Subtarget, DAG))
6257 return HorizontalOp;
6258 if (SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG))
6261 unsigned EVTBits = ExtVT.getSizeInBits();
6263 unsigned NumZero = 0;
6264 unsigned NumNonZero = 0;
6265 uint64_t NonZeros = 0;
6266 bool IsAllConstants = true;
6267 SmallSet<SDValue, 8> Values;
6268 for (unsigned i = 0; i < NumElems; ++i) {
6269 SDValue Elt = Op.getOperand(i);
6270 if (Elt.getOpcode() == ISD::UNDEF)
6273 if (Elt.getOpcode() != ISD::Constant &&
6274 Elt.getOpcode() != ISD::ConstantFP)
6275 IsAllConstants = false;
6276 if (X86::isZeroNode(Elt))
6279 assert(i < sizeof(NonZeros) * 8); // Make sure the shift is within range.
6280 NonZeros |= ((uint64_t)1 << i);
6285 // All undef vector. Return an UNDEF. All zero vectors were handled above.
6286 if (NumNonZero == 0)
6287 return DAG.getUNDEF(VT);
6289 // Special case for single non-zero, non-undef, element.
6290 if (NumNonZero == 1) {
6291 unsigned Idx = countTrailingZeros(NonZeros);
6292 SDValue Item = Op.getOperand(Idx);
6294 // If this is an insertion of an i64 value on x86-32, and if the top bits of
6295 // the value are obviously zero, truncate the value to i32 and do the
6296 // insertion that way. Only do this if the value is non-constant or if the
6297 // value is a constant being inserted into element 0. It is cheaper to do
6298 // a constant pool load than it is to do a movd + shuffle.
6299 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
6300 (!IsAllConstants || Idx == 0)) {
6301 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
6303 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
6304 MVT VecVT = MVT::v4i32;
6306 // Truncate the value (which may itself be a constant) to i32, and
6307 // convert it to a vector with movd (S2V+shuffle to zero extend).
6308 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
6309 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
6310 return DAG.getBitcast(VT, getShuffleVectorZeroOrUndef(
6311 Item, Idx * 2, true, Subtarget, DAG));
6315 // If we have a constant or non-constant insertion into the low element of
6316 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
6317 // the rest of the elements. This will be matched as movd/movq/movss/movsd
6318 // depending on what the source datatype is.
6321 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6323 if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
6324 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
6325 if (VT.is512BitVector()) {
6326 SDValue ZeroVec = getZeroVector(VT, Subtarget, DAG, dl);
6327 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, ZeroVec,
6328 Item, DAG.getIntPtrConstant(0, dl));
6330 assert((VT.is128BitVector() || VT.is256BitVector()) &&
6331 "Expected an SSE value type!");
6332 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6333 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
6334 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6337 // We can't directly insert an i8 or i16 into a vector, so zero extend
6339 if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
6340 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
6341 if (VT.is256BitVector()) {
6342 if (Subtarget->hasAVX()) {
6343 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v8i32, Item);
6344 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6346 // Without AVX, we need to extend to a 128-bit vector and then
6347 // insert into the 256-bit vector.
6348 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
6349 SDValue ZeroVec = getZeroVector(MVT::v8i32, Subtarget, DAG, dl);
6350 Item = Insert128BitVector(ZeroVec, Item, 0, DAG, dl);
6353 assert(VT.is128BitVector() && "Expected an SSE value type!");
6354 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
6355 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6357 return DAG.getBitcast(VT, Item);
6361 // Is it a vector logical left shift?
6362 if (NumElems == 2 && Idx == 1 &&
6363 X86::isZeroNode(Op.getOperand(0)) &&
6364 !X86::isZeroNode(Op.getOperand(1))) {
6365 unsigned NumBits = VT.getSizeInBits();
6366 return getVShift(true, VT,
6367 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6368 VT, Op.getOperand(1)),
6369 NumBits/2, DAG, *this, dl);
6372 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
6375 // Otherwise, if this is a vector with i32 or f32 elements, and the element
6376 // is a non-constant being inserted into an element other than the low one,
6377 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
6378 // movd/movss) to move this into the low element, then shuffle it into
6380 if (EVTBits == 32) {
6381 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6382 return getShuffleVectorZeroOrUndef(Item, Idx, NumZero > 0, Subtarget, DAG);
6386 // Splat is obviously ok. Let legalizer expand it to a shuffle.
6387 if (Values.size() == 1) {
6388 if (EVTBits == 32) {
6389 // Instead of a shuffle like this:
6390 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
6391 // Check if it's possible to issue this instead.
6392 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
6393 unsigned Idx = countTrailingZeros(NonZeros);
6394 SDValue Item = Op.getOperand(Idx);
6395 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
6396 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
6401 // A vector full of immediates; various special cases are already
6402 // handled, so this is best done with a single constant-pool load.
6406 // For AVX-length vectors, see if we can use a vector load to get all of the
6407 // elements, otherwise build the individual 128-bit pieces and use
6408 // shuffles to put them in place.
6409 if (VT.is256BitVector() || VT.is512BitVector()) {
6410 SmallVector<SDValue, 64> V(Op->op_begin(), Op->op_begin() + NumElems);
6412 // Check for a build vector of consecutive loads.
6413 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
6416 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
6418 // Build both the lower and upper subvector.
6419 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
6420 makeArrayRef(&V[0], NumElems/2));
6421 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
6422 makeArrayRef(&V[NumElems / 2], NumElems/2));
6424 // Recreate the wider vector with the lower and upper part.
6425 if (VT.is256BitVector())
6426 return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
6427 return Concat256BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
6430 // Let legalizer expand 2-wide build_vectors.
6431 if (EVTBits == 64) {
6432 if (NumNonZero == 1) {
6433 // One half is zero or undef.
6434 unsigned Idx = countTrailingZeros(NonZeros);
6435 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
6436 Op.getOperand(Idx));
6437 return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG);
6442 // If element VT is < 32 bits, convert it to inserts into a zero vector.
6443 if (EVTBits == 8 && NumElems == 16)
6444 if (SDValue V = LowerBuildVectorv16i8(Op, NonZeros, NumNonZero, NumZero,
6445 DAG, Subtarget, *this))
6448 if (EVTBits == 16 && NumElems == 8)
6449 if (SDValue V = LowerBuildVectorv8i16(Op, NonZeros, NumNonZero, NumZero,
6450 DAG, Subtarget, *this))
6453 // If element VT is == 32 bits and has 4 elems, try to generate an INSERTPS
6454 if (EVTBits == 32 && NumElems == 4)
6455 if (SDValue V = LowerBuildVectorv4x32(Op, DAG, Subtarget, *this))
6458 // If element VT is == 32 bits, turn it into a number of shuffles.
6459 SmallVector<SDValue, 8> V(NumElems);
6460 if (NumElems == 4 && NumZero > 0) {
6461 for (unsigned i = 0; i < 4; ++i) {
6462 bool isZero = !(NonZeros & (1ULL << i));
6464 V[i] = getZeroVector(VT, Subtarget, DAG, dl);
6466 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6469 for (unsigned i = 0; i < 2; ++i) {
6470 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
6473 V[i] = V[i*2]; // Must be a zero vector.
6476 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
6479 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
6482 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
6487 bool Reverse1 = (NonZeros & 0x3) == 2;
6488 bool Reverse2 = ((NonZeros & (0x3 << 2)) >> 2) == 2;
6492 static_cast<int>(Reverse2 ? NumElems+1 : NumElems),
6493 static_cast<int>(Reverse2 ? NumElems : NumElems+1)
6495 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
6498 if (Values.size() > 1 && VT.is128BitVector()) {
6499 // Check for a build vector of consecutive loads.
6500 for (unsigned i = 0; i < NumElems; ++i)
6501 V[i] = Op.getOperand(i);
6503 // Check for elements which are consecutive loads.
6504 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
6507 // Check for a build vector from mostly shuffle plus few inserting.
6508 if (SDValue Sh = buildFromShuffleMostly(Op, DAG))
6511 // For SSE 4.1, use insertps to put the high elements into the low element.
6512 if (Subtarget->hasSSE41()) {
6514 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
6515 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
6517 Result = DAG.getUNDEF(VT);
6519 for (unsigned i = 1; i < NumElems; ++i) {
6520 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
6521 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
6522 Op.getOperand(i), DAG.getIntPtrConstant(i, dl));
6527 // Otherwise, expand into a number of unpckl*, start by extending each of
6528 // our (non-undef) elements to the full vector width with the element in the
6529 // bottom slot of the vector (which generates no code for SSE).
6530 for (unsigned i = 0; i < NumElems; ++i) {
6531 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
6532 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6534 V[i] = DAG.getUNDEF(VT);
6537 // Next, we iteratively mix elements, e.g. for v4f32:
6538 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
6539 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
6540 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
6541 unsigned EltStride = NumElems >> 1;
6542 while (EltStride != 0) {
6543 for (unsigned i = 0; i < EltStride; ++i) {
6544 // If V[i+EltStride] is undef and this is the first round of mixing,
6545 // then it is safe to just drop this shuffle: V[i] is already in the
6546 // right place, the one element (since it's the first round) being
6547 // inserted as undef can be dropped. This isn't safe for successive
6548 // rounds because they will permute elements within both vectors.
6549 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
6550 EltStride == NumElems/2)
6553 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
6562 // 256-bit AVX can use the vinsertf128 instruction
6563 // to create 256-bit vectors from two other 128-bit ones.
6564 static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
6566 MVT ResVT = Op.getSimpleValueType();
6568 assert((ResVT.is256BitVector() ||
6569 ResVT.is512BitVector()) && "Value type must be 256-/512-bit wide");
6571 SDValue V1 = Op.getOperand(0);
6572 SDValue V2 = Op.getOperand(1);
6573 unsigned NumElems = ResVT.getVectorNumElements();
6574 if (ResVT.is256BitVector())
6575 return Concat128BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6577 if (Op.getNumOperands() == 4) {
6578 MVT HalfVT = MVT::getVectorVT(ResVT.getVectorElementType(),
6579 ResVT.getVectorNumElements()/2);
6580 SDValue V3 = Op.getOperand(2);
6581 SDValue V4 = Op.getOperand(3);
6582 return Concat256BitVectors(Concat128BitVectors(V1, V2, HalfVT, NumElems/2, DAG, dl),
6583 Concat128BitVectors(V3, V4, HalfVT, NumElems/2, DAG, dl), ResVT, NumElems, DAG, dl);
6585 return Concat256BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6588 static SDValue LowerCONCAT_VECTORSvXi1(SDValue Op,
6589 const X86Subtarget *Subtarget,
6590 SelectionDAG & DAG) {
6592 MVT ResVT = Op.getSimpleValueType();
6593 unsigned NumOfOperands = Op.getNumOperands();
6595 assert(isPowerOf2_32(NumOfOperands) &&
6596 "Unexpected number of operands in CONCAT_VECTORS");
6598 SDValue Undef = DAG.getUNDEF(ResVT);
6599 if (NumOfOperands > 2) {
6600 // Specialize the cases when all, or all but one, of the operands are undef.
6601 unsigned NumOfDefinedOps = 0;
6603 for (unsigned i = 0; i < NumOfOperands; i++)
6604 if (!Op.getOperand(i).isUndef()) {
6608 if (NumOfDefinedOps == 0)
6610 if (NumOfDefinedOps == 1) {
6611 unsigned SubVecNumElts =
6612 Op.getOperand(OpIdx).getValueType().getVectorNumElements();
6613 SDValue IdxVal = DAG.getIntPtrConstant(SubVecNumElts * OpIdx, dl);
6614 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef,
6615 Op.getOperand(OpIdx), IdxVal);
6618 MVT HalfVT = MVT::getVectorVT(ResVT.getVectorElementType(),
6619 ResVT.getVectorNumElements()/2);
6620 SmallVector<SDValue, 2> Ops;
6621 for (unsigned i = 0; i < NumOfOperands/2; i++)
6622 Ops.push_back(Op.getOperand(i));
6623 SDValue Lo = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6625 for (unsigned i = NumOfOperands/2; i < NumOfOperands; i++)
6626 Ops.push_back(Op.getOperand(i));
6627 SDValue Hi = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6628 return DAG.getNode(ISD::CONCAT_VECTORS, dl, ResVT, Lo, Hi);
6632 SDValue V1 = Op.getOperand(0);
6633 SDValue V2 = Op.getOperand(1);
6634 unsigned NumElems = ResVT.getVectorNumElements();
6635 assert(V1.getValueType() == V2.getValueType() &&
6636 V1.getValueType().getVectorNumElements() == NumElems/2 &&
6637 "Unexpected operands in CONCAT_VECTORS");
6639 if (ResVT.getSizeInBits() >= 16)
6640 return Op; // The operation is legal with KUNPCK
6642 bool IsZeroV1 = ISD::isBuildVectorAllZeros(V1.getNode());
6643 bool IsZeroV2 = ISD::isBuildVectorAllZeros(V2.getNode());
6644 SDValue ZeroVec = getZeroVector(ResVT, Subtarget, DAG, dl);
6645 if (IsZeroV1 && IsZeroV2)
6648 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
6650 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V1, ZeroIdx);
6652 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, ZeroVec, V1, ZeroIdx);
6654 SDValue IdxVal = DAG.getIntPtrConstant(NumElems/2, dl);
6656 V2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V2, IdxVal);
6659 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, ZeroVec, V2, IdxVal);
6661 V1 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V1, ZeroIdx);
6662 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, V1, V2, IdxVal);
6665 static SDValue LowerCONCAT_VECTORS(SDValue Op,
6666 const X86Subtarget *Subtarget,
6667 SelectionDAG &DAG) {
6668 MVT VT = Op.getSimpleValueType();
6669 if (VT.getVectorElementType() == MVT::i1)
6670 return LowerCONCAT_VECTORSvXi1(Op, Subtarget, DAG);
6672 assert((VT.is256BitVector() && Op.getNumOperands() == 2) ||
6673 (VT.is512BitVector() && (Op.getNumOperands() == 2 ||
6674 Op.getNumOperands() == 4)));
6676 // AVX can use the vinsertf128 instruction to create 256-bit vectors
6677 // from two other 128-bit ones.
6679 // 512-bit vector may contain 2 256-bit vectors or 4 128-bit vectors
6680 return LowerAVXCONCAT_VECTORS(Op, DAG);
6683 //===----------------------------------------------------------------------===//
6684 // Vector shuffle lowering
6686 // This is an experimental code path for lowering vector shuffles on x86. It is
6687 // designed to handle arbitrary vector shuffles and blends, gracefully
6688 // degrading performance as necessary. It works hard to recognize idiomatic
6689 // shuffles and lower them to optimal instruction patterns without leaving
6690 // a framework that allows reasonably efficient handling of all vector shuffle
6692 //===----------------------------------------------------------------------===//
6694 /// \brief Tiny helper function to identify a no-op mask.
6696 /// This is a somewhat boring predicate function. It checks whether the mask
6697 /// array input, which is assumed to be a single-input shuffle mask of the kind
6698 /// used by the X86 shuffle instructions (not a fully general
6699 /// ShuffleVectorSDNode mask) requires any shuffles to occur. Both undef and an
6700 /// in-place shuffle are 'no-op's.
6701 static bool isNoopShuffleMask(ArrayRef<int> Mask) {
6702 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6703 if (Mask[i] != -1 && Mask[i] != i)
6708 /// \brief Helper function to classify a mask as a single-input mask.
6710 /// This isn't a generic single-input test because in the vector shuffle
6711 /// lowering we canonicalize single inputs to be the first input operand. This
6712 /// means we can more quickly test for a single input by only checking whether
6713 /// an input from the second operand exists. We also assume that the size of
6714 /// mask corresponds to the size of the input vectors which isn't true in the
6715 /// fully general case.
6716 static bool isSingleInputShuffleMask(ArrayRef<int> Mask) {
6718 if (M >= (int)Mask.size())
6723 /// \brief Test whether there are elements crossing 128-bit lanes in this
6726 /// X86 divides up its shuffles into in-lane and cross-lane shuffle operations
6727 /// and we routinely test for these.
6728 static bool is128BitLaneCrossingShuffleMask(MVT VT, ArrayRef<int> Mask) {
6729 int LaneSize = 128 / VT.getScalarSizeInBits();
6730 int Size = Mask.size();
6731 for (int i = 0; i < Size; ++i)
6732 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
6737 /// \brief Test whether a shuffle mask is equivalent within each 128-bit lane.
6739 /// This checks a shuffle mask to see if it is performing the same
6740 /// 128-bit lane-relative shuffle in each 128-bit lane. This trivially implies
6741 /// that it is also not lane-crossing. It may however involve a blend from the
6742 /// same lane of a second vector.
6744 /// The specific repeated shuffle mask is populated in \p RepeatedMask, as it is
6745 /// non-trivial to compute in the face of undef lanes. The representation is
6746 /// *not* suitable for use with existing 128-bit shuffles as it will contain
6747 /// entries from both V1 and V2 inputs to the wider mask.
6749 is128BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask,
6750 SmallVectorImpl<int> &RepeatedMask) {
6751 int LaneSize = 128 / VT.getScalarSizeInBits();
6752 RepeatedMask.resize(LaneSize, -1);
6753 int Size = Mask.size();
6754 for (int i = 0; i < Size; ++i) {
6757 if ((Mask[i] % Size) / LaneSize != i / LaneSize)
6758 // This entry crosses lanes, so there is no way to model this shuffle.
6761 // Ok, handle the in-lane shuffles by detecting if and when they repeat.
6762 if (RepeatedMask[i % LaneSize] == -1)
6763 // This is the first non-undef entry in this slot of a 128-bit lane.
6764 RepeatedMask[i % LaneSize] =
6765 Mask[i] < Size ? Mask[i] % LaneSize : Mask[i] % LaneSize + Size;
6766 else if (RepeatedMask[i % LaneSize] + (i / LaneSize) * LaneSize != Mask[i])
6767 // Found a mismatch with the repeated mask.
6773 /// \brief Checks whether a shuffle mask is equivalent to an explicit list of
6776 /// This is a fast way to test a shuffle mask against a fixed pattern:
6778 /// if (isShuffleEquivalent(Mask, 3, 2, {1, 0})) { ... }
6780 /// It returns true if the mask is exactly as wide as the argument list, and
6781 /// each element of the mask is either -1 (signifying undef) or the value given
6782 /// in the argument.
6783 static bool isShuffleEquivalent(SDValue V1, SDValue V2, ArrayRef<int> Mask,
6784 ArrayRef<int> ExpectedMask) {
6785 if (Mask.size() != ExpectedMask.size())
6788 int Size = Mask.size();
6790 // If the values are build vectors, we can look through them to find
6791 // equivalent inputs that make the shuffles equivalent.
6792 auto *BV1 = dyn_cast<BuildVectorSDNode>(V1);
6793 auto *BV2 = dyn_cast<BuildVectorSDNode>(V2);
6795 for (int i = 0; i < Size; ++i)
6796 if (Mask[i] != -1 && Mask[i] != ExpectedMask[i]) {
6797 auto *MaskBV = Mask[i] < Size ? BV1 : BV2;
6798 auto *ExpectedBV = ExpectedMask[i] < Size ? BV1 : BV2;
6799 if (!MaskBV || !ExpectedBV ||
6800 MaskBV->getOperand(Mask[i] % Size) !=
6801 ExpectedBV->getOperand(ExpectedMask[i] % Size))
6808 /// \brief Get a 4-lane 8-bit shuffle immediate for a mask.
6810 /// This helper function produces an 8-bit shuffle immediate corresponding to
6811 /// the ubiquitous shuffle encoding scheme used in x86 instructions for
6812 /// shuffling 4 lanes. It can be used with most of the PSHUF instructions for
6815 /// NB: We rely heavily on "undef" masks preserving the input lane.
6816 static SDValue getV4X86ShuffleImm8ForMask(ArrayRef<int> Mask, SDLoc DL,
6817 SelectionDAG &DAG) {
6818 assert(Mask.size() == 4 && "Only 4-lane shuffle masks");
6819 assert(Mask[0] >= -1 && Mask[0] < 4 && "Out of bound mask element!");
6820 assert(Mask[1] >= -1 && Mask[1] < 4 && "Out of bound mask element!");
6821 assert(Mask[2] >= -1 && Mask[2] < 4 && "Out of bound mask element!");
6822 assert(Mask[3] >= -1 && Mask[3] < 4 && "Out of bound mask element!");
6825 Imm |= (Mask[0] == -1 ? 0 : Mask[0]) << 0;
6826 Imm |= (Mask[1] == -1 ? 1 : Mask[1]) << 2;
6827 Imm |= (Mask[2] == -1 ? 2 : Mask[2]) << 4;
6828 Imm |= (Mask[3] == -1 ? 3 : Mask[3]) << 6;
6829 return DAG.getConstant(Imm, DL, MVT::i8);
6832 /// \brief Compute whether each element of a shuffle is zeroable.
6834 /// A "zeroable" vector shuffle element is one which can be lowered to zero.
6835 /// Either it is an undef element in the shuffle mask, the element of the input
6836 /// referenced is undef, or the element of the input referenced is known to be
6837 /// zero. Many x86 shuffles can zero lanes cheaply and we often want to handle
6838 /// as many lanes with this technique as possible to simplify the remaining
6840 static SmallBitVector computeZeroableShuffleElements(ArrayRef<int> Mask,
6841 SDValue V1, SDValue V2) {
6842 SmallBitVector Zeroable(Mask.size(), false);
6844 while (V1.getOpcode() == ISD::BITCAST)
6845 V1 = V1->getOperand(0);
6846 while (V2.getOpcode() == ISD::BITCAST)
6847 V2 = V2->getOperand(0);
6849 bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
6850 bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
6852 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6854 // Handle the easy cases.
6855 if (M < 0 || (M >= 0 && M < Size && V1IsZero) || (M >= Size && V2IsZero)) {
6860 // If this is an index into a build_vector node (which has the same number
6861 // of elements), dig out the input value and use it.
6862 SDValue V = M < Size ? V1 : V2;
6863 if (V.getOpcode() != ISD::BUILD_VECTOR || Size != (int)V.getNumOperands())
6866 SDValue Input = V.getOperand(M % Size);
6867 // The UNDEF opcode check really should be dead code here, but not quite
6868 // worth asserting on (it isn't invalid, just unexpected).
6869 if (Input.getOpcode() == ISD::UNDEF || X86::isZeroNode(Input))
6876 // X86 has dedicated unpack instructions that can handle specific blend
6877 // operations: UNPCKH and UNPCKL.
6878 static SDValue lowerVectorShuffleWithUNPCK(SDLoc DL, MVT VT, ArrayRef<int> Mask,
6879 SDValue V1, SDValue V2,
6880 SelectionDAG &DAG) {
6881 int NumElts = VT.getVectorNumElements();
6882 int NumEltsInLane = 128 / VT.getScalarSizeInBits();
6883 SmallVector<int, 8> Unpckl;
6884 SmallVector<int, 8> Unpckh;
6886 for (int i = 0; i < NumElts; ++i) {
6887 unsigned LaneStart = (i / NumEltsInLane) * NumEltsInLane;
6888 int LoPos = (i % NumEltsInLane) / 2 + LaneStart + NumElts * (i % 2);
6889 int HiPos = LoPos + NumEltsInLane / 2;
6890 Unpckl.push_back(LoPos);
6891 Unpckh.push_back(HiPos);
6894 if (isShuffleEquivalent(V1, V2, Mask, Unpckl))
6895 return DAG.getNode(X86ISD::UNPCKL, DL, VT, V1, V2);
6896 if (isShuffleEquivalent(V1, V2, Mask, Unpckh))
6897 return DAG.getNode(X86ISD::UNPCKH, DL, VT, V1, V2);
6899 // Commute and try again.
6900 ShuffleVectorSDNode::commuteMask(Unpckl);
6901 if (isShuffleEquivalent(V1, V2, Mask, Unpckl))
6902 return DAG.getNode(X86ISD::UNPCKL, DL, VT, V2, V1);
6904 ShuffleVectorSDNode::commuteMask(Unpckh);
6905 if (isShuffleEquivalent(V1, V2, Mask, Unpckh))
6906 return DAG.getNode(X86ISD::UNPCKH, DL, VT, V2, V1);
6911 /// \brief Try to emit a bitmask instruction for a shuffle.
6913 /// This handles cases where we can model a blend exactly as a bitmask due to
6914 /// one of the inputs being zeroable.
6915 static SDValue lowerVectorShuffleAsBitMask(SDLoc DL, MVT VT, SDValue V1,
6916 SDValue V2, ArrayRef<int> Mask,
6917 SelectionDAG &DAG) {
6918 MVT EltVT = VT.getVectorElementType();
6919 int NumEltBits = EltVT.getSizeInBits();
6920 MVT IntEltVT = MVT::getIntegerVT(NumEltBits);
6921 SDValue Zero = DAG.getConstant(0, DL, IntEltVT);
6922 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
6924 if (EltVT.isFloatingPoint()) {
6925 Zero = DAG.getBitcast(EltVT, Zero);
6926 AllOnes = DAG.getBitcast(EltVT, AllOnes);
6928 SmallVector<SDValue, 16> VMaskOps(Mask.size(), Zero);
6929 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6931 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6934 if (Mask[i] % Size != i)
6935 return SDValue(); // Not a blend.
6937 V = Mask[i] < Size ? V1 : V2;
6938 else if (V != (Mask[i] < Size ? V1 : V2))
6939 return SDValue(); // Can only let one input through the mask.
6941 VMaskOps[i] = AllOnes;
6944 return SDValue(); // No non-zeroable elements!
6946 SDValue VMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, VMaskOps);
6947 V = DAG.getNode(VT.isFloatingPoint()
6948 ? (unsigned) X86ISD::FAND : (unsigned) ISD::AND,
6953 /// \brief Try to emit a blend instruction for a shuffle using bit math.
6955 /// This is used as a fallback approach when first class blend instructions are
6956 /// unavailable. Currently it is only suitable for integer vectors, but could
6957 /// be generalized for floating point vectors if desirable.
6958 static SDValue lowerVectorShuffleAsBitBlend(SDLoc DL, MVT VT, SDValue V1,
6959 SDValue V2, ArrayRef<int> Mask,
6960 SelectionDAG &DAG) {
6961 assert(VT.isInteger() && "Only supports integer vector types!");
6962 MVT EltVT = VT.getVectorElementType();
6963 int NumEltBits = EltVT.getSizeInBits();
6964 SDValue Zero = DAG.getConstant(0, DL, EltVT);
6965 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
6967 SmallVector<SDValue, 16> MaskOps;
6968 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6969 if (Mask[i] != -1 && Mask[i] != i && Mask[i] != i + Size)
6970 return SDValue(); // Shuffled input!
6971 MaskOps.push_back(Mask[i] < Size ? AllOnes : Zero);
6974 SDValue V1Mask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, MaskOps);
6975 V1 = DAG.getNode(ISD::AND, DL, VT, V1, V1Mask);
6976 // We have to cast V2 around.
6977 MVT MaskVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
6978 V2 = DAG.getBitcast(VT, DAG.getNode(X86ISD::ANDNP, DL, MaskVT,
6979 DAG.getBitcast(MaskVT, V1Mask),
6980 DAG.getBitcast(MaskVT, V2)));
6981 return DAG.getNode(ISD::OR, DL, VT, V1, V2);
6984 /// \brief Try to emit a blend instruction for a shuffle.
6986 /// This doesn't do any checks for the availability of instructions for blending
6987 /// these values. It relies on the availability of the X86ISD::BLENDI pattern to
6988 /// be matched in the backend with the type given. What it does check for is
6989 /// that the shuffle mask is a blend, or convertible into a blend with zero.
6990 static SDValue lowerVectorShuffleAsBlend(SDLoc DL, MVT VT, SDValue V1,
6991 SDValue V2, ArrayRef<int> Original,
6992 const X86Subtarget *Subtarget,
6993 SelectionDAG &DAG) {
6994 bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
6995 bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
6996 SmallVector<int, 8> Mask(Original.begin(), Original.end());
6997 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6998 bool ForceV1Zero = false, ForceV2Zero = false;
7000 // Attempt to generate the binary blend mask. If an input is zero then
7001 // we can use any lane.
7002 // TODO: generalize the zero matching to any scalar like isShuffleEquivalent.
7003 unsigned BlendMask = 0;
7004 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
7010 if (M == i + Size) {
7011 BlendMask |= 1u << i;
7022 BlendMask |= 1u << i;
7027 return SDValue(); // Shuffled input!
7030 // Create a REAL zero vector - ISD::isBuildVectorAllZeros allows UNDEFs.
7032 V1 = getZeroVector(VT, Subtarget, DAG, DL);
7034 V2 = getZeroVector(VT, Subtarget, DAG, DL);
7036 auto ScaleBlendMask = [](unsigned BlendMask, int Size, int Scale) {
7037 unsigned ScaledMask = 0;
7038 for (int i = 0; i != Size; ++i)
7039 if (BlendMask & (1u << i))
7040 for (int j = 0; j != Scale; ++j)
7041 ScaledMask |= 1u << (i * Scale + j);
7045 switch (VT.SimpleTy) {
7050 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V2,
7051 DAG.getConstant(BlendMask, DL, MVT::i8));
7055 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
7059 // If we have AVX2 it is faster to use VPBLENDD when the shuffle fits into
7060 // that instruction.
7061 if (Subtarget->hasAVX2()) {
7062 // Scale the blend by the number of 32-bit dwords per element.
7063 int Scale = VT.getScalarSizeInBits() / 32;
7064 BlendMask = ScaleBlendMask(BlendMask, Mask.size(), Scale);
7065 MVT BlendVT = VT.getSizeInBits() > 128 ? MVT::v8i32 : MVT::v4i32;
7066 V1 = DAG.getBitcast(BlendVT, V1);
7067 V2 = DAG.getBitcast(BlendVT, V2);
7068 return DAG.getBitcast(
7069 VT, DAG.getNode(X86ISD::BLENDI, DL, BlendVT, V1, V2,
7070 DAG.getConstant(BlendMask, DL, MVT::i8)));
7074 // For integer shuffles we need to expand the mask and cast the inputs to
7075 // v8i16s prior to blending.
7076 int Scale = 8 / VT.getVectorNumElements();
7077 BlendMask = ScaleBlendMask(BlendMask, Mask.size(), Scale);
7078 V1 = DAG.getBitcast(MVT::v8i16, V1);
7079 V2 = DAG.getBitcast(MVT::v8i16, V2);
7080 return DAG.getBitcast(VT,
7081 DAG.getNode(X86ISD::BLENDI, DL, MVT::v8i16, V1, V2,
7082 DAG.getConstant(BlendMask, DL, MVT::i8)));
7086 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
7087 SmallVector<int, 8> RepeatedMask;
7088 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
7089 // We can lower these with PBLENDW which is mirrored across 128-bit lanes.
7090 assert(RepeatedMask.size() == 8 && "Repeated mask size doesn't match!");
7092 for (int i = 0; i < 8; ++i)
7093 if (RepeatedMask[i] >= 16)
7094 BlendMask |= 1u << i;
7095 return DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2,
7096 DAG.getConstant(BlendMask, DL, MVT::i8));
7102 assert((VT.is128BitVector() || Subtarget->hasAVX2()) &&
7103 "256-bit byte-blends require AVX2 support!");
7105 // Attempt to lower to a bitmask if we can. VPAND is faster than VPBLENDVB.
7106 if (SDValue Masked = lowerVectorShuffleAsBitMask(DL, VT, V1, V2, Mask, DAG))
7109 // Scale the blend by the number of bytes per element.
7110 int Scale = VT.getScalarSizeInBits() / 8;
7112 // This form of blend is always done on bytes. Compute the byte vector
7114 MVT BlendVT = MVT::getVectorVT(MVT::i8, VT.getSizeInBits() / 8);
7116 // Compute the VSELECT mask. Note that VSELECT is really confusing in the
7117 // mix of LLVM's code generator and the x86 backend. We tell the code
7118 // generator that boolean values in the elements of an x86 vector register
7119 // are -1 for true and 0 for false. We then use the LLVM semantics of 'true'
7120 // mapping a select to operand #1, and 'false' mapping to operand #2. The
7121 // reality in x86 is that vector masks (pre-AVX-512) use only the high bit
7122 // of the element (the remaining are ignored) and 0 in that high bit would
7123 // mean operand #1 while 1 in the high bit would mean operand #2. So while
7124 // the LLVM model for boolean values in vector elements gets the relevant
7125 // bit set, it is set backwards and over constrained relative to x86's
7127 SmallVector<SDValue, 32> VSELECTMask;
7128 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7129 for (int j = 0; j < Scale; ++j)
7130 VSELECTMask.push_back(
7131 Mask[i] < 0 ? DAG.getUNDEF(MVT::i8)
7132 : DAG.getConstant(Mask[i] < Size ? -1 : 0, DL,
7135 V1 = DAG.getBitcast(BlendVT, V1);
7136 V2 = DAG.getBitcast(BlendVT, V2);
7137 return DAG.getBitcast(VT, DAG.getNode(ISD::VSELECT, DL, BlendVT,
7138 DAG.getNode(ISD::BUILD_VECTOR, DL,
7139 BlendVT, VSELECTMask),
7144 llvm_unreachable("Not a supported integer vector type!");
7148 /// \brief Try to lower as a blend of elements from two inputs followed by
7149 /// a single-input permutation.
7151 /// This matches the pattern where we can blend elements from two inputs and
7152 /// then reduce the shuffle to a single-input permutation.
7153 static SDValue lowerVectorShuffleAsBlendAndPermute(SDLoc DL, MVT VT, SDValue V1,
7156 SelectionDAG &DAG) {
7157 // We build up the blend mask while checking whether a blend is a viable way
7158 // to reduce the shuffle.
7159 SmallVector<int, 32> BlendMask(Mask.size(), -1);
7160 SmallVector<int, 32> PermuteMask(Mask.size(), -1);
7162 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
7166 assert(Mask[i] < Size * 2 && "Shuffle input is out of bounds.");
7168 if (BlendMask[Mask[i] % Size] == -1)
7169 BlendMask[Mask[i] % Size] = Mask[i];
7170 else if (BlendMask[Mask[i] % Size] != Mask[i])
7171 return SDValue(); // Can't blend in the needed input!
7173 PermuteMask[i] = Mask[i] % Size;
7176 SDValue V = DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
7177 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), PermuteMask);
7180 /// \brief Generic routine to decompose a shuffle and blend into indepndent
7181 /// blends and permutes.
7183 /// This matches the extremely common pattern for handling combined
7184 /// shuffle+blend operations on newer X86 ISAs where we have very fast blend
7185 /// operations. It will try to pick the best arrangement of shuffles and
7187 static SDValue lowerVectorShuffleAsDecomposedShuffleBlend(SDLoc DL, MVT VT,
7191 SelectionDAG &DAG) {
7192 // Shuffle the input elements into the desired positions in V1 and V2 and
7193 // blend them together.
7194 SmallVector<int, 32> V1Mask(Mask.size(), -1);
7195 SmallVector<int, 32> V2Mask(Mask.size(), -1);
7196 SmallVector<int, 32> BlendMask(Mask.size(), -1);
7197 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7198 if (Mask[i] >= 0 && Mask[i] < Size) {
7199 V1Mask[i] = Mask[i];
7201 } else if (Mask[i] >= Size) {
7202 V2Mask[i] = Mask[i] - Size;
7203 BlendMask[i] = i + Size;
7206 // Try to lower with the simpler initial blend strategy unless one of the
7207 // input shuffles would be a no-op. We prefer to shuffle inputs as the
7208 // shuffle may be able to fold with a load or other benefit. However, when
7209 // we'll have to do 2x as many shuffles in order to achieve this, blending
7210 // first is a better strategy.
7211 if (!isNoopShuffleMask(V1Mask) && !isNoopShuffleMask(V2Mask))
7212 if (SDValue BlendPerm =
7213 lowerVectorShuffleAsBlendAndPermute(DL, VT, V1, V2, Mask, DAG))
7216 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
7217 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
7218 return DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
7221 /// \brief Try to lower a vector shuffle as a byte rotation.
7223 /// SSSE3 has a generic PALIGNR instruction in x86 that will do an arbitrary
7224 /// byte-rotation of the concatenation of two vectors; pre-SSSE3 can use
7225 /// a PSRLDQ/PSLLDQ/POR pattern to get a similar effect. This routine will
7226 /// try to generically lower a vector shuffle through such an pattern. It
7227 /// does not check for the profitability of lowering either as PALIGNR or
7228 /// PSRLDQ/PSLLDQ/POR, only whether the mask is valid to lower in that form.
7229 /// This matches shuffle vectors that look like:
7231 /// v8i16 [11, 12, 13, 14, 15, 0, 1, 2]
7233 /// Essentially it concatenates V1 and V2, shifts right by some number of
7234 /// elements, and takes the low elements as the result. Note that while this is
7235 /// specified as a *right shift* because x86 is little-endian, it is a *left
7236 /// rotate* of the vector lanes.
7237 static SDValue lowerVectorShuffleAsByteRotate(SDLoc DL, MVT VT, SDValue V1,
7240 const X86Subtarget *Subtarget,
7241 SelectionDAG &DAG) {
7242 assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!");
7244 int NumElts = Mask.size();
7245 int NumLanes = VT.getSizeInBits() / 128;
7246 int NumLaneElts = NumElts / NumLanes;
7248 // We need to detect various ways of spelling a rotation:
7249 // [11, 12, 13, 14, 15, 0, 1, 2]
7250 // [-1, 12, 13, 14, -1, -1, 1, -1]
7251 // [-1, -1, -1, -1, -1, -1, 1, 2]
7252 // [ 3, 4, 5, 6, 7, 8, 9, 10]
7253 // [-1, 4, 5, 6, -1, -1, 9, -1]
7254 // [-1, 4, 5, 6, -1, -1, -1, -1]
7257 for (int l = 0; l < NumElts; l += NumLaneElts) {
7258 for (int i = 0; i < NumLaneElts; ++i) {
7259 if (Mask[l + i] == -1)
7261 assert(Mask[l + i] >= 0 && "Only -1 is a valid negative mask element!");
7263 // Get the mod-Size index and lane correct it.
7264 int LaneIdx = (Mask[l + i] % NumElts) - l;
7265 // Make sure it was in this lane.
7266 if (LaneIdx < 0 || LaneIdx >= NumLaneElts)
7269 // Determine where a rotated vector would have started.
7270 int StartIdx = i - LaneIdx;
7272 // The identity rotation isn't interesting, stop.
7275 // If we found the tail of a vector the rotation must be the missing
7276 // front. If we found the head of a vector, it must be how much of the
7278 int CandidateRotation = StartIdx < 0 ? -StartIdx : NumLaneElts - StartIdx;
7281 Rotation = CandidateRotation;
7282 else if (Rotation != CandidateRotation)
7283 // The rotations don't match, so we can't match this mask.
7286 // Compute which value this mask is pointing at.
7287 SDValue MaskV = Mask[l + i] < NumElts ? V1 : V2;
7289 // Compute which of the two target values this index should be assigned
7290 // to. This reflects whether the high elements are remaining or the low
7291 // elements are remaining.
7292 SDValue &TargetV = StartIdx < 0 ? Hi : Lo;
7294 // Either set up this value if we've not encountered it before, or check
7295 // that it remains consistent.
7298 else if (TargetV != MaskV)
7299 // This may be a rotation, but it pulls from the inputs in some
7300 // unsupported interleaving.
7305 // Check that we successfully analyzed the mask, and normalize the results.
7306 assert(Rotation != 0 && "Failed to locate a viable rotation!");
7307 assert((Lo || Hi) && "Failed to find a rotated input vector!");
7313 // The actual rotate instruction rotates bytes, so we need to scale the
7314 // rotation based on how many bytes are in the vector lane.
7315 int Scale = 16 / NumLaneElts;
7317 // SSSE3 targets can use the palignr instruction.
7318 if (Subtarget->hasSSSE3()) {
7319 // Cast the inputs to i8 vector of correct length to match PALIGNR.
7320 MVT AlignVT = MVT::getVectorVT(MVT::i8, 16 * NumLanes);
7321 Lo = DAG.getBitcast(AlignVT, Lo);
7322 Hi = DAG.getBitcast(AlignVT, Hi);
7324 return DAG.getBitcast(
7325 VT, DAG.getNode(X86ISD::PALIGNR, DL, AlignVT, Lo, Hi,
7326 DAG.getConstant(Rotation * Scale, DL, MVT::i8)));
7329 assert(VT.is128BitVector() &&
7330 "Rotate-based lowering only supports 128-bit lowering!");
7331 assert(Mask.size() <= 16 &&
7332 "Can shuffle at most 16 bytes in a 128-bit vector!");
7334 // Default SSE2 implementation
7335 int LoByteShift = 16 - Rotation * Scale;
7336 int HiByteShift = Rotation * Scale;
7338 // Cast the inputs to v2i64 to match PSLLDQ/PSRLDQ.
7339 Lo = DAG.getBitcast(MVT::v2i64, Lo);
7340 Hi = DAG.getBitcast(MVT::v2i64, Hi);
7342 SDValue LoShift = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v2i64, Lo,
7343 DAG.getConstant(LoByteShift, DL, MVT::i8));
7344 SDValue HiShift = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v2i64, Hi,
7345 DAG.getConstant(HiByteShift, DL, MVT::i8));
7346 return DAG.getBitcast(VT,
7347 DAG.getNode(ISD::OR, DL, MVT::v2i64, LoShift, HiShift));
7350 /// \brief Try to lower a vector shuffle as a bit shift (shifts in zeros).
7352 /// Attempts to match a shuffle mask against the PSLL(W/D/Q/DQ) and
7353 /// PSRL(W/D/Q/DQ) SSE2 and AVX2 logical bit-shift instructions. The function
7354 /// matches elements from one of the input vectors shuffled to the left or
7355 /// right with zeroable elements 'shifted in'. It handles both the strictly
7356 /// bit-wise element shifts and the byte shift across an entire 128-bit double
7359 /// PSHL : (little-endian) left bit shift.
7360 /// [ zz, 0, zz, 2 ]
7361 /// [ -1, 4, zz, -1 ]
7362 /// PSRL : (little-endian) right bit shift.
7364 /// [ -1, -1, 7, zz]
7365 /// PSLLDQ : (little-endian) left byte shift
7366 /// [ zz, 0, 1, 2, 3, 4, 5, 6]
7367 /// [ zz, zz, -1, -1, 2, 3, 4, -1]
7368 /// [ zz, zz, zz, zz, zz, zz, -1, 1]
7369 /// PSRLDQ : (little-endian) right byte shift
7370 /// [ 5, 6, 7, zz, zz, zz, zz, zz]
7371 /// [ -1, 5, 6, 7, zz, zz, zz, zz]
7372 /// [ 1, 2, -1, -1, -1, -1, zz, zz]
7373 static SDValue lowerVectorShuffleAsShift(SDLoc DL, MVT VT, SDValue V1,
7374 SDValue V2, ArrayRef<int> Mask,
7375 SelectionDAG &DAG) {
7376 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7378 int Size = Mask.size();
7379 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
7381 auto CheckZeros = [&](int Shift, int Scale, bool Left) {
7382 for (int i = 0; i < Size; i += Scale)
7383 for (int j = 0; j < Shift; ++j)
7384 if (!Zeroable[i + j + (Left ? 0 : (Scale - Shift))])
7390 auto MatchShift = [&](int Shift, int Scale, bool Left, SDValue V) {
7391 for (int i = 0; i != Size; i += Scale) {
7392 unsigned Pos = Left ? i + Shift : i;
7393 unsigned Low = Left ? i : i + Shift;
7394 unsigned Len = Scale - Shift;
7395 if (!isSequentialOrUndefInRange(Mask, Pos, Len,
7396 Low + (V == V1 ? 0 : Size)))
7400 int ShiftEltBits = VT.getScalarSizeInBits() * Scale;
7401 bool ByteShift = ShiftEltBits > 64;
7402 unsigned OpCode = Left ? (ByteShift ? X86ISD::VSHLDQ : X86ISD::VSHLI)
7403 : (ByteShift ? X86ISD::VSRLDQ : X86ISD::VSRLI);
7404 int ShiftAmt = Shift * VT.getScalarSizeInBits() / (ByteShift ? 8 : 1);
7406 // Normalize the scale for byte shifts to still produce an i64 element
7408 Scale = ByteShift ? Scale / 2 : Scale;
7410 // We need to round trip through the appropriate type for the shift.
7411 MVT ShiftSVT = MVT::getIntegerVT(VT.getScalarSizeInBits() * Scale);
7412 MVT ShiftVT = MVT::getVectorVT(ShiftSVT, Size / Scale);
7413 assert(DAG.getTargetLoweringInfo().isTypeLegal(ShiftVT) &&
7414 "Illegal integer vector type");
7415 V = DAG.getBitcast(ShiftVT, V);
7417 V = DAG.getNode(OpCode, DL, ShiftVT, V,
7418 DAG.getConstant(ShiftAmt, DL, MVT::i8));
7419 return DAG.getBitcast(VT, V);
7422 // SSE/AVX supports logical shifts up to 64-bit integers - so we can just
7423 // keep doubling the size of the integer elements up to that. We can
7424 // then shift the elements of the integer vector by whole multiples of
7425 // their width within the elements of the larger integer vector. Test each
7426 // multiple to see if we can find a match with the moved element indices
7427 // and that the shifted in elements are all zeroable.
7428 for (int Scale = 2; Scale * VT.getScalarSizeInBits() <= 128; Scale *= 2)
7429 for (int Shift = 1; Shift != Scale; ++Shift)
7430 for (bool Left : {true, false})
7431 if (CheckZeros(Shift, Scale, Left))
7432 for (SDValue V : {V1, V2})
7433 if (SDValue Match = MatchShift(Shift, Scale, Left, V))
7440 /// \brief Try to lower a vector shuffle using SSE4a EXTRQ/INSERTQ.
7441 static SDValue lowerVectorShuffleWithSSE4A(SDLoc DL, MVT VT, SDValue V1,
7442 SDValue V2, ArrayRef<int> Mask,
7443 SelectionDAG &DAG) {
7444 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7445 assert(!Zeroable.all() && "Fully zeroable shuffle mask");
7447 int Size = Mask.size();
7448 int HalfSize = Size / 2;
7449 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
7451 // Upper half must be undefined.
7452 if (!isUndefInRange(Mask, HalfSize, HalfSize))
7455 // EXTRQ: Extract Len elements from lower half of source, starting at Idx.
7456 // Remainder of lower half result is zero and upper half is all undef.
7457 auto LowerAsEXTRQ = [&]() {
7458 // Determine the extraction length from the part of the
7459 // lower half that isn't zeroable.
7461 for (; Len > 0; --Len)
7462 if (!Zeroable[Len - 1])
7464 assert(Len > 0 && "Zeroable shuffle mask");
7466 // Attempt to match first Len sequential elements from the lower half.
7469 for (int i = 0; i != Len; ++i) {
7473 SDValue &V = (M < Size ? V1 : V2);
7476 // The extracted elements must start at a valid index and all mask
7477 // elements must be in the lower half.
7478 if (i > M || M >= HalfSize)
7481 if (Idx < 0 || (Src == V && Idx == (M - i))) {
7492 assert((Idx + Len) <= HalfSize && "Illegal extraction mask");
7493 int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
7494 int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
7495 return DAG.getNode(X86ISD::EXTRQI, DL, VT, Src,
7496 DAG.getConstant(BitLen, DL, MVT::i8),
7497 DAG.getConstant(BitIdx, DL, MVT::i8));
7500 if (SDValue ExtrQ = LowerAsEXTRQ())
7503 // INSERTQ: Extract lowest Len elements from lower half of second source and
7504 // insert over first source, starting at Idx.
7505 // { A[0], .., A[Idx-1], B[0], .., B[Len-1], A[Idx+Len], .., UNDEF, ... }
7506 auto LowerAsInsertQ = [&]() {
7507 for (int Idx = 0; Idx != HalfSize; ++Idx) {
7510 // Attempt to match first source from mask before insertion point.
7511 if (isUndefInRange(Mask, 0, Idx)) {
7513 } else if (isSequentialOrUndefInRange(Mask, 0, Idx, 0)) {
7515 } else if (isSequentialOrUndefInRange(Mask, 0, Idx, Size)) {
7521 // Extend the extraction length looking to match both the insertion of
7522 // the second source and the remaining elements of the first.
7523 for (int Hi = Idx + 1; Hi <= HalfSize; ++Hi) {
7528 if (isSequentialOrUndefInRange(Mask, Idx, Len, 0)) {
7530 } else if (isSequentialOrUndefInRange(Mask, Idx, Len, Size)) {
7536 // Match the remaining elements of the lower half.
7537 if (isUndefInRange(Mask, Hi, HalfSize - Hi)) {
7539 } else if ((!Base || (Base == V1)) &&
7540 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi, Hi)) {
7542 } else if ((!Base || (Base == V2)) &&
7543 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi,
7550 // We may not have a base (first source) - this can safely be undefined.
7552 Base = DAG.getUNDEF(VT);
7554 int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
7555 int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
7556 return DAG.getNode(X86ISD::INSERTQI, DL, VT, Base, Insert,
7557 DAG.getConstant(BitLen, DL, MVT::i8),
7558 DAG.getConstant(BitIdx, DL, MVT::i8));
7565 if (SDValue InsertQ = LowerAsInsertQ())
7571 /// \brief Lower a vector shuffle as a zero or any extension.
7573 /// Given a specific number of elements, element bit width, and extension
7574 /// stride, produce either a zero or any extension based on the available
7575 /// features of the subtarget. The extended elements are consecutive and
7576 /// begin and can start from an offseted element index in the input; to
7577 /// avoid excess shuffling the offset must either being in the bottom lane
7578 /// or at the start of a higher lane. All extended elements must be from
7580 static SDValue lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7581 SDLoc DL, MVT VT, int Scale, int Offset, bool AnyExt, SDValue InputV,
7582 ArrayRef<int> Mask, const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7583 assert(Scale > 1 && "Need a scale to extend.");
7584 int EltBits = VT.getScalarSizeInBits();
7585 int NumElements = VT.getVectorNumElements();
7586 int NumEltsPerLane = 128 / EltBits;
7587 int OffsetLane = Offset / NumEltsPerLane;
7588 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
7589 "Only 8, 16, and 32 bit elements can be extended.");
7590 assert(Scale * EltBits <= 64 && "Cannot zero extend past 64 bits.");
7591 assert(0 <= Offset && "Extension offset must be positive.");
7592 assert((Offset < NumEltsPerLane || Offset % NumEltsPerLane == 0) &&
7593 "Extension offset must be in the first lane or start an upper lane.");
7595 // Check that an index is in same lane as the base offset.
7596 auto SafeOffset = [&](int Idx) {
7597 return OffsetLane == (Idx / NumEltsPerLane);
7600 // Shift along an input so that the offset base moves to the first element.
7601 auto ShuffleOffset = [&](SDValue V) {
7605 SmallVector<int, 8> ShMask((unsigned)NumElements, -1);
7606 for (int i = 0; i * Scale < NumElements; ++i) {
7607 int SrcIdx = i + Offset;
7608 ShMask[i] = SafeOffset(SrcIdx) ? SrcIdx : -1;
7610 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), ShMask);
7613 // Found a valid zext mask! Try various lowering strategies based on the
7614 // input type and available ISA extensions.
7615 if (Subtarget->hasSSE41()) {
7616 // Not worth offseting 128-bit vectors if scale == 2, a pattern using
7617 // PUNPCK will catch this in a later shuffle match.
7618 if (Offset && Scale == 2 && VT.is128BitVector())
7620 MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits * Scale),
7621 NumElements / Scale);
7622 InputV = DAG.getNode(X86ISD::VZEXT, DL, ExtVT, ShuffleOffset(InputV));
7623 return DAG.getBitcast(VT, InputV);
7626 assert(VT.is128BitVector() && "Only 128-bit vectors can be extended.");
7628 // For any extends we can cheat for larger element sizes and use shuffle
7629 // instructions that can fold with a load and/or copy.
7630 if (AnyExt && EltBits == 32) {
7631 int PSHUFDMask[4] = {Offset, -1, SafeOffset(Offset + 1) ? Offset + 1 : -1,
7633 return DAG.getBitcast(
7634 VT, DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7635 DAG.getBitcast(MVT::v4i32, InputV),
7636 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
7638 if (AnyExt && EltBits == 16 && Scale > 2) {
7639 int PSHUFDMask[4] = {Offset / 2, -1,
7640 SafeOffset(Offset + 1) ? (Offset + 1) / 2 : -1, -1};
7641 InputV = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7642 DAG.getBitcast(MVT::v4i32, InputV),
7643 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG));
7644 int PSHUFWMask[4] = {1, -1, -1, -1};
7645 unsigned OddEvenOp = (Offset & 1 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW);
7646 return DAG.getBitcast(
7647 VT, DAG.getNode(OddEvenOp, DL, MVT::v8i16,
7648 DAG.getBitcast(MVT::v8i16, InputV),
7649 getV4X86ShuffleImm8ForMask(PSHUFWMask, DL, DAG)));
7652 // The SSE4A EXTRQ instruction can efficiently extend the first 2 lanes
7654 if ((Scale * EltBits) == 64 && EltBits < 32 && Subtarget->hasSSE4A()) {
7655 assert(NumElements == (int)Mask.size() && "Unexpected shuffle mask size!");
7656 assert(VT.is128BitVector() && "Unexpected vector width!");
7658 int LoIdx = Offset * EltBits;
7659 SDValue Lo = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7660 DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
7661 DAG.getConstant(EltBits, DL, MVT::i8),
7662 DAG.getConstant(LoIdx, DL, MVT::i8)));
7664 if (isUndefInRange(Mask, NumElements / 2, NumElements / 2) ||
7665 !SafeOffset(Offset + 1))
7666 return DAG.getNode(ISD::BITCAST, DL, VT, Lo);
7668 int HiIdx = (Offset + 1) * EltBits;
7669 SDValue Hi = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7670 DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
7671 DAG.getConstant(EltBits, DL, MVT::i8),
7672 DAG.getConstant(HiIdx, DL, MVT::i8)));
7673 return DAG.getNode(ISD::BITCAST, DL, VT,
7674 DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, Lo, Hi));
7677 // If this would require more than 2 unpack instructions to expand, use
7678 // pshufb when available. We can only use more than 2 unpack instructions
7679 // when zero extending i8 elements which also makes it easier to use pshufb.
7680 if (Scale > 4 && EltBits == 8 && Subtarget->hasSSSE3()) {
7681 assert(NumElements == 16 && "Unexpected byte vector width!");
7682 SDValue PSHUFBMask[16];
7683 for (int i = 0; i < 16; ++i) {
7684 int Idx = Offset + (i / Scale);
7685 PSHUFBMask[i] = DAG.getConstant(
7686 (i % Scale == 0 && SafeOffset(Idx)) ? Idx : 0x80, DL, MVT::i8);
7688 InputV = DAG.getBitcast(MVT::v16i8, InputV);
7689 return DAG.getBitcast(VT,
7690 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, InputV,
7691 DAG.getNode(ISD::BUILD_VECTOR, DL,
7692 MVT::v16i8, PSHUFBMask)));
7695 // If we are extending from an offset, ensure we start on a boundary that
7696 // we can unpack from.
7697 int AlignToUnpack = Offset % (NumElements / Scale);
7698 if (AlignToUnpack) {
7699 SmallVector<int, 8> ShMask((unsigned)NumElements, -1);
7700 for (int i = AlignToUnpack; i < NumElements; ++i)
7701 ShMask[i - AlignToUnpack] = i;
7702 InputV = DAG.getVectorShuffle(VT, DL, InputV, DAG.getUNDEF(VT), ShMask);
7703 Offset -= AlignToUnpack;
7706 // Otherwise emit a sequence of unpacks.
7708 unsigned UnpackLoHi = X86ISD::UNPCKL;
7709 if (Offset >= (NumElements / 2)) {
7710 UnpackLoHi = X86ISD::UNPCKH;
7711 Offset -= (NumElements / 2);
7714 MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements);
7715 SDValue Ext = AnyExt ? DAG.getUNDEF(InputVT)
7716 : getZeroVector(InputVT, Subtarget, DAG, DL);
7717 InputV = DAG.getBitcast(InputVT, InputV);
7718 InputV = DAG.getNode(UnpackLoHi, DL, InputVT, InputV, Ext);
7722 } while (Scale > 1);
7723 return DAG.getBitcast(VT, InputV);
7726 /// \brief Try to lower a vector shuffle as a zero extension on any microarch.
7728 /// This routine will try to do everything in its power to cleverly lower
7729 /// a shuffle which happens to match the pattern of a zero extend. It doesn't
7730 /// check for the profitability of this lowering, it tries to aggressively
7731 /// match this pattern. It will use all of the micro-architectural details it
7732 /// can to emit an efficient lowering. It handles both blends with all-zero
7733 /// inputs to explicitly zero-extend and undef-lanes (sometimes undef due to
7734 /// masking out later).
7736 /// The reason we have dedicated lowering for zext-style shuffles is that they
7737 /// are both incredibly common and often quite performance sensitive.
7738 static SDValue lowerVectorShuffleAsZeroOrAnyExtend(
7739 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7740 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7741 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7743 int Bits = VT.getSizeInBits();
7744 int NumLanes = Bits / 128;
7745 int NumElements = VT.getVectorNumElements();
7746 int NumEltsPerLane = NumElements / NumLanes;
7747 assert(VT.getScalarSizeInBits() <= 32 &&
7748 "Exceeds 32-bit integer zero extension limit");
7749 assert((int)Mask.size() == NumElements && "Unexpected shuffle mask size");
7751 // Define a helper function to check a particular ext-scale and lower to it if
7753 auto Lower = [&](int Scale) -> SDValue {
7758 for (int i = 0; i < NumElements; ++i) {
7761 continue; // Valid anywhere but doesn't tell us anything.
7762 if (i % Scale != 0) {
7763 // Each of the extended elements need to be zeroable.
7767 // We no longer are in the anyext case.
7772 // Each of the base elements needs to be consecutive indices into the
7773 // same input vector.
7774 SDValue V = M < NumElements ? V1 : V2;
7775 M = M % NumElements;
7778 Offset = M - (i / Scale);
7779 } else if (InputV != V)
7780 return SDValue(); // Flip-flopping inputs.
7782 // Offset must start in the lowest 128-bit lane or at the start of an
7784 // FIXME: Is it ever worth allowing a negative base offset?
7785 if (!((0 <= Offset && Offset < NumEltsPerLane) ||
7786 (Offset % NumEltsPerLane) == 0))
7789 // If we are offsetting, all referenced entries must come from the same
7791 if (Offset && (Offset / NumEltsPerLane) != (M / NumEltsPerLane))
7794 if ((M % NumElements) != (Offset + (i / Scale)))
7795 return SDValue(); // Non-consecutive strided elements.
7799 // If we fail to find an input, we have a zero-shuffle which should always
7800 // have already been handled.
7801 // FIXME: Maybe handle this here in case during blending we end up with one?
7805 // If we are offsetting, don't extend if we only match a single input, we
7806 // can always do better by using a basic PSHUF or PUNPCK.
7807 if (Offset != 0 && Matches < 2)
7810 return lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7811 DL, VT, Scale, Offset, AnyExt, InputV, Mask, Subtarget, DAG);
7814 // The widest scale possible for extending is to a 64-bit integer.
7815 assert(Bits % 64 == 0 &&
7816 "The number of bits in a vector must be divisible by 64 on x86!");
7817 int NumExtElements = Bits / 64;
7819 // Each iteration, try extending the elements half as much, but into twice as
7821 for (; NumExtElements < NumElements; NumExtElements *= 2) {
7822 assert(NumElements % NumExtElements == 0 &&
7823 "The input vector size must be divisible by the extended size.");
7824 if (SDValue V = Lower(NumElements / NumExtElements))
7828 // General extends failed, but 128-bit vectors may be able to use MOVQ.
7832 // Returns one of the source operands if the shuffle can be reduced to a
7833 // MOVQ, copying the lower 64-bits and zero-extending to the upper 64-bits.
7834 auto CanZExtLowHalf = [&]() {
7835 for (int i = NumElements / 2; i != NumElements; ++i)
7838 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, 0))
7840 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, NumElements))
7845 if (SDValue V = CanZExtLowHalf()) {
7846 V = DAG.getBitcast(MVT::v2i64, V);
7847 V = DAG.getNode(X86ISD::VZEXT_MOVL, DL, MVT::v2i64, V);
7848 return DAG.getBitcast(VT, V);
7851 // No viable ext lowering found.
7855 /// \brief Try to get a scalar value for a specific element of a vector.
7857 /// Looks through BUILD_VECTOR and SCALAR_TO_VECTOR nodes to find a scalar.
7858 static SDValue getScalarValueForVectorElement(SDValue V, int Idx,
7859 SelectionDAG &DAG) {
7860 MVT VT = V.getSimpleValueType();
7861 MVT EltVT = VT.getVectorElementType();
7862 while (V.getOpcode() == ISD::BITCAST)
7863 V = V.getOperand(0);
7864 // If the bitcasts shift the element size, we can't extract an equivalent
7866 MVT NewVT = V.getSimpleValueType();
7867 if (!NewVT.isVector() || NewVT.getScalarSizeInBits() != VT.getScalarSizeInBits())
7870 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7871 (Idx == 0 && V.getOpcode() == ISD::SCALAR_TO_VECTOR)) {
7872 // Ensure the scalar operand is the same size as the destination.
7873 // FIXME: Add support for scalar truncation where possible.
7874 SDValue S = V.getOperand(Idx);
7875 if (EltVT.getSizeInBits() == S.getSimpleValueType().getSizeInBits())
7876 return DAG.getNode(ISD::BITCAST, SDLoc(V), EltVT, S);
7882 /// \brief Helper to test for a load that can be folded with x86 shuffles.
7884 /// This is particularly important because the set of instructions varies
7885 /// significantly based on whether the operand is a load or not.
7886 static bool isShuffleFoldableLoad(SDValue V) {
7887 while (V.getOpcode() == ISD::BITCAST)
7888 V = V.getOperand(0);
7890 return ISD::isNON_EXTLoad(V.getNode());
7893 /// \brief Try to lower insertion of a single element into a zero vector.
7895 /// This is a common pattern that we have especially efficient patterns to lower
7896 /// across all subtarget feature sets.
7897 static SDValue lowerVectorShuffleAsElementInsertion(
7898 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7899 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7900 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7902 MVT EltVT = VT.getVectorElementType();
7904 int V2Index = std::find_if(Mask.begin(), Mask.end(),
7905 [&Mask](int M) { return M >= (int)Mask.size(); }) -
7907 bool IsV1Zeroable = true;
7908 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7909 if (i != V2Index && !Zeroable[i]) {
7910 IsV1Zeroable = false;
7914 // Check for a single input from a SCALAR_TO_VECTOR node.
7915 // FIXME: All of this should be canonicalized into INSERT_VECTOR_ELT and
7916 // all the smarts here sunk into that routine. However, the current
7917 // lowering of BUILD_VECTOR makes that nearly impossible until the old
7918 // vector shuffle lowering is dead.
7919 SDValue V2S = getScalarValueForVectorElement(V2, Mask[V2Index] - Mask.size(),
7921 if (V2S && DAG.getTargetLoweringInfo().isTypeLegal(V2S.getValueType())) {
7922 // We need to zext the scalar if it is smaller than an i32.
7923 V2S = DAG.getBitcast(EltVT, V2S);
7924 if (EltVT == MVT::i8 || EltVT == MVT::i16) {
7925 // Using zext to expand a narrow element won't work for non-zero
7930 // Zero-extend directly to i32.
7932 V2S = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, V2S);
7934 V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtVT, V2S);
7935 } else if (Mask[V2Index] != (int)Mask.size() || EltVT == MVT::i8 ||
7936 EltVT == MVT::i16) {
7937 // Either not inserting from the low element of the input or the input
7938 // element size is too small to use VZEXT_MOVL to clear the high bits.
7942 if (!IsV1Zeroable) {
7943 // If V1 can't be treated as a zero vector we have fewer options to lower
7944 // this. We can't support integer vectors or non-zero targets cheaply, and
7945 // the V1 elements can't be permuted in any way.
7946 assert(VT == ExtVT && "Cannot change extended type when non-zeroable!");
7947 if (!VT.isFloatingPoint() || V2Index != 0)
7949 SmallVector<int, 8> V1Mask(Mask.begin(), Mask.end());
7950 V1Mask[V2Index] = -1;
7951 if (!isNoopShuffleMask(V1Mask))
7953 // This is essentially a special case blend operation, but if we have
7954 // general purpose blend operations, they are always faster. Bail and let
7955 // the rest of the lowering handle these as blends.
7956 if (Subtarget->hasSSE41())
7959 // Otherwise, use MOVSD or MOVSS.
7960 assert((EltVT == MVT::f32 || EltVT == MVT::f64) &&
7961 "Only two types of floating point element types to handle!");
7962 return DAG.getNode(EltVT == MVT::f32 ? X86ISD::MOVSS : X86ISD::MOVSD, DL,
7966 // This lowering only works for the low element with floating point vectors.
7967 if (VT.isFloatingPoint() && V2Index != 0)
7970 V2 = DAG.getNode(X86ISD::VZEXT_MOVL, DL, ExtVT, V2);
7972 V2 = DAG.getBitcast(VT, V2);
7975 // If we have 4 or fewer lanes we can cheaply shuffle the element into
7976 // the desired position. Otherwise it is more efficient to do a vector
7977 // shift left. We know that we can do a vector shift left because all
7978 // the inputs are zero.
7979 if (VT.isFloatingPoint() || VT.getVectorNumElements() <= 4) {
7980 SmallVector<int, 4> V2Shuffle(Mask.size(), 1);
7981 V2Shuffle[V2Index] = 0;
7982 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Shuffle);
7984 V2 = DAG.getBitcast(MVT::v2i64, V2);
7986 X86ISD::VSHLDQ, DL, MVT::v2i64, V2,
7987 DAG.getConstant(V2Index * EltVT.getSizeInBits() / 8, DL,
7988 DAG.getTargetLoweringInfo().getScalarShiftAmountTy(
7989 DAG.getDataLayout(), VT)));
7990 V2 = DAG.getBitcast(VT, V2);
7996 /// \brief Try to lower broadcast of a single - truncated - integer element,
7997 /// coming from a scalar_to_vector/build_vector node \p V0 with larger elements.
7999 /// This assumes we have AVX2.
8000 static SDValue lowerVectorShuffleAsTruncBroadcast(SDLoc DL, MVT VT, SDValue V0,
8002 const X86Subtarget *Subtarget,
8003 SelectionDAG &DAG) {
8004 assert(Subtarget->hasAVX2() &&
8005 "We can only lower integer broadcasts with AVX2!");
8007 EVT EltVT = VT.getVectorElementType();
8008 EVT V0VT = V0.getValueType();
8010 assert(VT.isInteger() && "Unexpected non-integer trunc broadcast!");
8011 assert(V0VT.isVector() && "Unexpected non-vector vector-sized value!");
8013 EVT V0EltVT = V0VT.getVectorElementType();
8014 if (!V0EltVT.isInteger())
8017 const unsigned EltSize = EltVT.getSizeInBits();
8018 const unsigned V0EltSize = V0EltVT.getSizeInBits();
8020 // This is only a truncation if the original element type is larger.
8021 if (V0EltSize <= EltSize)
8024 assert(((V0EltSize % EltSize) == 0) &&
8025 "Scalar type sizes must all be powers of 2 on x86!");
8027 const unsigned V0Opc = V0.getOpcode();
8028 const unsigned Scale = V0EltSize / EltSize;
8029 const unsigned V0BroadcastIdx = BroadcastIdx / Scale;
8031 if ((V0Opc != ISD::SCALAR_TO_VECTOR || V0BroadcastIdx != 0) &&
8032 V0Opc != ISD::BUILD_VECTOR)
8035 SDValue Scalar = V0.getOperand(V0BroadcastIdx);
8037 // If we're extracting non-least-significant bits, shift so we can truncate.
8038 // Hopefully, we can fold away the trunc/srl/load into the broadcast.
8039 // Even if we can't (and !isShuffleFoldableLoad(Scalar)), prefer
8040 // vpbroadcast+vmovd+shr to vpshufb(m)+vmovd.
8041 if (const int OffsetIdx = BroadcastIdx % Scale)
8042 Scalar = DAG.getNode(ISD::SRL, DL, Scalar.getValueType(), Scalar,
8043 DAG.getConstant(OffsetIdx * EltSize, DL, Scalar.getValueType()));
8045 return DAG.getNode(X86ISD::VBROADCAST, DL, VT,
8046 DAG.getNode(ISD::TRUNCATE, DL, EltVT, Scalar));
8049 /// \brief Try to lower broadcast of a single element.
8051 /// For convenience, this code also bundles all of the subtarget feature set
8052 /// filtering. While a little annoying to re-dispatch on type here, there isn't
8053 /// a convenient way to factor it out.
8054 /// FIXME: This is very similar to LowerVectorBroadcast - can we merge them?
8055 static SDValue lowerVectorShuffleAsBroadcast(SDLoc DL, MVT VT, SDValue V,
8057 const X86Subtarget *Subtarget,
8058 SelectionDAG &DAG) {
8059 if (!Subtarget->hasAVX())
8061 if (VT.isInteger() && !Subtarget->hasAVX2())
8064 // Check that the mask is a broadcast.
8065 int BroadcastIdx = -1;
8067 if (M >= 0 && BroadcastIdx == -1)
8069 else if (M >= 0 && M != BroadcastIdx)
8072 assert(BroadcastIdx < (int)Mask.size() && "We only expect to be called with "
8073 "a sorted mask where the broadcast "
8076 // Go up the chain of (vector) values to find a scalar load that we can
8077 // combine with the broadcast.
8079 switch (V.getOpcode()) {
8080 case ISD::CONCAT_VECTORS: {
8081 int OperandSize = Mask.size() / V.getNumOperands();
8082 V = V.getOperand(BroadcastIdx / OperandSize);
8083 BroadcastIdx %= OperandSize;
8087 case ISD::INSERT_SUBVECTOR: {
8088 SDValue VOuter = V.getOperand(0), VInner = V.getOperand(1);
8089 auto ConstantIdx = dyn_cast<ConstantSDNode>(V.getOperand(2));
8093 int BeginIdx = (int)ConstantIdx->getZExtValue();
8095 BeginIdx + (int)VInner.getSimpleValueType().getVectorNumElements();
8096 if (BroadcastIdx >= BeginIdx && BroadcastIdx < EndIdx) {
8097 BroadcastIdx -= BeginIdx;
8108 // Check if this is a broadcast of a scalar. We special case lowering
8109 // for scalars so that we can more effectively fold with loads.
8110 // First, look through bitcast: if the original value has a larger element
8111 // type than the shuffle, the broadcast element is in essence truncated.
8112 // Make that explicit to ease folding.
8113 if (V.getOpcode() == ISD::BITCAST && VT.isInteger())
8114 if (SDValue TruncBroadcast = lowerVectorShuffleAsTruncBroadcast(
8115 DL, VT, V.getOperand(0), BroadcastIdx, Subtarget, DAG))
8116 return TruncBroadcast;
8118 // Also check the simpler case, where we can directly reuse the scalar.
8119 if (V.getOpcode() == ISD::BUILD_VECTOR ||
8120 (V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)) {
8121 V = V.getOperand(BroadcastIdx);
8123 // If the scalar isn't a load, we can't broadcast from it in AVX1.
8124 // Only AVX2 has register broadcasts.
8125 if (!Subtarget->hasAVX2() && !isShuffleFoldableLoad(V))
8127 } else if (MayFoldLoad(V) && !cast<LoadSDNode>(V)->isVolatile()) {
8128 // If we are broadcasting a load that is only used by the shuffle
8129 // then we can reduce the vector load to the broadcasted scalar load.
8130 LoadSDNode *Ld = cast<LoadSDNode>(V);
8131 SDValue BaseAddr = Ld->getOperand(1);
8132 EVT AddrVT = BaseAddr.getValueType();
8133 EVT SVT = VT.getScalarType();
8134 unsigned Offset = BroadcastIdx * SVT.getStoreSize();
8135 SDValue NewAddr = DAG.getNode(
8136 ISD::ADD, DL, AddrVT, BaseAddr,
8137 DAG.getConstant(Offset, DL, AddrVT));
8138 V = DAG.getLoad(SVT, DL, Ld->getChain(), NewAddr,
8139 DAG.getMachineFunction().getMachineMemOperand(
8140 Ld->getMemOperand(), Offset, SVT.getStoreSize()));
8141 } else if (BroadcastIdx != 0 || !Subtarget->hasAVX2()) {
8142 // We can't broadcast from a vector register without AVX2, and we can only
8143 // broadcast from the zero-element of a vector register.
8147 return DAG.getNode(X86ISD::VBROADCAST, DL, VT, V);
8150 // Check for whether we can use INSERTPS to perform the shuffle. We only use
8151 // INSERTPS when the V1 elements are already in the correct locations
8152 // because otherwise we can just always use two SHUFPS instructions which
8153 // are much smaller to encode than a SHUFPS and an INSERTPS. We can also
8154 // perform INSERTPS if a single V1 element is out of place and all V2
8155 // elements are zeroable.
8156 static SDValue lowerVectorShuffleAsInsertPS(SDValue Op, SDValue V1, SDValue V2,
8158 SelectionDAG &DAG) {
8159 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
8160 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8161 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8162 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8164 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
8167 int V1DstIndex = -1;
8168 int V2DstIndex = -1;
8169 bool V1UsedInPlace = false;
8171 for (int i = 0; i < 4; ++i) {
8172 // Synthesize a zero mask from the zeroable elements (includes undefs).
8178 // Flag if we use any V1 inputs in place.
8180 V1UsedInPlace = true;
8184 // We can only insert a single non-zeroable element.
8185 if (V1DstIndex != -1 || V2DstIndex != -1)
8189 // V1 input out of place for insertion.
8192 // V2 input for insertion.
8197 // Don't bother if we have no (non-zeroable) element for insertion.
8198 if (V1DstIndex == -1 && V2DstIndex == -1)
8201 // Determine element insertion src/dst indices. The src index is from the
8202 // start of the inserted vector, not the start of the concatenated vector.
8203 unsigned V2SrcIndex = 0;
8204 if (V1DstIndex != -1) {
8205 // If we have a V1 input out of place, we use V1 as the V2 element insertion
8206 // and don't use the original V2 at all.
8207 V2SrcIndex = Mask[V1DstIndex];
8208 V2DstIndex = V1DstIndex;
8211 V2SrcIndex = Mask[V2DstIndex] - 4;
8214 // If no V1 inputs are used in place, then the result is created only from
8215 // the zero mask and the V2 insertion - so remove V1 dependency.
8217 V1 = DAG.getUNDEF(MVT::v4f32);
8219 unsigned InsertPSMask = V2SrcIndex << 6 | V2DstIndex << 4 | ZMask;
8220 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
8222 // Insert the V2 element into the desired position.
8224 return DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
8225 DAG.getConstant(InsertPSMask, DL, MVT::i8));
8228 /// \brief Try to lower a shuffle as a permute of the inputs followed by an
8229 /// UNPCK instruction.
8231 /// This specifically targets cases where we end up with alternating between
8232 /// the two inputs, and so can permute them into something that feeds a single
8233 /// UNPCK instruction. Note that this routine only targets integer vectors
8234 /// because for floating point vectors we have a generalized SHUFPS lowering
8235 /// strategy that handles everything that doesn't *exactly* match an unpack,
8236 /// making this clever lowering unnecessary.
8237 static SDValue lowerVectorShuffleAsPermuteAndUnpack(SDLoc DL, MVT VT,
8238 SDValue V1, SDValue V2,
8240 SelectionDAG &DAG) {
8241 assert(!VT.isFloatingPoint() &&
8242 "This routine only supports integer vectors.");
8243 assert(!isSingleInputShuffleMask(Mask) &&
8244 "This routine should only be used when blending two inputs.");
8245 assert(Mask.size() >= 2 && "Single element masks are invalid.");
8247 int Size = Mask.size();
8249 int NumLoInputs = std::count_if(Mask.begin(), Mask.end(), [Size](int M) {
8250 return M >= 0 && M % Size < Size / 2;
8252 int NumHiInputs = std::count_if(
8253 Mask.begin(), Mask.end(), [Size](int M) { return M % Size >= Size / 2; });
8255 bool UnpackLo = NumLoInputs >= NumHiInputs;
8257 auto TryUnpack = [&](MVT UnpackVT, int Scale) {
8258 SmallVector<int, 32> V1Mask(Mask.size(), -1);
8259 SmallVector<int, 32> V2Mask(Mask.size(), -1);
8261 for (int i = 0; i < Size; ++i) {
8265 // Each element of the unpack contains Scale elements from this mask.
8266 int UnpackIdx = i / Scale;
8268 // We only handle the case where V1 feeds the first slots of the unpack.
8269 // We rely on canonicalization to ensure this is the case.
8270 if ((UnpackIdx % 2 == 0) != (Mask[i] < Size))
8273 // Setup the mask for this input. The indexing is tricky as we have to
8274 // handle the unpack stride.
8275 SmallVectorImpl<int> &VMask = (UnpackIdx % 2 == 0) ? V1Mask : V2Mask;
8276 VMask[(UnpackIdx / 2) * Scale + i % Scale + (UnpackLo ? 0 : Size / 2)] =
8280 // If we will have to shuffle both inputs to use the unpack, check whether
8281 // we can just unpack first and shuffle the result. If so, skip this unpack.
8282 if ((NumLoInputs == 0 || NumHiInputs == 0) && !isNoopShuffleMask(V1Mask) &&
8283 !isNoopShuffleMask(V2Mask))
8286 // Shuffle the inputs into place.
8287 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
8288 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
8290 // Cast the inputs to the type we will use to unpack them.
8291 V1 = DAG.getBitcast(UnpackVT, V1);
8292 V2 = DAG.getBitcast(UnpackVT, V2);
8294 // Unpack the inputs and cast the result back to the desired type.
8295 return DAG.getBitcast(
8296 VT, DAG.getNode(UnpackLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
8300 // We try each unpack from the largest to the smallest to try and find one
8301 // that fits this mask.
8302 int OrigNumElements = VT.getVectorNumElements();
8303 int OrigScalarSize = VT.getScalarSizeInBits();
8304 for (int ScalarSize = 64; ScalarSize >= OrigScalarSize; ScalarSize /= 2) {
8305 int Scale = ScalarSize / OrigScalarSize;
8306 int NumElements = OrigNumElements / Scale;
8307 MVT UnpackVT = MVT::getVectorVT(MVT::getIntegerVT(ScalarSize), NumElements);
8308 if (SDValue Unpack = TryUnpack(UnpackVT, Scale))
8312 // If none of the unpack-rooted lowerings worked (or were profitable) try an
8314 if (NumLoInputs == 0 || NumHiInputs == 0) {
8315 assert((NumLoInputs > 0 || NumHiInputs > 0) &&
8316 "We have to have *some* inputs!");
8317 int HalfOffset = NumLoInputs == 0 ? Size / 2 : 0;
8319 // FIXME: We could consider the total complexity of the permute of each
8320 // possible unpacking. Or at the least we should consider how many
8321 // half-crossings are created.
8322 // FIXME: We could consider commuting the unpacks.
8324 SmallVector<int, 32> PermMask;
8325 PermMask.assign(Size, -1);
8326 for (int i = 0; i < Size; ++i) {
8330 assert(Mask[i] % Size >= HalfOffset && "Found input from wrong half!");
8333 2 * ((Mask[i] % Size) - HalfOffset) + (Mask[i] < Size ? 0 : 1);
8335 return DAG.getVectorShuffle(
8336 VT, DL, DAG.getNode(NumLoInputs == 0 ? X86ISD::UNPCKH : X86ISD::UNPCKL,
8338 DAG.getUNDEF(VT), PermMask);
8344 /// \brief Handle lowering of 2-lane 64-bit floating point shuffles.
8346 /// This is the basis function for the 2-lane 64-bit shuffles as we have full
8347 /// support for floating point shuffles but not integer shuffles. These
8348 /// instructions will incur a domain crossing penalty on some chips though so
8349 /// it is better to avoid lowering through this for integer vectors where
8351 static SDValue lowerV2F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8352 const X86Subtarget *Subtarget,
8353 SelectionDAG &DAG) {
8355 assert(Op.getSimpleValueType() == MVT::v2f64 && "Bad shuffle type!");
8356 assert(V1.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
8357 assert(V2.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
8358 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8359 ArrayRef<int> Mask = SVOp->getMask();
8360 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
8362 if (isSingleInputShuffleMask(Mask)) {
8363 // Use low duplicate instructions for masks that match their pattern.
8364 if (Subtarget->hasSSE3())
8365 if (isShuffleEquivalent(V1, V2, Mask, {0, 0}))
8366 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v2f64, V1);
8368 // Straight shuffle of a single input vector. Simulate this by using the
8369 // single input as both of the "inputs" to this instruction..
8370 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1);
8372 if (Subtarget->hasAVX()) {
8373 // If we have AVX, we can use VPERMILPS which will allow folding a load
8374 // into the shuffle.
8375 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v2f64, V1,
8376 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8379 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V1,
8380 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8382 assert(Mask[0] >= 0 && Mask[0] < 2 && "Non-canonicalized blend!");
8383 assert(Mask[1] >= 2 && "Non-canonicalized blend!");
8385 // If we have a single input, insert that into V1 if we can do so cheaply.
8386 if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1) {
8387 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8388 DL, MVT::v2f64, V1, V2, Mask, Subtarget, DAG))
8390 // Try inverting the insertion since for v2 masks it is easy to do and we
8391 // can't reliably sort the mask one way or the other.
8392 int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2),
8393 Mask[1] < 0 ? -1 : (Mask[1] ^ 2)};
8394 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8395 DL, MVT::v2f64, V2, V1, InverseMask, Subtarget, DAG))
8399 // Try to use one of the special instruction patterns to handle two common
8400 // blend patterns if a zero-blend above didn't work.
8401 if (isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
8402 isShuffleEquivalent(V1, V2, Mask, {1, 3}))
8403 if (SDValue V1S = getScalarValueForVectorElement(V1, Mask[0], DAG))
8404 // We can either use a special instruction to load over the low double or
8405 // to move just the low double.
8407 isShuffleFoldableLoad(V1S) ? X86ISD::MOVLPD : X86ISD::MOVSD,
8409 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, V1S));
8411 if (Subtarget->hasSSE41())
8412 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2f64, V1, V2, Mask,
8416 // Use dedicated unpack instructions for masks that match their pattern.
8418 lowerVectorShuffleWithUNPCK(DL, MVT::v2f64, Mask, V1, V2, DAG))
8421 unsigned SHUFPDMask = (Mask[0] == 1) | (((Mask[1] - 2) == 1) << 1);
8422 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V2,
8423 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8426 /// \brief Handle lowering of 2-lane 64-bit integer shuffles.
8428 /// Tries to lower a 2-lane 64-bit shuffle using shuffle operations provided by
8429 /// the integer unit to minimize domain crossing penalties. However, for blends
8430 /// it falls back to the floating point shuffle operation with appropriate bit
8432 static SDValue lowerV2I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8433 const X86Subtarget *Subtarget,
8434 SelectionDAG &DAG) {
8436 assert(Op.getSimpleValueType() == MVT::v2i64 && "Bad shuffle type!");
8437 assert(V1.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
8438 assert(V2.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
8439 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8440 ArrayRef<int> Mask = SVOp->getMask();
8441 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
8443 if (isSingleInputShuffleMask(Mask)) {
8444 // Check for being able to broadcast a single element.
8445 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v2i64, V1,
8446 Mask, Subtarget, DAG))
8449 // Straight shuffle of a single input vector. For everything from SSE2
8450 // onward this has a single fast instruction with no scary immediates.
8451 // We have to map the mask as it is actually a v4i32 shuffle instruction.
8452 V1 = DAG.getBitcast(MVT::v4i32, V1);
8453 int WidenedMask[4] = {
8454 std::max(Mask[0], 0) * 2, std::max(Mask[0], 0) * 2 + 1,
8455 std::max(Mask[1], 0) * 2, std::max(Mask[1], 0) * 2 + 1};
8456 return DAG.getBitcast(
8458 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
8459 getV4X86ShuffleImm8ForMask(WidenedMask, DL, DAG)));
8461 assert(Mask[0] != -1 && "No undef lanes in multi-input v2 shuffles!");
8462 assert(Mask[1] != -1 && "No undef lanes in multi-input v2 shuffles!");
8463 assert(Mask[0] < 2 && "We sort V1 to be the first input.");
8464 assert(Mask[1] >= 2 && "We sort V2 to be the second input.");
8466 // If we have a blend of two PACKUS operations an the blend aligns with the
8467 // low and half halves, we can just merge the PACKUS operations. This is
8468 // particularly important as it lets us merge shuffles that this routine itself
8470 auto GetPackNode = [](SDValue V) {
8471 while (V.getOpcode() == ISD::BITCAST)
8472 V = V.getOperand(0);
8474 return V.getOpcode() == X86ISD::PACKUS ? V : SDValue();
8476 if (SDValue V1Pack = GetPackNode(V1))
8477 if (SDValue V2Pack = GetPackNode(V2))
8478 return DAG.getBitcast(MVT::v2i64,
8479 DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8,
8480 Mask[0] == 0 ? V1Pack.getOperand(0)
8481 : V1Pack.getOperand(1),
8482 Mask[1] == 2 ? V2Pack.getOperand(0)
8483 : V2Pack.getOperand(1)));
8485 // Try to use shift instructions.
8487 lowerVectorShuffleAsShift(DL, MVT::v2i64, V1, V2, Mask, DAG))
8490 // When loading a scalar and then shuffling it into a vector we can often do
8491 // the insertion cheaply.
8492 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8493 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
8495 // Try inverting the insertion since for v2 masks it is easy to do and we
8496 // can't reliably sort the mask one way or the other.
8497 int InverseMask[2] = {Mask[0] ^ 2, Mask[1] ^ 2};
8498 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8499 DL, MVT::v2i64, V2, V1, InverseMask, Subtarget, DAG))
8502 // We have different paths for blend lowering, but they all must use the
8503 // *exact* same predicate.
8504 bool IsBlendSupported = Subtarget->hasSSE41();
8505 if (IsBlendSupported)
8506 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2i64, V1, V2, Mask,
8510 // Use dedicated unpack instructions for masks that match their pattern.
8512 lowerVectorShuffleWithUNPCK(DL, MVT::v2i64, Mask, V1, V2, DAG))
8515 // Try to use byte rotation instructions.
8516 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8517 if (Subtarget->hasSSSE3())
8518 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8519 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
8522 // If we have direct support for blends, we should lower by decomposing into
8523 // a permute. That will be faster than the domain cross.
8524 if (IsBlendSupported)
8525 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v2i64, V1, V2,
8528 // We implement this with SHUFPD which is pretty lame because it will likely
8529 // incur 2 cycles of stall for integer vectors on Nehalem and older chips.
8530 // However, all the alternatives are still more cycles and newer chips don't
8531 // have this problem. It would be really nice if x86 had better shuffles here.
8532 V1 = DAG.getBitcast(MVT::v2f64, V1);
8533 V2 = DAG.getBitcast(MVT::v2f64, V2);
8534 return DAG.getBitcast(MVT::v2i64,
8535 DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask));
8538 /// \brief Test whether this can be lowered with a single SHUFPS instruction.
8540 /// This is used to disable more specialized lowerings when the shufps lowering
8541 /// will happen to be efficient.
8542 static bool isSingleSHUFPSMask(ArrayRef<int> Mask) {
8543 // This routine only handles 128-bit shufps.
8544 assert(Mask.size() == 4 && "Unsupported mask size!");
8546 // To lower with a single SHUFPS we need to have the low half and high half
8547 // each requiring a single input.
8548 if (Mask[0] != -1 && Mask[1] != -1 && (Mask[0] < 4) != (Mask[1] < 4))
8550 if (Mask[2] != -1 && Mask[3] != -1 && (Mask[2] < 4) != (Mask[3] < 4))
8556 /// \brief Lower a vector shuffle using the SHUFPS instruction.
8558 /// This is a helper routine dedicated to lowering vector shuffles using SHUFPS.
8559 /// It makes no assumptions about whether this is the *best* lowering, it simply
8561 static SDValue lowerVectorShuffleWithSHUFPS(SDLoc DL, MVT VT,
8562 ArrayRef<int> Mask, SDValue V1,
8563 SDValue V2, SelectionDAG &DAG) {
8564 SDValue LowV = V1, HighV = V2;
8565 int NewMask[4] = {Mask[0], Mask[1], Mask[2], Mask[3]};
8568 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8570 if (NumV2Elements == 1) {
8572 std::find_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }) -
8575 // Compute the index adjacent to V2Index and in the same half by toggling
8577 int V2AdjIndex = V2Index ^ 1;
8579 if (Mask[V2AdjIndex] == -1) {
8580 // Handles all the cases where we have a single V2 element and an undef.
8581 // This will only ever happen in the high lanes because we commute the
8582 // vector otherwise.
8584 std::swap(LowV, HighV);
8585 NewMask[V2Index] -= 4;
8587 // Handle the case where the V2 element ends up adjacent to a V1 element.
8588 // To make this work, blend them together as the first step.
8589 int V1Index = V2AdjIndex;
8590 int BlendMask[4] = {Mask[V2Index] - 4, 0, Mask[V1Index], 0};
8591 V2 = DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
8592 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
8594 // Now proceed to reconstruct the final blend as we have the necessary
8595 // high or low half formed.
8602 NewMask[V1Index] = 2; // We put the V1 element in V2[2].
8603 NewMask[V2Index] = 0; // We shifted the V2 element into V2[0].
8605 } else if (NumV2Elements == 2) {
8606 if (Mask[0] < 4 && Mask[1] < 4) {
8607 // Handle the easy case where we have V1 in the low lanes and V2 in the
8611 } else if (Mask[2] < 4 && Mask[3] < 4) {
8612 // We also handle the reversed case because this utility may get called
8613 // when we detect a SHUFPS pattern but can't easily commute the shuffle to
8614 // arrange things in the right direction.
8620 // We have a mixture of V1 and V2 in both low and high lanes. Rather than
8621 // trying to place elements directly, just blend them and set up the final
8622 // shuffle to place them.
8624 // The first two blend mask elements are for V1, the second two are for
8626 int BlendMask[4] = {Mask[0] < 4 ? Mask[0] : Mask[1],
8627 Mask[2] < 4 ? Mask[2] : Mask[3],
8628 (Mask[0] >= 4 ? Mask[0] : Mask[1]) - 4,
8629 (Mask[2] >= 4 ? Mask[2] : Mask[3]) - 4};
8630 V1 = DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
8631 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
8633 // Now we do a normal shuffle of V1 by giving V1 as both operands to
8636 NewMask[0] = Mask[0] < 4 ? 0 : 2;
8637 NewMask[1] = Mask[0] < 4 ? 2 : 0;
8638 NewMask[2] = Mask[2] < 4 ? 1 : 3;
8639 NewMask[3] = Mask[2] < 4 ? 3 : 1;
8642 return DAG.getNode(X86ISD::SHUFP, DL, VT, LowV, HighV,
8643 getV4X86ShuffleImm8ForMask(NewMask, DL, DAG));
8646 /// \brief Lower 4-lane 32-bit floating point shuffles.
8648 /// Uses instructions exclusively from the floating point unit to minimize
8649 /// domain crossing penalties, as these are sufficient to implement all v4f32
8651 static SDValue lowerV4F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8652 const X86Subtarget *Subtarget,
8653 SelectionDAG &DAG) {
8655 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
8656 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8657 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8658 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8659 ArrayRef<int> Mask = SVOp->getMask();
8660 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8663 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8665 if (NumV2Elements == 0) {
8666 // Check for being able to broadcast a single element.
8667 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f32, V1,
8668 Mask, Subtarget, DAG))
8671 // Use even/odd duplicate instructions for masks that match their pattern.
8672 if (Subtarget->hasSSE3()) {
8673 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
8674 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v4f32, V1);
8675 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3}))
8676 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v4f32, V1);
8679 if (Subtarget->hasAVX()) {
8680 // If we have AVX, we can use VPERMILPS which will allow folding a load
8681 // into the shuffle.
8682 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f32, V1,
8683 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8686 // Otherwise, use a straight shuffle of a single input vector. We pass the
8687 // input vector to both operands to simulate this with a SHUFPS.
8688 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1,
8689 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8692 // There are special ways we can lower some single-element blends. However, we
8693 // have custom ways we can lower more complex single-element blends below that
8694 // we defer to if both this and BLENDPS fail to match, so restrict this to
8695 // when the V2 input is targeting element 0 of the mask -- that is the fast
8697 if (NumV2Elements == 1 && Mask[0] >= 4)
8698 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4f32, V1, V2,
8699 Mask, Subtarget, DAG))
8702 if (Subtarget->hasSSE41()) {
8703 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f32, V1, V2, Mask,
8707 // Use INSERTPS if we can complete the shuffle efficiently.
8708 if (SDValue V = lowerVectorShuffleAsInsertPS(Op, V1, V2, Mask, DAG))
8711 if (!isSingleSHUFPSMask(Mask))
8712 if (SDValue BlendPerm = lowerVectorShuffleAsBlendAndPermute(
8713 DL, MVT::v4f32, V1, V2, Mask, DAG))
8717 // Use dedicated unpack instructions for masks that match their pattern.
8719 lowerVectorShuffleWithUNPCK(DL, MVT::v4f32, Mask, V1, V2, DAG))
8722 // Otherwise fall back to a SHUFPS lowering strategy.
8723 return lowerVectorShuffleWithSHUFPS(DL, MVT::v4f32, Mask, V1, V2, DAG);
8726 /// \brief Lower 4-lane i32 vector shuffles.
8728 /// We try to handle these with integer-domain shuffles where we can, but for
8729 /// blends we use the floating point domain blend instructions.
8730 static SDValue lowerV4I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8731 const X86Subtarget *Subtarget,
8732 SelectionDAG &DAG) {
8734 assert(Op.getSimpleValueType() == MVT::v4i32 && "Bad shuffle type!");
8735 assert(V1.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8736 assert(V2.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8737 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8738 ArrayRef<int> Mask = SVOp->getMask();
8739 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8741 // Whenever we can lower this as a zext, that instruction is strictly faster
8742 // than any alternative. It also allows us to fold memory operands into the
8743 // shuffle in many cases.
8744 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v4i32, V1, V2,
8745 Mask, Subtarget, DAG))
8749 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8751 if (NumV2Elements == 0) {
8752 // Check for being able to broadcast a single element.
8753 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i32, V1,
8754 Mask, Subtarget, DAG))
8757 // Straight shuffle of a single input vector. For everything from SSE2
8758 // onward this has a single fast instruction with no scary immediates.
8759 // We coerce the shuffle pattern to be compatible with UNPCK instructions
8760 // but we aren't actually going to use the UNPCK instruction because doing
8761 // so prevents folding a load into this instruction or making a copy.
8762 const int UnpackLoMask[] = {0, 0, 1, 1};
8763 const int UnpackHiMask[] = {2, 2, 3, 3};
8764 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 1, 1}))
8765 Mask = UnpackLoMask;
8766 else if (isShuffleEquivalent(V1, V2, Mask, {2, 2, 3, 3}))
8767 Mask = UnpackHiMask;
8769 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
8770 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8773 // Try to use shift instructions.
8775 lowerVectorShuffleAsShift(DL, MVT::v4i32, V1, V2, Mask, DAG))
8778 // There are special ways we can lower some single-element blends.
8779 if (NumV2Elements == 1)
8780 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4i32, V1, V2,
8781 Mask, Subtarget, DAG))
8784 // We have different paths for blend lowering, but they all must use the
8785 // *exact* same predicate.
8786 bool IsBlendSupported = Subtarget->hasSSE41();
8787 if (IsBlendSupported)
8788 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i32, V1, V2, Mask,
8792 if (SDValue Masked =
8793 lowerVectorShuffleAsBitMask(DL, MVT::v4i32, V1, V2, Mask, DAG))
8796 // Use dedicated unpack instructions for masks that match their pattern.
8798 lowerVectorShuffleWithUNPCK(DL, MVT::v4i32, Mask, V1, V2, DAG))
8801 // Try to use byte rotation instructions.
8802 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8803 if (Subtarget->hasSSSE3())
8804 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8805 DL, MVT::v4i32, V1, V2, Mask, Subtarget, DAG))
8808 // If we have direct support for blends, we should lower by decomposing into
8809 // a permute. That will be faster than the domain cross.
8810 if (IsBlendSupported)
8811 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i32, V1, V2,
8814 // Try to lower by permuting the inputs into an unpack instruction.
8815 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(DL, MVT::v4i32, V1,
8819 // We implement this with SHUFPS because it can blend from two vectors.
8820 // Because we're going to eventually use SHUFPS, we use SHUFPS even to build
8821 // up the inputs, bypassing domain shift penalties that we would encur if we
8822 // directly used PSHUFD on Nehalem and older. For newer chips, this isn't
8824 return DAG.getBitcast(
8826 DAG.getVectorShuffle(MVT::v4f32, DL, DAG.getBitcast(MVT::v4f32, V1),
8827 DAG.getBitcast(MVT::v4f32, V2), Mask));
8830 /// \brief Lowering of single-input v8i16 shuffles is the cornerstone of SSE2
8831 /// shuffle lowering, and the most complex part.
8833 /// The lowering strategy is to try to form pairs of input lanes which are
8834 /// targeted at the same half of the final vector, and then use a dword shuffle
8835 /// to place them onto the right half, and finally unpack the paired lanes into
8836 /// their final position.
8838 /// The exact breakdown of how to form these dword pairs and align them on the
8839 /// correct sides is really tricky. See the comments within the function for
8840 /// more of the details.
8842 /// This code also handles repeated 128-bit lanes of v8i16 shuffles, but each
8843 /// lane must shuffle the *exact* same way. In fact, you must pass a v8 Mask to
8844 /// this routine for it to work correctly. To shuffle a 256-bit or 512-bit i16
8845 /// vector, form the analogous 128-bit 8-element Mask.
8846 static SDValue lowerV8I16GeneralSingleInputVectorShuffle(
8847 SDLoc DL, MVT VT, SDValue V, MutableArrayRef<int> Mask,
8848 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
8849 assert(VT.getVectorElementType() == MVT::i16 && "Bad input type!");
8850 MVT PSHUFDVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
8852 assert(Mask.size() == 8 && "Shuffle mask length doen't match!");
8853 MutableArrayRef<int> LoMask = Mask.slice(0, 4);
8854 MutableArrayRef<int> HiMask = Mask.slice(4, 4);
8856 SmallVector<int, 4> LoInputs;
8857 std::copy_if(LoMask.begin(), LoMask.end(), std::back_inserter(LoInputs),
8858 [](int M) { return M >= 0; });
8859 std::sort(LoInputs.begin(), LoInputs.end());
8860 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), LoInputs.end());
8861 SmallVector<int, 4> HiInputs;
8862 std::copy_if(HiMask.begin(), HiMask.end(), std::back_inserter(HiInputs),
8863 [](int M) { return M >= 0; });
8864 std::sort(HiInputs.begin(), HiInputs.end());
8865 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), HiInputs.end());
8867 std::lower_bound(LoInputs.begin(), LoInputs.end(), 4) - LoInputs.begin();
8868 int NumHToL = LoInputs.size() - NumLToL;
8870 std::lower_bound(HiInputs.begin(), HiInputs.end(), 4) - HiInputs.begin();
8871 int NumHToH = HiInputs.size() - NumLToH;
8872 MutableArrayRef<int> LToLInputs(LoInputs.data(), NumLToL);
8873 MutableArrayRef<int> LToHInputs(HiInputs.data(), NumLToH);
8874 MutableArrayRef<int> HToLInputs(LoInputs.data() + NumLToL, NumHToL);
8875 MutableArrayRef<int> HToHInputs(HiInputs.data() + NumLToH, NumHToH);
8877 // Simplify the 1-into-3 and 3-into-1 cases with a single pshufd. For all
8878 // such inputs we can swap two of the dwords across the half mark and end up
8879 // with <=2 inputs to each half in each half. Once there, we can fall through
8880 // to the generic code below. For example:
8882 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8883 // Mask: [0, 1, 2, 7, 4, 5, 6, 3] -----------------> [0, 1, 4, 7, 2, 3, 6, 5]
8885 // However in some very rare cases we have a 1-into-3 or 3-into-1 on one half
8886 // and an existing 2-into-2 on the other half. In this case we may have to
8887 // pre-shuffle the 2-into-2 half to avoid turning it into a 3-into-1 or
8888 // 1-into-3 which could cause us to cycle endlessly fixing each side in turn.
8889 // Fortunately, we don't have to handle anything but a 2-into-2 pattern
8890 // because any other situation (including a 3-into-1 or 1-into-3 in the other
8891 // half than the one we target for fixing) will be fixed when we re-enter this
8892 // path. We will also combine away any sequence of PSHUFD instructions that
8893 // result into a single instruction. Here is an example of the tricky case:
8895 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8896 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -THIS-IS-BAD!!!!-> [5, 7, 1, 0, 4, 7, 5, 3]
8898 // This now has a 1-into-3 in the high half! Instead, we do two shuffles:
8900 // Input: [a, b, c, d, e, f, g, h] PSHUFHW[0,2,1,3]-> [a, b, c, d, e, g, f, h]
8901 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -----------------> [3, 7, 1, 0, 2, 7, 3, 6]
8903 // Input: [a, b, c, d, e, g, f, h] -PSHUFD[0,2,1,3]-> [a, b, e, g, c, d, f, h]
8904 // Mask: [3, 7, 1, 0, 2, 7, 3, 6] -----------------> [5, 7, 1, 0, 4, 7, 5, 6]
8906 // The result is fine to be handled by the generic logic.
8907 auto balanceSides = [&](ArrayRef<int> AToAInputs, ArrayRef<int> BToAInputs,
8908 ArrayRef<int> BToBInputs, ArrayRef<int> AToBInputs,
8909 int AOffset, int BOffset) {
8910 assert((AToAInputs.size() == 3 || AToAInputs.size() == 1) &&
8911 "Must call this with A having 3 or 1 inputs from the A half.");
8912 assert((BToAInputs.size() == 1 || BToAInputs.size() == 3) &&
8913 "Must call this with B having 1 or 3 inputs from the B half.");
8914 assert(AToAInputs.size() + BToAInputs.size() == 4 &&
8915 "Must call this with either 3:1 or 1:3 inputs (summing to 4).");
8917 bool ThreeAInputs = AToAInputs.size() == 3;
8919 // Compute the index of dword with only one word among the three inputs in
8920 // a half by taking the sum of the half with three inputs and subtracting
8921 // the sum of the actual three inputs. The difference is the remaining
8924 int &TripleDWord = ThreeAInputs ? ADWord : BDWord;
8925 int &OneInputDWord = ThreeAInputs ? BDWord : ADWord;
8926 int TripleInputOffset = ThreeAInputs ? AOffset : BOffset;
8927 ArrayRef<int> TripleInputs = ThreeAInputs ? AToAInputs : BToAInputs;
8928 int OneInput = ThreeAInputs ? BToAInputs[0] : AToAInputs[0];
8929 int TripleInputSum = 0 + 1 + 2 + 3 + (4 * TripleInputOffset);
8930 int TripleNonInputIdx =
8931 TripleInputSum - std::accumulate(TripleInputs.begin(), TripleInputs.end(), 0);
8932 TripleDWord = TripleNonInputIdx / 2;
8934 // We use xor with one to compute the adjacent DWord to whichever one the
8936 OneInputDWord = (OneInput / 2) ^ 1;
8938 // Check for one tricky case: We're fixing a 3<-1 or a 1<-3 shuffle for AToA
8939 // and BToA inputs. If there is also such a problem with the BToB and AToB
8940 // inputs, we don't try to fix it necessarily -- we'll recurse and see it in
8941 // the next pass. However, if we have a 2<-2 in the BToB and AToB inputs, it
8942 // is essential that we don't *create* a 3<-1 as then we might oscillate.
8943 if (BToBInputs.size() == 2 && AToBInputs.size() == 2) {
8944 // Compute how many inputs will be flipped by swapping these DWords. We
8946 // to balance this to ensure we don't form a 3-1 shuffle in the other
8948 int NumFlippedAToBInputs =
8949 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord) +
8950 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord + 1);
8951 int NumFlippedBToBInputs =
8952 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord) +
8953 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord + 1);
8954 if ((NumFlippedAToBInputs == 1 &&
8955 (NumFlippedBToBInputs == 0 || NumFlippedBToBInputs == 2)) ||
8956 (NumFlippedBToBInputs == 1 &&
8957 (NumFlippedAToBInputs == 0 || NumFlippedAToBInputs == 2))) {
8958 // We choose whether to fix the A half or B half based on whether that
8959 // half has zero flipped inputs. At zero, we may not be able to fix it
8960 // with that half. We also bias towards fixing the B half because that
8961 // will more commonly be the high half, and we have to bias one way.
8962 auto FixFlippedInputs = [&V, &DL, &Mask, &DAG](int PinnedIdx, int DWord,
8963 ArrayRef<int> Inputs) {
8964 int FixIdx = PinnedIdx ^ 1; // The adjacent slot to the pinned slot.
8965 bool IsFixIdxInput = std::find(Inputs.begin(), Inputs.end(),
8966 PinnedIdx ^ 1) != Inputs.end();
8967 // Determine whether the free index is in the flipped dword or the
8968 // unflipped dword based on where the pinned index is. We use this bit
8969 // in an xor to conditionally select the adjacent dword.
8970 int FixFreeIdx = 2 * (DWord ^ (PinnedIdx / 2 == DWord));
8971 bool IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8972 FixFreeIdx) != Inputs.end();
8973 if (IsFixIdxInput == IsFixFreeIdxInput)
8975 IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8976 FixFreeIdx) != Inputs.end();
8977 assert(IsFixIdxInput != IsFixFreeIdxInput &&
8978 "We need to be changing the number of flipped inputs!");
8979 int PSHUFHalfMask[] = {0, 1, 2, 3};
8980 std::swap(PSHUFHalfMask[FixFreeIdx % 4], PSHUFHalfMask[FixIdx % 4]);
8981 V = DAG.getNode(FixIdx < 4 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW, DL,
8983 getV4X86ShuffleImm8ForMask(PSHUFHalfMask, DL, DAG));
8986 if (M != -1 && M == FixIdx)
8988 else if (M != -1 && M == FixFreeIdx)
8991 if (NumFlippedBToBInputs != 0) {
8993 BToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
8994 FixFlippedInputs(BPinnedIdx, BDWord, BToBInputs);
8996 assert(NumFlippedAToBInputs != 0 && "Impossible given predicates!");
8997 int APinnedIdx = ThreeAInputs ? TripleNonInputIdx : OneInput;
8998 FixFlippedInputs(APinnedIdx, ADWord, AToBInputs);
9003 int PSHUFDMask[] = {0, 1, 2, 3};
9004 PSHUFDMask[ADWord] = BDWord;
9005 PSHUFDMask[BDWord] = ADWord;
9008 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
9009 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
9011 // Adjust the mask to match the new locations of A and B.
9013 if (M != -1 && M/2 == ADWord)
9014 M = 2 * BDWord + M % 2;
9015 else if (M != -1 && M/2 == BDWord)
9016 M = 2 * ADWord + M % 2;
9018 // Recurse back into this routine to re-compute state now that this isn't
9019 // a 3 and 1 problem.
9020 return lowerV8I16GeneralSingleInputVectorShuffle(DL, VT, V, Mask, Subtarget,
9023 if ((NumLToL == 3 && NumHToL == 1) || (NumLToL == 1 && NumHToL == 3))
9024 return balanceSides(LToLInputs, HToLInputs, HToHInputs, LToHInputs, 0, 4);
9025 else if ((NumHToH == 3 && NumLToH == 1) || (NumHToH == 1 && NumLToH == 3))
9026 return balanceSides(HToHInputs, LToHInputs, LToLInputs, HToLInputs, 4, 0);
9028 // At this point there are at most two inputs to the low and high halves from
9029 // each half. That means the inputs can always be grouped into dwords and
9030 // those dwords can then be moved to the correct half with a dword shuffle.
9031 // We use at most one low and one high word shuffle to collect these paired
9032 // inputs into dwords, and finally a dword shuffle to place them.
9033 int PSHUFLMask[4] = {-1, -1, -1, -1};
9034 int PSHUFHMask[4] = {-1, -1, -1, -1};
9035 int PSHUFDMask[4] = {-1, -1, -1, -1};
9037 // First fix the masks for all the inputs that are staying in their
9038 // original halves. This will then dictate the targets of the cross-half
9040 auto fixInPlaceInputs =
9041 [&PSHUFDMask](ArrayRef<int> InPlaceInputs, ArrayRef<int> IncomingInputs,
9042 MutableArrayRef<int> SourceHalfMask,
9043 MutableArrayRef<int> HalfMask, int HalfOffset) {
9044 if (InPlaceInputs.empty())
9046 if (InPlaceInputs.size() == 1) {
9047 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
9048 InPlaceInputs[0] - HalfOffset;
9049 PSHUFDMask[InPlaceInputs[0] / 2] = InPlaceInputs[0] / 2;
9052 if (IncomingInputs.empty()) {
9053 // Just fix all of the in place inputs.
9054 for (int Input : InPlaceInputs) {
9055 SourceHalfMask[Input - HalfOffset] = Input - HalfOffset;
9056 PSHUFDMask[Input / 2] = Input / 2;
9061 assert(InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!");
9062 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
9063 InPlaceInputs[0] - HalfOffset;
9064 // Put the second input next to the first so that they are packed into
9065 // a dword. We find the adjacent index by toggling the low bit.
9066 int AdjIndex = InPlaceInputs[0] ^ 1;
9067 SourceHalfMask[AdjIndex - HalfOffset] = InPlaceInputs[1] - HalfOffset;
9068 std::replace(HalfMask.begin(), HalfMask.end(), InPlaceInputs[1], AdjIndex);
9069 PSHUFDMask[AdjIndex / 2] = AdjIndex / 2;
9071 fixInPlaceInputs(LToLInputs, HToLInputs, PSHUFLMask, LoMask, 0);
9072 fixInPlaceInputs(HToHInputs, LToHInputs, PSHUFHMask, HiMask, 4);
9074 // Now gather the cross-half inputs and place them into a free dword of
9075 // their target half.
9076 // FIXME: This operation could almost certainly be simplified dramatically to
9077 // look more like the 3-1 fixing operation.
9078 auto moveInputsToRightHalf = [&PSHUFDMask](
9079 MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs,
9080 MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask,
9081 MutableArrayRef<int> FinalSourceHalfMask, int SourceOffset,
9083 auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) {
9084 return SourceHalfMask[Word] != -1 && SourceHalfMask[Word] != Word;
9086 auto isDWordClobbered = [&isWordClobbered](ArrayRef<int> SourceHalfMask,
9088 int LowWord = Word & ~1;
9089 int HighWord = Word | 1;
9090 return isWordClobbered(SourceHalfMask, LowWord) ||
9091 isWordClobbered(SourceHalfMask, HighWord);
9094 if (IncomingInputs.empty())
9097 if (ExistingInputs.empty()) {
9098 // Map any dwords with inputs from them into the right half.
9099 for (int Input : IncomingInputs) {
9100 // If the source half mask maps over the inputs, turn those into
9101 // swaps and use the swapped lane.
9102 if (isWordClobbered(SourceHalfMask, Input - SourceOffset)) {
9103 if (SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == -1) {
9104 SourceHalfMask[SourceHalfMask[Input - SourceOffset]] =
9105 Input - SourceOffset;
9106 // We have to swap the uses in our half mask in one sweep.
9107 for (int &M : HalfMask)
9108 if (M == SourceHalfMask[Input - SourceOffset] + SourceOffset)
9110 else if (M == Input)
9111 M = SourceHalfMask[Input - SourceOffset] + SourceOffset;
9113 assert(SourceHalfMask[SourceHalfMask[Input - SourceOffset]] ==
9114 Input - SourceOffset &&
9115 "Previous placement doesn't match!");
9117 // Note that this correctly re-maps both when we do a swap and when
9118 // we observe the other side of the swap above. We rely on that to
9119 // avoid swapping the members of the input list directly.
9120 Input = SourceHalfMask[Input - SourceOffset] + SourceOffset;
9123 // Map the input's dword into the correct half.
9124 if (PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == -1)
9125 PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] = Input / 2;
9127 assert(PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] ==
9129 "Previous placement doesn't match!");
9132 // And just directly shift any other-half mask elements to be same-half
9133 // as we will have mirrored the dword containing the element into the
9134 // same position within that half.
9135 for (int &M : HalfMask)
9136 if (M >= SourceOffset && M < SourceOffset + 4) {
9137 M = M - SourceOffset + DestOffset;
9138 assert(M >= 0 && "This should never wrap below zero!");
9143 // Ensure we have the input in a viable dword of its current half. This
9144 // is particularly tricky because the original position may be clobbered
9145 // by inputs being moved and *staying* in that half.
9146 if (IncomingInputs.size() == 1) {
9147 if (isWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
9148 int InputFixed = std::find(std::begin(SourceHalfMask),
9149 std::end(SourceHalfMask), -1) -
9150 std::begin(SourceHalfMask) + SourceOffset;
9151 SourceHalfMask[InputFixed - SourceOffset] =
9152 IncomingInputs[0] - SourceOffset;
9153 std::replace(HalfMask.begin(), HalfMask.end(), IncomingInputs[0],
9155 IncomingInputs[0] = InputFixed;
9157 } else if (IncomingInputs.size() == 2) {
9158 if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 ||
9159 isDWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
9160 // We have two non-adjacent or clobbered inputs we need to extract from
9161 // the source half. To do this, we need to map them into some adjacent
9162 // dword slot in the source mask.
9163 int InputsFixed[2] = {IncomingInputs[0] - SourceOffset,
9164 IncomingInputs[1] - SourceOffset};
9166 // If there is a free slot in the source half mask adjacent to one of
9167 // the inputs, place the other input in it. We use (Index XOR 1) to
9168 // compute an adjacent index.
9169 if (!isWordClobbered(SourceHalfMask, InputsFixed[0]) &&
9170 SourceHalfMask[InputsFixed[0] ^ 1] == -1) {
9171 SourceHalfMask[InputsFixed[0]] = InputsFixed[0];
9172 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
9173 InputsFixed[1] = InputsFixed[0] ^ 1;
9174 } else if (!isWordClobbered(SourceHalfMask, InputsFixed[1]) &&
9175 SourceHalfMask[InputsFixed[1] ^ 1] == -1) {
9176 SourceHalfMask[InputsFixed[1]] = InputsFixed[1];
9177 SourceHalfMask[InputsFixed[1] ^ 1] = InputsFixed[0];
9178 InputsFixed[0] = InputsFixed[1] ^ 1;
9179 } else if (SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] == -1 &&
9180 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] == -1) {
9181 // The two inputs are in the same DWord but it is clobbered and the
9182 // adjacent DWord isn't used at all. Move both inputs to the free
9184 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] = InputsFixed[0];
9185 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] = InputsFixed[1];
9186 InputsFixed[0] = 2 * ((InputsFixed[0] / 2) ^ 1);
9187 InputsFixed[1] = 2 * ((InputsFixed[0] / 2) ^ 1) + 1;
9189 // The only way we hit this point is if there is no clobbering
9190 // (because there are no off-half inputs to this half) and there is no
9191 // free slot adjacent to one of the inputs. In this case, we have to
9192 // swap an input with a non-input.
9193 for (int i = 0; i < 4; ++i)
9194 assert((SourceHalfMask[i] == -1 || SourceHalfMask[i] == i) &&
9195 "We can't handle any clobbers here!");
9196 assert(InputsFixed[1] != (InputsFixed[0] ^ 1) &&
9197 "Cannot have adjacent inputs here!");
9199 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
9200 SourceHalfMask[InputsFixed[1]] = InputsFixed[0] ^ 1;
9202 // We also have to update the final source mask in this case because
9203 // it may need to undo the above swap.
9204 for (int &M : FinalSourceHalfMask)
9205 if (M == (InputsFixed[0] ^ 1) + SourceOffset)
9206 M = InputsFixed[1] + SourceOffset;
9207 else if (M == InputsFixed[1] + SourceOffset)
9208 M = (InputsFixed[0] ^ 1) + SourceOffset;
9210 InputsFixed[1] = InputsFixed[0] ^ 1;
9213 // Point everything at the fixed inputs.
9214 for (int &M : HalfMask)
9215 if (M == IncomingInputs[0])
9216 M = InputsFixed[0] + SourceOffset;
9217 else if (M == IncomingInputs[1])
9218 M = InputsFixed[1] + SourceOffset;
9220 IncomingInputs[0] = InputsFixed[0] + SourceOffset;
9221 IncomingInputs[1] = InputsFixed[1] + SourceOffset;
9224 llvm_unreachable("Unhandled input size!");
9227 // Now hoist the DWord down to the right half.
9228 int FreeDWord = (PSHUFDMask[DestOffset / 2] == -1 ? 0 : 1) + DestOffset / 2;
9229 assert(PSHUFDMask[FreeDWord] == -1 && "DWord not free");
9230 PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2;
9231 for (int &M : HalfMask)
9232 for (int Input : IncomingInputs)
9234 M = FreeDWord * 2 + Input % 2;
9236 moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask, HiMask,
9237 /*SourceOffset*/ 4, /*DestOffset*/ 0);
9238 moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask, LoMask,
9239 /*SourceOffset*/ 0, /*DestOffset*/ 4);
9241 // Now enact all the shuffles we've computed to move the inputs into their
9243 if (!isNoopShuffleMask(PSHUFLMask))
9244 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
9245 getV4X86ShuffleImm8ForMask(PSHUFLMask, DL, DAG));
9246 if (!isNoopShuffleMask(PSHUFHMask))
9247 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
9248 getV4X86ShuffleImm8ForMask(PSHUFHMask, DL, DAG));
9249 if (!isNoopShuffleMask(PSHUFDMask))
9252 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
9253 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
9255 // At this point, each half should contain all its inputs, and we can then
9256 // just shuffle them into their final position.
9257 assert(std::count_if(LoMask.begin(), LoMask.end(),
9258 [](int M) { return M >= 4; }) == 0 &&
9259 "Failed to lift all the high half inputs to the low mask!");
9260 assert(std::count_if(HiMask.begin(), HiMask.end(),
9261 [](int M) { return M >= 0 && M < 4; }) == 0 &&
9262 "Failed to lift all the low half inputs to the high mask!");
9264 // Do a half shuffle for the low mask.
9265 if (!isNoopShuffleMask(LoMask))
9266 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
9267 getV4X86ShuffleImm8ForMask(LoMask, DL, DAG));
9269 // Do a half shuffle with the high mask after shifting its values down.
9270 for (int &M : HiMask)
9273 if (!isNoopShuffleMask(HiMask))
9274 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
9275 getV4X86ShuffleImm8ForMask(HiMask, DL, DAG));
9280 /// \brief Helper to form a PSHUFB-based shuffle+blend.
9281 static SDValue lowerVectorShuffleAsPSHUFB(SDLoc DL, MVT VT, SDValue V1,
9282 SDValue V2, ArrayRef<int> Mask,
9283 SelectionDAG &DAG, bool &V1InUse,
9285 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
9291 int Size = Mask.size();
9292 int Scale = 16 / Size;
9293 for (int i = 0; i < 16; ++i) {
9294 if (Mask[i / Scale] == -1) {
9295 V1Mask[i] = V2Mask[i] = DAG.getUNDEF(MVT::i8);
9297 const int ZeroMask = 0x80;
9298 int V1Idx = Mask[i / Scale] < Size ? Mask[i / Scale] * Scale + i % Scale
9300 int V2Idx = Mask[i / Scale] < Size
9302 : (Mask[i / Scale] - Size) * Scale + i % Scale;
9303 if (Zeroable[i / Scale])
9304 V1Idx = V2Idx = ZeroMask;
9305 V1Mask[i] = DAG.getConstant(V1Idx, DL, MVT::i8);
9306 V2Mask[i] = DAG.getConstant(V2Idx, DL, MVT::i8);
9307 V1InUse |= (ZeroMask != V1Idx);
9308 V2InUse |= (ZeroMask != V2Idx);
9313 V1 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
9314 DAG.getBitcast(MVT::v16i8, V1),
9315 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V1Mask));
9317 V2 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
9318 DAG.getBitcast(MVT::v16i8, V2),
9319 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V2Mask));
9321 // If we need shuffled inputs from both, blend the two.
9323 if (V1InUse && V2InUse)
9324 V = DAG.getNode(ISD::OR, DL, MVT::v16i8, V1, V2);
9326 V = V1InUse ? V1 : V2;
9328 // Cast the result back to the correct type.
9329 return DAG.getBitcast(VT, V);
9332 /// \brief Generic lowering of 8-lane i16 shuffles.
9334 /// This handles both single-input shuffles and combined shuffle/blends with
9335 /// two inputs. The single input shuffles are immediately delegated to
9336 /// a dedicated lowering routine.
9338 /// The blends are lowered in one of three fundamental ways. If there are few
9339 /// enough inputs, it delegates to a basic UNPCK-based strategy. If the shuffle
9340 /// of the input is significantly cheaper when lowered as an interleaving of
9341 /// the two inputs, try to interleave them. Otherwise, blend the low and high
9342 /// halves of the inputs separately (making them have relatively few inputs)
9343 /// and then concatenate them.
9344 static SDValue lowerV8I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9345 const X86Subtarget *Subtarget,
9346 SelectionDAG &DAG) {
9348 assert(Op.getSimpleValueType() == MVT::v8i16 && "Bad shuffle type!");
9349 assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
9350 assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
9351 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9352 ArrayRef<int> OrigMask = SVOp->getMask();
9353 int MaskStorage[8] = {OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
9354 OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7]};
9355 MutableArrayRef<int> Mask(MaskStorage);
9357 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9359 // Whenever we can lower this as a zext, that instruction is strictly faster
9360 // than any alternative.
9361 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
9362 DL, MVT::v8i16, V1, V2, OrigMask, Subtarget, DAG))
9365 auto isV1 = [](int M) { return M >= 0 && M < 8; };
9367 auto isV2 = [](int M) { return M >= 8; };
9369 int NumV2Inputs = std::count_if(Mask.begin(), Mask.end(), isV2);
9371 if (NumV2Inputs == 0) {
9372 // Check for being able to broadcast a single element.
9373 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i16, V1,
9374 Mask, Subtarget, DAG))
9377 // Try to use shift instructions.
9379 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V1, Mask, DAG))
9382 // Use dedicated unpack instructions for masks that match their pattern.
9384 lowerVectorShuffleWithUNPCK(DL, MVT::v8i16, Mask, V1, V2, DAG))
9387 // Try to use byte rotation instructions.
9388 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(DL, MVT::v8i16, V1, V1,
9389 Mask, Subtarget, DAG))
9392 return lowerV8I16GeneralSingleInputVectorShuffle(DL, MVT::v8i16, V1, Mask,
9396 assert(std::any_of(Mask.begin(), Mask.end(), isV1) &&
9397 "All single-input shuffles should be canonicalized to be V1-input "
9400 // Try to use shift instructions.
9402 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V2, Mask, DAG))
9405 // See if we can use SSE4A Extraction / Insertion.
9406 if (Subtarget->hasSSE4A())
9407 if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v8i16, V1, V2, Mask, DAG))
9410 // There are special ways we can lower some single-element blends.
9411 if (NumV2Inputs == 1)
9412 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v8i16, V1, V2,
9413 Mask, Subtarget, DAG))
9416 // We have different paths for blend lowering, but they all must use the
9417 // *exact* same predicate.
9418 bool IsBlendSupported = Subtarget->hasSSE41();
9419 if (IsBlendSupported)
9420 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i16, V1, V2, Mask,
9424 if (SDValue Masked =
9425 lowerVectorShuffleAsBitMask(DL, MVT::v8i16, V1, V2, Mask, DAG))
9428 // Use dedicated unpack instructions for masks that match their pattern.
9430 lowerVectorShuffleWithUNPCK(DL, MVT::v8i16, Mask, V1, V2, DAG))
9433 // Try to use byte rotation instructions.
9434 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9435 DL, MVT::v8i16, V1, V2, Mask, Subtarget, DAG))
9438 if (SDValue BitBlend =
9439 lowerVectorShuffleAsBitBlend(DL, MVT::v8i16, V1, V2, Mask, DAG))
9442 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(DL, MVT::v8i16, V1,
9446 // If we can't directly blend but can use PSHUFB, that will be better as it
9447 // can both shuffle and set up the inefficient blend.
9448 if (!IsBlendSupported && Subtarget->hasSSSE3()) {
9449 bool V1InUse, V2InUse;
9450 return lowerVectorShuffleAsPSHUFB(DL, MVT::v8i16, V1, V2, Mask, DAG,
9454 // We can always bit-blend if we have to so the fallback strategy is to
9455 // decompose into single-input permutes and blends.
9456 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i16, V1, V2,
9460 /// \brief Check whether a compaction lowering can be done by dropping even
9461 /// elements and compute how many times even elements must be dropped.
9463 /// This handles shuffles which take every Nth element where N is a power of
9464 /// two. Example shuffle masks:
9466 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 0, 2, 4, 6, 8, 10, 12, 14
9467 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30
9468 /// N = 2: 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12
9469 /// N = 2: 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28
9470 /// N = 3: 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8
9471 /// N = 3: 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24
9473 /// Any of these lanes can of course be undef.
9475 /// This routine only supports N <= 3.
9476 /// FIXME: Evaluate whether either AVX or AVX-512 have any opportunities here
9479 /// \returns N above, or the number of times even elements must be dropped if
9480 /// there is such a number. Otherwise returns zero.
9481 static int canLowerByDroppingEvenElements(ArrayRef<int> Mask) {
9482 // Figure out whether we're looping over two inputs or just one.
9483 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9485 // The modulus for the shuffle vector entries is based on whether this is
9486 // a single input or not.
9487 int ShuffleModulus = Mask.size() * (IsSingleInput ? 1 : 2);
9488 assert(isPowerOf2_32((uint32_t)ShuffleModulus) &&
9489 "We should only be called with masks with a power-of-2 size!");
9491 uint64_t ModMask = (uint64_t)ShuffleModulus - 1;
9493 // We track whether the input is viable for all power-of-2 strides 2^1, 2^2,
9494 // and 2^3 simultaneously. This is because we may have ambiguity with
9495 // partially undef inputs.
9496 bool ViableForN[3] = {true, true, true};
9498 for (int i = 0, e = Mask.size(); i < e; ++i) {
9499 // Ignore undef lanes, we'll optimistically collapse them to the pattern we
9504 bool IsAnyViable = false;
9505 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
9506 if (ViableForN[j]) {
9509 // The shuffle mask must be equal to (i * 2^N) % M.
9510 if ((uint64_t)Mask[i] == (((uint64_t)i << N) & ModMask))
9513 ViableForN[j] = false;
9515 // Early exit if we exhaust the possible powers of two.
9520 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
9524 // Return 0 as there is no viable power of two.
9528 /// \brief Generic lowering of v16i8 shuffles.
9530 /// This is a hybrid strategy to lower v16i8 vectors. It first attempts to
9531 /// detect any complexity reducing interleaving. If that doesn't help, it uses
9532 /// UNPCK to spread the i8 elements across two i16-element vectors, and uses
9533 /// the existing lowering for v8i16 blends on each half, finally PACK-ing them
9535 static SDValue lowerV16I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9536 const X86Subtarget *Subtarget,
9537 SelectionDAG &DAG) {
9539 assert(Op.getSimpleValueType() == MVT::v16i8 && "Bad shuffle type!");
9540 assert(V1.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
9541 assert(V2.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
9542 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9543 ArrayRef<int> Mask = SVOp->getMask();
9544 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
9546 // Try to use shift instructions.
9548 lowerVectorShuffleAsShift(DL, MVT::v16i8, V1, V2, Mask, DAG))
9551 // Try to use byte rotation instructions.
9552 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9553 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
9556 // Try to use a zext lowering.
9557 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
9558 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
9561 // See if we can use SSE4A Extraction / Insertion.
9562 if (Subtarget->hasSSE4A())
9563 if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v16i8, V1, V2, Mask, DAG))
9567 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 16; });
9569 // For single-input shuffles, there are some nicer lowering tricks we can use.
9570 if (NumV2Elements == 0) {
9571 // Check for being able to broadcast a single element.
9572 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i8, V1,
9573 Mask, Subtarget, DAG))
9576 // Check whether we can widen this to an i16 shuffle by duplicating bytes.
9577 // Notably, this handles splat and partial-splat shuffles more efficiently.
9578 // However, it only makes sense if the pre-duplication shuffle simplifies
9579 // things significantly. Currently, this means we need to be able to
9580 // express the pre-duplication shuffle as an i16 shuffle.
9582 // FIXME: We should check for other patterns which can be widened into an
9583 // i16 shuffle as well.
9584 auto canWidenViaDuplication = [](ArrayRef<int> Mask) {
9585 for (int i = 0; i < 16; i += 2)
9586 if (Mask[i] != -1 && Mask[i + 1] != -1 && Mask[i] != Mask[i + 1])
9591 auto tryToWidenViaDuplication = [&]() -> SDValue {
9592 if (!canWidenViaDuplication(Mask))
9594 SmallVector<int, 4> LoInputs;
9595 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(LoInputs),
9596 [](int M) { return M >= 0 && M < 8; });
9597 std::sort(LoInputs.begin(), LoInputs.end());
9598 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()),
9600 SmallVector<int, 4> HiInputs;
9601 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(HiInputs),
9602 [](int M) { return M >= 8; });
9603 std::sort(HiInputs.begin(), HiInputs.end());
9604 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()),
9607 bool TargetLo = LoInputs.size() >= HiInputs.size();
9608 ArrayRef<int> InPlaceInputs = TargetLo ? LoInputs : HiInputs;
9609 ArrayRef<int> MovingInputs = TargetLo ? HiInputs : LoInputs;
9611 int PreDupI16Shuffle[] = {-1, -1, -1, -1, -1, -1, -1, -1};
9612 SmallDenseMap<int, int, 8> LaneMap;
9613 for (int I : InPlaceInputs) {
9614 PreDupI16Shuffle[I/2] = I/2;
9617 int j = TargetLo ? 0 : 4, je = j + 4;
9618 for (int i = 0, ie = MovingInputs.size(); i < ie; ++i) {
9619 // Check if j is already a shuffle of this input. This happens when
9620 // there are two adjacent bytes after we move the low one.
9621 if (PreDupI16Shuffle[j] != MovingInputs[i] / 2) {
9622 // If we haven't yet mapped the input, search for a slot into which
9624 while (j < je && PreDupI16Shuffle[j] != -1)
9628 // We can't place the inputs into a single half with a simple i16 shuffle, so bail.
9631 // Map this input with the i16 shuffle.
9632 PreDupI16Shuffle[j] = MovingInputs[i] / 2;
9635 // Update the lane map based on the mapping we ended up with.
9636 LaneMap[MovingInputs[i]] = 2 * j + MovingInputs[i] % 2;
9638 V1 = DAG.getBitcast(
9640 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
9641 DAG.getUNDEF(MVT::v8i16), PreDupI16Shuffle));
9643 // Unpack the bytes to form the i16s that will be shuffled into place.
9644 V1 = DAG.getNode(TargetLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
9645 MVT::v16i8, V1, V1);
9647 int PostDupI16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9648 for (int i = 0; i < 16; ++i)
9649 if (Mask[i] != -1) {
9650 int MappedMask = LaneMap[Mask[i]] - (TargetLo ? 0 : 8);
9651 assert(MappedMask < 8 && "Invalid v8 shuffle mask!");
9652 if (PostDupI16Shuffle[i / 2] == -1)
9653 PostDupI16Shuffle[i / 2] = MappedMask;
9655 assert(PostDupI16Shuffle[i / 2] == MappedMask &&
9656 "Conflicting entrties in the original shuffle!");
9658 return DAG.getBitcast(
9660 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
9661 DAG.getUNDEF(MVT::v8i16), PostDupI16Shuffle));
9663 if (SDValue V = tryToWidenViaDuplication())
9667 if (SDValue Masked =
9668 lowerVectorShuffleAsBitMask(DL, MVT::v16i8, V1, V2, Mask, DAG))
9671 // Use dedicated unpack instructions for masks that match their pattern.
9673 lowerVectorShuffleWithUNPCK(DL, MVT::v16i8, Mask, V1, V2, DAG))
9676 // Check for SSSE3 which lets us lower all v16i8 shuffles much more directly
9677 // with PSHUFB. It is important to do this before we attempt to generate any
9678 // blends but after all of the single-input lowerings. If the single input
9679 // lowerings can find an instruction sequence that is faster than a PSHUFB, we
9680 // want to preserve that and we can DAG combine any longer sequences into
9681 // a PSHUFB in the end. But once we start blending from multiple inputs,
9682 // the complexity of DAG combining bad patterns back into PSHUFB is too high,
9683 // and there are *very* few patterns that would actually be faster than the
9684 // PSHUFB approach because of its ability to zero lanes.
9686 // FIXME: The only exceptions to the above are blends which are exact
9687 // interleavings with direct instructions supporting them. We currently don't
9688 // handle those well here.
9689 if (Subtarget->hasSSSE3()) {
9690 bool V1InUse = false;
9691 bool V2InUse = false;
9693 SDValue PSHUFB = lowerVectorShuffleAsPSHUFB(DL, MVT::v16i8, V1, V2, Mask,
9694 DAG, V1InUse, V2InUse);
9696 // If both V1 and V2 are in use and we can use a direct blend or an unpack,
9697 // do so. This avoids using them to handle blends-with-zero which is
9698 // important as a single pshufb is significantly faster for that.
9699 if (V1InUse && V2InUse) {
9700 if (Subtarget->hasSSE41())
9701 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i8, V1, V2,
9702 Mask, Subtarget, DAG))
9705 // We can use an unpack to do the blending rather than an or in some
9706 // cases. Even though the or may be (very minorly) more efficient, we
9707 // preference this lowering because there are common cases where part of
9708 // the complexity of the shuffles goes away when we do the final blend as
9710 // FIXME: It might be worth trying to detect if the unpack-feeding
9711 // shuffles will both be pshufb, in which case we shouldn't bother with
9713 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(
9714 DL, MVT::v16i8, V1, V2, Mask, DAG))
9721 // There are special ways we can lower some single-element blends.
9722 if (NumV2Elements == 1)
9723 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v16i8, V1, V2,
9724 Mask, Subtarget, DAG))
9727 if (SDValue BitBlend =
9728 lowerVectorShuffleAsBitBlend(DL, MVT::v16i8, V1, V2, Mask, DAG))
9731 // Check whether a compaction lowering can be done. This handles shuffles
9732 // which take every Nth element for some even N. See the helper function for
9735 // We special case these as they can be particularly efficiently handled with
9736 // the PACKUSB instruction on x86 and they show up in common patterns of
9737 // rearranging bytes to truncate wide elements.
9738 if (int NumEvenDrops = canLowerByDroppingEvenElements(Mask)) {
9739 // NumEvenDrops is the power of two stride of the elements. Another way of
9740 // thinking about it is that we need to drop the even elements this many
9741 // times to get the original input.
9742 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9744 // First we need to zero all the dropped bytes.
9745 assert(NumEvenDrops <= 3 &&
9746 "No support for dropping even elements more than 3 times.");
9747 // We use the mask type to pick which bytes are preserved based on how many
9748 // elements are dropped.
9749 MVT MaskVTs[] = { MVT::v8i16, MVT::v4i32, MVT::v2i64 };
9750 SDValue ByteClearMask = DAG.getBitcast(
9751 MVT::v16i8, DAG.getConstant(0xFF, DL, MaskVTs[NumEvenDrops - 1]));
9752 V1 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V1, ByteClearMask);
9754 V2 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V2, ByteClearMask);
9756 // Now pack things back together.
9757 V1 = DAG.getBitcast(MVT::v8i16, V1);
9758 V2 = IsSingleInput ? V1 : DAG.getBitcast(MVT::v8i16, V2);
9759 SDValue Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, V1, V2);
9760 for (int i = 1; i < NumEvenDrops; ++i) {
9761 Result = DAG.getBitcast(MVT::v8i16, Result);
9762 Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, Result, Result);
9768 // Handle multi-input cases by blending single-input shuffles.
9769 if (NumV2Elements > 0)
9770 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v16i8, V1, V2,
9773 // The fallback path for single-input shuffles widens this into two v8i16
9774 // vectors with unpacks, shuffles those, and then pulls them back together
9778 int LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9779 int HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9780 for (int i = 0; i < 16; ++i)
9782 (i < 8 ? LoBlendMask[i] : HiBlendMask[i % 8]) = Mask[i];
9784 SDValue Zero = getZeroVector(MVT::v8i16, Subtarget, DAG, DL);
9786 SDValue VLoHalf, VHiHalf;
9787 // Check if any of the odd lanes in the v16i8 are used. If not, we can mask
9788 // them out and avoid using UNPCK{L,H} to extract the elements of V as
9790 if (std::none_of(std::begin(LoBlendMask), std::end(LoBlendMask),
9791 [](int M) { return M >= 0 && M % 2 == 1; }) &&
9792 std::none_of(std::begin(HiBlendMask), std::end(HiBlendMask),
9793 [](int M) { return M >= 0 && M % 2 == 1; })) {
9794 // Use a mask to drop the high bytes.
9795 VLoHalf = DAG.getBitcast(MVT::v8i16, V);
9796 VLoHalf = DAG.getNode(ISD::AND, DL, MVT::v8i16, VLoHalf,
9797 DAG.getConstant(0x00FF, DL, MVT::v8i16));
9799 // This will be a single vector shuffle instead of a blend so nuke VHiHalf.
9800 VHiHalf = DAG.getUNDEF(MVT::v8i16);
9802 // Squash the masks to point directly into VLoHalf.
9803 for (int &M : LoBlendMask)
9806 for (int &M : HiBlendMask)
9810 // Otherwise just unpack the low half of V into VLoHalf and the high half into
9811 // VHiHalf so that we can blend them as i16s.
9812 VLoHalf = DAG.getBitcast(
9813 MVT::v8i16, DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V, Zero));
9814 VHiHalf = DAG.getBitcast(
9815 MVT::v8i16, DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V, Zero));
9818 SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, LoBlendMask);
9819 SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, HiBlendMask);
9821 return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, LoV, HiV);
9824 /// \brief Dispatching routine to lower various 128-bit x86 vector shuffles.
9826 /// This routine breaks down the specific type of 128-bit shuffle and
9827 /// dispatches to the lowering routines accordingly.
9828 static SDValue lower128BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9829 MVT VT, const X86Subtarget *Subtarget,
9830 SelectionDAG &DAG) {
9831 switch (VT.SimpleTy) {
9833 return lowerV2I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9835 return lowerV2F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9837 return lowerV4I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9839 return lowerV4F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9841 return lowerV8I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
9843 return lowerV16I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
9846 llvm_unreachable("Unimplemented!");
9850 /// \brief Helper function to test whether a shuffle mask could be
9851 /// simplified by widening the elements being shuffled.
9853 /// Appends the mask for wider elements in WidenedMask if valid. Otherwise
9854 /// leaves it in an unspecified state.
9856 /// NOTE: This must handle normal vector shuffle masks and *target* vector
9857 /// shuffle masks. The latter have the special property of a '-2' representing
9858 /// a zero-ed lane of a vector.
9859 static bool canWidenShuffleElements(ArrayRef<int> Mask,
9860 SmallVectorImpl<int> &WidenedMask) {
9861 for (int i = 0, Size = Mask.size(); i < Size; i += 2) {
9862 // If both elements are undef, its trivial.
9863 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] == SM_SentinelUndef) {
9864 WidenedMask.push_back(SM_SentinelUndef);
9868 // Check for an undef mask and a mask value properly aligned to fit with
9869 // a pair of values. If we find such a case, use the non-undef mask's value.
9870 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] >= 0 && Mask[i + 1] % 2 == 1) {
9871 WidenedMask.push_back(Mask[i + 1] / 2);
9874 if (Mask[i + 1] == SM_SentinelUndef && Mask[i] >= 0 && Mask[i] % 2 == 0) {
9875 WidenedMask.push_back(Mask[i] / 2);
9879 // When zeroing, we need to spread the zeroing across both lanes to widen.
9880 if (Mask[i] == SM_SentinelZero || Mask[i + 1] == SM_SentinelZero) {
9881 if ((Mask[i] == SM_SentinelZero || Mask[i] == SM_SentinelUndef) &&
9882 (Mask[i + 1] == SM_SentinelZero || Mask[i + 1] == SM_SentinelUndef)) {
9883 WidenedMask.push_back(SM_SentinelZero);
9889 // Finally check if the two mask values are adjacent and aligned with
9891 if (Mask[i] != SM_SentinelUndef && Mask[i] % 2 == 0 && Mask[i] + 1 == Mask[i + 1]) {
9892 WidenedMask.push_back(Mask[i] / 2);
9896 // Otherwise we can't safely widen the elements used in this shuffle.
9899 assert(WidenedMask.size() == Mask.size() / 2 &&
9900 "Incorrect size of mask after widening the elements!");
9905 /// \brief Generic routine to split vector shuffle into half-sized shuffles.
9907 /// This routine just extracts two subvectors, shuffles them independently, and
9908 /// then concatenates them back together. This should work effectively with all
9909 /// AVX vector shuffle types.
9910 static SDValue splitAndLowerVectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9911 SDValue V2, ArrayRef<int> Mask,
9912 SelectionDAG &DAG) {
9913 assert(VT.getSizeInBits() >= 256 &&
9914 "Only for 256-bit or wider vector shuffles!");
9915 assert(V1.getSimpleValueType() == VT && "Bad operand type!");
9916 assert(V2.getSimpleValueType() == VT && "Bad operand type!");
9918 ArrayRef<int> LoMask = Mask.slice(0, Mask.size() / 2);
9919 ArrayRef<int> HiMask = Mask.slice(Mask.size() / 2);
9921 int NumElements = VT.getVectorNumElements();
9922 int SplitNumElements = NumElements / 2;
9923 MVT ScalarVT = VT.getVectorElementType();
9924 MVT SplitVT = MVT::getVectorVT(ScalarVT, NumElements / 2);
9926 // Rather than splitting build-vectors, just build two narrower build
9927 // vectors. This helps shuffling with splats and zeros.
9928 auto SplitVector = [&](SDValue V) {
9929 while (V.getOpcode() == ISD::BITCAST)
9930 V = V->getOperand(0);
9932 MVT OrigVT = V.getSimpleValueType();
9933 int OrigNumElements = OrigVT.getVectorNumElements();
9934 int OrigSplitNumElements = OrigNumElements / 2;
9935 MVT OrigScalarVT = OrigVT.getVectorElementType();
9936 MVT OrigSplitVT = MVT::getVectorVT(OrigScalarVT, OrigNumElements / 2);
9940 auto *BV = dyn_cast<BuildVectorSDNode>(V);
9942 LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9943 DAG.getIntPtrConstant(0, DL));
9944 HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9945 DAG.getIntPtrConstant(OrigSplitNumElements, DL));
9948 SmallVector<SDValue, 16> LoOps, HiOps;
9949 for (int i = 0; i < OrigSplitNumElements; ++i) {
9950 LoOps.push_back(BV->getOperand(i));
9951 HiOps.push_back(BV->getOperand(i + OrigSplitNumElements));
9953 LoV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, LoOps);
9954 HiV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, HiOps);
9956 return std::make_pair(DAG.getBitcast(SplitVT, LoV),
9957 DAG.getBitcast(SplitVT, HiV));
9960 SDValue LoV1, HiV1, LoV2, HiV2;
9961 std::tie(LoV1, HiV1) = SplitVector(V1);
9962 std::tie(LoV2, HiV2) = SplitVector(V2);
9964 // Now create two 4-way blends of these half-width vectors.
9965 auto HalfBlend = [&](ArrayRef<int> HalfMask) {
9966 bool UseLoV1 = false, UseHiV1 = false, UseLoV2 = false, UseHiV2 = false;
9967 SmallVector<int, 32> V1BlendMask, V2BlendMask, BlendMask;
9968 for (int i = 0; i < SplitNumElements; ++i) {
9969 int M = HalfMask[i];
9970 if (M >= NumElements) {
9971 if (M >= NumElements + SplitNumElements)
9975 V2BlendMask.push_back(M - NumElements);
9976 V1BlendMask.push_back(-1);
9977 BlendMask.push_back(SplitNumElements + i);
9978 } else if (M >= 0) {
9979 if (M >= SplitNumElements)
9983 V2BlendMask.push_back(-1);
9984 V1BlendMask.push_back(M);
9985 BlendMask.push_back(i);
9987 V2BlendMask.push_back(-1);
9988 V1BlendMask.push_back(-1);
9989 BlendMask.push_back(-1);
9993 // Because the lowering happens after all combining takes place, we need to
9994 // manually combine these blend masks as much as possible so that we create
9995 // a minimal number of high-level vector shuffle nodes.
9997 // First try just blending the halves of V1 or V2.
9998 if (!UseLoV1 && !UseHiV1 && !UseLoV2 && !UseHiV2)
9999 return DAG.getUNDEF(SplitVT);
10000 if (!UseLoV2 && !UseHiV2)
10001 return DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
10002 if (!UseLoV1 && !UseHiV1)
10003 return DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
10005 SDValue V1Blend, V2Blend;
10006 if (UseLoV1 && UseHiV1) {
10008 DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
10010 // We only use half of V1 so map the usage down into the final blend mask.
10011 V1Blend = UseLoV1 ? LoV1 : HiV1;
10012 for (int i = 0; i < SplitNumElements; ++i)
10013 if (BlendMask[i] >= 0 && BlendMask[i] < SplitNumElements)
10014 BlendMask[i] = V1BlendMask[i] - (UseLoV1 ? 0 : SplitNumElements);
10016 if (UseLoV2 && UseHiV2) {
10018 DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
10020 // We only use half of V2 so map the usage down into the final blend mask.
10021 V2Blend = UseLoV2 ? LoV2 : HiV2;
10022 for (int i = 0; i < SplitNumElements; ++i)
10023 if (BlendMask[i] >= SplitNumElements)
10024 BlendMask[i] = V2BlendMask[i] + (UseLoV2 ? SplitNumElements : 0);
10026 return DAG.getVectorShuffle(SplitVT, DL, V1Blend, V2Blend, BlendMask);
10028 SDValue Lo = HalfBlend(LoMask);
10029 SDValue Hi = HalfBlend(HiMask);
10030 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
10033 /// \brief Either split a vector in halves or decompose the shuffles and the
10036 /// This is provided as a good fallback for many lowerings of non-single-input
10037 /// shuffles with more than one 128-bit lane. In those cases, we want to select
10038 /// between splitting the shuffle into 128-bit components and stitching those
10039 /// back together vs. extracting the single-input shuffles and blending those
10041 static SDValue lowerVectorShuffleAsSplitOrBlend(SDLoc DL, MVT VT, SDValue V1,
10042 SDValue V2, ArrayRef<int> Mask,
10043 SelectionDAG &DAG) {
10044 assert(!isSingleInputShuffleMask(Mask) && "This routine must not be used to "
10045 "lower single-input shuffles as it "
10046 "could then recurse on itself.");
10047 int Size = Mask.size();
10049 // If this can be modeled as a broadcast of two elements followed by a blend,
10050 // prefer that lowering. This is especially important because broadcasts can
10051 // often fold with memory operands.
10052 auto DoBothBroadcast = [&] {
10053 int V1BroadcastIdx = -1, V2BroadcastIdx = -1;
10056 if (V2BroadcastIdx == -1)
10057 V2BroadcastIdx = M - Size;
10058 else if (M - Size != V2BroadcastIdx)
10060 } else if (M >= 0) {
10061 if (V1BroadcastIdx == -1)
10062 V1BroadcastIdx = M;
10063 else if (M != V1BroadcastIdx)
10068 if (DoBothBroadcast())
10069 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask,
10072 // If the inputs all stem from a single 128-bit lane of each input, then we
10073 // split them rather than blending because the split will decompose to
10074 // unusually few instructions.
10075 int LaneCount = VT.getSizeInBits() / 128;
10076 int LaneSize = Size / LaneCount;
10077 SmallBitVector LaneInputs[2];
10078 LaneInputs[0].resize(LaneCount, false);
10079 LaneInputs[1].resize(LaneCount, false);
10080 for (int i = 0; i < Size; ++i)
10082 LaneInputs[Mask[i] / Size][(Mask[i] % Size) / LaneSize] = true;
10083 if (LaneInputs[0].count() <= 1 && LaneInputs[1].count() <= 1)
10084 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10086 // Otherwise, just fall back to decomposed shuffles and a blend. This requires
10087 // that the decomposed single-input shuffles don't end up here.
10088 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
10091 /// \brief Lower a vector shuffle crossing multiple 128-bit lanes as
10092 /// a permutation and blend of those lanes.
10094 /// This essentially blends the out-of-lane inputs to each lane into the lane
10095 /// from a permuted copy of the vector. This lowering strategy results in four
10096 /// instructions in the worst case for a single-input cross lane shuffle which
10097 /// is lower than any other fully general cross-lane shuffle strategy I'm aware
10098 /// of. Special cases for each particular shuffle pattern should be handled
10099 /// prior to trying this lowering.
10100 static SDValue lowerVectorShuffleAsLanePermuteAndBlend(SDLoc DL, MVT VT,
10101 SDValue V1, SDValue V2,
10102 ArrayRef<int> Mask,
10103 SelectionDAG &DAG) {
10104 // FIXME: This should probably be generalized for 512-bit vectors as well.
10105 assert(VT.is256BitVector() && "Only for 256-bit vector shuffles!");
10106 int LaneSize = Mask.size() / 2;
10108 // If there are only inputs from one 128-bit lane, splitting will in fact be
10109 // less expensive. The flags track whether the given lane contains an element
10110 // that crosses to another lane.
10111 bool LaneCrossing[2] = {false, false};
10112 for (int i = 0, Size = Mask.size(); i < Size; ++i)
10113 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
10114 LaneCrossing[(Mask[i] % Size) / LaneSize] = true;
10115 if (!LaneCrossing[0] || !LaneCrossing[1])
10116 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10118 if (isSingleInputShuffleMask(Mask)) {
10119 SmallVector<int, 32> FlippedBlendMask;
10120 for (int i = 0, Size = Mask.size(); i < Size; ++i)
10121 FlippedBlendMask.push_back(
10122 Mask[i] < 0 ? -1 : (((Mask[i] % Size) / LaneSize == i / LaneSize)
10124 : Mask[i] % LaneSize +
10125 (i / LaneSize) * LaneSize + Size));
10127 // Flip the vector, and blend the results which should now be in-lane. The
10128 // VPERM2X128 mask uses the low 2 bits for the low source and bits 4 and
10129 // 5 for the high source. The value 3 selects the high half of source 2 and
10130 // the value 2 selects the low half of source 2. We only use source 2 to
10131 // allow folding it into a memory operand.
10132 unsigned PERMMask = 3 | 2 << 4;
10133 SDValue Flipped = DAG.getNode(X86ISD::VPERM2X128, DL, VT, DAG.getUNDEF(VT),
10134 V1, DAG.getConstant(PERMMask, DL, MVT::i8));
10135 return DAG.getVectorShuffle(VT, DL, V1, Flipped, FlippedBlendMask);
10138 // This now reduces to two single-input shuffles of V1 and V2 which at worst
10139 // will be handled by the above logic and a blend of the results, much like
10140 // other patterns in AVX.
10141 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
10144 /// \brief Handle lowering 2-lane 128-bit shuffles.
10145 static SDValue lowerV2X128VectorShuffle(SDLoc DL, MVT VT, SDValue V1,
10146 SDValue V2, ArrayRef<int> Mask,
10147 const X86Subtarget *Subtarget,
10148 SelectionDAG &DAG) {
10149 // TODO: If minimizing size and one of the inputs is a zero vector and the
10150 // the zero vector has only one use, we could use a VPERM2X128 to save the
10151 // instruction bytes needed to explicitly generate the zero vector.
10153 // Blends are faster and handle all the non-lane-crossing cases.
10154 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, VT, V1, V2, Mask,
10158 bool IsV1Zero = ISD::isBuildVectorAllZeros(V1.getNode());
10159 bool IsV2Zero = ISD::isBuildVectorAllZeros(V2.getNode());
10161 // If either input operand is a zero vector, use VPERM2X128 because its mask
10162 // allows us to replace the zero input with an implicit zero.
10163 if (!IsV1Zero && !IsV2Zero) {
10164 // Check for patterns which can be matched with a single insert of a 128-bit
10166 bool OnlyUsesV1 = isShuffleEquivalent(V1, V2, Mask, {0, 1, 0, 1});
10167 if (OnlyUsesV1 || isShuffleEquivalent(V1, V2, Mask, {0, 1, 4, 5})) {
10168 MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(),
10169 VT.getVectorNumElements() / 2);
10170 SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
10171 DAG.getIntPtrConstant(0, DL));
10172 SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT,
10173 OnlyUsesV1 ? V1 : V2,
10174 DAG.getIntPtrConstant(0, DL));
10175 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
10179 // Otherwise form a 128-bit permutation. After accounting for undefs,
10180 // convert the 64-bit shuffle mask selection values into 128-bit
10181 // selection bits by dividing the indexes by 2 and shifting into positions
10182 // defined by a vperm2*128 instruction's immediate control byte.
10184 // The immediate permute control byte looks like this:
10185 // [1:0] - select 128 bits from sources for low half of destination
10187 // [3] - zero low half of destination
10188 // [5:4] - select 128 bits from sources for high half of destination
10190 // [7] - zero high half of destination
10192 int MaskLO = Mask[0];
10193 if (MaskLO == SM_SentinelUndef)
10194 MaskLO = Mask[1] == SM_SentinelUndef ? 0 : Mask[1];
10196 int MaskHI = Mask[2];
10197 if (MaskHI == SM_SentinelUndef)
10198 MaskHI = Mask[3] == SM_SentinelUndef ? 0 : Mask[3];
10200 unsigned PermMask = MaskLO / 2 | (MaskHI / 2) << 4;
10202 // If either input is a zero vector, replace it with an undef input.
10203 // Shuffle mask values < 4 are selecting elements of V1.
10204 // Shuffle mask values >= 4 are selecting elements of V2.
10205 // Adjust each half of the permute mask by clearing the half that was
10206 // selecting the zero vector and setting the zero mask bit.
10208 V1 = DAG.getUNDEF(VT);
10210 PermMask = (PermMask & 0xf0) | 0x08;
10212 PermMask = (PermMask & 0x0f) | 0x80;
10215 V2 = DAG.getUNDEF(VT);
10217 PermMask = (PermMask & 0xf0) | 0x08;
10219 PermMask = (PermMask & 0x0f) | 0x80;
10222 return DAG.getNode(X86ISD::VPERM2X128, DL, VT, V1, V2,
10223 DAG.getConstant(PermMask, DL, MVT::i8));
10226 /// \brief Lower a vector shuffle by first fixing the 128-bit lanes and then
10227 /// shuffling each lane.
10229 /// This will only succeed when the result of fixing the 128-bit lanes results
10230 /// in a single-input non-lane-crossing shuffle with a repeating shuffle mask in
10231 /// each 128-bit lanes. This handles many cases where we can quickly blend away
10232 /// the lane crosses early and then use simpler shuffles within each lane.
10234 /// FIXME: It might be worthwhile at some point to support this without
10235 /// requiring the 128-bit lane-relative shuffles to be repeating, but currently
10236 /// in x86 only floating point has interesting non-repeating shuffles, and even
10237 /// those are still *marginally* more expensive.
10238 static SDValue lowerVectorShuffleByMerging128BitLanes(
10239 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
10240 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
10241 assert(!isSingleInputShuffleMask(Mask) &&
10242 "This is only useful with multiple inputs.");
10244 int Size = Mask.size();
10245 int LaneSize = 128 / VT.getScalarSizeInBits();
10246 int NumLanes = Size / LaneSize;
10247 assert(NumLanes > 1 && "Only handles 256-bit and wider shuffles.");
10249 // See if we can build a hypothetical 128-bit lane-fixing shuffle mask. Also
10250 // check whether the in-128-bit lane shuffles share a repeating pattern.
10251 SmallVector<int, 4> Lanes;
10252 Lanes.resize(NumLanes, -1);
10253 SmallVector<int, 4> InLaneMask;
10254 InLaneMask.resize(LaneSize, -1);
10255 for (int i = 0; i < Size; ++i) {
10259 int j = i / LaneSize;
10261 if (Lanes[j] < 0) {
10262 // First entry we've seen for this lane.
10263 Lanes[j] = Mask[i] / LaneSize;
10264 } else if (Lanes[j] != Mask[i] / LaneSize) {
10265 // This doesn't match the lane selected previously!
10269 // Check that within each lane we have a consistent shuffle mask.
10270 int k = i % LaneSize;
10271 if (InLaneMask[k] < 0) {
10272 InLaneMask[k] = Mask[i] % LaneSize;
10273 } else if (InLaneMask[k] != Mask[i] % LaneSize) {
10274 // This doesn't fit a repeating in-lane mask.
10279 // First shuffle the lanes into place.
10280 MVT LaneVT = MVT::getVectorVT(VT.isFloatingPoint() ? MVT::f64 : MVT::i64,
10281 VT.getSizeInBits() / 64);
10282 SmallVector<int, 8> LaneMask;
10283 LaneMask.resize(NumLanes * 2, -1);
10284 for (int i = 0; i < NumLanes; ++i)
10285 if (Lanes[i] >= 0) {
10286 LaneMask[2 * i + 0] = 2*Lanes[i] + 0;
10287 LaneMask[2 * i + 1] = 2*Lanes[i] + 1;
10290 V1 = DAG.getBitcast(LaneVT, V1);
10291 V2 = DAG.getBitcast(LaneVT, V2);
10292 SDValue LaneShuffle = DAG.getVectorShuffle(LaneVT, DL, V1, V2, LaneMask);
10294 // Cast it back to the type we actually want.
10295 LaneShuffle = DAG.getBitcast(VT, LaneShuffle);
10297 // Now do a simple shuffle that isn't lane crossing.
10298 SmallVector<int, 8> NewMask;
10299 NewMask.resize(Size, -1);
10300 for (int i = 0; i < Size; ++i)
10302 NewMask[i] = (i / LaneSize) * LaneSize + Mask[i] % LaneSize;
10303 assert(!is128BitLaneCrossingShuffleMask(VT, NewMask) &&
10304 "Must not introduce lane crosses at this point!");
10306 return DAG.getVectorShuffle(VT, DL, LaneShuffle, DAG.getUNDEF(VT), NewMask);
10309 /// \brief Test whether the specified input (0 or 1) is in-place blended by the
10312 /// This returns true if the elements from a particular input are already in the
10313 /// slot required by the given mask and require no permutation.
10314 static bool isShuffleMaskInputInPlace(int Input, ArrayRef<int> Mask) {
10315 assert((Input == 0 || Input == 1) && "Only two inputs to shuffles.");
10316 int Size = Mask.size();
10317 for (int i = 0; i < Size; ++i)
10318 if (Mask[i] >= 0 && Mask[i] / Size == Input && Mask[i] % Size != i)
10324 static SDValue lowerVectorShuffleWithSHUFPD(SDLoc DL, MVT VT,
10325 ArrayRef<int> Mask, SDValue V1,
10326 SDValue V2, SelectionDAG &DAG) {
10328 // Mask for V8F64: 0/1, 8/9, 2/3, 10/11, 4/5, ..
10329 // Mask for V4F64; 0/1, 4/5, 2/3, 6/7..
10330 assert(VT.getScalarSizeInBits() == 64 && "Unexpected data type for VSHUFPD");
10331 int NumElts = VT.getVectorNumElements();
10332 bool ShufpdMask = true;
10333 bool CommutableMask = true;
10334 unsigned Immediate = 0;
10335 for (int i = 0; i < NumElts; ++i) {
10338 int Val = (i & 6) + NumElts * (i & 1);
10339 int CommutVal = (i & 0xe) + NumElts * ((i & 1)^1);
10340 if (Mask[i] < Val || Mask[i] > Val + 1)
10341 ShufpdMask = false;
10342 if (Mask[i] < CommutVal || Mask[i] > CommutVal + 1)
10343 CommutableMask = false;
10344 Immediate |= (Mask[i] % 2) << i;
10347 return DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
10348 DAG.getConstant(Immediate, DL, MVT::i8));
10349 if (CommutableMask)
10350 return DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
10351 DAG.getConstant(Immediate, DL, MVT::i8));
10355 /// \brief Handle lowering of 4-lane 64-bit floating point shuffles.
10357 /// Also ends up handling lowering of 4-lane 64-bit integer shuffles when AVX2
10358 /// isn't available.
10359 static SDValue lowerV4F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10360 const X86Subtarget *Subtarget,
10361 SelectionDAG &DAG) {
10363 assert(V1.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
10364 assert(V2.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
10365 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10366 ArrayRef<int> Mask = SVOp->getMask();
10367 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
10369 SmallVector<int, 4> WidenedMask;
10370 if (canWidenShuffleElements(Mask, WidenedMask))
10371 return lowerV2X128VectorShuffle(DL, MVT::v4f64, V1, V2, Mask, Subtarget,
10374 if (isSingleInputShuffleMask(Mask)) {
10375 // Check for being able to broadcast a single element.
10376 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f64, V1,
10377 Mask, Subtarget, DAG))
10380 // Use low duplicate instructions for masks that match their pattern.
10381 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
10382 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v4f64, V1);
10384 if (!is128BitLaneCrossingShuffleMask(MVT::v4f64, Mask)) {
10385 // Non-half-crossing single input shuffles can be lowerid with an
10386 // interleaved permutation.
10387 unsigned VPERMILPMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1) |
10388 ((Mask[2] == 3) << 2) | ((Mask[3] == 3) << 3);
10389 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f64, V1,
10390 DAG.getConstant(VPERMILPMask, DL, MVT::i8));
10393 // With AVX2 we have direct support for this permutation.
10394 if (Subtarget->hasAVX2())
10395 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4f64, V1,
10396 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
10398 // Otherwise, fall back.
10399 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v4f64, V1, V2, Mask,
10403 // Use dedicated unpack instructions for masks that match their pattern.
10405 lowerVectorShuffleWithUNPCK(DL, MVT::v4f64, Mask, V1, V2, DAG))
10408 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f64, V1, V2, Mask,
10412 // Check if the blend happens to exactly fit that of SHUFPD.
10414 lowerVectorShuffleWithSHUFPD(DL, MVT::v4f64, Mask, V1, V2, DAG))
10417 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10418 // shuffle. However, if we have AVX2 and either inputs are already in place,
10419 // we will be able to shuffle even across lanes the other input in a single
10420 // instruction so skip this pattern.
10421 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
10422 isShuffleMaskInputInPlace(1, Mask))))
10423 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10424 DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG))
10427 // If we have AVX2 then we always want to lower with a blend because an v4 we
10428 // can fully permute the elements.
10429 if (Subtarget->hasAVX2())
10430 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4f64, V1, V2,
10433 // Otherwise fall back on generic lowering.
10434 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v4f64, V1, V2, Mask, DAG);
10437 /// \brief Handle lowering of 4-lane 64-bit integer shuffles.
10439 /// This routine is only called when we have AVX2 and thus a reasonable
10440 /// instruction set for v4i64 shuffling..
10441 static SDValue lowerV4I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10442 const X86Subtarget *Subtarget,
10443 SelectionDAG &DAG) {
10445 assert(V1.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
10446 assert(V2.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
10447 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10448 ArrayRef<int> Mask = SVOp->getMask();
10449 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
10450 assert(Subtarget->hasAVX2() && "We can only lower v4i64 with AVX2!");
10452 SmallVector<int, 4> WidenedMask;
10453 if (canWidenShuffleElements(Mask, WidenedMask))
10454 return lowerV2X128VectorShuffle(DL, MVT::v4i64, V1, V2, Mask, Subtarget,
10457 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i64, V1, V2, Mask,
10461 // Check for being able to broadcast a single element.
10462 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i64, V1,
10463 Mask, Subtarget, DAG))
10466 // When the shuffle is mirrored between the 128-bit lanes of the unit, we can
10467 // use lower latency instructions that will operate on both 128-bit lanes.
10468 SmallVector<int, 2> RepeatedMask;
10469 if (is128BitLaneRepeatedShuffleMask(MVT::v4i64, Mask, RepeatedMask)) {
10470 if (isSingleInputShuffleMask(Mask)) {
10471 int PSHUFDMask[] = {-1, -1, -1, -1};
10472 for (int i = 0; i < 2; ++i)
10473 if (RepeatedMask[i] >= 0) {
10474 PSHUFDMask[2 * i] = 2 * RepeatedMask[i];
10475 PSHUFDMask[2 * i + 1] = 2 * RepeatedMask[i] + 1;
10477 return DAG.getBitcast(
10479 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32,
10480 DAG.getBitcast(MVT::v8i32, V1),
10481 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
10485 // AVX2 provides a direct instruction for permuting a single input across
10487 if (isSingleInputShuffleMask(Mask))
10488 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4i64, V1,
10489 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
10491 // Try to use shift instructions.
10492 if (SDValue Shift =
10493 lowerVectorShuffleAsShift(DL, MVT::v4i64, V1, V2, Mask, DAG))
10496 // Use dedicated unpack instructions for masks that match their pattern.
10498 lowerVectorShuffleWithUNPCK(DL, MVT::v4i64, Mask, V1, V2, DAG))
10501 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10502 // shuffle. However, if we have AVX2 and either inputs are already in place,
10503 // we will be able to shuffle even across lanes the other input in a single
10504 // instruction so skip this pattern.
10505 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
10506 isShuffleMaskInputInPlace(1, Mask))))
10507 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10508 DL, MVT::v4i64, V1, V2, Mask, Subtarget, DAG))
10511 // Otherwise fall back on generic blend lowering.
10512 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i64, V1, V2,
10516 /// \brief Handle lowering of 8-lane 32-bit floating point shuffles.
10518 /// Also ends up handling lowering of 8-lane 32-bit integer shuffles when AVX2
10519 /// isn't available.
10520 static SDValue lowerV8F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10521 const X86Subtarget *Subtarget,
10522 SelectionDAG &DAG) {
10524 assert(V1.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
10525 assert(V2.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
10526 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10527 ArrayRef<int> Mask = SVOp->getMask();
10528 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10530 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8f32, V1, V2, Mask,
10534 // Check for being able to broadcast a single element.
10535 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8f32, V1,
10536 Mask, Subtarget, DAG))
10539 // If the shuffle mask is repeated in each 128-bit lane, we have many more
10540 // options to efficiently lower the shuffle.
10541 SmallVector<int, 4> RepeatedMask;
10542 if (is128BitLaneRepeatedShuffleMask(MVT::v8f32, Mask, RepeatedMask)) {
10543 assert(RepeatedMask.size() == 4 &&
10544 "Repeated masks must be half the mask width!");
10546 // Use even/odd duplicate instructions for masks that match their pattern.
10547 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2, 4, 4, 6, 6}))
10548 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v8f32, V1);
10549 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3, 5, 5, 7, 7}))
10550 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v8f32, V1);
10552 if (isSingleInputShuffleMask(Mask))
10553 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f32, V1,
10554 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
10556 // Use dedicated unpack instructions for masks that match their pattern.
10558 lowerVectorShuffleWithUNPCK(DL, MVT::v8f32, Mask, V1, V2, DAG))
10561 // Otherwise, fall back to a SHUFPS sequence. Here it is important that we
10562 // have already handled any direct blends. We also need to squash the
10563 // repeated mask into a simulated v4f32 mask.
10564 for (int i = 0; i < 4; ++i)
10565 if (RepeatedMask[i] >= 8)
10566 RepeatedMask[i] -= 4;
10567 return lowerVectorShuffleWithSHUFPS(DL, MVT::v8f32, RepeatedMask, V1, V2, DAG);
10570 // If we have a single input shuffle with different shuffle patterns in the
10571 // two 128-bit lanes use the variable mask to VPERMILPS.
10572 if (isSingleInputShuffleMask(Mask)) {
10573 SDValue VPermMask[8];
10574 for (int i = 0; i < 8; ++i)
10575 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10576 : DAG.getConstant(Mask[i], DL, MVT::i32);
10577 if (!is128BitLaneCrossingShuffleMask(MVT::v8f32, Mask))
10578 return DAG.getNode(
10579 X86ISD::VPERMILPV, DL, MVT::v8f32, V1,
10580 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask));
10582 if (Subtarget->hasAVX2())
10583 return DAG.getNode(
10584 X86ISD::VPERMV, DL, MVT::v8f32,
10585 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1);
10587 // Otherwise, fall back.
10588 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v8f32, V1, V2, Mask,
10592 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10594 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10595 DL, MVT::v8f32, V1, V2, Mask, Subtarget, DAG))
10598 // If we have AVX2 then we always want to lower with a blend because at v8 we
10599 // can fully permute the elements.
10600 if (Subtarget->hasAVX2())
10601 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8f32, V1, V2,
10604 // Otherwise fall back on generic lowering.
10605 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v8f32, V1, V2, Mask, DAG);
10608 /// \brief Handle lowering of 8-lane 32-bit integer shuffles.
10610 /// This routine is only called when we have AVX2 and thus a reasonable
10611 /// instruction set for v8i32 shuffling..
10612 static SDValue lowerV8I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10613 const X86Subtarget *Subtarget,
10614 SelectionDAG &DAG) {
10616 assert(V1.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10617 assert(V2.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10618 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10619 ArrayRef<int> Mask = SVOp->getMask();
10620 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10621 assert(Subtarget->hasAVX2() && "We can only lower v8i32 with AVX2!");
10623 // Whenever we can lower this as a zext, that instruction is strictly faster
10624 // than any alternative. It also allows us to fold memory operands into the
10625 // shuffle in many cases.
10626 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v8i32, V1, V2,
10627 Mask, Subtarget, DAG))
10630 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i32, V1, V2, Mask,
10634 // Check for being able to broadcast a single element.
10635 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i32, V1,
10636 Mask, Subtarget, DAG))
10639 // If the shuffle mask is repeated in each 128-bit lane we can use more
10640 // efficient instructions that mirror the shuffles across the two 128-bit
10642 SmallVector<int, 4> RepeatedMask;
10643 if (is128BitLaneRepeatedShuffleMask(MVT::v8i32, Mask, RepeatedMask)) {
10644 assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!");
10645 if (isSingleInputShuffleMask(Mask))
10646 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, V1,
10647 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
10649 // Use dedicated unpack instructions for masks that match their pattern.
10651 lowerVectorShuffleWithUNPCK(DL, MVT::v8i32, Mask, V1, V2, DAG))
10655 // Try to use shift instructions.
10656 if (SDValue Shift =
10657 lowerVectorShuffleAsShift(DL, MVT::v8i32, V1, V2, Mask, DAG))
10660 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10661 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10664 // If the shuffle patterns aren't repeated but it is a single input, directly
10665 // generate a cross-lane VPERMD instruction.
10666 if (isSingleInputShuffleMask(Mask)) {
10667 SDValue VPermMask[8];
10668 for (int i = 0; i < 8; ++i)
10669 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10670 : DAG.getConstant(Mask[i], DL, MVT::i32);
10671 return DAG.getNode(
10672 X86ISD::VPERMV, DL, MVT::v8i32,
10673 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1);
10676 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10678 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10679 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10682 // Otherwise fall back on generic blend lowering.
10683 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i32, V1, V2,
10687 /// \brief Handle lowering of 16-lane 16-bit integer shuffles.
10689 /// This routine is only called when we have AVX2 and thus a reasonable
10690 /// instruction set for v16i16 shuffling..
10691 static SDValue lowerV16I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10692 const X86Subtarget *Subtarget,
10693 SelectionDAG &DAG) {
10695 assert(V1.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10696 assert(V2.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10697 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10698 ArrayRef<int> Mask = SVOp->getMask();
10699 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10700 assert(Subtarget->hasAVX2() && "We can only lower v16i16 with AVX2!");
10702 // Whenever we can lower this as a zext, that instruction is strictly faster
10703 // than any alternative. It also allows us to fold memory operands into the
10704 // shuffle in many cases.
10705 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v16i16, V1, V2,
10706 Mask, Subtarget, DAG))
10709 // Check for being able to broadcast a single element.
10710 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i16, V1,
10711 Mask, Subtarget, DAG))
10714 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i16, V1, V2, Mask,
10718 // Use dedicated unpack instructions for masks that match their pattern.
10720 lowerVectorShuffleWithUNPCK(DL, MVT::v16i16, Mask, V1, V2, DAG))
10723 // Try to use shift instructions.
10724 if (SDValue Shift =
10725 lowerVectorShuffleAsShift(DL, MVT::v16i16, V1, V2, Mask, DAG))
10728 // Try to use byte rotation instructions.
10729 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10730 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10733 if (isSingleInputShuffleMask(Mask)) {
10734 // There are no generalized cross-lane shuffle operations available on i16
10736 if (is128BitLaneCrossingShuffleMask(MVT::v16i16, Mask))
10737 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v16i16, V1, V2,
10740 SmallVector<int, 8> RepeatedMask;
10741 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
10742 // As this is a single-input shuffle, the repeated mask should be
10743 // a strictly valid v8i16 mask that we can pass through to the v8i16
10744 // lowering to handle even the v16 case.
10745 return lowerV8I16GeneralSingleInputVectorShuffle(
10746 DL, MVT::v16i16, V1, RepeatedMask, Subtarget, DAG);
10749 SDValue PSHUFBMask[32];
10750 for (int i = 0; i < 16; ++i) {
10751 if (Mask[i] == -1) {
10752 PSHUFBMask[2 * i] = PSHUFBMask[2 * i + 1] = DAG.getUNDEF(MVT::i8);
10756 int M = i < 8 ? Mask[i] : Mask[i] - 8;
10757 assert(M >= 0 && M < 8 && "Invalid single-input mask!");
10758 PSHUFBMask[2 * i] = DAG.getConstant(2 * M, DL, MVT::i8);
10759 PSHUFBMask[2 * i + 1] = DAG.getConstant(2 * M + 1, DL, MVT::i8);
10761 return DAG.getBitcast(MVT::v16i16,
10762 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8,
10763 DAG.getBitcast(MVT::v32i8, V1),
10764 DAG.getNode(ISD::BUILD_VECTOR, DL,
10765 MVT::v32i8, PSHUFBMask)));
10768 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10770 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10771 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10774 // Otherwise fall back on generic lowering.
10775 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v16i16, V1, V2, Mask, DAG);
10778 /// \brief Handle lowering of 32-lane 8-bit integer shuffles.
10780 /// This routine is only called when we have AVX2 and thus a reasonable
10781 /// instruction set for v32i8 shuffling..
10782 static SDValue lowerV32I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10783 const X86Subtarget *Subtarget,
10784 SelectionDAG &DAG) {
10786 assert(V1.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10787 assert(V2.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10788 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10789 ArrayRef<int> Mask = SVOp->getMask();
10790 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10791 assert(Subtarget->hasAVX2() && "We can only lower v32i8 with AVX2!");
10793 // Whenever we can lower this as a zext, that instruction is strictly faster
10794 // than any alternative. It also allows us to fold memory operands into the
10795 // shuffle in many cases.
10796 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v32i8, V1, V2,
10797 Mask, Subtarget, DAG))
10800 // Check for being able to broadcast a single element.
10801 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v32i8, V1,
10802 Mask, Subtarget, DAG))
10805 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v32i8, V1, V2, Mask,
10809 // Use dedicated unpack instructions for masks that match their pattern.
10811 lowerVectorShuffleWithUNPCK(DL, MVT::v32i8, Mask, V1, V2, DAG))
10814 // Try to use shift instructions.
10815 if (SDValue Shift =
10816 lowerVectorShuffleAsShift(DL, MVT::v32i8, V1, V2, Mask, DAG))
10819 // Try to use byte rotation instructions.
10820 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10821 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10824 if (isSingleInputShuffleMask(Mask)) {
10825 // There are no generalized cross-lane shuffle operations available on i8
10827 if (is128BitLaneCrossingShuffleMask(MVT::v32i8, Mask))
10828 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v32i8, V1, V2,
10831 SDValue PSHUFBMask[32];
10832 for (int i = 0; i < 32; ++i)
10835 ? DAG.getUNDEF(MVT::i8)
10836 : DAG.getConstant(Mask[i] < 16 ? Mask[i] : Mask[i] - 16, DL,
10839 return DAG.getNode(
10840 X86ISD::PSHUFB, DL, MVT::v32i8, V1,
10841 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask));
10844 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10846 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10847 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10850 // Otherwise fall back on generic lowering.
10851 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v32i8, V1, V2, Mask, DAG);
10854 /// \brief High-level routine to lower various 256-bit x86 vector shuffles.
10856 /// This routine either breaks down the specific type of a 256-bit x86 vector
10857 /// shuffle or splits it into two 128-bit shuffles and fuses the results back
10858 /// together based on the available instructions.
10859 static SDValue lower256BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10860 MVT VT, const X86Subtarget *Subtarget,
10861 SelectionDAG &DAG) {
10863 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10864 ArrayRef<int> Mask = SVOp->getMask();
10866 // If we have a single input to the zero element, insert that into V1 if we
10867 // can do so cheaply.
10868 int NumElts = VT.getVectorNumElements();
10869 int NumV2Elements = std::count_if(Mask.begin(), Mask.end(), [NumElts](int M) {
10870 return M >= NumElts;
10873 if (NumV2Elements == 1 && Mask[0] >= NumElts)
10874 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
10875 DL, VT, V1, V2, Mask, Subtarget, DAG))
10878 // There is a really nice hard cut-over between AVX1 and AVX2 that means we
10879 // can check for those subtargets here and avoid much of the subtarget
10880 // querying in the per-vector-type lowering routines. With AVX1 we have
10881 // essentially *zero* ability to manipulate a 256-bit vector with integer
10882 // types. Since we'll use floating point types there eventually, just
10883 // immediately cast everything to a float and operate entirely in that domain.
10884 if (VT.isInteger() && !Subtarget->hasAVX2()) {
10885 int ElementBits = VT.getScalarSizeInBits();
10886 if (ElementBits < 32)
10887 // No floating point type available, decompose into 128-bit vectors.
10888 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10890 MVT FpVT = MVT::getVectorVT(MVT::getFloatingPointVT(ElementBits),
10891 VT.getVectorNumElements());
10892 V1 = DAG.getBitcast(FpVT, V1);
10893 V2 = DAG.getBitcast(FpVT, V2);
10894 return DAG.getBitcast(VT, DAG.getVectorShuffle(FpVT, DL, V1, V2, Mask));
10897 switch (VT.SimpleTy) {
10899 return lowerV4F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10901 return lowerV4I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10903 return lowerV8F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10905 return lowerV8I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10907 return lowerV16I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10909 return lowerV32I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10912 llvm_unreachable("Not a valid 256-bit x86 vector type!");
10916 /// \brief Try to lower a vector shuffle as a 128-bit shuffles.
10917 static SDValue lowerV4X128VectorShuffle(SDLoc DL, MVT VT,
10918 ArrayRef<int> Mask,
10919 SDValue V1, SDValue V2,
10920 SelectionDAG &DAG) {
10921 assert(VT.getScalarSizeInBits() == 64 &&
10922 "Unexpected element type size for 128bit shuffle.");
10924 // To handle 256 bit vector requires VLX and most probably
10925 // function lowerV2X128VectorShuffle() is better solution.
10926 assert(VT.is512BitVector() && "Unexpected vector size for 128bit shuffle.");
10928 SmallVector<int, 4> WidenedMask;
10929 if (!canWidenShuffleElements(Mask, WidenedMask))
10932 // Form a 128-bit permutation.
10933 // Convert the 64-bit shuffle mask selection values into 128-bit selection
10934 // bits defined by a vshuf64x2 instruction's immediate control byte.
10935 unsigned PermMask = 0, Imm = 0;
10936 unsigned ControlBitsNum = WidenedMask.size() / 2;
10938 for (int i = 0, Size = WidenedMask.size(); i < Size; ++i) {
10939 if (WidenedMask[i] == SM_SentinelZero)
10942 // Use first element in place of undef mask.
10943 Imm = (WidenedMask[i] == SM_SentinelUndef) ? 0 : WidenedMask[i];
10944 PermMask |= (Imm % WidenedMask.size()) << (i * ControlBitsNum);
10947 return DAG.getNode(X86ISD::SHUF128, DL, VT, V1, V2,
10948 DAG.getConstant(PermMask, DL, MVT::i8));
10951 static SDValue lowerVectorShuffleWithPERMV(SDLoc DL, MVT VT,
10952 ArrayRef<int> Mask, SDValue V1,
10953 SDValue V2, SelectionDAG &DAG) {
10955 assert(VT.getScalarSizeInBits() >= 16 && "Unexpected data type for PERMV");
10957 MVT MaskEltVT = MVT::getIntegerVT(VT.getScalarSizeInBits());
10958 MVT MaskVecVT = MVT::getVectorVT(MaskEltVT, VT.getVectorNumElements());
10960 SDValue MaskNode = getConstVector(Mask, MaskVecVT, DAG, DL, true);
10961 if (isSingleInputShuffleMask(Mask))
10962 return DAG.getNode(X86ISD::VPERMV, DL, VT, MaskNode, V1);
10964 return DAG.getNode(X86ISD::VPERMV3, DL, VT, V1, MaskNode, V2);
10967 /// \brief Handle lowering of 8-lane 64-bit floating point shuffles.
10968 static SDValue lowerV8F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10969 const X86Subtarget *Subtarget,
10970 SelectionDAG &DAG) {
10972 assert(V1.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10973 assert(V2.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10974 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10975 ArrayRef<int> Mask = SVOp->getMask();
10976 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10978 if (SDValue Shuf128 =
10979 lowerV4X128VectorShuffle(DL, MVT::v8f64, Mask, V1, V2, DAG))
10982 if (SDValue Unpck =
10983 lowerVectorShuffleWithUNPCK(DL, MVT::v8f64, Mask, V1, V2, DAG))
10986 return lowerVectorShuffleWithPERMV(DL, MVT::v8f64, Mask, V1, V2, DAG);
10989 /// \brief Handle lowering of 16-lane 32-bit floating point shuffles.
10990 static SDValue lowerV16F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10991 const X86Subtarget *Subtarget,
10992 SelectionDAG &DAG) {
10994 assert(V1.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10995 assert(V2.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10996 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10997 ArrayRef<int> Mask = SVOp->getMask();
10998 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
11000 if (SDValue Unpck =
11001 lowerVectorShuffleWithUNPCK(DL, MVT::v16f32, Mask, V1, V2, DAG))
11004 return lowerVectorShuffleWithPERMV(DL, MVT::v16f32, Mask, V1, V2, DAG);
11007 /// \brief Handle lowering of 8-lane 64-bit integer shuffles.
11008 static SDValue lowerV8I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11009 const X86Subtarget *Subtarget,
11010 SelectionDAG &DAG) {
11012 assert(V1.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
11013 assert(V2.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
11014 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11015 ArrayRef<int> Mask = SVOp->getMask();
11016 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
11018 if (SDValue Shuf128 =
11019 lowerV4X128VectorShuffle(DL, MVT::v8i64, Mask, V1, V2, DAG))
11022 if (SDValue Unpck =
11023 lowerVectorShuffleWithUNPCK(DL, MVT::v8i64, Mask, V1, V2, DAG))
11026 return lowerVectorShuffleWithPERMV(DL, MVT::v8i64, Mask, V1, V2, DAG);
11029 /// \brief Handle lowering of 16-lane 32-bit integer shuffles.
11030 static SDValue lowerV16I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11031 const X86Subtarget *Subtarget,
11032 SelectionDAG &DAG) {
11034 assert(V1.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
11035 assert(V2.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
11036 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11037 ArrayRef<int> Mask = SVOp->getMask();
11038 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
11040 if (SDValue Unpck =
11041 lowerVectorShuffleWithUNPCK(DL, MVT::v16i32, Mask, V1, V2, DAG))
11044 return lowerVectorShuffleWithPERMV(DL, MVT::v16i32, Mask, V1, V2, DAG);
11047 /// \brief Handle lowering of 32-lane 16-bit integer shuffles.
11048 static SDValue lowerV32I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11049 const X86Subtarget *Subtarget,
11050 SelectionDAG &DAG) {
11052 assert(V1.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
11053 assert(V2.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
11054 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11055 ArrayRef<int> Mask = SVOp->getMask();
11056 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
11057 assert(Subtarget->hasBWI() && "We can only lower v32i16 with AVX-512-BWI!");
11059 return lowerVectorShuffleWithPERMV(DL, MVT::v32i16, Mask, V1, V2, DAG);
11062 /// \brief Handle lowering of 64-lane 8-bit integer shuffles.
11063 static SDValue lowerV64I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11064 const X86Subtarget *Subtarget,
11065 SelectionDAG &DAG) {
11067 assert(V1.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
11068 assert(V2.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
11069 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11070 ArrayRef<int> Mask = SVOp->getMask();
11071 assert(Mask.size() == 64 && "Unexpected mask size for v64 shuffle!");
11072 assert(Subtarget->hasBWI() && "We can only lower v64i8 with AVX-512-BWI!");
11074 // FIXME: Implement direct support for this type!
11075 return splitAndLowerVectorShuffle(DL, MVT::v64i8, V1, V2, Mask, DAG);
11078 /// \brief High-level routine to lower various 512-bit x86 vector shuffles.
11080 /// This routine either breaks down the specific type of a 512-bit x86 vector
11081 /// shuffle or splits it into two 256-bit shuffles and fuses the results back
11082 /// together based on the available instructions.
11083 static SDValue lower512BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11084 MVT VT, const X86Subtarget *Subtarget,
11085 SelectionDAG &DAG) {
11087 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11088 ArrayRef<int> Mask = SVOp->getMask();
11089 assert(Subtarget->hasAVX512() &&
11090 "Cannot lower 512-bit vectors w/ basic ISA!");
11092 // Check for being able to broadcast a single element.
11093 if (SDValue Broadcast =
11094 lowerVectorShuffleAsBroadcast(DL, VT, V1, Mask, Subtarget, DAG))
11097 // Dispatch to each element type for lowering. If we don't have supprot for
11098 // specific element type shuffles at 512 bits, immediately split them and
11099 // lower them. Each lowering routine of a given type is allowed to assume that
11100 // the requisite ISA extensions for that element type are available.
11101 switch (VT.SimpleTy) {
11103 return lowerV8F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
11105 return lowerV16F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
11107 return lowerV8I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
11109 return lowerV16I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
11111 if (Subtarget->hasBWI())
11112 return lowerV32I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
11115 if (Subtarget->hasBWI())
11116 return lowerV64I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
11120 llvm_unreachable("Not a valid 512-bit x86 vector type!");
11123 // Otherwise fall back on splitting.
11124 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
11127 // Lower vXi1 vector shuffles.
11128 // There is no a dedicated instruction on AVX-512 that shuffles the masks.
11129 // The only way to shuffle bits is to sign-extend the mask vector to SIMD
11130 // vector, shuffle and then truncate it back.
11131 static SDValue lower1BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11132 MVT VT, const X86Subtarget *Subtarget,
11133 SelectionDAG &DAG) {
11135 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11136 ArrayRef<int> Mask = SVOp->getMask();
11137 assert(Subtarget->hasAVX512() &&
11138 "Cannot lower 512-bit vectors w/o basic ISA!");
11140 switch (VT.SimpleTy) {
11142 llvm_unreachable("Expected a vector of i1 elements");
11144 ExtVT = MVT::v2i64;
11147 ExtVT = MVT::v4i32;
11150 ExtVT = MVT::v8i64; // Take 512-bit type, more shuffles on KNL
11153 ExtVT = MVT::v16i32;
11156 ExtVT = MVT::v32i16;
11159 ExtVT = MVT::v64i8;
11163 if (ISD::isBuildVectorAllZeros(V1.getNode()))
11164 V1 = getZeroVector(ExtVT, Subtarget, DAG, DL);
11165 else if (ISD::isBuildVectorAllOnes(V1.getNode()))
11166 V1 = getOnesVector(ExtVT, Subtarget, DAG, DL);
11168 V1 = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, V1);
11171 V2 = DAG.getUNDEF(ExtVT);
11172 else if (ISD::isBuildVectorAllZeros(V2.getNode()))
11173 V2 = getZeroVector(ExtVT, Subtarget, DAG, DL);
11174 else if (ISD::isBuildVectorAllOnes(V2.getNode()))
11175 V2 = getOnesVector(ExtVT, Subtarget, DAG, DL);
11177 V2 = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, V2);
11178 return DAG.getNode(ISD::TRUNCATE, DL, VT,
11179 DAG.getVectorShuffle(ExtVT, DL, V1, V2, Mask));
11181 /// \brief Top-level lowering for x86 vector shuffles.
11183 /// This handles decomposition, canonicalization, and lowering of all x86
11184 /// vector shuffles. Most of the specific lowering strategies are encapsulated
11185 /// above in helper routines. The canonicalization attempts to widen shuffles
11186 /// to involve fewer lanes of wider elements, consolidate symmetric patterns
11187 /// s.t. only one of the two inputs needs to be tested, etc.
11188 static SDValue lowerVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
11189 SelectionDAG &DAG) {
11190 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11191 ArrayRef<int> Mask = SVOp->getMask();
11192 SDValue V1 = Op.getOperand(0);
11193 SDValue V2 = Op.getOperand(1);
11194 MVT VT = Op.getSimpleValueType();
11195 int NumElements = VT.getVectorNumElements();
11197 bool Is1BitVector = (VT.getVectorElementType() == MVT::i1);
11199 assert((VT.getSizeInBits() != 64 || Is1BitVector) &&
11200 "Can't lower MMX shuffles");
11202 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
11203 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
11204 if (V1IsUndef && V2IsUndef)
11205 return DAG.getUNDEF(VT);
11207 // When we create a shuffle node we put the UNDEF node to second operand,
11208 // but in some cases the first operand may be transformed to UNDEF.
11209 // In this case we should just commute the node.
11211 return DAG.getCommutedVectorShuffle(*SVOp);
11213 // Check for non-undef masks pointing at an undef vector and make the masks
11214 // undef as well. This makes it easier to match the shuffle based solely on
11218 if (M >= NumElements) {
11219 SmallVector<int, 8> NewMask(Mask.begin(), Mask.end());
11220 for (int &M : NewMask)
11221 if (M >= NumElements)
11223 return DAG.getVectorShuffle(VT, dl, V1, V2, NewMask);
11226 // We actually see shuffles that are entirely re-arrangements of a set of
11227 // zero inputs. This mostly happens while decomposing complex shuffles into
11228 // simple ones. Directly lower these as a buildvector of zeros.
11229 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
11230 if (Zeroable.all())
11231 return getZeroVector(VT, Subtarget, DAG, dl);
11233 // Try to collapse shuffles into using a vector type with fewer elements but
11234 // wider element types. We cap this to not form integers or floating point
11235 // elements wider than 64 bits, but it might be interesting to form i128
11236 // integers to handle flipping the low and high halves of AVX 256-bit vectors.
11237 SmallVector<int, 16> WidenedMask;
11238 if (VT.getScalarSizeInBits() < 64 && !Is1BitVector &&
11239 canWidenShuffleElements(Mask, WidenedMask)) {
11240 MVT NewEltVT = VT.isFloatingPoint()
11241 ? MVT::getFloatingPointVT(VT.getScalarSizeInBits() * 2)
11242 : MVT::getIntegerVT(VT.getScalarSizeInBits() * 2);
11243 MVT NewVT = MVT::getVectorVT(NewEltVT, VT.getVectorNumElements() / 2);
11244 // Make sure that the new vector type is legal. For example, v2f64 isn't
11246 if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) {
11247 V1 = DAG.getBitcast(NewVT, V1);
11248 V2 = DAG.getBitcast(NewVT, V2);
11249 return DAG.getBitcast(
11250 VT, DAG.getVectorShuffle(NewVT, dl, V1, V2, WidenedMask));
11254 int NumV1Elements = 0, NumUndefElements = 0, NumV2Elements = 0;
11255 for (int M : SVOp->getMask())
11257 ++NumUndefElements;
11258 else if (M < NumElements)
11263 // Commute the shuffle as needed such that more elements come from V1 than
11264 // V2. This allows us to match the shuffle pattern strictly on how many
11265 // elements come from V1 without handling the symmetric cases.
11266 if (NumV2Elements > NumV1Elements)
11267 return DAG.getCommutedVectorShuffle(*SVOp);
11269 // When the number of V1 and V2 elements are the same, try to minimize the
11270 // number of uses of V2 in the low half of the vector. When that is tied,
11271 // ensure that the sum of indices for V1 is equal to or lower than the sum
11272 // indices for V2. When those are equal, try to ensure that the number of odd
11273 // indices for V1 is lower than the number of odd indices for V2.
11274 if (NumV1Elements == NumV2Elements) {
11275 int LowV1Elements = 0, LowV2Elements = 0;
11276 for (int M : SVOp->getMask().slice(0, NumElements / 2))
11277 if (M >= NumElements)
11281 if (LowV2Elements > LowV1Elements) {
11282 return DAG.getCommutedVectorShuffle(*SVOp);
11283 } else if (LowV2Elements == LowV1Elements) {
11284 int SumV1Indices = 0, SumV2Indices = 0;
11285 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
11286 if (SVOp->getMask()[i] >= NumElements)
11288 else if (SVOp->getMask()[i] >= 0)
11290 if (SumV2Indices < SumV1Indices) {
11291 return DAG.getCommutedVectorShuffle(*SVOp);
11292 } else if (SumV2Indices == SumV1Indices) {
11293 int NumV1OddIndices = 0, NumV2OddIndices = 0;
11294 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
11295 if (SVOp->getMask()[i] >= NumElements)
11296 NumV2OddIndices += i % 2;
11297 else if (SVOp->getMask()[i] >= 0)
11298 NumV1OddIndices += i % 2;
11299 if (NumV2OddIndices < NumV1OddIndices)
11300 return DAG.getCommutedVectorShuffle(*SVOp);
11305 // For each vector width, delegate to a specialized lowering routine.
11306 if (VT.is128BitVector())
11307 return lower128BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11309 if (VT.is256BitVector())
11310 return lower256BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11312 if (VT.is512BitVector())
11313 return lower512BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11316 return lower1BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11317 llvm_unreachable("Unimplemented!");
11320 // This function assumes its argument is a BUILD_VECTOR of constants or
11321 // undef SDNodes. i.e: ISD::isBuildVectorOfConstantSDNodes(BuildVector) is
11323 static bool BUILD_VECTORtoBlendMask(BuildVectorSDNode *BuildVector,
11324 unsigned &MaskValue) {
11326 unsigned NumElems = BuildVector->getNumOperands();
11328 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
11329 // We don't handle the >2 lanes case right now.
11330 unsigned NumLanes = (NumElems - 1) / 8 + 1;
11334 unsigned NumElemsInLane = NumElems / NumLanes;
11336 // Blend for v16i16 should be symmetric for the both lanes.
11337 for (unsigned i = 0; i < NumElemsInLane; ++i) {
11338 SDValue EltCond = BuildVector->getOperand(i);
11339 SDValue SndLaneEltCond =
11340 (NumLanes == 2) ? BuildVector->getOperand(i + NumElemsInLane) : EltCond;
11342 int Lane1Cond = -1, Lane2Cond = -1;
11343 if (isa<ConstantSDNode>(EltCond))
11344 Lane1Cond = !isNullConstant(EltCond);
11345 if (isa<ConstantSDNode>(SndLaneEltCond))
11346 Lane2Cond = !isNullConstant(SndLaneEltCond);
11348 unsigned LaneMask = 0;
11349 if (Lane1Cond == Lane2Cond || Lane2Cond < 0)
11350 // Lane1Cond != 0, means we want the first argument.
11351 // Lane1Cond == 0, means we want the second argument.
11352 // The encoding of this argument is 0 for the first argument, 1
11353 // for the second. Therefore, invert the condition.
11354 LaneMask = !Lane1Cond << i;
11355 else if (Lane1Cond < 0)
11356 LaneMask = !Lane2Cond << i;
11360 MaskValue |= LaneMask;
11362 MaskValue |= LaneMask << NumElemsInLane;
11367 /// \brief Try to lower a VSELECT instruction to a vector shuffle.
11368 static SDValue lowerVSELECTtoVectorShuffle(SDValue Op,
11369 const X86Subtarget *Subtarget,
11370 SelectionDAG &DAG) {
11371 SDValue Cond = Op.getOperand(0);
11372 SDValue LHS = Op.getOperand(1);
11373 SDValue RHS = Op.getOperand(2);
11375 MVT VT = Op.getSimpleValueType();
11377 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
11379 auto *CondBV = cast<BuildVectorSDNode>(Cond);
11381 // Only non-legal VSELECTs reach this lowering, convert those into generic
11382 // shuffles and re-use the shuffle lowering path for blends.
11383 SmallVector<int, 32> Mask;
11384 for (int i = 0, Size = VT.getVectorNumElements(); i < Size; ++i) {
11385 SDValue CondElt = CondBV->getOperand(i);
11387 isa<ConstantSDNode>(CondElt) ? i + (isNullConstant(CondElt) ? Size : 0)
11390 return DAG.getVectorShuffle(VT, dl, LHS, RHS, Mask);
11393 SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const {
11394 // A vselect where all conditions and data are constants can be optimized into
11395 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
11396 if (ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(0).getNode()) &&
11397 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(1).getNode()) &&
11398 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(2).getNode()))
11401 // Try to lower this to a blend-style vector shuffle. This can handle all
11402 // constant condition cases.
11403 if (SDValue BlendOp = lowerVSELECTtoVectorShuffle(Op, Subtarget, DAG))
11406 // Variable blends are only legal from SSE4.1 onward.
11407 if (!Subtarget->hasSSE41())
11410 // Only some types will be legal on some subtargets. If we can emit a legal
11411 // VSELECT-matching blend, return Op, and but if we need to expand, return
11413 switch (Op.getSimpleValueType().SimpleTy) {
11415 // Most of the vector types have blends past SSE4.1.
11419 // The byte blends for AVX vectors were introduced only in AVX2.
11420 if (Subtarget->hasAVX2())
11427 // AVX-512 BWI and VLX features support VSELECT with i16 elements.
11428 if (Subtarget->hasBWI() && Subtarget->hasVLX())
11431 // FIXME: We should custom lower this by fixing the condition and using i8
11437 static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
11438 MVT VT = Op.getSimpleValueType();
11441 if (!Op.getOperand(0).getSimpleValueType().is128BitVector())
11444 if (VT.getSizeInBits() == 8) {
11445 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
11446 Op.getOperand(0), Op.getOperand(1));
11447 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
11448 DAG.getValueType(VT));
11449 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11452 if (VT.getSizeInBits() == 16) {
11453 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
11454 if (isNullConstant(Op.getOperand(1)))
11455 return DAG.getNode(
11456 ISD::TRUNCATE, dl, MVT::i16,
11457 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11458 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
11459 Op.getOperand(1)));
11460 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
11461 Op.getOperand(0), Op.getOperand(1));
11462 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
11463 DAG.getValueType(VT));
11464 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11467 if (VT == MVT::f32) {
11468 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
11469 // the result back to FR32 register. It's only worth matching if the
11470 // result has a single use which is a store or a bitcast to i32. And in
11471 // the case of a store, it's not worth it if the index is a constant 0,
11472 // because a MOVSSmr can be used instead, which is smaller and faster.
11473 if (!Op.hasOneUse())
11475 SDNode *User = *Op.getNode()->use_begin();
11476 if ((User->getOpcode() != ISD::STORE ||
11477 isNullConstant(Op.getOperand(1))) &&
11478 (User->getOpcode() != ISD::BITCAST ||
11479 User->getValueType(0) != MVT::i32))
11481 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11482 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
11484 return DAG.getBitcast(MVT::f32, Extract);
11487 if (VT == MVT::i32 || VT == MVT::i64) {
11488 // ExtractPS/pextrq works with constant index.
11489 if (isa<ConstantSDNode>(Op.getOperand(1)))
11495 /// Extract one bit from mask vector, like v16i1 or v8i1.
11496 /// AVX-512 feature.
11498 X86TargetLowering::ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const {
11499 SDValue Vec = Op.getOperand(0);
11501 MVT VecVT = Vec.getSimpleValueType();
11502 SDValue Idx = Op.getOperand(1);
11503 MVT EltVT = Op.getSimpleValueType();
11505 assert((EltVT == MVT::i1) && "Unexpected operands in ExtractBitFromMaskVector");
11506 assert((VecVT.getVectorNumElements() <= 16 || Subtarget->hasBWI()) &&
11507 "Unexpected vector type in ExtractBitFromMaskVector");
11509 // variable index can't be handled in mask registers,
11510 // extend vector to VR512
11511 if (!isa<ConstantSDNode>(Idx)) {
11512 MVT ExtVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
11513 SDValue Ext = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVT, Vec);
11514 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
11515 ExtVT.getVectorElementType(), Ext, Idx);
11516 return DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
11519 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11520 const TargetRegisterClass* rc = getRegClassFor(VecVT);
11521 if (!Subtarget->hasDQI() && (VecVT.getVectorNumElements() <= 8))
11522 rc = getRegClassFor(MVT::v16i1);
11523 unsigned MaxSift = rc->getSize()*8 - 1;
11524 Vec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, Vec,
11525 DAG.getConstant(MaxSift - IdxVal, dl, MVT::i8));
11526 Vec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, Vec,
11527 DAG.getConstant(MaxSift, dl, MVT::i8));
11528 return DAG.getNode(X86ISD::VEXTRACT, dl, MVT::i1, Vec,
11529 DAG.getIntPtrConstant(0, dl));
11533 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
11534 SelectionDAG &DAG) const {
11536 SDValue Vec = Op.getOperand(0);
11537 MVT VecVT = Vec.getSimpleValueType();
11538 SDValue Idx = Op.getOperand(1);
11540 if (Op.getSimpleValueType() == MVT::i1)
11541 return ExtractBitFromMaskVector(Op, DAG);
11543 if (!isa<ConstantSDNode>(Idx)) {
11544 if (VecVT.is512BitVector() ||
11545 (VecVT.is256BitVector() && Subtarget->hasInt256() &&
11546 VecVT.getVectorElementType().getSizeInBits() == 32)) {
11549 MVT::getIntegerVT(VecVT.getVectorElementType().getSizeInBits());
11550 MVT MaskVT = MVT::getVectorVT(MaskEltVT, VecVT.getSizeInBits() /
11551 MaskEltVT.getSizeInBits());
11553 Idx = DAG.getZExtOrTrunc(Idx, dl, MaskEltVT);
11554 auto PtrVT = getPointerTy(DAG.getDataLayout());
11555 SDValue Mask = DAG.getNode(X86ISD::VINSERT, dl, MaskVT,
11556 getZeroVector(MaskVT, Subtarget, DAG, dl), Idx,
11557 DAG.getConstant(0, dl, PtrVT));
11558 SDValue Perm = DAG.getNode(X86ISD::VPERMV, dl, VecVT, Mask, Vec);
11559 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Perm,
11560 DAG.getConstant(0, dl, PtrVT));
11565 // If this is a 256-bit vector result, first extract the 128-bit vector and
11566 // then extract the element from the 128-bit vector.
11567 if (VecVT.is256BitVector() || VecVT.is512BitVector()) {
11569 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11570 // Get the 128-bit vector.
11571 Vec = Extract128BitVector(Vec, IdxVal, DAG, dl);
11572 MVT EltVT = VecVT.getVectorElementType();
11574 unsigned ElemsPerChunk = 128 / EltVT.getSizeInBits();
11575 assert(isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2");
11577 // Find IdxVal modulo ElemsPerChunk. Since ElemsPerChunk is a power of 2
11578 // this can be done with a mask.
11579 IdxVal &= ElemsPerChunk - 1;
11580 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
11581 DAG.getConstant(IdxVal, dl, MVT::i32));
11584 assert(VecVT.is128BitVector() && "Unexpected vector length");
11586 if (Subtarget->hasSSE41())
11587 if (SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG))
11590 MVT VT = Op.getSimpleValueType();
11591 // TODO: handle v16i8.
11592 if (VT.getSizeInBits() == 16) {
11593 SDValue Vec = Op.getOperand(0);
11594 if (isNullConstant(Op.getOperand(1)))
11595 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
11596 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11597 DAG.getBitcast(MVT::v4i32, Vec),
11598 Op.getOperand(1)));
11599 // Transform it so it match pextrw which produces a 32-bit result.
11600 MVT EltVT = MVT::i32;
11601 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
11602 Op.getOperand(0), Op.getOperand(1));
11603 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
11604 DAG.getValueType(VT));
11605 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11608 if (VT.getSizeInBits() == 32) {
11609 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11613 // SHUFPS the element to the lowest double word, then movss.
11614 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
11615 MVT VVT = Op.getOperand(0).getSimpleValueType();
11616 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
11617 DAG.getUNDEF(VVT), Mask);
11618 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
11619 DAG.getIntPtrConstant(0, dl));
11622 if (VT.getSizeInBits() == 64) {
11623 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
11624 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
11625 // to match extract_elt for f64.
11626 if (isNullConstant(Op.getOperand(1)))
11629 // UNPCKHPD the element to the lowest double word, then movsd.
11630 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
11631 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
11632 int Mask[2] = { 1, -1 };
11633 MVT VVT = Op.getOperand(0).getSimpleValueType();
11634 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
11635 DAG.getUNDEF(VVT), Mask);
11636 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
11637 DAG.getIntPtrConstant(0, dl));
11643 /// Insert one bit to mask vector, like v16i1 or v8i1.
11644 /// AVX-512 feature.
11646 X86TargetLowering::InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const {
11648 SDValue Vec = Op.getOperand(0);
11649 SDValue Elt = Op.getOperand(1);
11650 SDValue Idx = Op.getOperand(2);
11651 MVT VecVT = Vec.getSimpleValueType();
11653 if (!isa<ConstantSDNode>(Idx)) {
11654 // Non constant index. Extend source and destination,
11655 // insert element and then truncate the result.
11656 MVT ExtVecVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
11657 MVT ExtEltVT = (VecVT == MVT::v8i1 ? MVT::i64 : MVT::i32);
11658 SDValue ExtOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ExtVecVT,
11659 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVecVT, Vec),
11660 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtEltVT, Elt), Idx);
11661 return DAG.getNode(ISD::TRUNCATE, dl, VecVT, ExtOp);
11664 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11665 SDValue EltInVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Elt);
11667 EltInVec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
11668 DAG.getConstant(IdxVal, dl, MVT::i8));
11669 if (Vec.getOpcode() == ISD::UNDEF)
11671 return DAG.getNode(ISD::OR, dl, VecVT, Vec, EltInVec);
11674 SDValue X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
11675 SelectionDAG &DAG) const {
11676 MVT VT = Op.getSimpleValueType();
11677 MVT EltVT = VT.getVectorElementType();
11679 if (EltVT == MVT::i1)
11680 return InsertBitToMaskVector(Op, DAG);
11683 SDValue N0 = Op.getOperand(0);
11684 SDValue N1 = Op.getOperand(1);
11685 SDValue N2 = Op.getOperand(2);
11686 if (!isa<ConstantSDNode>(N2))
11688 auto *N2C = cast<ConstantSDNode>(N2);
11689 unsigned IdxVal = N2C->getZExtValue();
11691 // If the vector is wider than 128 bits, extract the 128-bit subvector, insert
11692 // into that, and then insert the subvector back into the result.
11693 if (VT.is256BitVector() || VT.is512BitVector()) {
11694 // With a 256-bit vector, we can insert into the zero element efficiently
11695 // using a blend if we have AVX or AVX2 and the right data type.
11696 if (VT.is256BitVector() && IdxVal == 0) {
11697 // TODO: It is worthwhile to cast integer to floating point and back
11698 // and incur a domain crossing penalty if that's what we'll end up
11699 // doing anyway after extracting to a 128-bit vector.
11700 if ((Subtarget->hasAVX() && (EltVT == MVT::f64 || EltVT == MVT::f32)) ||
11701 (Subtarget->hasAVX2() && EltVT == MVT::i32)) {
11702 SDValue N1Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, N1);
11703 N2 = DAG.getIntPtrConstant(1, dl);
11704 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1Vec, N2);
11708 // Get the desired 128-bit vector chunk.
11709 SDValue V = Extract128BitVector(N0, IdxVal, DAG, dl);
11711 // Insert the element into the desired chunk.
11712 unsigned NumEltsIn128 = 128 / EltVT.getSizeInBits();
11713 assert(isPowerOf2_32(NumEltsIn128));
11714 // Since NumEltsIn128 is a power of 2 we can use mask instead of modulo.
11715 unsigned IdxIn128 = IdxVal & (NumEltsIn128 - 1);
11717 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1,
11718 DAG.getConstant(IdxIn128, dl, MVT::i32));
11720 // Insert the changed part back into the bigger vector
11721 return Insert128BitVector(N0, V, IdxVal, DAG, dl);
11723 assert(VT.is128BitVector() && "Only 128-bit vector types should be left!");
11725 if (Subtarget->hasSSE41()) {
11726 if (EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) {
11728 if (VT == MVT::v8i16) {
11729 Opc = X86ISD::PINSRW;
11731 assert(VT == MVT::v16i8);
11732 Opc = X86ISD::PINSRB;
11735 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
11737 if (N1.getValueType() != MVT::i32)
11738 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11739 if (N2.getValueType() != MVT::i32)
11740 N2 = DAG.getIntPtrConstant(IdxVal, dl);
11741 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
11744 if (EltVT == MVT::f32) {
11745 // Bits [7:6] of the constant are the source select. This will always be
11746 // zero here. The DAG Combiner may combine an extract_elt index into
11747 // these bits. For example (insert (extract, 3), 2) could be matched by
11748 // putting the '3' into bits [7:6] of X86ISD::INSERTPS.
11749 // Bits [5:4] of the constant are the destination select. This is the
11750 // value of the incoming immediate.
11751 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
11752 // combine either bitwise AND or insert of float 0.0 to set these bits.
11754 bool MinSize = DAG.getMachineFunction().getFunction()->optForMinSize();
11755 if (IdxVal == 0 && (!MinSize || !MayFoldLoad(N1))) {
11756 // If this is an insertion of 32-bits into the low 32-bits of
11757 // a vector, we prefer to generate a blend with immediate rather
11758 // than an insertps. Blends are simpler operations in hardware and so
11759 // will always have equal or better performance than insertps.
11760 // But if optimizing for size and there's a load folding opportunity,
11761 // generate insertps because blendps does not have a 32-bit memory
11763 N2 = DAG.getIntPtrConstant(1, dl);
11764 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
11765 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1, N2);
11767 N2 = DAG.getIntPtrConstant(IdxVal << 4, dl);
11768 // Create this as a scalar to vector..
11769 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
11770 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
11773 if (EltVT == MVT::i32 || EltVT == MVT::i64) {
11774 // PINSR* works with constant index.
11779 if (EltVT == MVT::i8)
11782 if (EltVT.getSizeInBits() == 16) {
11783 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
11784 // as its second argument.
11785 if (N1.getValueType() != MVT::i32)
11786 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11787 if (N2.getValueType() != MVT::i32)
11788 N2 = DAG.getIntPtrConstant(IdxVal, dl);
11789 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
11794 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
11796 MVT OpVT = Op.getSimpleValueType();
11798 // If this is a 256-bit vector result, first insert into a 128-bit
11799 // vector and then insert into the 256-bit vector.
11800 if (!OpVT.is128BitVector()) {
11801 // Insert into a 128-bit vector.
11802 unsigned SizeFactor = OpVT.getSizeInBits()/128;
11803 MVT VT128 = MVT::getVectorVT(OpVT.getVectorElementType(),
11804 OpVT.getVectorNumElements() / SizeFactor);
11806 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
11808 // Insert the 128-bit vector.
11809 return Insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl);
11812 if (OpVT == MVT::v1i64 &&
11813 Op.getOperand(0).getValueType() == MVT::i64)
11814 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
11816 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
11817 assert(OpVT.is128BitVector() && "Expected an SSE type!");
11818 return DAG.getBitcast(
11819 OpVT, DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, AnyExt));
11822 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
11823 // a simple subregister reference or explicit instructions to grab
11824 // upper bits of a vector.
11825 static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
11826 SelectionDAG &DAG) {
11828 SDValue In = Op.getOperand(0);
11829 SDValue Idx = Op.getOperand(1);
11830 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11831 MVT ResVT = Op.getSimpleValueType();
11832 MVT InVT = In.getSimpleValueType();
11834 if (Subtarget->hasFp256()) {
11835 if (ResVT.is128BitVector() &&
11836 (InVT.is256BitVector() || InVT.is512BitVector()) &&
11837 isa<ConstantSDNode>(Idx)) {
11838 return Extract128BitVector(In, IdxVal, DAG, dl);
11840 if (ResVT.is256BitVector() && InVT.is512BitVector() &&
11841 isa<ConstantSDNode>(Idx)) {
11842 return Extract256BitVector(In, IdxVal, DAG, dl);
11848 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
11849 // simple superregister reference or explicit instructions to insert
11850 // the upper bits of a vector.
11851 static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
11852 SelectionDAG &DAG) {
11853 if (!Subtarget->hasAVX())
11857 SDValue Vec = Op.getOperand(0);
11858 SDValue SubVec = Op.getOperand(1);
11859 SDValue Idx = Op.getOperand(2);
11861 if (!isa<ConstantSDNode>(Idx))
11864 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11865 MVT OpVT = Op.getSimpleValueType();
11866 MVT SubVecVT = SubVec.getSimpleValueType();
11868 // Fold two 16-byte subvector loads into one 32-byte load:
11869 // (insert_subvector (insert_subvector undef, (load addr), 0),
11870 // (load addr + 16), Elts/2)
11872 if ((IdxVal == OpVT.getVectorNumElements() / 2) &&
11873 Vec.getOpcode() == ISD::INSERT_SUBVECTOR &&
11874 OpVT.is256BitVector() && SubVecVT.is128BitVector()) {
11875 auto *Idx2 = dyn_cast<ConstantSDNode>(Vec.getOperand(2));
11876 if (Idx2 && Idx2->getZExtValue() == 0) {
11877 SDValue SubVec2 = Vec.getOperand(1);
11878 // If needed, look through a bitcast to get to the load.
11879 if (SubVec2.getNode() && SubVec2.getOpcode() == ISD::BITCAST)
11880 SubVec2 = SubVec2.getOperand(0);
11882 if (auto *FirstLd = dyn_cast<LoadSDNode>(SubVec2)) {
11884 unsigned Alignment = FirstLd->getAlignment();
11885 unsigned AS = FirstLd->getAddressSpace();
11886 const X86TargetLowering *TLI = Subtarget->getTargetLowering();
11887 if (TLI->allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(),
11888 OpVT, AS, Alignment, &Fast) && Fast) {
11889 SDValue Ops[] = { SubVec2, SubVec };
11890 if (SDValue Ld = EltsFromConsecutiveLoads(OpVT, Ops, dl, DAG, false))
11897 if ((OpVT.is256BitVector() || OpVT.is512BitVector()) &&
11898 SubVecVT.is128BitVector())
11899 return Insert128BitVector(Vec, SubVec, IdxVal, DAG, dl);
11901 if (OpVT.is512BitVector() && SubVecVT.is256BitVector())
11902 return Insert256BitVector(Vec, SubVec, IdxVal, DAG, dl);
11904 if (OpVT.getVectorElementType() == MVT::i1)
11905 return Insert1BitVector(Op, DAG);
11910 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
11911 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
11912 // one of the above mentioned nodes. It has to be wrapped because otherwise
11913 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
11914 // be used to form addressing mode. These wrapped nodes will be selected
11917 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
11918 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
11920 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11921 // global base reg.
11922 unsigned char OpFlag = 0;
11923 unsigned WrapperKind = X86ISD::Wrapper;
11924 CodeModel::Model M = DAG.getTarget().getCodeModel();
11926 if (Subtarget->isPICStyleRIPRel() &&
11927 (M == CodeModel::Small || M == CodeModel::Kernel))
11928 WrapperKind = X86ISD::WrapperRIP;
11929 else if (Subtarget->isPICStyleGOT())
11930 OpFlag = X86II::MO_GOTOFF;
11931 else if (Subtarget->isPICStyleStubPIC())
11932 OpFlag = X86II::MO_PIC_BASE_OFFSET;
11934 auto PtrVT = getPointerTy(DAG.getDataLayout());
11935 SDValue Result = DAG.getTargetConstantPool(
11936 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(), OpFlag);
11938 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11939 // With PIC, the address is actually $g + Offset.
11942 DAG.getNode(ISD::ADD, DL, PtrVT,
11943 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11949 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
11950 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
11952 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11953 // global base reg.
11954 unsigned char OpFlag = 0;
11955 unsigned WrapperKind = X86ISD::Wrapper;
11956 CodeModel::Model M = DAG.getTarget().getCodeModel();
11958 if (Subtarget->isPICStyleRIPRel() &&
11959 (M == CodeModel::Small || M == CodeModel::Kernel))
11960 WrapperKind = X86ISD::WrapperRIP;
11961 else if (Subtarget->isPICStyleGOT())
11962 OpFlag = X86II::MO_GOTOFF;
11963 else if (Subtarget->isPICStyleStubPIC())
11964 OpFlag = X86II::MO_PIC_BASE_OFFSET;
11966 auto PtrVT = getPointerTy(DAG.getDataLayout());
11967 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, OpFlag);
11969 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11971 // With PIC, the address is actually $g + Offset.
11974 DAG.getNode(ISD::ADD, DL, PtrVT,
11975 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11981 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
11982 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
11984 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11985 // global base reg.
11986 unsigned char OpFlag = 0;
11987 unsigned WrapperKind = X86ISD::Wrapper;
11988 CodeModel::Model M = DAG.getTarget().getCodeModel();
11990 if (Subtarget->isPICStyleRIPRel() &&
11991 (M == CodeModel::Small || M == CodeModel::Kernel)) {
11992 if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF())
11993 OpFlag = X86II::MO_GOTPCREL;
11994 WrapperKind = X86ISD::WrapperRIP;
11995 } else if (Subtarget->isPICStyleGOT()) {
11996 OpFlag = X86II::MO_GOT;
11997 } else if (Subtarget->isPICStyleStubPIC()) {
11998 OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE;
11999 } else if (Subtarget->isPICStyleStubNoDynamic()) {
12000 OpFlag = X86II::MO_DARWIN_NONLAZY;
12003 auto PtrVT = getPointerTy(DAG.getDataLayout());
12004 SDValue Result = DAG.getTargetExternalSymbol(Sym, PtrVT, OpFlag);
12007 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
12009 // With PIC, the address is actually $g + Offset.
12010 if (DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
12011 !Subtarget->is64Bit()) {
12013 DAG.getNode(ISD::ADD, DL, PtrVT,
12014 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
12017 // For symbols that require a load from a stub to get the address, emit the
12019 if (isGlobalStubReference(OpFlag))
12020 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
12021 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
12022 false, false, false, 0);
12028 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
12029 // Create the TargetBlockAddressAddress node.
12030 unsigned char OpFlags =
12031 Subtarget->ClassifyBlockAddressReference();
12032 CodeModel::Model M = DAG.getTarget().getCodeModel();
12033 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
12034 int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset();
12036 auto PtrVT = getPointerTy(DAG.getDataLayout());
12037 SDValue Result = DAG.getTargetBlockAddress(BA, PtrVT, Offset, OpFlags);
12039 if (Subtarget->isPICStyleRIPRel() &&
12040 (M == CodeModel::Small || M == CodeModel::Kernel))
12041 Result = DAG.getNode(X86ISD::WrapperRIP, dl, PtrVT, Result);
12043 Result = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, Result);
12045 // With PIC, the address is actually $g + Offset.
12046 if (isGlobalRelativeToPICBase(OpFlags)) {
12047 Result = DAG.getNode(ISD::ADD, dl, PtrVT,
12048 DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
12055 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
12056 int64_t Offset, SelectionDAG &DAG) const {
12057 // Create the TargetGlobalAddress node, folding in the constant
12058 // offset if it is legal.
12059 unsigned char OpFlags =
12060 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget());
12061 CodeModel::Model M = DAG.getTarget().getCodeModel();
12062 auto PtrVT = getPointerTy(DAG.getDataLayout());
12064 if (OpFlags == X86II::MO_NO_FLAG &&
12065 X86::isOffsetSuitableForCodeModel(Offset, M)) {
12066 // A direct static reference to a global.
12067 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, Offset);
12070 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, OpFlags);
12073 if (Subtarget->isPICStyleRIPRel() &&
12074 (M == CodeModel::Small || M == CodeModel::Kernel))
12075 Result = DAG.getNode(X86ISD::WrapperRIP, dl, PtrVT, Result);
12077 Result = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, Result);
12079 // With PIC, the address is actually $g + Offset.
12080 if (isGlobalRelativeToPICBase(OpFlags)) {
12081 Result = DAG.getNode(ISD::ADD, dl, PtrVT,
12082 DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
12085 // For globals that require a load from a stub to get the address, emit the
12087 if (isGlobalStubReference(OpFlags))
12088 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
12089 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
12090 false, false, false, 0);
12092 // If there was a non-zero offset that we didn't fold, create an explicit
12093 // addition for it.
12095 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result,
12096 DAG.getConstant(Offset, dl, PtrVT));
12102 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
12103 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
12104 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
12105 return LowerGlobalAddress(GV, SDLoc(Op), Offset, DAG);
12109 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
12110 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
12111 unsigned char OperandFlags, bool LocalDynamic = false) {
12112 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
12113 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
12115 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
12116 GA->getValueType(0),
12120 X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR
12124 SDValue Ops[] = { Chain, TGA, *InFlag };
12125 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
12127 SDValue Ops[] = { Chain, TGA };
12128 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
12131 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
12132 MFI->setAdjustsStack(true);
12133 MFI->setHasCalls(true);
12135 SDValue Flag = Chain.getValue(1);
12136 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
12139 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
12141 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
12144 SDLoc dl(GA); // ? function entry point might be better
12145 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
12146 DAG.getNode(X86ISD::GlobalBaseReg,
12147 SDLoc(), PtrVT), InFlag);
12148 InFlag = Chain.getValue(1);
12150 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
12153 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
12155 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
12157 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT,
12158 X86::RAX, X86II::MO_TLSGD);
12161 static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA,
12167 // Get the start address of the TLS block for this module.
12168 X86MachineFunctionInfo* MFI = DAG.getMachineFunction()
12169 .getInfo<X86MachineFunctionInfo>();
12170 MFI->incNumLocalDynamicTLSAccesses();
12174 Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, X86::RAX,
12175 X86II::MO_TLSLD, /*LocalDynamic=*/true);
12178 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
12179 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), InFlag);
12180 InFlag = Chain.getValue(1);
12181 Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX,
12182 X86II::MO_TLSLDM, /*LocalDynamic=*/true);
12185 // Note: the CleanupLocalDynamicTLSPass will remove redundant computations
12189 unsigned char OperandFlags = X86II::MO_DTPOFF;
12190 unsigned WrapperKind = X86ISD::Wrapper;
12191 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
12192 GA->getValueType(0),
12193 GA->getOffset(), OperandFlags);
12194 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
12196 // Add x@dtpoff with the base.
12197 return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base);
12200 // Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model.
12201 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
12202 const EVT PtrVT, TLSModel::Model model,
12203 bool is64Bit, bool isPIC) {
12206 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
12207 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
12208 is64Bit ? 257 : 256));
12210 SDValue ThreadPointer =
12211 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), DAG.getIntPtrConstant(0, dl),
12212 MachinePointerInfo(Ptr), false, false, false, 0);
12214 unsigned char OperandFlags = 0;
12215 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
12217 unsigned WrapperKind = X86ISD::Wrapper;
12218 if (model == TLSModel::LocalExec) {
12219 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
12220 } else if (model == TLSModel::InitialExec) {
12222 OperandFlags = X86II::MO_GOTTPOFF;
12223 WrapperKind = X86ISD::WrapperRIP;
12225 OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF;
12228 llvm_unreachable("Unexpected model");
12231 // emit "addl x@ntpoff,%eax" (local exec)
12232 // or "addl x@indntpoff,%eax" (initial exec)
12233 // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic)
12235 DAG.getTargetGlobalAddress(GA->getGlobal(), dl, GA->getValueType(0),
12236 GA->getOffset(), OperandFlags);
12237 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
12239 if (model == TLSModel::InitialExec) {
12240 if (isPIC && !is64Bit) {
12241 Offset = DAG.getNode(ISD::ADD, dl, PtrVT,
12242 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
12246 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
12247 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
12248 false, false, false, 0);
12251 // The address of the thread local variable is the add of the thread
12252 // pointer with the offset of the variable.
12253 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
12257 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
12259 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
12260 const GlobalValue *GV = GA->getGlobal();
12261 auto PtrVT = getPointerTy(DAG.getDataLayout());
12263 if (Subtarget->isTargetELF()) {
12264 if (DAG.getTarget().Options.EmulatedTLS)
12265 return LowerToTLSEmulatedModel(GA, DAG);
12266 TLSModel::Model model = DAG.getTarget().getTLSModel(GV);
12268 case TLSModel::GeneralDynamic:
12269 if (Subtarget->is64Bit())
12270 return LowerToTLSGeneralDynamicModel64(GA, DAG, PtrVT);
12271 return LowerToTLSGeneralDynamicModel32(GA, DAG, PtrVT);
12272 case TLSModel::LocalDynamic:
12273 return LowerToTLSLocalDynamicModel(GA, DAG, PtrVT,
12274 Subtarget->is64Bit());
12275 case TLSModel::InitialExec:
12276 case TLSModel::LocalExec:
12277 return LowerToTLSExecModel(GA, DAG, PtrVT, model, Subtarget->is64Bit(),
12278 DAG.getTarget().getRelocationModel() ==
12281 llvm_unreachable("Unknown TLS model.");
12284 if (Subtarget->isTargetDarwin()) {
12285 // Darwin only has one model of TLS. Lower to that.
12286 unsigned char OpFlag = 0;
12287 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
12288 X86ISD::WrapperRIP : X86ISD::Wrapper;
12290 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
12291 // global base reg.
12292 bool PIC32 = (DAG.getTarget().getRelocationModel() == Reloc::PIC_) &&
12293 !Subtarget->is64Bit();
12295 OpFlag = X86II::MO_TLVP_PIC_BASE;
12297 OpFlag = X86II::MO_TLVP;
12299 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
12300 GA->getValueType(0),
12301 GA->getOffset(), OpFlag);
12302 SDValue Offset = DAG.getNode(WrapperKind, DL, PtrVT, Result);
12304 // With PIC32, the address is actually $g + Offset.
12306 Offset = DAG.getNode(ISD::ADD, DL, PtrVT,
12307 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
12310 // Lowering the machine isd will make sure everything is in the right
12312 SDValue Chain = DAG.getEntryNode();
12313 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
12314 SDValue Args[] = { Chain, Offset };
12315 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args);
12317 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
12318 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
12319 MFI->setAdjustsStack(true);
12321 // And our return value (tls address) is in the standard call return value
12323 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
12324 return DAG.getCopyFromReg(Chain, DL, Reg, PtrVT, Chain.getValue(1));
12327 if (Subtarget->isTargetKnownWindowsMSVC() ||
12328 Subtarget->isTargetWindowsGNU()) {
12329 // Just use the implicit TLS architecture
12330 // Need to generate someting similar to:
12331 // mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage
12333 // mov ecx, dword [rel _tls_index]: Load index (from C runtime)
12334 // mov rcx, qword [rdx+rcx*8]
12335 // mov eax, .tls$:tlsvar
12336 // [rax+rcx] contains the address
12337 // Windows 64bit: gs:0x58
12338 // Windows 32bit: fs:__tls_array
12341 SDValue Chain = DAG.getEntryNode();
12343 // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or
12344 // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly
12345 // use its literal value of 0x2C.
12346 Value *Ptr = Constant::getNullValue(Subtarget->is64Bit()
12347 ? Type::getInt8PtrTy(*DAG.getContext(),
12349 : Type::getInt32PtrTy(*DAG.getContext(),
12352 SDValue TlsArray = Subtarget->is64Bit()
12353 ? DAG.getIntPtrConstant(0x58, dl)
12354 : (Subtarget->isTargetWindowsGNU()
12355 ? DAG.getIntPtrConstant(0x2C, dl)
12356 : DAG.getExternalSymbol("_tls_array", PtrVT));
12358 SDValue ThreadPointer =
12359 DAG.getLoad(PtrVT, dl, Chain, TlsArray, MachinePointerInfo(Ptr), false,
12363 if (GV->getThreadLocalMode() == GlobalVariable::LocalExecTLSModel) {
12364 res = ThreadPointer;
12366 // Load the _tls_index variable
12367 SDValue IDX = DAG.getExternalSymbol("_tls_index", PtrVT);
12368 if (Subtarget->is64Bit())
12369 IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, PtrVT, Chain, IDX,
12370 MachinePointerInfo(), MVT::i32, false, false,
12373 IDX = DAG.getLoad(PtrVT, dl, Chain, IDX, MachinePointerInfo(), false,
12376 auto &DL = DAG.getDataLayout();
12378 DAG.getConstant(Log2_64_Ceil(DL.getPointerSize()), dl, PtrVT);
12379 IDX = DAG.getNode(ISD::SHL, dl, PtrVT, IDX, Scale);
12381 res = DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, IDX);
12384 res = DAG.getLoad(PtrVT, dl, Chain, res, MachinePointerInfo(), false, false,
12387 // Get the offset of start of .tls section
12388 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
12389 GA->getValueType(0),
12390 GA->getOffset(), X86II::MO_SECREL);
12391 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, TGA);
12393 // The address of the thread local variable is the add of the thread
12394 // pointer with the offset of the variable.
12395 return DAG.getNode(ISD::ADD, dl, PtrVT, res, Offset);
12398 llvm_unreachable("TLS not implemented for this target.");
12401 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values
12402 /// and take a 2 x i32 value to shift plus a shift amount.
12403 static SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) {
12404 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
12405 MVT VT = Op.getSimpleValueType();
12406 unsigned VTBits = VT.getSizeInBits();
12408 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
12409 SDValue ShOpLo = Op.getOperand(0);
12410 SDValue ShOpHi = Op.getOperand(1);
12411 SDValue ShAmt = Op.getOperand(2);
12412 // X86ISD::SHLD and X86ISD::SHRD have defined overflow behavior but the
12413 // generic ISD nodes haven't. Insert an AND to be safe, it's optimized away
12415 SDValue SafeShAmt = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
12416 DAG.getConstant(VTBits - 1, dl, MVT::i8));
12417 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
12418 DAG.getConstant(VTBits - 1, dl, MVT::i8))
12419 : DAG.getConstant(0, dl, VT);
12421 SDValue Tmp2, Tmp3;
12422 if (Op.getOpcode() == ISD::SHL_PARTS) {
12423 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
12424 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, SafeShAmt);
12426 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
12427 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, SafeShAmt);
12430 // If the shift amount is larger or equal than the width of a part we can't
12431 // rely on the results of shld/shrd. Insert a test and select the appropriate
12432 // values for large shift amounts.
12433 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
12434 DAG.getConstant(VTBits, dl, MVT::i8));
12435 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
12436 AndNode, DAG.getConstant(0, dl, MVT::i8));
12439 SDValue CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
12440 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
12441 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
12443 if (Op.getOpcode() == ISD::SHL_PARTS) {
12444 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
12445 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
12447 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
12448 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
12451 SDValue Ops[2] = { Lo, Hi };
12452 return DAG.getMergeValues(Ops, dl);
12455 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
12456 SelectionDAG &DAG) const {
12457 SDValue Src = Op.getOperand(0);
12458 MVT SrcVT = Src.getSimpleValueType();
12459 MVT VT = Op.getSimpleValueType();
12462 if (SrcVT.isVector()) {
12463 if (SrcVT == MVT::v2i32 && VT == MVT::v2f64) {
12464 return DAG.getNode(X86ISD::CVTDQ2PD, dl, VT,
12465 DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4i32, Src,
12466 DAG.getUNDEF(SrcVT)));
12468 if (SrcVT.getVectorElementType() == MVT::i1) {
12469 MVT IntegerVT = MVT::getVectorVT(MVT::i32, SrcVT.getVectorNumElements());
12470 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
12471 DAG.getNode(ISD::SIGN_EXTEND, dl, IntegerVT, Src));
12476 assert(SrcVT <= MVT::i64 && SrcVT >= MVT::i16 &&
12477 "Unknown SINT_TO_FP to lower!");
12479 // These are really Legal; return the operand so the caller accepts it as
12481 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
12483 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
12484 Subtarget->is64Bit()) {
12488 unsigned Size = SrcVT.getSizeInBits()/8;
12489 MachineFunction &MF = DAG.getMachineFunction();
12490 auto PtrVT = getPointerTy(MF.getDataLayout());
12491 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
12492 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12493 SDValue Chain = DAG.getStore(
12494 DAG.getEntryNode(), dl, Op.getOperand(0), StackSlot,
12495 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI), false,
12497 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
12500 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
12502 SelectionDAG &DAG) const {
12506 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
12508 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
12510 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
12512 unsigned ByteSize = SrcVT.getSizeInBits()/8;
12514 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
12515 MachineMemOperand *MMO;
12517 int SSFI = FI->getIndex();
12518 MMO = DAG.getMachineFunction().getMachineMemOperand(
12519 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12520 MachineMemOperand::MOLoad, ByteSize, ByteSize);
12522 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
12523 StackSlot = StackSlot.getOperand(1);
12525 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
12526 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
12528 Tys, Ops, SrcVT, MMO);
12531 Chain = Result.getValue(1);
12532 SDValue InFlag = Result.getValue(2);
12534 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
12535 // shouldn't be necessary except that RFP cannot be live across
12536 // multiple blocks. When stackifier is fixed, they can be uncoupled.
12537 MachineFunction &MF = DAG.getMachineFunction();
12538 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
12539 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
12540 auto PtrVT = getPointerTy(MF.getDataLayout());
12541 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12542 Tys = DAG.getVTList(MVT::Other);
12544 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
12546 MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand(
12547 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12548 MachineMemOperand::MOStore, SSFISize, SSFISize);
12550 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
12551 Ops, Op.getValueType(), MMO);
12552 Result = DAG.getLoad(
12553 Op.getValueType(), DL, Chain, StackSlot,
12554 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12555 false, false, false, 0);
12561 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
12562 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
12563 SelectionDAG &DAG) const {
12564 // This algorithm is not obvious. Here it is what we're trying to output:
12567 punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U }
12568 subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 }
12570 haddpd %xmm0, %xmm0
12572 pshufd $0x4e, %xmm0, %xmm1
12578 LLVMContext *Context = DAG.getContext();
12580 // Build some magic constants.
12581 static const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 };
12582 Constant *C0 = ConstantDataVector::get(*Context, CV0);
12583 auto PtrVT = getPointerTy(DAG.getDataLayout());
12584 SDValue CPIdx0 = DAG.getConstantPool(C0, PtrVT, 16);
12586 SmallVector<Constant*,2> CV1;
12588 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
12589 APInt(64, 0x4330000000000000ULL))));
12591 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
12592 APInt(64, 0x4530000000000000ULL))));
12593 Constant *C1 = ConstantVector::get(CV1);
12594 SDValue CPIdx1 = DAG.getConstantPool(C1, PtrVT, 16);
12596 // Load the 64-bit value into an XMM register.
12597 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
12600 DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
12601 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12602 false, false, false, 16);
12604 getUnpackl(DAG, dl, MVT::v4i32, DAG.getBitcast(MVT::v4i32, XR1), CLod0);
12607 DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
12608 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12609 false, false, false, 16);
12610 SDValue XR2F = DAG.getBitcast(MVT::v2f64, Unpck1);
12611 // TODO: Are there any fast-math-flags to propagate here?
12612 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
12615 if (Subtarget->hasSSE3()) {
12616 // FIXME: The 'haddpd' instruction may be slower than 'movhlps + addsd'.
12617 Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub);
12619 SDValue S2F = DAG.getBitcast(MVT::v4i32, Sub);
12620 SDValue Shuffle = getTargetShuffleNode(X86ISD::PSHUFD, dl, MVT::v4i32,
12622 Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64,
12623 DAG.getBitcast(MVT::v2f64, Shuffle), Sub);
12626 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result,
12627 DAG.getIntPtrConstant(0, dl));
12630 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
12631 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
12632 SelectionDAG &DAG) const {
12634 // FP constant to bias correct the final result.
12635 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
12638 // Load the 32-bit value into an XMM register.
12639 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
12642 // Zero out the upper parts of the register.
12643 Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG);
12645 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12646 DAG.getBitcast(MVT::v2f64, Load),
12647 DAG.getIntPtrConstant(0, dl));
12649 // Or the load with the bias.
12650 SDValue Or = DAG.getNode(
12651 ISD::OR, dl, MVT::v2i64,
12652 DAG.getBitcast(MVT::v2i64,
12653 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Load)),
12654 DAG.getBitcast(MVT::v2i64,
12655 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Bias)));
12657 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12658 DAG.getBitcast(MVT::v2f64, Or), DAG.getIntPtrConstant(0, dl));
12660 // Subtract the bias.
12661 // TODO: Are there any fast-math-flags to propagate here?
12662 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
12664 // Handle final rounding.
12665 MVT DestVT = Op.getSimpleValueType();
12667 if (DestVT.bitsLT(MVT::f64))
12668 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
12669 DAG.getIntPtrConstant(0, dl));
12670 if (DestVT.bitsGT(MVT::f64))
12671 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
12673 // Handle final rounding.
12677 static SDValue lowerUINT_TO_FP_vXi32(SDValue Op, SelectionDAG &DAG,
12678 const X86Subtarget &Subtarget) {
12679 // The algorithm is the following:
12680 // #ifdef __SSE4_1__
12681 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
12682 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
12683 // (uint4) 0x53000000, 0xaa);
12685 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
12686 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
12688 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
12689 // return (float4) lo + fhi;
12691 // We shouldn't use it when unsafe-fp-math is enabled though: we might later
12692 // reassociate the two FADDs, and if we do that, the algorithm fails
12693 // spectacularly (PR24512).
12694 // FIXME: If we ever have some kind of Machine FMF, this should be marked
12695 // as non-fast and always be enabled. Why isn't SDAG FMF enough? Because
12696 // there's also the MachineCombiner reassociations happening on Machine IR.
12697 if (DAG.getTarget().Options.UnsafeFPMath)
12701 SDValue V = Op->getOperand(0);
12702 MVT VecIntVT = V.getSimpleValueType();
12703 bool Is128 = VecIntVT == MVT::v4i32;
12704 MVT VecFloatVT = Is128 ? MVT::v4f32 : MVT::v8f32;
12705 // If we convert to something else than the supported type, e.g., to v4f64,
12707 if (VecFloatVT != Op->getSimpleValueType(0))
12710 unsigned NumElts = VecIntVT.getVectorNumElements();
12711 assert((VecIntVT == MVT::v4i32 || VecIntVT == MVT::v8i32) &&
12712 "Unsupported custom type");
12713 assert(NumElts <= 8 && "The size of the constant array must be fixed");
12715 // In the #idef/#else code, we have in common:
12716 // - The vector of constants:
12722 // Create the splat vector for 0x4b000000.
12723 SDValue CstLow = DAG.getConstant(0x4b000000, DL, MVT::i32);
12724 SDValue CstLowArray[] = {CstLow, CstLow, CstLow, CstLow,
12725 CstLow, CstLow, CstLow, CstLow};
12726 SDValue VecCstLow = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12727 makeArrayRef(&CstLowArray[0], NumElts));
12728 // Create the splat vector for 0x53000000.
12729 SDValue CstHigh = DAG.getConstant(0x53000000, DL, MVT::i32);
12730 SDValue CstHighArray[] = {CstHigh, CstHigh, CstHigh, CstHigh,
12731 CstHigh, CstHigh, CstHigh, CstHigh};
12732 SDValue VecCstHigh = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12733 makeArrayRef(&CstHighArray[0], NumElts));
12735 // Create the right shift.
12736 SDValue CstShift = DAG.getConstant(16, DL, MVT::i32);
12737 SDValue CstShiftArray[] = {CstShift, CstShift, CstShift, CstShift,
12738 CstShift, CstShift, CstShift, CstShift};
12739 SDValue VecCstShift = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12740 makeArrayRef(&CstShiftArray[0], NumElts));
12741 SDValue HighShift = DAG.getNode(ISD::SRL, DL, VecIntVT, V, VecCstShift);
12744 if (Subtarget.hasSSE41()) {
12745 MVT VecI16VT = Is128 ? MVT::v8i16 : MVT::v16i16;
12746 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
12747 SDValue VecCstLowBitcast = DAG.getBitcast(VecI16VT, VecCstLow);
12748 SDValue VecBitcast = DAG.getBitcast(VecI16VT, V);
12749 // Low will be bitcasted right away, so do not bother bitcasting back to its
12751 Low = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecBitcast,
12752 VecCstLowBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
12753 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
12754 // (uint4) 0x53000000, 0xaa);
12755 SDValue VecCstHighBitcast = DAG.getBitcast(VecI16VT, VecCstHigh);
12756 SDValue VecShiftBitcast = DAG.getBitcast(VecI16VT, HighShift);
12757 // High will be bitcasted right away, so do not bother bitcasting back to
12758 // its original type.
12759 High = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecShiftBitcast,
12760 VecCstHighBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
12762 SDValue CstMask = DAG.getConstant(0xffff, DL, MVT::i32);
12763 SDValue VecCstMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT, CstMask,
12764 CstMask, CstMask, CstMask);
12765 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
12766 SDValue LowAnd = DAG.getNode(ISD::AND, DL, VecIntVT, V, VecCstMask);
12767 Low = DAG.getNode(ISD::OR, DL, VecIntVT, LowAnd, VecCstLow);
12769 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
12770 High = DAG.getNode(ISD::OR, DL, VecIntVT, HighShift, VecCstHigh);
12773 // Create the vector constant for -(0x1.0p39f + 0x1.0p23f).
12774 SDValue CstFAdd = DAG.getConstantFP(
12775 APFloat(APFloat::IEEEsingle, APInt(32, 0xD3000080)), DL, MVT::f32);
12776 SDValue CstFAddArray[] = {CstFAdd, CstFAdd, CstFAdd, CstFAdd,
12777 CstFAdd, CstFAdd, CstFAdd, CstFAdd};
12778 SDValue VecCstFAdd = DAG.getNode(ISD::BUILD_VECTOR, DL, VecFloatVT,
12779 makeArrayRef(&CstFAddArray[0], NumElts));
12781 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
12782 SDValue HighBitcast = DAG.getBitcast(VecFloatVT, High);
12783 // TODO: Are there any fast-math-flags to propagate here?
12785 DAG.getNode(ISD::FADD, DL, VecFloatVT, HighBitcast, VecCstFAdd);
12786 // return (float4) lo + fhi;
12787 SDValue LowBitcast = DAG.getBitcast(VecFloatVT, Low);
12788 return DAG.getNode(ISD::FADD, DL, VecFloatVT, LowBitcast, FHigh);
12791 SDValue X86TargetLowering::lowerUINT_TO_FP_vec(SDValue Op,
12792 SelectionDAG &DAG) const {
12793 SDValue N0 = Op.getOperand(0);
12794 MVT SVT = N0.getSimpleValueType();
12797 switch (SVT.SimpleTy) {
12799 llvm_unreachable("Custom UINT_TO_FP is not supported!");
12804 MVT NVT = MVT::getVectorVT(MVT::i32, SVT.getVectorNumElements());
12805 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
12806 DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0));
12810 return lowerUINT_TO_FP_vXi32(Op, DAG, *Subtarget);
12813 assert(Subtarget->hasAVX512());
12814 return DAG.getNode(ISD::UINT_TO_FP, dl, Op.getValueType(),
12815 DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v16i32, N0));
12819 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
12820 SelectionDAG &DAG) const {
12821 SDValue N0 = Op.getOperand(0);
12823 auto PtrVT = getPointerTy(DAG.getDataLayout());
12825 if (Op.getSimpleValueType().isVector())
12826 return lowerUINT_TO_FP_vec(Op, DAG);
12828 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
12829 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
12830 // the optimization here.
12831 if (DAG.SignBitIsZero(N0))
12832 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
12834 MVT SrcVT = N0.getSimpleValueType();
12835 MVT DstVT = Op.getSimpleValueType();
12837 if (Subtarget->hasAVX512() && isScalarFPTypeInSSEReg(DstVT) &&
12838 (SrcVT == MVT::i32 || (SrcVT == MVT::i64 && Subtarget->is64Bit()))) {
12839 // Conversions from unsigned i32 to f32/f64 are legal,
12840 // using VCVTUSI2SS/SD. Same for i64 in 64-bit mode.
12844 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
12845 return LowerUINT_TO_FP_i64(Op, DAG);
12846 if (SrcVT == MVT::i32 && X86ScalarSSEf64)
12847 return LowerUINT_TO_FP_i32(Op, DAG);
12848 if (Subtarget->is64Bit() && SrcVT == MVT::i64 && DstVT == MVT::f32)
12851 // Make a 64-bit buffer, and use it to build an FILD.
12852 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
12853 if (SrcVT == MVT::i32) {
12854 SDValue WordOff = DAG.getConstant(4, dl, PtrVT);
12855 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, WordOff);
12856 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
12857 StackSlot, MachinePointerInfo(),
12859 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, dl, MVT::i32),
12860 OffsetSlot, MachinePointerInfo(),
12862 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
12866 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
12867 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
12868 StackSlot, MachinePointerInfo(),
12870 // For i64 source, we need to add the appropriate power of 2 if the input
12871 // was negative. This is the same as the optimization in
12872 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
12873 // we must be careful to do the computation in x87 extended precision, not
12874 // in SSE. (The generic code can't know it's OK to do this, or how to.)
12875 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
12876 MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand(
12877 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12878 MachineMemOperand::MOLoad, 8, 8);
12880 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
12881 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
12882 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops,
12885 APInt FF(32, 0x5F800000ULL);
12887 // Check whether the sign bit is set.
12888 SDValue SignSet = DAG.getSetCC(
12889 dl, getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i64),
12890 Op.getOperand(0), DAG.getConstant(0, dl, MVT::i64), ISD::SETLT);
12892 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
12893 SDValue FudgePtr = DAG.getConstantPool(
12894 ConstantInt::get(*DAG.getContext(), FF.zext(64)), PtrVT);
12896 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
12897 SDValue Zero = DAG.getIntPtrConstant(0, dl);
12898 SDValue Four = DAG.getIntPtrConstant(4, dl);
12899 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
12901 FudgePtr = DAG.getNode(ISD::ADD, dl, PtrVT, FudgePtr, Offset);
12903 // Load the value out, extending it from f32 to f80.
12904 // FIXME: Avoid the extend by constructing the right constant pool?
12905 SDValue Fudge = DAG.getExtLoad(
12906 ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(), FudgePtr,
12907 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), MVT::f32,
12908 false, false, false, 4);
12909 // Extend everything to 80 bits to force it to be done on x87.
12910 // TODO: Are there any fast-math-flags to propagate here?
12911 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
12912 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add,
12913 DAG.getIntPtrConstant(0, dl));
12916 // If the given FP_TO_SINT (IsSigned) or FP_TO_UINT (!IsSigned) operation
12917 // is legal, or has an fp128 or f16 source (which needs to be promoted to f32),
12918 // just return an <SDValue(), SDValue()> pair.
12919 // Otherwise it is assumed to be a conversion from one of f32, f64 or f80
12920 // to i16, i32 or i64, and we lower it to a legal sequence.
12921 // If lowered to the final integer result we return a <result, SDValue()> pair.
12922 // Otherwise we lower it to a sequence ending with a FIST, return a
12923 // <FIST, StackSlot> pair, and the caller is responsible for loading
12924 // the final integer result from StackSlot.
12925 std::pair<SDValue,SDValue>
12926 X86TargetLowering::FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
12927 bool IsSigned, bool IsReplace) const {
12930 EVT DstTy = Op.getValueType();
12931 EVT TheVT = Op.getOperand(0).getValueType();
12932 auto PtrVT = getPointerTy(DAG.getDataLayout());
12934 if (TheVT != MVT::f32 && TheVT != MVT::f64 && TheVT != MVT::f80) {
12935 // f16 must be promoted before using the lowering in this routine.
12936 // fp128 does not use this lowering.
12937 return std::make_pair(SDValue(), SDValue());
12940 // If using FIST to compute an unsigned i64, we'll need some fixup
12941 // to handle values above the maximum signed i64. A FIST is always
12942 // used for the 32-bit subtarget, but also for f80 on a 64-bit target.
12943 bool UnsignedFixup = !IsSigned &&
12944 DstTy == MVT::i64 &&
12945 (!Subtarget->is64Bit() ||
12946 !isScalarFPTypeInSSEReg(TheVT));
12948 if (!IsSigned && DstTy != MVT::i64 && !Subtarget->hasAVX512()) {
12949 // Replace the fp-to-uint32 operation with an fp-to-sint64 FIST.
12950 // The low 32 bits of the fist result will have the correct uint32 result.
12951 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
12955 assert(DstTy.getSimpleVT() <= MVT::i64 &&
12956 DstTy.getSimpleVT() >= MVT::i16 &&
12957 "Unknown FP_TO_INT to lower!");
12959 // These are really Legal.
12960 if (DstTy == MVT::i32 &&
12961 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
12962 return std::make_pair(SDValue(), SDValue());
12963 if (Subtarget->is64Bit() &&
12964 DstTy == MVT::i64 &&
12965 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
12966 return std::make_pair(SDValue(), SDValue());
12968 // We lower FP->int64 into FISTP64 followed by a load from a temporary
12970 MachineFunction &MF = DAG.getMachineFunction();
12971 unsigned MemSize = DstTy.getSizeInBits()/8;
12972 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
12973 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12976 switch (DstTy.getSimpleVT().SimpleTy) {
12977 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
12978 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
12979 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
12980 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
12983 SDValue Chain = DAG.getEntryNode();
12984 SDValue Value = Op.getOperand(0);
12985 SDValue Adjust; // 0x0 or 0x80000000, for result sign bit adjustment.
12987 if (UnsignedFixup) {
12989 // Conversion to unsigned i64 is implemented with a select,
12990 // depending on whether the source value fits in the range
12991 // of a signed i64. Let Thresh be the FP equivalent of
12992 // 0x8000000000000000ULL.
12994 // Adjust i32 = (Value < Thresh) ? 0 : 0x80000000;
12995 // FistSrc = (Value < Thresh) ? Value : (Value - Thresh);
12996 // Fist-to-mem64 FistSrc
12997 // Add 0 or 0x800...0ULL to the 64-bit result, which is equivalent
12998 // to XOR'ing the high 32 bits with Adjust.
13000 // Being a power of 2, Thresh is exactly representable in all FP formats.
13001 // For X87 we'd like to use the smallest FP type for this constant, but
13002 // for DAG type consistency we have to match the FP operand type.
13004 APFloat Thresh(APFloat::IEEEsingle, APInt(32, 0x5f000000));
13005 LLVM_ATTRIBUTE_UNUSED APFloat::opStatus Status = APFloat::opOK;
13006 bool LosesInfo = false;
13007 if (TheVT == MVT::f64)
13008 // The rounding mode is irrelevant as the conversion should be exact.
13009 Status = Thresh.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven,
13011 else if (TheVT == MVT::f80)
13012 Status = Thresh.convert(APFloat::x87DoubleExtended,
13013 APFloat::rmNearestTiesToEven, &LosesInfo);
13015 assert(Status == APFloat::opOK && !LosesInfo &&
13016 "FP conversion should have been exact");
13018 SDValue ThreshVal = DAG.getConstantFP(Thresh, DL, TheVT);
13020 SDValue Cmp = DAG.getSetCC(DL,
13021 getSetCCResultType(DAG.getDataLayout(),
13022 *DAG.getContext(), TheVT),
13023 Value, ThreshVal, ISD::SETLT);
13024 Adjust = DAG.getSelect(DL, MVT::i32, Cmp,
13025 DAG.getConstant(0, DL, MVT::i32),
13026 DAG.getConstant(0x80000000, DL, MVT::i32));
13027 SDValue Sub = DAG.getNode(ISD::FSUB, DL, TheVT, Value, ThreshVal);
13028 Cmp = DAG.getSetCC(DL, getSetCCResultType(DAG.getDataLayout(),
13029 *DAG.getContext(), TheVT),
13030 Value, ThreshVal, ISD::SETLT);
13031 Value = DAG.getSelect(DL, TheVT, Cmp, Value, Sub);
13034 // FIXME This causes a redundant load/store if the SSE-class value is already
13035 // in memory, such as if it is on the callstack.
13036 if (isScalarFPTypeInSSEReg(TheVT)) {
13037 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
13038 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
13039 MachinePointerInfo::getFixedStack(MF, SSFI), false,
13041 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
13043 Chain, StackSlot, DAG.getValueType(TheVT)
13046 MachineMemOperand *MMO =
13047 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
13048 MachineMemOperand::MOLoad, MemSize, MemSize);
13049 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, DstTy, MMO);
13050 Chain = Value.getValue(1);
13051 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
13052 StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
13055 MachineMemOperand *MMO =
13056 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
13057 MachineMemOperand::MOStore, MemSize, MemSize);
13059 if (UnsignedFixup) {
13061 // Insert the FIST, load its result as two i32's,
13062 // and XOR the high i32 with Adjust.
13064 SDValue FistOps[] = { Chain, Value, StackSlot };
13065 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
13066 FistOps, DstTy, MMO);
13068 SDValue Low32 = DAG.getLoad(MVT::i32, DL, FIST, StackSlot,
13069 MachinePointerInfo(),
13070 false, false, false, 0);
13071 SDValue HighAddr = DAG.getNode(ISD::ADD, DL, PtrVT, StackSlot,
13072 DAG.getConstant(4, DL, PtrVT));
13074 SDValue High32 = DAG.getLoad(MVT::i32, DL, FIST, HighAddr,
13075 MachinePointerInfo(),
13076 false, false, false, 0);
13077 High32 = DAG.getNode(ISD::XOR, DL, MVT::i32, High32, Adjust);
13079 if (Subtarget->is64Bit()) {
13080 // Join High32 and Low32 into a 64-bit result.
13081 // (High32 << 32) | Low32
13082 Low32 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, Low32);
13083 High32 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, High32);
13084 High32 = DAG.getNode(ISD::SHL, DL, MVT::i64, High32,
13085 DAG.getConstant(32, DL, MVT::i8));
13086 SDValue Result = DAG.getNode(ISD::OR, DL, MVT::i64, High32, Low32);
13087 return std::make_pair(Result, SDValue());
13090 SDValue ResultOps[] = { Low32, High32 };
13092 SDValue pair = IsReplace
13093 ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, ResultOps)
13094 : DAG.getMergeValues(ResultOps, DL);
13095 return std::make_pair(pair, SDValue());
13097 // Build the FP_TO_INT*_IN_MEM
13098 SDValue Ops[] = { Chain, Value, StackSlot };
13099 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
13101 return std::make_pair(FIST, StackSlot);
13105 static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
13106 const X86Subtarget *Subtarget) {
13107 MVT VT = Op->getSimpleValueType(0);
13108 SDValue In = Op->getOperand(0);
13109 MVT InVT = In.getSimpleValueType();
13112 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
13113 return DAG.getNode(ISD::ZERO_EXTEND, dl, VT, In);
13115 // Optimize vectors in AVX mode:
13118 // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32.
13119 // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32.
13120 // Concat upper and lower parts.
13123 // Use vpunpckldq for 4 lower elements v4i32 -> v2i64.
13124 // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64.
13125 // Concat upper and lower parts.
13128 if (((VT != MVT::v16i16) || (InVT != MVT::v16i8)) &&
13129 ((VT != MVT::v8i32) || (InVT != MVT::v8i16)) &&
13130 ((VT != MVT::v4i64) || (InVT != MVT::v4i32)))
13133 if (Subtarget->hasInt256())
13134 return DAG.getNode(X86ISD::VZEXT, dl, VT, In);
13136 SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl);
13137 SDValue Undef = DAG.getUNDEF(InVT);
13138 bool NeedZero = Op.getOpcode() == ISD::ZERO_EXTEND;
13139 SDValue OpLo = getUnpackl(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
13140 SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
13142 MVT HVT = MVT::getVectorVT(VT.getVectorElementType(),
13143 VT.getVectorNumElements()/2);
13145 OpLo = DAG.getBitcast(HVT, OpLo);
13146 OpHi = DAG.getBitcast(HVT, OpHi);
13148 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
13151 static SDValue LowerZERO_EXTEND_AVX512(SDValue Op,
13152 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
13153 MVT VT = Op->getSimpleValueType(0);
13154 SDValue In = Op->getOperand(0);
13155 MVT InVT = In.getSimpleValueType();
13157 unsigned int NumElts = VT.getVectorNumElements();
13158 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
13161 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1)
13162 return DAG.getNode(X86ISD::VZEXT, DL, VT, In);
13164 assert(InVT.getVectorElementType() == MVT::i1);
13165 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
13167 DAG.getConstant(APInt(ExtVT.getScalarSizeInBits(), 1), DL, ExtVT);
13169 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), DL, ExtVT);
13171 SDValue V = DAG.getNode(ISD::VSELECT, DL, ExtVT, In, One, Zero);
13172 if (VT.is512BitVector())
13174 return DAG.getNode(X86ISD::VTRUNC, DL, VT, V);
13177 static SDValue LowerANY_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
13178 SelectionDAG &DAG) {
13179 if (Subtarget->hasFp256())
13180 if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget))
13186 static SDValue LowerZERO_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
13187 SelectionDAG &DAG) {
13189 MVT VT = Op.getSimpleValueType();
13190 SDValue In = Op.getOperand(0);
13191 MVT SVT = In.getSimpleValueType();
13193 if (VT.is512BitVector() || SVT.getVectorElementType() == MVT::i1)
13194 return LowerZERO_EXTEND_AVX512(Op, Subtarget, DAG);
13196 if (Subtarget->hasFp256())
13197 if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget))
13200 assert(!VT.is256BitVector() || !SVT.is128BitVector() ||
13201 VT.getVectorNumElements() != SVT.getVectorNumElements());
13205 SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
13207 MVT VT = Op.getSimpleValueType();
13208 SDValue In = Op.getOperand(0);
13209 MVT InVT = In.getSimpleValueType();
13211 if (VT == MVT::i1) {
13212 assert((InVT.isInteger() && (InVT.getSizeInBits() <= 64)) &&
13213 "Invalid scalar TRUNCATE operation");
13214 if (InVT.getSizeInBits() >= 32)
13216 In = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, In);
13217 return DAG.getNode(ISD::TRUNCATE, DL, VT, In);
13219 assert(VT.getVectorNumElements() == InVT.getVectorNumElements() &&
13220 "Invalid TRUNCATE operation");
13222 // move vector to mask - truncate solution for SKX
13223 if (VT.getVectorElementType() == MVT::i1) {
13224 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() <= 16 &&
13225 Subtarget->hasBWI())
13226 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
13227 if ((InVT.is256BitVector() || InVT.is128BitVector())
13228 && InVT.getScalarSizeInBits() <= 16 &&
13229 Subtarget->hasBWI() && Subtarget->hasVLX())
13230 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
13231 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() >= 32 &&
13232 Subtarget->hasDQI())
13233 return Op; // legal, will go to VPMOVD2M, VPMOVQ2M
13234 if ((InVT.is256BitVector() || InVT.is128BitVector())
13235 && InVT.getScalarSizeInBits() >= 32 &&
13236 Subtarget->hasDQI() && Subtarget->hasVLX())
13237 return Op; // legal, will go to VPMOVB2M, VPMOVQ2M
13240 if (VT.getVectorElementType() == MVT::i1) {
13241 assert(VT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
13242 unsigned NumElts = InVT.getVectorNumElements();
13243 assert ((NumElts == 8 || NumElts == 16) && "Unexpected vector type");
13244 if (InVT.getSizeInBits() < 512) {
13245 MVT ExtVT = (NumElts == 16)? MVT::v16i32 : MVT::v8i64;
13246 In = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, In);
13251 DAG.getConstant(APInt::getSignBit(InVT.getScalarSizeInBits()), DL, InVT);
13252 SDValue And = DAG.getNode(ISD::AND, DL, InVT, OneV, In);
13253 return DAG.getNode(X86ISD::TESTM, DL, VT, And, And);
13256 // vpmovqb/w/d, vpmovdb/w, vpmovwb
13257 if (Subtarget->hasAVX512()) {
13258 // word to byte only under BWI
13259 if (InVT == MVT::v16i16 && !Subtarget->hasBWI()) // v16i16 -> v16i8
13260 return DAG.getNode(X86ISD::VTRUNC, DL, VT,
13261 DAG.getNode(X86ISD::VSEXT, DL, MVT::v16i32, In));
13262 return DAG.getNode(X86ISD::VTRUNC, DL, VT, In);
13264 if ((VT == MVT::v4i32) && (InVT == MVT::v4i64)) {
13265 // On AVX2, v4i64 -> v4i32 becomes VPERMD.
13266 if (Subtarget->hasInt256()) {
13267 static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1};
13268 In = DAG.getBitcast(MVT::v8i32, In);
13269 In = DAG.getVectorShuffle(MVT::v8i32, DL, In, DAG.getUNDEF(MVT::v8i32),
13271 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, In,
13272 DAG.getIntPtrConstant(0, DL));
13275 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13276 DAG.getIntPtrConstant(0, DL));
13277 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13278 DAG.getIntPtrConstant(2, DL));
13279 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
13280 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
13281 static const int ShufMask[] = {0, 2, 4, 6};
13282 return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask);
13285 if ((VT == MVT::v8i16) && (InVT == MVT::v8i32)) {
13286 // On AVX2, v8i32 -> v8i16 becomed PSHUFB.
13287 if (Subtarget->hasInt256()) {
13288 In = DAG.getBitcast(MVT::v32i8, In);
13290 SmallVector<SDValue,32> pshufbMask;
13291 for (unsigned i = 0; i < 2; ++i) {
13292 pshufbMask.push_back(DAG.getConstant(0x0, DL, MVT::i8));
13293 pshufbMask.push_back(DAG.getConstant(0x1, DL, MVT::i8));
13294 pshufbMask.push_back(DAG.getConstant(0x4, DL, MVT::i8));
13295 pshufbMask.push_back(DAG.getConstant(0x5, DL, MVT::i8));
13296 pshufbMask.push_back(DAG.getConstant(0x8, DL, MVT::i8));
13297 pshufbMask.push_back(DAG.getConstant(0x9, DL, MVT::i8));
13298 pshufbMask.push_back(DAG.getConstant(0xc, DL, MVT::i8));
13299 pshufbMask.push_back(DAG.getConstant(0xd, DL, MVT::i8));
13300 for (unsigned j = 0; j < 8; ++j)
13301 pshufbMask.push_back(DAG.getConstant(0x80, DL, MVT::i8));
13303 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, pshufbMask);
13304 In = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8, In, BV);
13305 In = DAG.getBitcast(MVT::v4i64, In);
13307 static const int ShufMask[] = {0, 2, -1, -1};
13308 In = DAG.getVectorShuffle(MVT::v4i64, DL, In, DAG.getUNDEF(MVT::v4i64),
13310 In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13311 DAG.getIntPtrConstant(0, DL));
13312 return DAG.getBitcast(VT, In);
13315 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
13316 DAG.getIntPtrConstant(0, DL));
13318 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
13319 DAG.getIntPtrConstant(4, DL));
13321 OpLo = DAG.getBitcast(MVT::v16i8, OpLo);
13322 OpHi = DAG.getBitcast(MVT::v16i8, OpHi);
13324 // The PSHUFB mask:
13325 static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13,
13326 -1, -1, -1, -1, -1, -1, -1, -1};
13328 SDValue Undef = DAG.getUNDEF(MVT::v16i8);
13329 OpLo = DAG.getVectorShuffle(MVT::v16i8, DL, OpLo, Undef, ShufMask1);
13330 OpHi = DAG.getVectorShuffle(MVT::v16i8, DL, OpHi, Undef, ShufMask1);
13332 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
13333 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
13335 // The MOVLHPS Mask:
13336 static const int ShufMask2[] = {0, 1, 4, 5};
13337 SDValue res = DAG.getVectorShuffle(MVT::v4i32, DL, OpLo, OpHi, ShufMask2);
13338 return DAG.getBitcast(MVT::v8i16, res);
13341 // Handle truncation of V256 to V128 using shuffles.
13342 if (!VT.is128BitVector() || !InVT.is256BitVector())
13345 assert(Subtarget->hasFp256() && "256-bit vector without AVX!");
13347 unsigned NumElems = VT.getVectorNumElements();
13348 MVT NVT = MVT::getVectorVT(VT.getVectorElementType(), NumElems * 2);
13350 SmallVector<int, 16> MaskVec(NumElems * 2, -1);
13351 // Prepare truncation shuffle mask
13352 for (unsigned i = 0; i != NumElems; ++i)
13353 MaskVec[i] = i * 2;
13354 SDValue V = DAG.getVectorShuffle(NVT, DL, DAG.getBitcast(NVT, In),
13355 DAG.getUNDEF(NVT), &MaskVec[0]);
13356 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V,
13357 DAG.getIntPtrConstant(0, DL));
13360 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
13361 SelectionDAG &DAG) const {
13362 assert(!Op.getSimpleValueType().isVector());
13364 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
13365 /*IsSigned=*/ true, /*IsReplace=*/ false);
13366 SDValue FIST = Vals.first, StackSlot = Vals.second;
13367 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
13368 if (!FIST.getNode())
13371 if (StackSlot.getNode())
13372 // Load the result.
13373 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
13374 FIST, StackSlot, MachinePointerInfo(),
13375 false, false, false, 0);
13377 // The node is the result.
13381 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
13382 SelectionDAG &DAG) const {
13383 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
13384 /*IsSigned=*/ false, /*IsReplace=*/ false);
13385 SDValue FIST = Vals.first, StackSlot = Vals.second;
13386 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
13387 if (!FIST.getNode())
13390 if (StackSlot.getNode())
13391 // Load the result.
13392 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
13393 FIST, StackSlot, MachinePointerInfo(),
13394 false, false, false, 0);
13396 // The node is the result.
13400 static SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) {
13402 MVT VT = Op.getSimpleValueType();
13403 SDValue In = Op.getOperand(0);
13404 MVT SVT = In.getSimpleValueType();
13406 assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!");
13408 return DAG.getNode(X86ISD::VFPEXT, DL, VT,
13409 DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32,
13410 In, DAG.getUNDEF(SVT)));
13413 /// The only differences between FABS and FNEG are the mask and the logic op.
13414 /// FNEG also has a folding opportunity for FNEG(FABS(x)).
13415 static SDValue LowerFABSorFNEG(SDValue Op, SelectionDAG &DAG) {
13416 assert((Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG) &&
13417 "Wrong opcode for lowering FABS or FNEG.");
13419 bool IsFABS = (Op.getOpcode() == ISD::FABS);
13421 // If this is a FABS and it has an FNEG user, bail out to fold the combination
13422 // into an FNABS. We'll lower the FABS after that if it is still in use.
13424 for (SDNode *User : Op->uses())
13425 if (User->getOpcode() == ISD::FNEG)
13429 MVT VT = Op.getSimpleValueType();
13431 bool IsF128 = (VT == MVT::f128);
13433 // FIXME: Use function attribute "OptimizeForSize" and/or CodeGenOpt::Level to
13434 // decide if we should generate a 16-byte constant mask when we only need 4 or
13435 // 8 bytes for the scalar case.
13441 if (VT.isVector()) {
13443 EltVT = VT.getVectorElementType();
13444 NumElts = VT.getVectorNumElements();
13445 } else if (IsF128) {
13446 // SSE instructions are used for optimized f128 logical operations.
13447 LogicVT = MVT::f128;
13451 // There are no scalar bitwise logical SSE/AVX instructions, so we
13452 // generate a 16-byte vector constant and logic op even for the scalar case.
13453 // Using a 16-byte mask allows folding the load of the mask with
13454 // the logic op, so it can save (~4 bytes) on code size.
13455 LogicVT = (VT == MVT::f64) ? MVT::v2f64 : MVT::v4f32;
13457 NumElts = (VT == MVT::f64) ? 2 : 4;
13460 unsigned EltBits = EltVT.getSizeInBits();
13461 LLVMContext *Context = DAG.getContext();
13462 // For FABS, mask is 0x7f...; for FNEG, mask is 0x80...
13464 IsFABS ? APInt::getSignedMaxValue(EltBits) : APInt::getSignBit(EltBits);
13465 Constant *C = ConstantInt::get(*Context, MaskElt);
13466 C = ConstantVector::getSplat(NumElts, C);
13467 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13468 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout()));
13469 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
13471 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13472 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13473 false, false, false, Alignment);
13475 SDValue Op0 = Op.getOperand(0);
13476 bool IsFNABS = !IsFABS && (Op0.getOpcode() == ISD::FABS);
13478 IsFABS ? X86ISD::FAND : IsFNABS ? X86ISD::FOR : X86ISD::FXOR;
13479 SDValue Operand = IsFNABS ? Op0.getOperand(0) : Op0;
13481 if (VT.isVector() || IsF128)
13482 return DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask);
13484 // For the scalar case extend to a 128-bit vector, perform the logic op,
13485 // and extract the scalar result back out.
13486 Operand = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Operand);
13487 SDValue LogicNode = DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask);
13488 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, LogicNode,
13489 DAG.getIntPtrConstant(0, dl));
13492 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
13493 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13494 LLVMContext *Context = DAG.getContext();
13495 SDValue Op0 = Op.getOperand(0);
13496 SDValue Op1 = Op.getOperand(1);
13498 MVT VT = Op.getSimpleValueType();
13499 MVT SrcVT = Op1.getSimpleValueType();
13500 bool IsF128 = (VT == MVT::f128);
13502 // If second operand is smaller, extend it first.
13503 if (SrcVT.bitsLT(VT)) {
13504 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
13507 // And if it is bigger, shrink it first.
13508 if (SrcVT.bitsGT(VT)) {
13509 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1, dl));
13513 // At this point the operands and the result should have the same
13514 // type, and that won't be f80 since that is not custom lowered.
13515 assert((VT == MVT::f64 || VT == MVT::f32 || IsF128) &&
13516 "Unexpected type in LowerFCOPYSIGN");
13518 const fltSemantics &Sem =
13519 VT == MVT::f64 ? APFloat::IEEEdouble :
13520 (IsF128 ? APFloat::IEEEquad : APFloat::IEEEsingle);
13521 const unsigned SizeInBits = VT.getSizeInBits();
13523 SmallVector<Constant *, 4> CV(
13524 VT == MVT::f64 ? 2 : (IsF128 ? 1 : 4),
13525 ConstantFP::get(*Context, APFloat(Sem, APInt(SizeInBits, 0))));
13527 // First, clear all bits but the sign bit from the second operand (sign).
13528 CV[0] = ConstantFP::get(*Context,
13529 APFloat(Sem, APInt::getHighBitsSet(SizeInBits, 1)));
13530 Constant *C = ConstantVector::get(CV);
13531 auto PtrVT = TLI.getPointerTy(DAG.getDataLayout());
13532 SDValue CPIdx = DAG.getConstantPool(C, PtrVT, 16);
13534 // Perform all logic operations as 16-byte vectors because there are no
13535 // scalar FP logic instructions in SSE. This allows load folding of the
13536 // constants into the logic instructions.
13537 MVT LogicVT = (VT == MVT::f64) ? MVT::v2f64 : (IsF128 ? MVT::f128 : MVT::v4f32);
13539 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13540 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13541 false, false, false, 16);
13543 Op1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Op1);
13544 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, LogicVT, Op1, Mask1);
13546 // Next, clear the sign bit from the first operand (magnitude).
13547 // If it's a constant, we can clear it here.
13548 if (ConstantFPSDNode *Op0CN = dyn_cast<ConstantFPSDNode>(Op0)) {
13549 APFloat APF = Op0CN->getValueAPF();
13550 // If the magnitude is a positive zero, the sign bit alone is enough.
13551 if (APF.isPosZero())
13552 return IsF128 ? SignBit :
13553 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcVT, SignBit,
13554 DAG.getIntPtrConstant(0, dl));
13556 CV[0] = ConstantFP::get(*Context, APF);
13558 CV[0] = ConstantFP::get(
13560 APFloat(Sem, APInt::getLowBitsSet(SizeInBits, SizeInBits - 1)));
13562 C = ConstantVector::get(CV);
13563 CPIdx = DAG.getConstantPool(C, PtrVT, 16);
13565 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13566 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13567 false, false, false, 16);
13568 // If the magnitude operand wasn't a constant, we need to AND out the sign.
13569 if (!isa<ConstantFPSDNode>(Op0)) {
13571 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Op0);
13572 Val = DAG.getNode(X86ISD::FAND, dl, LogicVT, Op0, Val);
13574 // OR the magnitude value with the sign bit.
13575 Val = DAG.getNode(X86ISD::FOR, dl, LogicVT, Val, SignBit);
13576 return IsF128 ? Val :
13577 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcVT, Val,
13578 DAG.getIntPtrConstant(0, dl));
13581 static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) {
13582 SDValue N0 = Op.getOperand(0);
13584 MVT VT = Op.getSimpleValueType();
13586 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
13587 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
13588 DAG.getConstant(1, dl, VT));
13589 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, dl, VT));
13592 // Check whether an OR'd tree is PTEST-able.
13593 static SDValue LowerVectorAllZeroTest(SDValue Op, const X86Subtarget *Subtarget,
13594 SelectionDAG &DAG) {
13595 assert(Op.getOpcode() == ISD::OR && "Only check OR'd tree.");
13597 if (!Subtarget->hasSSE41())
13600 if (!Op->hasOneUse())
13603 SDNode *N = Op.getNode();
13606 SmallVector<SDValue, 8> Opnds;
13607 DenseMap<SDValue, unsigned> VecInMap;
13608 SmallVector<SDValue, 8> VecIns;
13609 EVT VT = MVT::Other;
13611 // Recognize a special case where a vector is casted into wide integer to
13613 Opnds.push_back(N->getOperand(0));
13614 Opnds.push_back(N->getOperand(1));
13616 for (unsigned Slot = 0, e = Opnds.size(); Slot < e; ++Slot) {
13617 SmallVectorImpl<SDValue>::const_iterator I = Opnds.begin() + Slot;
13618 // BFS traverse all OR'd operands.
13619 if (I->getOpcode() == ISD::OR) {
13620 Opnds.push_back(I->getOperand(0));
13621 Opnds.push_back(I->getOperand(1));
13622 // Re-evaluate the number of nodes to be traversed.
13623 e += 2; // 2 more nodes (LHS and RHS) are pushed.
13627 // Quit if a non-EXTRACT_VECTOR_ELT
13628 if (I->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
13631 // Quit if without a constant index.
13632 SDValue Idx = I->getOperand(1);
13633 if (!isa<ConstantSDNode>(Idx))
13636 SDValue ExtractedFromVec = I->getOperand(0);
13637 DenseMap<SDValue, unsigned>::iterator M = VecInMap.find(ExtractedFromVec);
13638 if (M == VecInMap.end()) {
13639 VT = ExtractedFromVec.getValueType();
13640 // Quit if not 128/256-bit vector.
13641 if (!VT.is128BitVector() && !VT.is256BitVector())
13643 // Quit if not the same type.
13644 if (VecInMap.begin() != VecInMap.end() &&
13645 VT != VecInMap.begin()->first.getValueType())
13647 M = VecInMap.insert(std::make_pair(ExtractedFromVec, 0)).first;
13648 VecIns.push_back(ExtractedFromVec);
13650 M->second |= 1U << cast<ConstantSDNode>(Idx)->getZExtValue();
13653 assert((VT.is128BitVector() || VT.is256BitVector()) &&
13654 "Not extracted from 128-/256-bit vector.");
13656 unsigned FullMask = (1U << VT.getVectorNumElements()) - 1U;
13658 for (DenseMap<SDValue, unsigned>::const_iterator
13659 I = VecInMap.begin(), E = VecInMap.end(); I != E; ++I) {
13660 // Quit if not all elements are used.
13661 if (I->second != FullMask)
13665 MVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
13667 // Cast all vectors into TestVT for PTEST.
13668 for (unsigned i = 0, e = VecIns.size(); i < e; ++i)
13669 VecIns[i] = DAG.getBitcast(TestVT, VecIns[i]);
13671 // If more than one full vectors are evaluated, OR them first before PTEST.
13672 for (unsigned Slot = 0, e = VecIns.size(); e - Slot > 1; Slot += 2, e += 1) {
13673 // Each iteration will OR 2 nodes and append the result until there is only
13674 // 1 node left, i.e. the final OR'd value of all vectors.
13675 SDValue LHS = VecIns[Slot];
13676 SDValue RHS = VecIns[Slot + 1];
13677 VecIns.push_back(DAG.getNode(ISD::OR, DL, TestVT, LHS, RHS));
13680 return DAG.getNode(X86ISD::PTEST, DL, MVT::i32,
13681 VecIns.back(), VecIns.back());
13684 /// \brief return true if \c Op has a use that doesn't just read flags.
13685 static bool hasNonFlagsUse(SDValue Op) {
13686 for (SDNode::use_iterator UI = Op->use_begin(), UE = Op->use_end(); UI != UE;
13688 SDNode *User = *UI;
13689 unsigned UOpNo = UI.getOperandNo();
13690 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
13691 // Look pass truncate.
13692 UOpNo = User->use_begin().getOperandNo();
13693 User = *User->use_begin();
13696 if (User->getOpcode() != ISD::BRCOND && User->getOpcode() != ISD::SETCC &&
13697 !(User->getOpcode() == ISD::SELECT && UOpNo == 0))
13703 /// Emit nodes that will be selected as "test Op0,Op0", or something
13705 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC, SDLoc dl,
13706 SelectionDAG &DAG) const {
13707 if (Op.getValueType() == MVT::i1) {
13708 SDValue ExtOp = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i8, Op);
13709 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, ExtOp,
13710 DAG.getConstant(0, dl, MVT::i8));
13712 // CF and OF aren't always set the way we want. Determine which
13713 // of these we need.
13714 bool NeedCF = false;
13715 bool NeedOF = false;
13718 case X86::COND_A: case X86::COND_AE:
13719 case X86::COND_B: case X86::COND_BE:
13722 case X86::COND_G: case X86::COND_GE:
13723 case X86::COND_L: case X86::COND_LE:
13724 case X86::COND_O: case X86::COND_NO: {
13725 // Check if we really need to set the
13726 // Overflow flag. If NoSignedWrap is present
13727 // that is not actually needed.
13728 switch (Op->getOpcode()) {
13733 const auto *BinNode = cast<BinaryWithFlagsSDNode>(Op.getNode());
13734 if (BinNode->Flags.hasNoSignedWrap())
13744 // See if we can use the EFLAGS value from the operand instead of
13745 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
13746 // we prove that the arithmetic won't overflow, we can't use OF or CF.
13747 if (Op.getResNo() != 0 || NeedOF || NeedCF) {
13748 // Emit a CMP with 0, which is the TEST pattern.
13749 //if (Op.getValueType() == MVT::i1)
13750 // return DAG.getNode(X86ISD::CMP, dl, MVT::i1, Op,
13751 // DAG.getConstant(0, MVT::i1));
13752 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
13753 DAG.getConstant(0, dl, Op.getValueType()));
13755 unsigned Opcode = 0;
13756 unsigned NumOperands = 0;
13758 // Truncate operations may prevent the merge of the SETCC instruction
13759 // and the arithmetic instruction before it. Attempt to truncate the operands
13760 // of the arithmetic instruction and use a reduced bit-width instruction.
13761 bool NeedTruncation = false;
13762 SDValue ArithOp = Op;
13763 if (Op->getOpcode() == ISD::TRUNCATE && Op->hasOneUse()) {
13764 SDValue Arith = Op->getOperand(0);
13765 // Both the trunc and the arithmetic op need to have one user each.
13766 if (Arith->hasOneUse())
13767 switch (Arith.getOpcode()) {
13774 NeedTruncation = true;
13780 // NOTICE: In the code below we use ArithOp to hold the arithmetic operation
13781 // which may be the result of a CAST. We use the variable 'Op', which is the
13782 // non-casted variable when we check for possible users.
13783 switch (ArithOp.getOpcode()) {
13785 // Due to an isel shortcoming, be conservative if this add is likely to be
13786 // selected as part of a load-modify-store instruction. When the root node
13787 // in a match is a store, isel doesn't know how to remap non-chain non-flag
13788 // uses of other nodes in the match, such as the ADD in this case. This
13789 // leads to the ADD being left around and reselected, with the result being
13790 // two adds in the output. Alas, even if none our users are stores, that
13791 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
13792 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
13793 // climbing the DAG back to the root, and it doesn't seem to be worth the
13795 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13796 UE = Op.getNode()->use_end(); UI != UE; ++UI)
13797 if (UI->getOpcode() != ISD::CopyToReg &&
13798 UI->getOpcode() != ISD::SETCC &&
13799 UI->getOpcode() != ISD::STORE)
13802 if (ConstantSDNode *C =
13803 dyn_cast<ConstantSDNode>(ArithOp.getNode()->getOperand(1))) {
13804 // An add of one will be selected as an INC.
13805 if (C->isOne() && !Subtarget->slowIncDec()) {
13806 Opcode = X86ISD::INC;
13811 // An add of negative one (subtract of one) will be selected as a DEC.
13812 if (C->isAllOnesValue() && !Subtarget->slowIncDec()) {
13813 Opcode = X86ISD::DEC;
13819 // Otherwise use a regular EFLAGS-setting add.
13820 Opcode = X86ISD::ADD;
13825 // If we have a constant logical shift that's only used in a comparison
13826 // against zero turn it into an equivalent AND. This allows turning it into
13827 // a TEST instruction later.
13828 if ((X86CC == X86::COND_E || X86CC == X86::COND_NE) && Op->hasOneUse() &&
13829 isa<ConstantSDNode>(Op->getOperand(1)) && !hasNonFlagsUse(Op)) {
13830 EVT VT = Op.getValueType();
13831 unsigned BitWidth = VT.getSizeInBits();
13832 unsigned ShAmt = Op->getConstantOperandVal(1);
13833 if (ShAmt >= BitWidth) // Avoid undefined shifts.
13835 APInt Mask = ArithOp.getOpcode() == ISD::SRL
13836 ? APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt)
13837 : APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt);
13838 if (!Mask.isSignedIntN(32)) // Avoid large immediates.
13840 SDValue New = DAG.getNode(ISD::AND, dl, VT, Op->getOperand(0),
13841 DAG.getConstant(Mask, dl, VT));
13842 DAG.ReplaceAllUsesWith(Op, New);
13848 // If the primary and result isn't used, don't bother using X86ISD::AND,
13849 // because a TEST instruction will be better.
13850 if (!hasNonFlagsUse(Op))
13856 // Due to the ISEL shortcoming noted above, be conservative if this op is
13857 // likely to be selected as part of a load-modify-store instruction.
13858 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13859 UE = Op.getNode()->use_end(); UI != UE; ++UI)
13860 if (UI->getOpcode() == ISD::STORE)
13863 // Otherwise use a regular EFLAGS-setting instruction.
13864 switch (ArithOp.getOpcode()) {
13865 default: llvm_unreachable("unexpected operator!");
13866 case ISD::SUB: Opcode = X86ISD::SUB; break;
13867 case ISD::XOR: Opcode = X86ISD::XOR; break;
13868 case ISD::AND: Opcode = X86ISD::AND; break;
13870 if (!NeedTruncation && (X86CC == X86::COND_E || X86CC == X86::COND_NE)) {
13871 SDValue EFLAGS = LowerVectorAllZeroTest(Op, Subtarget, DAG);
13872 if (EFLAGS.getNode())
13875 Opcode = X86ISD::OR;
13889 return SDValue(Op.getNode(), 1);
13895 // If we found that truncation is beneficial, perform the truncation and
13897 if (NeedTruncation) {
13898 EVT VT = Op.getValueType();
13899 SDValue WideVal = Op->getOperand(0);
13900 EVT WideVT = WideVal.getValueType();
13901 unsigned ConvertedOp = 0;
13902 // Use a target machine opcode to prevent further DAGCombine
13903 // optimizations that may separate the arithmetic operations
13904 // from the setcc node.
13905 switch (WideVal.getOpcode()) {
13907 case ISD::ADD: ConvertedOp = X86ISD::ADD; break;
13908 case ISD::SUB: ConvertedOp = X86ISD::SUB; break;
13909 case ISD::AND: ConvertedOp = X86ISD::AND; break;
13910 case ISD::OR: ConvertedOp = X86ISD::OR; break;
13911 case ISD::XOR: ConvertedOp = X86ISD::XOR; break;
13915 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13916 if (TLI.isOperationLegal(WideVal.getOpcode(), WideVT)) {
13917 SDValue V0 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(0));
13918 SDValue V1 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(1));
13919 Op = DAG.getNode(ConvertedOp, dl, VT, V0, V1);
13925 // Emit a CMP with 0, which is the TEST pattern.
13926 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
13927 DAG.getConstant(0, dl, Op.getValueType()));
13929 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
13930 SmallVector<SDValue, 4> Ops(Op->op_begin(), Op->op_begin() + NumOperands);
13932 SDValue New = DAG.getNode(Opcode, dl, VTs, Ops);
13933 DAG.ReplaceAllUsesWith(Op, New);
13934 return SDValue(New.getNode(), 1);
13937 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
13939 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
13940 SDLoc dl, SelectionDAG &DAG) const {
13941 if (isNullConstant(Op1))
13942 return EmitTest(Op0, X86CC, dl, DAG);
13944 assert(!(isa<ConstantSDNode>(Op1) && Op0.getValueType() == MVT::i1) &&
13945 "Unexpected comparison operation for MVT::i1 operands");
13947 if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 ||
13948 Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) {
13949 // Do the comparison at i32 if it's smaller, besides the Atom case.
13950 // This avoids subregister aliasing issues. Keep the smaller reference
13951 // if we're optimizing for size, however, as that'll allow better folding
13952 // of memory operations.
13953 if (Op0.getValueType() != MVT::i32 && Op0.getValueType() != MVT::i64 &&
13954 !DAG.getMachineFunction().getFunction()->optForMinSize() &&
13955 !Subtarget->isAtom()) {
13956 unsigned ExtendOp =
13957 isX86CCUnsigned(X86CC) ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
13958 Op0 = DAG.getNode(ExtendOp, dl, MVT::i32, Op0);
13959 Op1 = DAG.getNode(ExtendOp, dl, MVT::i32, Op1);
13961 // Use SUB instead of CMP to enable CSE between SUB and CMP.
13962 SDVTList VTs = DAG.getVTList(Op0.getValueType(), MVT::i32);
13963 SDValue Sub = DAG.getNode(X86ISD::SUB, dl, VTs,
13965 return SDValue(Sub.getNode(), 1);
13967 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
13970 /// Convert a comparison if required by the subtarget.
13971 SDValue X86TargetLowering::ConvertCmpIfNecessary(SDValue Cmp,
13972 SelectionDAG &DAG) const {
13973 // If the subtarget does not support the FUCOMI instruction, floating-point
13974 // comparisons have to be converted.
13975 if (Subtarget->hasCMov() ||
13976 Cmp.getOpcode() != X86ISD::CMP ||
13977 !Cmp.getOperand(0).getValueType().isFloatingPoint() ||
13978 !Cmp.getOperand(1).getValueType().isFloatingPoint())
13981 // The instruction selector will select an FUCOM instruction instead of
13982 // FUCOMI, which writes the comparison result to FPSW instead of EFLAGS. Hence
13983 // build an SDNode sequence that transfers the result from FPSW into EFLAGS:
13984 // (X86sahf (trunc (srl (X86fp_stsw (trunc (X86cmp ...)), 8))))
13986 SDValue TruncFPSW = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, Cmp);
13987 SDValue FNStSW = DAG.getNode(X86ISD::FNSTSW16r, dl, MVT::i16, TruncFPSW);
13988 SDValue Srl = DAG.getNode(ISD::SRL, dl, MVT::i16, FNStSW,
13989 DAG.getConstant(8, dl, MVT::i8));
13990 SDValue TruncSrl = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Srl);
13992 // Some 64-bit targets lack SAHF support, but they do support FCOMI.
13993 assert(Subtarget->hasLAHFSAHF() && "Target doesn't support SAHF or FCOMI?");
13994 return DAG.getNode(X86ISD::SAHF, dl, MVT::i32, TruncSrl);
13997 /// The minimum architected relative accuracy is 2^-12. We need one
13998 /// Newton-Raphson step to have a good float result (24 bits of precision).
13999 SDValue X86TargetLowering::getRsqrtEstimate(SDValue Op,
14000 DAGCombinerInfo &DCI,
14001 unsigned &RefinementSteps,
14002 bool &UseOneConstNR) const {
14003 EVT VT = Op.getValueType();
14004 const char *RecipOp;
14006 // SSE1 has rsqrtss and rsqrtps. AVX adds a 256-bit variant for rsqrtps.
14007 // TODO: Add support for AVX512 (v16f32).
14008 // It is likely not profitable to do this for f64 because a double-precision
14009 // rsqrt estimate with refinement on x86 prior to FMA requires at least 16
14010 // instructions: convert to single, rsqrtss, convert back to double, refine
14011 // (3 steps = at least 13 insts). If an 'rsqrtsd' variant was added to the ISA
14012 // along with FMA, this could be a throughput win.
14013 if (VT == MVT::f32 && Subtarget->hasSSE1())
14015 else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
14016 (VT == MVT::v8f32 && Subtarget->hasAVX()))
14017 RecipOp = "vec-sqrtf";
14021 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
14022 if (!Recips.isEnabled(RecipOp))
14025 RefinementSteps = Recips.getRefinementSteps(RecipOp);
14026 UseOneConstNR = false;
14027 return DCI.DAG.getNode(X86ISD::FRSQRT, SDLoc(Op), VT, Op);
14030 /// The minimum architected relative accuracy is 2^-12. We need one
14031 /// Newton-Raphson step to have a good float result (24 bits of precision).
14032 SDValue X86TargetLowering::getRecipEstimate(SDValue Op,
14033 DAGCombinerInfo &DCI,
14034 unsigned &RefinementSteps) const {
14035 EVT VT = Op.getValueType();
14036 const char *RecipOp;
14038 // SSE1 has rcpss and rcpps. AVX adds a 256-bit variant for rcpps.
14039 // TODO: Add support for AVX512 (v16f32).
14040 // It is likely not profitable to do this for f64 because a double-precision
14041 // reciprocal estimate with refinement on x86 prior to FMA requires
14042 // 15 instructions: convert to single, rcpss, convert back to double, refine
14043 // (3 steps = 12 insts). If an 'rcpsd' variant was added to the ISA
14044 // along with FMA, this could be a throughput win.
14045 if (VT == MVT::f32 && Subtarget->hasSSE1())
14047 else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
14048 (VT == MVT::v8f32 && Subtarget->hasAVX()))
14049 RecipOp = "vec-divf";
14053 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
14054 if (!Recips.isEnabled(RecipOp))
14057 RefinementSteps = Recips.getRefinementSteps(RecipOp);
14058 return DCI.DAG.getNode(X86ISD::FRCP, SDLoc(Op), VT, Op);
14061 /// If we have at least two divisions that use the same divisor, convert to
14062 /// multplication by a reciprocal. This may need to be adjusted for a given
14063 /// CPU if a division's cost is not at least twice the cost of a multiplication.
14064 /// This is because we still need one division to calculate the reciprocal and
14065 /// then we need two multiplies by that reciprocal as replacements for the
14066 /// original divisions.
14067 unsigned X86TargetLowering::combineRepeatedFPDivisors() const {
14071 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
14072 /// if it's possible.
14073 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
14074 SDLoc dl, SelectionDAG &DAG) const {
14075 SDValue Op0 = And.getOperand(0);
14076 SDValue Op1 = And.getOperand(1);
14077 if (Op0.getOpcode() == ISD::TRUNCATE)
14078 Op0 = Op0.getOperand(0);
14079 if (Op1.getOpcode() == ISD::TRUNCATE)
14080 Op1 = Op1.getOperand(0);
14083 if (Op1.getOpcode() == ISD::SHL)
14084 std::swap(Op0, Op1);
14085 if (Op0.getOpcode() == ISD::SHL) {
14086 if (isOneConstant(Op0.getOperand(0))) {
14087 // If we looked past a truncate, check that it's only truncating away
14089 unsigned BitWidth = Op0.getValueSizeInBits();
14090 unsigned AndBitWidth = And.getValueSizeInBits();
14091 if (BitWidth > AndBitWidth) {
14093 DAG.computeKnownBits(Op0, Zeros, Ones);
14094 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
14098 RHS = Op0.getOperand(1);
14100 } else if (Op1.getOpcode() == ISD::Constant) {
14101 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
14102 uint64_t AndRHSVal = AndRHS->getZExtValue();
14103 SDValue AndLHS = Op0;
14105 if (AndRHSVal == 1 && AndLHS.getOpcode() == ISD::SRL) {
14106 LHS = AndLHS.getOperand(0);
14107 RHS = AndLHS.getOperand(1);
14110 // Use BT if the immediate can't be encoded in a TEST instruction.
14111 if (!isUInt<32>(AndRHSVal) && isPowerOf2_64(AndRHSVal)) {
14113 RHS = DAG.getConstant(Log2_64_Ceil(AndRHSVal), dl, LHS.getValueType());
14117 if (LHS.getNode()) {
14118 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
14119 // instruction. Since the shift amount is in-range-or-undefined, we know
14120 // that doing a bittest on the i32 value is ok. We extend to i32 because
14121 // the encoding for the i16 version is larger than the i32 version.
14122 // Also promote i16 to i32 for performance / code size reason.
14123 if (LHS.getValueType() == MVT::i8 ||
14124 LHS.getValueType() == MVT::i16)
14125 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
14127 // If the operand types disagree, extend the shift amount to match. Since
14128 // BT ignores high bits (like shifts) we can use anyextend.
14129 if (LHS.getValueType() != RHS.getValueType())
14130 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
14132 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
14133 X86::CondCode Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
14134 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14135 DAG.getConstant(Cond, dl, MVT::i8), BT);
14141 /// \brief - Turns an ISD::CondCode into a value suitable for SSE floating point
14143 static int translateX86FSETCC(ISD::CondCode SetCCOpcode, SDValue &Op0,
14148 // SSE Condition code mapping:
14157 switch (SetCCOpcode) {
14158 default: llvm_unreachable("Unexpected SETCC condition");
14160 case ISD::SETEQ: SSECC = 0; break;
14162 case ISD::SETGT: Swap = true; // Fallthrough
14164 case ISD::SETOLT: SSECC = 1; break;
14166 case ISD::SETGE: Swap = true; // Fallthrough
14168 case ISD::SETOLE: SSECC = 2; break;
14169 case ISD::SETUO: SSECC = 3; break;
14171 case ISD::SETNE: SSECC = 4; break;
14172 case ISD::SETULE: Swap = true; // Fallthrough
14173 case ISD::SETUGE: SSECC = 5; break;
14174 case ISD::SETULT: Swap = true; // Fallthrough
14175 case ISD::SETUGT: SSECC = 6; break;
14176 case ISD::SETO: SSECC = 7; break;
14178 case ISD::SETONE: SSECC = 8; break;
14181 std::swap(Op0, Op1);
14186 // Lower256IntVSETCC - Break a VSETCC 256-bit integer VSETCC into two new 128
14187 // ones, and then concatenate the result back.
14188 static SDValue Lower256IntVSETCC(SDValue Op, SelectionDAG &DAG) {
14189 MVT VT = Op.getSimpleValueType();
14191 assert(VT.is256BitVector() && Op.getOpcode() == ISD::SETCC &&
14192 "Unsupported value type for operation");
14194 unsigned NumElems = VT.getVectorNumElements();
14196 SDValue CC = Op.getOperand(2);
14198 // Extract the LHS vectors
14199 SDValue LHS = Op.getOperand(0);
14200 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
14201 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
14203 // Extract the RHS vectors
14204 SDValue RHS = Op.getOperand(1);
14205 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
14206 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
14208 // Issue the operation on the smaller types and concatenate the result back
14209 MVT EltVT = VT.getVectorElementType();
14210 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
14211 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
14212 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC),
14213 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
14216 static SDValue LowerBoolVSETCC_AVX512(SDValue Op, SelectionDAG &DAG) {
14217 SDValue Op0 = Op.getOperand(0);
14218 SDValue Op1 = Op.getOperand(1);
14219 SDValue CC = Op.getOperand(2);
14220 MVT VT = Op.getSimpleValueType();
14223 assert(Op0.getSimpleValueType().getVectorElementType() == MVT::i1 &&
14224 "Unexpected type for boolean compare operation");
14225 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14226 SDValue NotOp0 = DAG.getNode(ISD::XOR, dl, VT, Op0,
14227 DAG.getConstant(-1, dl, VT));
14228 SDValue NotOp1 = DAG.getNode(ISD::XOR, dl, VT, Op1,
14229 DAG.getConstant(-1, dl, VT));
14230 switch (SetCCOpcode) {
14231 default: llvm_unreachable("Unexpected SETCC condition");
14233 // (x == y) -> ~(x ^ y)
14234 return DAG.getNode(ISD::XOR, dl, VT,
14235 DAG.getNode(ISD::XOR, dl, VT, Op0, Op1),
14236 DAG.getConstant(-1, dl, VT));
14238 // (x != y) -> (x ^ y)
14239 return DAG.getNode(ISD::XOR, dl, VT, Op0, Op1);
14242 // (x > y) -> (x & ~y)
14243 return DAG.getNode(ISD::AND, dl, VT, Op0, NotOp1);
14246 // (x < y) -> (~x & y)
14247 return DAG.getNode(ISD::AND, dl, VT, NotOp0, Op1);
14250 // (x <= y) -> (~x | y)
14251 return DAG.getNode(ISD::OR, dl, VT, NotOp0, Op1);
14254 // (x >=y) -> (x | ~y)
14255 return DAG.getNode(ISD::OR, dl, VT, Op0, NotOp1);
14259 static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG,
14260 const X86Subtarget *Subtarget) {
14261 SDValue Op0 = Op.getOperand(0);
14262 SDValue Op1 = Op.getOperand(1);
14263 SDValue CC = Op.getOperand(2);
14264 MVT VT = Op.getSimpleValueType();
14267 assert(Op0.getSimpleValueType().getVectorElementType().getSizeInBits() >= 8 &&
14268 Op.getSimpleValueType().getVectorElementType() == MVT::i1 &&
14269 "Cannot set masked compare for this operation");
14271 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14273 bool Unsigned = false;
14276 switch (SetCCOpcode) {
14277 default: llvm_unreachable("Unexpected SETCC condition");
14278 case ISD::SETNE: SSECC = 4; break;
14279 case ISD::SETEQ: Opc = X86ISD::PCMPEQM; break;
14280 case ISD::SETUGT: SSECC = 6; Unsigned = true; break;
14281 case ISD::SETLT: Swap = true; //fall-through
14282 case ISD::SETGT: Opc = X86ISD::PCMPGTM; break;
14283 case ISD::SETULT: SSECC = 1; Unsigned = true; break;
14284 case ISD::SETUGE: SSECC = 5; Unsigned = true; break; //NLT
14285 case ISD::SETGE: Swap = true; SSECC = 2; break; // LE + swap
14286 case ISD::SETULE: Unsigned = true; //fall-through
14287 case ISD::SETLE: SSECC = 2; break;
14291 std::swap(Op0, Op1);
14293 return DAG.getNode(Opc, dl, VT, Op0, Op1);
14294 Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM;
14295 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14296 DAG.getConstant(SSECC, dl, MVT::i8));
14299 /// \brief Try to turn a VSETULT into a VSETULE by modifying its second
14300 /// operand \p Op1. If non-trivial (for example because it's not constant)
14301 /// return an empty value.
14302 static SDValue ChangeVSETULTtoVSETULE(SDLoc dl, SDValue Op1, SelectionDAG &DAG)
14304 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Op1.getNode());
14308 MVT VT = Op1.getSimpleValueType();
14309 MVT EVT = VT.getVectorElementType();
14310 unsigned n = VT.getVectorNumElements();
14311 SmallVector<SDValue, 8> ULTOp1;
14313 for (unsigned i = 0; i < n; ++i) {
14314 ConstantSDNode *Elt = dyn_cast<ConstantSDNode>(BV->getOperand(i));
14315 if (!Elt || Elt->isOpaque() || Elt->getSimpleValueType(0) != EVT)
14318 // Avoid underflow.
14319 APInt Val = Elt->getAPIntValue();
14323 ULTOp1.push_back(DAG.getConstant(Val - 1, dl, EVT));
14326 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, ULTOp1);
14329 static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
14330 SelectionDAG &DAG) {
14331 SDValue Op0 = Op.getOperand(0);
14332 SDValue Op1 = Op.getOperand(1);
14333 SDValue CC = Op.getOperand(2);
14334 MVT VT = Op.getSimpleValueType();
14335 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14336 bool isFP = Op.getOperand(1).getSimpleValueType().isFloatingPoint();
14341 MVT EltVT = Op0.getSimpleValueType().getVectorElementType();
14342 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
14345 unsigned SSECC = translateX86FSETCC(SetCCOpcode, Op0, Op1);
14346 unsigned Opc = X86ISD::CMPP;
14347 if (Subtarget->hasAVX512() && VT.getVectorElementType() == MVT::i1) {
14348 assert(VT.getVectorNumElements() <= 16);
14349 Opc = X86ISD::CMPM;
14351 // In the two special cases we can't handle, emit two comparisons.
14354 unsigned CombineOpc;
14355 if (SetCCOpcode == ISD::SETUEQ) {
14356 CC0 = 3; CC1 = 0; CombineOpc = ISD::OR;
14358 assert(SetCCOpcode == ISD::SETONE);
14359 CC0 = 7; CC1 = 4; CombineOpc = ISD::AND;
14362 SDValue Cmp0 = DAG.getNode(Opc, dl, VT, Op0, Op1,
14363 DAG.getConstant(CC0, dl, MVT::i8));
14364 SDValue Cmp1 = DAG.getNode(Opc, dl, VT, Op0, Op1,
14365 DAG.getConstant(CC1, dl, MVT::i8));
14366 return DAG.getNode(CombineOpc, dl, VT, Cmp0, Cmp1);
14368 // Handle all other FP comparisons here.
14369 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14370 DAG.getConstant(SSECC, dl, MVT::i8));
14373 MVT VTOp0 = Op0.getSimpleValueType();
14374 assert(VTOp0 == Op1.getSimpleValueType() &&
14375 "Expected operands with same type!");
14376 assert(VT.getVectorNumElements() == VTOp0.getVectorNumElements() &&
14377 "Invalid number of packed elements for source and destination!");
14379 if (VT.is128BitVector() && VTOp0.is256BitVector()) {
14380 // On non-AVX512 targets, a vector of MVT::i1 is promoted by the type
14381 // legalizer to a wider vector type. In the case of 'vsetcc' nodes, the
14382 // legalizer firstly checks if the first operand in input to the setcc has
14383 // a legal type. If so, then it promotes the return type to that same type.
14384 // Otherwise, the return type is promoted to the 'next legal type' which,
14385 // for a vector of MVT::i1 is always a 128-bit integer vector type.
14387 // We reach this code only if the following two conditions are met:
14388 // 1. Both return type and operand type have been promoted to wider types
14389 // by the type legalizer.
14390 // 2. The original operand type has been promoted to a 256-bit vector.
14392 // Note that condition 2. only applies for AVX targets.
14393 SDValue NewOp = DAG.getSetCC(dl, VTOp0, Op0, Op1, SetCCOpcode);
14394 return DAG.getZExtOrTrunc(NewOp, dl, VT);
14397 // The non-AVX512 code below works under the assumption that source and
14398 // destination types are the same.
14399 assert((Subtarget->hasAVX512() || (VT == VTOp0)) &&
14400 "Value types for source and destination must be the same!");
14402 // Break 256-bit integer vector compare into smaller ones.
14403 if (VT.is256BitVector() && !Subtarget->hasInt256())
14404 return Lower256IntVSETCC(Op, DAG);
14406 MVT OpVT = Op1.getSimpleValueType();
14407 if (OpVT.getVectorElementType() == MVT::i1)
14408 return LowerBoolVSETCC_AVX512(Op, DAG);
14410 bool MaskResult = (VT.getVectorElementType() == MVT::i1);
14411 if (Subtarget->hasAVX512()) {
14412 if (Op1.getSimpleValueType().is512BitVector() ||
14413 (Subtarget->hasBWI() && Subtarget->hasVLX()) ||
14414 (MaskResult && OpVT.getVectorElementType().getSizeInBits() >= 32))
14415 return LowerIntVSETCC_AVX512(Op, DAG, Subtarget);
14417 // In AVX-512 architecture setcc returns mask with i1 elements,
14418 // But there is no compare instruction for i8 and i16 elements in KNL.
14419 // We are not talking about 512-bit operands in this case, these
14420 // types are illegal.
14422 (OpVT.getVectorElementType().getSizeInBits() < 32 &&
14423 OpVT.getVectorElementType().getSizeInBits() >= 8))
14424 return DAG.getNode(ISD::TRUNCATE, dl, VT,
14425 DAG.getNode(ISD::SETCC, dl, OpVT, Op0, Op1, CC));
14428 // Lower using XOP integer comparisons.
14429 if ((VT == MVT::v16i8 || VT == MVT::v8i16 ||
14430 VT == MVT::v4i32 || VT == MVT::v2i64) && Subtarget->hasXOP()) {
14431 // Translate compare code to XOP PCOM compare mode.
14432 unsigned CmpMode = 0;
14433 switch (SetCCOpcode) {
14434 default: llvm_unreachable("Unexpected SETCC condition");
14436 case ISD::SETLT: CmpMode = 0x00; break;
14438 case ISD::SETLE: CmpMode = 0x01; break;
14440 case ISD::SETGT: CmpMode = 0x02; break;
14442 case ISD::SETGE: CmpMode = 0x03; break;
14443 case ISD::SETEQ: CmpMode = 0x04; break;
14444 case ISD::SETNE: CmpMode = 0x05; break;
14447 // Are we comparing unsigned or signed integers?
14448 unsigned Opc = ISD::isUnsignedIntSetCC(SetCCOpcode)
14449 ? X86ISD::VPCOMU : X86ISD::VPCOM;
14451 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14452 DAG.getConstant(CmpMode, dl, MVT::i8));
14455 // We are handling one of the integer comparisons here. Since SSE only has
14456 // GT and EQ comparisons for integer, swapping operands and multiple
14457 // operations may be required for some comparisons.
14459 bool Swap = false, Invert = false, FlipSigns = false, MinMax = false;
14460 bool Subus = false;
14462 switch (SetCCOpcode) {
14463 default: llvm_unreachable("Unexpected SETCC condition");
14464 case ISD::SETNE: Invert = true;
14465 case ISD::SETEQ: Opc = X86ISD::PCMPEQ; break;
14466 case ISD::SETLT: Swap = true;
14467 case ISD::SETGT: Opc = X86ISD::PCMPGT; break;
14468 case ISD::SETGE: Swap = true;
14469 case ISD::SETLE: Opc = X86ISD::PCMPGT;
14470 Invert = true; break;
14471 case ISD::SETULT: Swap = true;
14472 case ISD::SETUGT: Opc = X86ISD::PCMPGT;
14473 FlipSigns = true; break;
14474 case ISD::SETUGE: Swap = true;
14475 case ISD::SETULE: Opc = X86ISD::PCMPGT;
14476 FlipSigns = true; Invert = true; break;
14479 // Special case: Use min/max operations for SETULE/SETUGE
14480 MVT VET = VT.getVectorElementType();
14482 (Subtarget->hasSSE41() && (VET >= MVT::i8 && VET <= MVT::i32))
14483 || (Subtarget->hasSSE2() && (VET == MVT::i8));
14486 switch (SetCCOpcode) {
14488 case ISD::SETULE: Opc = ISD::UMIN; MinMax = true; break;
14489 case ISD::SETUGE: Opc = ISD::UMAX; MinMax = true; break;
14492 if (MinMax) { Swap = false; Invert = false; FlipSigns = false; }
14495 bool hasSubus = Subtarget->hasSSE2() && (VET == MVT::i8 || VET == MVT::i16);
14496 if (!MinMax && hasSubus) {
14497 // As another special case, use PSUBUS[BW] when it's profitable. E.g. for
14499 // t = psubus Op0, Op1
14500 // pcmpeq t, <0..0>
14501 switch (SetCCOpcode) {
14503 case ISD::SETULT: {
14504 // If the comparison is against a constant we can turn this into a
14505 // setule. With psubus, setule does not require a swap. This is
14506 // beneficial because the constant in the register is no longer
14507 // destructed as the destination so it can be hoisted out of a loop.
14508 // Only do this pre-AVX since vpcmp* is no longer destructive.
14509 if (Subtarget->hasAVX())
14511 SDValue ULEOp1 = ChangeVSETULTtoVSETULE(dl, Op1, DAG);
14512 if (ULEOp1.getNode()) {
14514 Subus = true; Invert = false; Swap = false;
14518 // Psubus is better than flip-sign because it requires no inversion.
14519 case ISD::SETUGE: Subus = true; Invert = false; Swap = true; break;
14520 case ISD::SETULE: Subus = true; Invert = false; Swap = false; break;
14524 Opc = X86ISD::SUBUS;
14530 std::swap(Op0, Op1);
14532 // Check that the operation in question is available (most are plain SSE2,
14533 // but PCMPGTQ and PCMPEQQ have different requirements).
14534 if (VT == MVT::v2i64) {
14535 if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42()) {
14536 assert(Subtarget->hasSSE2() && "Don't know how to lower!");
14538 // First cast everything to the right type.
14539 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
14540 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
14542 // Since SSE has no unsigned integer comparisons, we need to flip the sign
14543 // bits of the inputs before performing those operations. The lower
14544 // compare is always unsigned.
14547 SB = DAG.getConstant(0x80000000U, dl, MVT::v4i32);
14549 SDValue Sign = DAG.getConstant(0x80000000U, dl, MVT::i32);
14550 SDValue Zero = DAG.getConstant(0x00000000U, dl, MVT::i32);
14551 SB = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
14552 Sign, Zero, Sign, Zero);
14554 Op0 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op0, SB);
14555 Op1 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op1, SB);
14557 // Emulate PCMPGTQ with (hi1 > hi2) | ((hi1 == hi2) & (lo1 > lo2))
14558 SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1);
14559 SDValue EQ = DAG.getNode(X86ISD::PCMPEQ, dl, MVT::v4i32, Op0, Op1);
14561 // Create masks for only the low parts/high parts of the 64 bit integers.
14562 static const int MaskHi[] = { 1, 1, 3, 3 };
14563 static const int MaskLo[] = { 0, 0, 2, 2 };
14564 SDValue EQHi = DAG.getVectorShuffle(MVT::v4i32, dl, EQ, EQ, MaskHi);
14565 SDValue GTLo = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskLo);
14566 SDValue GTHi = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi);
14568 SDValue Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, EQHi, GTLo);
14569 Result = DAG.getNode(ISD::OR, dl, MVT::v4i32, Result, GTHi);
14572 Result = DAG.getNOT(dl, Result, MVT::v4i32);
14574 return DAG.getBitcast(VT, Result);
14577 if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) {
14578 // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with
14579 // pcmpeqd + pshufd + pand.
14580 assert(Subtarget->hasSSE2() && !FlipSigns && "Don't know how to lower!");
14582 // First cast everything to the right type.
14583 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
14584 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
14587 SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1);
14589 // Make sure the lower and upper halves are both all-ones.
14590 static const int Mask[] = { 1, 0, 3, 2 };
14591 SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask);
14592 Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf);
14595 Result = DAG.getNOT(dl, Result, MVT::v4i32);
14597 return DAG.getBitcast(VT, Result);
14601 // Since SSE has no unsigned integer comparisons, we need to flip the sign
14602 // bits of the inputs before performing those operations.
14604 MVT EltVT = VT.getVectorElementType();
14605 SDValue SB = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()), dl,
14607 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SB);
14608 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SB);
14611 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
14613 // If the logical-not of the result is required, perform that now.
14615 Result = DAG.getNOT(dl, Result, VT);
14618 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Op0, Result);
14621 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Result,
14622 getZeroVector(VT, Subtarget, DAG, dl));
14627 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
14629 MVT VT = Op.getSimpleValueType();
14631 if (VT.isVector()) return LowerVSETCC(Op, Subtarget, DAG);
14633 assert(((!Subtarget->hasAVX512() && VT == MVT::i8) || (VT == MVT::i1))
14634 && "SetCC type must be 8-bit or 1-bit integer");
14635 SDValue Op0 = Op.getOperand(0);
14636 SDValue Op1 = Op.getOperand(1);
14638 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
14640 // Optimize to BT if possible.
14641 // Lower (X & (1 << N)) == 0 to BT(X, N).
14642 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
14643 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
14644 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
14645 isNullConstant(Op1) &&
14646 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14647 if (SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG)) {
14649 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewSetCC);
14654 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
14656 if ((isOneConstant(Op1) || isNullConstant(Op1)) &&
14657 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14659 // If the input is a setcc, then reuse the input setcc or use a new one with
14660 // the inverted condition.
14661 if (Op0.getOpcode() == X86ISD::SETCC) {
14662 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
14663 bool Invert = (CC == ISD::SETNE) ^ isNullConstant(Op1);
14667 CCode = X86::GetOppositeBranchCondition(CCode);
14668 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14669 DAG.getConstant(CCode, dl, MVT::i8),
14670 Op0.getOperand(1));
14672 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
14676 if ((Op0.getValueType() == MVT::i1) && isOneConstant(Op1) &&
14677 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14679 ISD::CondCode NewCC = ISD::getSetCCInverse(CC, true);
14680 return DAG.getSetCC(dl, VT, Op0, DAG.getConstant(0, dl, MVT::i1), NewCC);
14683 bool isFP = Op1.getSimpleValueType().isFloatingPoint();
14684 unsigned X86CC = TranslateX86CC(CC, dl, isFP, Op0, Op1, DAG);
14685 if (X86CC == X86::COND_INVALID)
14688 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, dl, DAG);
14689 EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
14690 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14691 DAG.getConstant(X86CC, dl, MVT::i8), EFLAGS);
14693 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
14697 SDValue X86TargetLowering::LowerSETCCE(SDValue Op, SelectionDAG &DAG) const {
14698 SDValue LHS = Op.getOperand(0);
14699 SDValue RHS = Op.getOperand(1);
14700 SDValue Carry = Op.getOperand(2);
14701 SDValue Cond = Op.getOperand(3);
14704 assert(LHS.getSimpleValueType().isInteger() && "SETCCE is integer only.");
14705 X86::CondCode CC = TranslateIntegerX86CC(cast<CondCodeSDNode>(Cond)->get());
14707 assert(Carry.getOpcode() != ISD::CARRY_FALSE);
14708 SDVTList VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
14709 SDValue Cmp = DAG.getNode(X86ISD::SBB, DL, VTs, LHS, RHS, Carry);
14710 return DAG.getNode(X86ISD::SETCC, DL, Op.getValueType(),
14711 DAG.getConstant(CC, DL, MVT::i8), Cmp.getValue(1));
14714 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
14715 static bool isX86LogicalCmp(SDValue Op) {
14716 unsigned Opc = Op.getNode()->getOpcode();
14717 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI ||
14718 Opc == X86ISD::SAHF)
14720 if (Op.getResNo() == 1 &&
14721 (Opc == X86ISD::ADD ||
14722 Opc == X86ISD::SUB ||
14723 Opc == X86ISD::ADC ||
14724 Opc == X86ISD::SBB ||
14725 Opc == X86ISD::SMUL ||
14726 Opc == X86ISD::UMUL ||
14727 Opc == X86ISD::INC ||
14728 Opc == X86ISD::DEC ||
14729 Opc == X86ISD::OR ||
14730 Opc == X86ISD::XOR ||
14731 Opc == X86ISD::AND))
14734 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
14740 static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) {
14741 if (V.getOpcode() != ISD::TRUNCATE)
14744 SDValue VOp0 = V.getOperand(0);
14745 unsigned InBits = VOp0.getValueSizeInBits();
14746 unsigned Bits = V.getValueSizeInBits();
14747 return DAG.MaskedValueIsZero(VOp0, APInt::getHighBitsSet(InBits,InBits-Bits));
14750 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
14751 bool addTest = true;
14752 SDValue Cond = Op.getOperand(0);
14753 SDValue Op1 = Op.getOperand(1);
14754 SDValue Op2 = Op.getOperand(2);
14756 MVT VT = Op1.getSimpleValueType();
14759 // Lower FP selects into a CMP/AND/ANDN/OR sequence when the necessary SSE ops
14760 // are available or VBLENDV if AVX is available.
14761 // Otherwise FP cmovs get lowered into a less efficient branch sequence later.
14762 if (Cond.getOpcode() == ISD::SETCC &&
14763 ((Subtarget->hasSSE2() && (VT == MVT::f32 || VT == MVT::f64)) ||
14764 (Subtarget->hasSSE1() && VT == MVT::f32)) &&
14765 VT == Cond.getOperand(0).getSimpleValueType() && Cond->hasOneUse()) {
14766 SDValue CondOp0 = Cond.getOperand(0), CondOp1 = Cond.getOperand(1);
14767 int SSECC = translateX86FSETCC(
14768 cast<CondCodeSDNode>(Cond.getOperand(2))->get(), CondOp0, CondOp1);
14771 if (Subtarget->hasAVX512()) {
14772 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CondOp0, CondOp1,
14773 DAG.getConstant(SSECC, DL, MVT::i8));
14774 return DAG.getNode(X86ISD::SELECT, DL, VT, Cmp, Op1, Op2);
14777 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, VT, CondOp0, CondOp1,
14778 DAG.getConstant(SSECC, DL, MVT::i8));
14780 // If we have AVX, we can use a variable vector select (VBLENDV) instead
14781 // of 3 logic instructions for size savings and potentially speed.
14782 // Unfortunately, there is no scalar form of VBLENDV.
14784 // If either operand is a constant, don't try this. We can expect to
14785 // optimize away at least one of the logic instructions later in that
14786 // case, so that sequence would be faster than a variable blend.
14788 // BLENDV was introduced with SSE 4.1, but the 2 register form implicitly
14789 // uses XMM0 as the selection register. That may need just as many
14790 // instructions as the AND/ANDN/OR sequence due to register moves, so
14793 if (Subtarget->hasAVX() &&
14794 !isa<ConstantFPSDNode>(Op1) && !isa<ConstantFPSDNode>(Op2)) {
14796 // Convert to vectors, do a VSELECT, and convert back to scalar.
14797 // All of the conversions should be optimized away.
14799 MVT VecVT = VT == MVT::f32 ? MVT::v4f32 : MVT::v2f64;
14800 SDValue VOp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op1);
14801 SDValue VOp2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op2);
14802 SDValue VCmp = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Cmp);
14804 MVT VCmpVT = VT == MVT::f32 ? MVT::v4i32 : MVT::v2i64;
14805 VCmp = DAG.getBitcast(VCmpVT, VCmp);
14807 SDValue VSel = DAG.getNode(ISD::VSELECT, DL, VecVT, VCmp, VOp1, VOp2);
14809 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT,
14810 VSel, DAG.getIntPtrConstant(0, DL));
14812 SDValue AndN = DAG.getNode(X86ISD::FANDN, DL, VT, Cmp, Op2);
14813 SDValue And = DAG.getNode(X86ISD::FAND, DL, VT, Cmp, Op1);
14814 return DAG.getNode(X86ISD::FOR, DL, VT, AndN, And);
14818 if (VT.isVector() && VT.getVectorElementType() == MVT::i1) {
14820 if (ISD::isBuildVectorOfConstantSDNodes(Op1.getNode()))
14821 Op1Scalar = ConvertI1VectorToInteger(Op1, DAG);
14822 else if (Op1.getOpcode() == ISD::BITCAST && Op1.getOperand(0))
14823 Op1Scalar = Op1.getOperand(0);
14825 if (ISD::isBuildVectorOfConstantSDNodes(Op2.getNode()))
14826 Op2Scalar = ConvertI1VectorToInteger(Op2, DAG);
14827 else if (Op2.getOpcode() == ISD::BITCAST && Op2.getOperand(0))
14828 Op2Scalar = Op2.getOperand(0);
14829 if (Op1Scalar.getNode() && Op2Scalar.getNode()) {
14830 SDValue newSelect = DAG.getNode(ISD::SELECT, DL,
14831 Op1Scalar.getValueType(),
14832 Cond, Op1Scalar, Op2Scalar);
14833 if (newSelect.getValueSizeInBits() == VT.getSizeInBits())
14834 return DAG.getBitcast(VT, newSelect);
14835 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, newSelect);
14836 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, ExtVec,
14837 DAG.getIntPtrConstant(0, DL));
14841 if (VT == MVT::v4i1 || VT == MVT::v2i1) {
14842 SDValue zeroConst = DAG.getIntPtrConstant(0, DL);
14843 Op1 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
14844 DAG.getUNDEF(MVT::v8i1), Op1, zeroConst);
14845 Op2 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
14846 DAG.getUNDEF(MVT::v8i1), Op2, zeroConst);
14847 SDValue newSelect = DAG.getNode(ISD::SELECT, DL, MVT::v8i1,
14849 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, newSelect, zeroConst);
14852 if (Cond.getOpcode() == ISD::SETCC) {
14853 SDValue NewCond = LowerSETCC(Cond, DAG);
14854 if (NewCond.getNode())
14858 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
14859 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
14860 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
14861 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
14862 if (Cond.getOpcode() == X86ISD::SETCC &&
14863 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
14864 isNullConstant(Cond.getOperand(1).getOperand(1))) {
14865 SDValue Cmp = Cond.getOperand(1);
14867 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
14869 if ((isAllOnesConstant(Op1) || isAllOnesConstant(Op2)) &&
14870 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
14871 SDValue Y = isAllOnesConstant(Op2) ? Op1 : Op2;
14873 SDValue CmpOp0 = Cmp.getOperand(0);
14874 // Apply further optimizations for special cases
14875 // (select (x != 0), -1, 0) -> neg & sbb
14876 // (select (x == 0), 0, -1) -> neg & sbb
14877 if (isNullConstant(Y) &&
14878 (isAllOnesConstant(Op1) == (CondCode == X86::COND_NE))) {
14879 SDVTList VTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32);
14880 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, VTs,
14881 DAG.getConstant(0, DL,
14882 CmpOp0.getValueType()),
14884 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14885 DAG.getConstant(X86::COND_B, DL, MVT::i8),
14886 SDValue(Neg.getNode(), 1));
14890 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
14891 CmpOp0, DAG.getConstant(1, DL, CmpOp0.getValueType()));
14892 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
14894 SDValue Res = // Res = 0 or -1.
14895 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14896 DAG.getConstant(X86::COND_B, DL, MVT::i8), Cmp);
14898 if (isAllOnesConstant(Op1) != (CondCode == X86::COND_E))
14899 Res = DAG.getNOT(DL, Res, Res.getValueType());
14901 if (!isNullConstant(Op2))
14902 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
14907 // Look past (and (setcc_carry (cmp ...)), 1).
14908 if (Cond.getOpcode() == ISD::AND &&
14909 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY &&
14910 isOneConstant(Cond.getOperand(1)))
14911 Cond = Cond.getOperand(0);
14913 // If condition flag is set by a X86ISD::CMP, then use it as the condition
14914 // setting operand in place of the X86ISD::SETCC.
14915 unsigned CondOpcode = Cond.getOpcode();
14916 if (CondOpcode == X86ISD::SETCC ||
14917 CondOpcode == X86ISD::SETCC_CARRY) {
14918 CC = Cond.getOperand(0);
14920 SDValue Cmp = Cond.getOperand(1);
14921 unsigned Opc = Cmp.getOpcode();
14922 MVT VT = Op.getSimpleValueType();
14924 bool IllegalFPCMov = false;
14925 if (VT.isFloatingPoint() && !VT.isVector() &&
14926 !isScalarFPTypeInSSEReg(VT)) // FPStack?
14927 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
14929 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
14930 Opc == X86ISD::BT) { // FIXME
14934 } else if (CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
14935 CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
14936 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
14937 Cond.getOperand(0).getValueType() != MVT::i8)) {
14938 SDValue LHS = Cond.getOperand(0);
14939 SDValue RHS = Cond.getOperand(1);
14940 unsigned X86Opcode;
14943 switch (CondOpcode) {
14944 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
14945 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
14946 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
14947 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
14948 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
14949 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
14950 default: llvm_unreachable("unexpected overflowing operator");
14952 if (CondOpcode == ISD::UMULO)
14953 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
14956 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
14958 SDValue X86Op = DAG.getNode(X86Opcode, DL, VTs, LHS, RHS);
14960 if (CondOpcode == ISD::UMULO)
14961 Cond = X86Op.getValue(2);
14963 Cond = X86Op.getValue(1);
14965 CC = DAG.getConstant(X86Cond, DL, MVT::i8);
14970 // Look past the truncate if the high bits are known zero.
14971 if (isTruncWithZeroHighBitsInput(Cond, DAG))
14972 Cond = Cond.getOperand(0);
14974 // We know the result of AND is compared against zero. Try to match
14976 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
14977 if (SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG)) {
14978 CC = NewSetCC.getOperand(0);
14979 Cond = NewSetCC.getOperand(1);
14986 CC = DAG.getConstant(X86::COND_NE, DL, MVT::i8);
14987 Cond = EmitTest(Cond, X86::COND_NE, DL, DAG);
14990 // a < b ? -1 : 0 -> RES = ~setcc_carry
14991 // a < b ? 0 : -1 -> RES = setcc_carry
14992 // a >= b ? -1 : 0 -> RES = setcc_carry
14993 // a >= b ? 0 : -1 -> RES = ~setcc_carry
14994 if (Cond.getOpcode() == X86ISD::SUB) {
14995 Cond = ConvertCmpIfNecessary(Cond, DAG);
14996 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
14998 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
14999 (isAllOnesConstant(Op1) || isAllOnesConstant(Op2)) &&
15000 (isNullConstant(Op1) || isNullConstant(Op2))) {
15001 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
15002 DAG.getConstant(X86::COND_B, DL, MVT::i8),
15004 if (isAllOnesConstant(Op1) != (CondCode == X86::COND_B))
15005 return DAG.getNOT(DL, Res, Res.getValueType());
15010 // X86 doesn't have an i8 cmov. If both operands are the result of a truncate
15011 // widen the cmov and push the truncate through. This avoids introducing a new
15012 // branch during isel and doesn't add any extensions.
15013 if (Op.getValueType() == MVT::i8 &&
15014 Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) {
15015 SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0);
15016 if (T1.getValueType() == T2.getValueType() &&
15017 // Blacklist CopyFromReg to avoid partial register stalls.
15018 T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){
15019 SDVTList VTs = DAG.getVTList(T1.getValueType(), MVT::Glue);
15020 SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, VTs, T2, T1, CC, Cond);
15021 return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov);
15025 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
15026 // condition is true.
15027 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
15028 SDValue Ops[] = { Op2, Op1, CC, Cond };
15029 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops);
15032 static SDValue LowerSIGN_EXTEND_AVX512(SDValue Op,
15033 const X86Subtarget *Subtarget,
15034 SelectionDAG &DAG) {
15035 MVT VT = Op->getSimpleValueType(0);
15036 SDValue In = Op->getOperand(0);
15037 MVT InVT = In.getSimpleValueType();
15038 MVT VTElt = VT.getVectorElementType();
15039 MVT InVTElt = InVT.getVectorElementType();
15043 if ((InVTElt == MVT::i1) &&
15044 (((Subtarget->hasBWI() && Subtarget->hasVLX() &&
15045 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() <= 16)) ||
15047 ((Subtarget->hasBWI() && VT.is512BitVector() &&
15048 VTElt.getSizeInBits() <= 16)) ||
15050 ((Subtarget->hasDQI() && Subtarget->hasVLX() &&
15051 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() >= 32)) ||
15053 ((Subtarget->hasDQI() && VT.is512BitVector() &&
15054 VTElt.getSizeInBits() >= 32))))
15055 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
15057 unsigned int NumElts = VT.getVectorNumElements();
15059 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
15062 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1) {
15063 if (In.getOpcode() == X86ISD::VSEXT || In.getOpcode() == X86ISD::VZEXT)
15064 return DAG.getNode(In.getOpcode(), dl, VT, In.getOperand(0));
15065 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
15068 assert (InVT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
15069 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
15071 DAG.getConstant(APInt::getAllOnesValue(ExtVT.getScalarSizeInBits()), dl,
15074 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), dl, ExtVT);
15076 SDValue V = DAG.getNode(ISD::VSELECT, dl, ExtVT, In, NegOne, Zero);
15077 if (VT.is512BitVector())
15079 return DAG.getNode(X86ISD::VTRUNC, dl, VT, V);
15082 static SDValue LowerSIGN_EXTEND_VECTOR_INREG(SDValue Op,
15083 const X86Subtarget *Subtarget,
15084 SelectionDAG &DAG) {
15085 SDValue In = Op->getOperand(0);
15086 MVT VT = Op->getSimpleValueType(0);
15087 MVT InVT = In.getSimpleValueType();
15088 assert(VT.getSizeInBits() == InVT.getSizeInBits());
15090 MVT InSVT = InVT.getVectorElementType();
15091 assert(VT.getVectorElementType().getSizeInBits() > InSVT.getSizeInBits());
15093 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
15095 if (InSVT != MVT::i32 && InSVT != MVT::i16 && InSVT != MVT::i8)
15100 // SSE41 targets can use the pmovsx* instructions directly.
15101 if (Subtarget->hasSSE41())
15102 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
15104 // pre-SSE41 targets unpack lower lanes and then sign-extend using SRAI.
15108 // As SRAI is only available on i16/i32 types, we expand only up to i32
15109 // and handle i64 separately.
15110 while (CurrVT != VT && CurrVT.getVectorElementType() != MVT::i32) {
15111 Curr = DAG.getNode(X86ISD::UNPCKL, dl, CurrVT, DAG.getUNDEF(CurrVT), Curr);
15112 MVT CurrSVT = MVT::getIntegerVT(CurrVT.getScalarSizeInBits() * 2);
15113 CurrVT = MVT::getVectorVT(CurrSVT, CurrVT.getVectorNumElements() / 2);
15114 Curr = DAG.getBitcast(CurrVT, Curr);
15117 SDValue SignExt = Curr;
15118 if (CurrVT != InVT) {
15119 unsigned SignExtShift =
15120 CurrVT.getVectorElementType().getSizeInBits() - InSVT.getSizeInBits();
15121 SignExt = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
15122 DAG.getConstant(SignExtShift, dl, MVT::i8));
15128 if (VT == MVT::v2i64 && CurrVT == MVT::v4i32) {
15129 SDValue Sign = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
15130 DAG.getConstant(31, dl, MVT::i8));
15131 SDValue Ext = DAG.getVectorShuffle(CurrVT, dl, SignExt, Sign, {0, 4, 1, 5});
15132 return DAG.getBitcast(VT, Ext);
15138 static SDValue LowerSIGN_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
15139 SelectionDAG &DAG) {
15140 MVT VT = Op->getSimpleValueType(0);
15141 SDValue In = Op->getOperand(0);
15142 MVT InVT = In.getSimpleValueType();
15145 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
15146 return LowerSIGN_EXTEND_AVX512(Op, Subtarget, DAG);
15148 if ((VT != MVT::v4i64 || InVT != MVT::v4i32) &&
15149 (VT != MVT::v8i32 || InVT != MVT::v8i16) &&
15150 (VT != MVT::v16i16 || InVT != MVT::v16i8))
15153 if (Subtarget->hasInt256())
15154 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
15156 // Optimize vectors in AVX mode
15157 // Sign extend v8i16 to v8i32 and
15160 // Divide input vector into two parts
15161 // for v4i32 the shuffle mask will be { 0, 1, -1, -1} {2, 3, -1, -1}
15162 // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32
15163 // concat the vectors to original VT
15165 unsigned NumElems = InVT.getVectorNumElements();
15166 SDValue Undef = DAG.getUNDEF(InVT);
15168 SmallVector<int,8> ShufMask1(NumElems, -1);
15169 for (unsigned i = 0; i != NumElems/2; ++i)
15172 SDValue OpLo = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask1[0]);
15174 SmallVector<int,8> ShufMask2(NumElems, -1);
15175 for (unsigned i = 0; i != NumElems/2; ++i)
15176 ShufMask2[i] = i + NumElems/2;
15178 SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask2[0]);
15180 MVT HalfVT = MVT::getVectorVT(VT.getVectorElementType(),
15181 VT.getVectorNumElements()/2);
15183 OpLo = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpLo);
15184 OpHi = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpHi);
15186 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
15189 // Lower vector extended loads using a shuffle. If SSSE3 is not available we
15190 // may emit an illegal shuffle but the expansion is still better than scalar
15191 // code. We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise
15192 // we'll emit a shuffle and a arithmetic shift.
15193 // FIXME: Is the expansion actually better than scalar code? It doesn't seem so.
15194 // TODO: It is possible to support ZExt by zeroing the undef values during
15195 // the shuffle phase or after the shuffle.
15196 static SDValue LowerExtendedLoad(SDValue Op, const X86Subtarget *Subtarget,
15197 SelectionDAG &DAG) {
15198 MVT RegVT = Op.getSimpleValueType();
15199 assert(RegVT.isVector() && "We only custom lower vector sext loads.");
15200 assert(RegVT.isInteger() &&
15201 "We only custom lower integer vector sext loads.");
15203 // Nothing useful we can do without SSE2 shuffles.
15204 assert(Subtarget->hasSSE2() && "We only custom lower sext loads with SSE2.");
15206 LoadSDNode *Ld = cast<LoadSDNode>(Op.getNode());
15208 EVT MemVT = Ld->getMemoryVT();
15209 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15210 unsigned RegSz = RegVT.getSizeInBits();
15212 ISD::LoadExtType Ext = Ld->getExtensionType();
15214 assert((Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)
15215 && "Only anyext and sext are currently implemented.");
15216 assert(MemVT != RegVT && "Cannot extend to the same type");
15217 assert(MemVT.isVector() && "Must load a vector from memory");
15219 unsigned NumElems = RegVT.getVectorNumElements();
15220 unsigned MemSz = MemVT.getSizeInBits();
15221 assert(RegSz > MemSz && "Register size must be greater than the mem size");
15223 if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256()) {
15224 // The only way in which we have a legal 256-bit vector result but not the
15225 // integer 256-bit operations needed to directly lower a sextload is if we
15226 // have AVX1 but not AVX2. In that case, we can always emit a sextload to
15227 // a 128-bit vector and a normal sign_extend to 256-bits that should get
15228 // correctly legalized. We do this late to allow the canonical form of
15229 // sextload to persist throughout the rest of the DAG combiner -- it wants
15230 // to fold together any extensions it can, and so will fuse a sign_extend
15231 // of an sextload into a sextload targeting a wider value.
15233 if (MemSz == 128) {
15234 // Just switch this to a normal load.
15235 assert(TLI.isTypeLegal(MemVT) && "If the memory type is a 128-bit type, "
15236 "it must be a legal 128-bit vector "
15238 Load = DAG.getLoad(MemVT, dl, Ld->getChain(), Ld->getBasePtr(),
15239 Ld->getPointerInfo(), Ld->isVolatile(), Ld->isNonTemporal(),
15240 Ld->isInvariant(), Ld->getAlignment());
15242 assert(MemSz < 128 &&
15243 "Can't extend a type wider than 128 bits to a 256 bit vector!");
15244 // Do an sext load to a 128-bit vector type. We want to use the same
15245 // number of elements, but elements half as wide. This will end up being
15246 // recursively lowered by this routine, but will succeed as we definitely
15247 // have all the necessary features if we're using AVX1.
15249 EVT::getIntegerVT(*DAG.getContext(), RegVT.getScalarSizeInBits() / 2);
15250 EVT HalfVecVT = EVT::getVectorVT(*DAG.getContext(), HalfEltVT, NumElems);
15252 DAG.getExtLoad(Ext, dl, HalfVecVT, Ld->getChain(), Ld->getBasePtr(),
15253 Ld->getPointerInfo(), MemVT, Ld->isVolatile(),
15254 Ld->isNonTemporal(), Ld->isInvariant(),
15255 Ld->getAlignment());
15258 // Replace chain users with the new chain.
15259 assert(Load->getNumValues() == 2 && "Loads must carry a chain!");
15260 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Load.getValue(1));
15262 // Finally, do a normal sign-extend to the desired register.
15263 return DAG.getSExtOrTrunc(Load, dl, RegVT);
15266 // All sizes must be a power of two.
15267 assert(isPowerOf2_32(RegSz * MemSz * NumElems) &&
15268 "Non-power-of-two elements are not custom lowered!");
15270 // Attempt to load the original value using scalar loads.
15271 // Find the largest scalar type that divides the total loaded size.
15272 MVT SclrLoadTy = MVT::i8;
15273 for (MVT Tp : MVT::integer_valuetypes()) {
15274 if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
15279 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
15280 if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
15282 SclrLoadTy = MVT::f64;
15284 // Calculate the number of scalar loads that we need to perform
15285 // in order to load our vector from memory.
15286 unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
15288 assert((Ext != ISD::SEXTLOAD || NumLoads == 1) &&
15289 "Can only lower sext loads with a single scalar load!");
15291 unsigned loadRegZize = RegSz;
15292 if (Ext == ISD::SEXTLOAD && RegSz >= 256)
15295 // Represent our vector as a sequence of elements which are the
15296 // largest scalar that we can load.
15297 EVT LoadUnitVecVT = EVT::getVectorVT(
15298 *DAG.getContext(), SclrLoadTy, loadRegZize / SclrLoadTy.getSizeInBits());
15300 // Represent the data using the same element type that is stored in
15301 // memory. In practice, we ''widen'' MemVT.
15303 EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
15304 loadRegZize / MemVT.getScalarSizeInBits());
15306 assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
15307 "Invalid vector type");
15309 // We can't shuffle using an illegal type.
15310 assert(TLI.isTypeLegal(WideVecVT) &&
15311 "We only lower types that form legal widened vector types");
15313 SmallVector<SDValue, 8> Chains;
15314 SDValue Ptr = Ld->getBasePtr();
15315 SDValue Increment = DAG.getConstant(SclrLoadTy.getSizeInBits() / 8, dl,
15316 TLI.getPointerTy(DAG.getDataLayout()));
15317 SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
15319 for (unsigned i = 0; i < NumLoads; ++i) {
15320 // Perform a single load.
15321 SDValue ScalarLoad =
15322 DAG.getLoad(SclrLoadTy, dl, Ld->getChain(), Ptr, Ld->getPointerInfo(),
15323 Ld->isVolatile(), Ld->isNonTemporal(), Ld->isInvariant(),
15324 Ld->getAlignment());
15325 Chains.push_back(ScalarLoad.getValue(1));
15326 // Create the first element type using SCALAR_TO_VECTOR in order to avoid
15327 // another round of DAGCombining.
15329 Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
15331 Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
15332 ScalarLoad, DAG.getIntPtrConstant(i, dl));
15334 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
15337 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
15339 // Bitcast the loaded value to a vector of the original element type, in
15340 // the size of the target vector type.
15341 SDValue SlicedVec = DAG.getBitcast(WideVecVT, Res);
15342 unsigned SizeRatio = RegSz / MemSz;
15344 if (Ext == ISD::SEXTLOAD) {
15345 // If we have SSE4.1, we can directly emit a VSEXT node.
15346 if (Subtarget->hasSSE41()) {
15347 SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
15348 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
15352 // Otherwise we'll use SIGN_EXTEND_VECTOR_INREG to sign extend the lowest
15354 assert(TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND_VECTOR_INREG, RegVT) &&
15355 "We can't implement a sext load without SIGN_EXTEND_VECTOR_INREG!");
15357 SDValue Shuff = DAG.getSignExtendVectorInReg(SlicedVec, dl, RegVT);
15358 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
15362 // Redistribute the loaded elements into the different locations.
15363 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
15364 for (unsigned i = 0; i != NumElems; ++i)
15365 ShuffleVec[i * SizeRatio] = i;
15367 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
15368 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
15370 // Bitcast to the requested type.
15371 Shuff = DAG.getBitcast(RegVT, Shuff);
15372 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
15376 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
15377 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
15378 // from the AND / OR.
15379 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
15380 Opc = Op.getOpcode();
15381 if (Opc != ISD::OR && Opc != ISD::AND)
15383 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
15384 Op.getOperand(0).hasOneUse() &&
15385 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
15386 Op.getOperand(1).hasOneUse());
15389 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
15390 // 1 and that the SETCC node has a single use.
15391 static bool isXor1OfSetCC(SDValue Op) {
15392 if (Op.getOpcode() != ISD::XOR)
15394 if (isOneConstant(Op.getOperand(1)))
15395 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
15396 Op.getOperand(0).hasOneUse();
15400 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
15401 bool addTest = true;
15402 SDValue Chain = Op.getOperand(0);
15403 SDValue Cond = Op.getOperand(1);
15404 SDValue Dest = Op.getOperand(2);
15407 bool Inverted = false;
15409 if (Cond.getOpcode() == ISD::SETCC) {
15410 // Check for setcc([su]{add,sub,mul}o == 0).
15411 if (cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETEQ &&
15412 isNullConstant(Cond.getOperand(1)) &&
15413 Cond.getOperand(0).getResNo() == 1 &&
15414 (Cond.getOperand(0).getOpcode() == ISD::SADDO ||
15415 Cond.getOperand(0).getOpcode() == ISD::UADDO ||
15416 Cond.getOperand(0).getOpcode() == ISD::SSUBO ||
15417 Cond.getOperand(0).getOpcode() == ISD::USUBO ||
15418 Cond.getOperand(0).getOpcode() == ISD::SMULO ||
15419 Cond.getOperand(0).getOpcode() == ISD::UMULO)) {
15421 Cond = Cond.getOperand(0);
15423 SDValue NewCond = LowerSETCC(Cond, DAG);
15424 if (NewCond.getNode())
15429 // FIXME: LowerXALUO doesn't handle these!!
15430 else if (Cond.getOpcode() == X86ISD::ADD ||
15431 Cond.getOpcode() == X86ISD::SUB ||
15432 Cond.getOpcode() == X86ISD::SMUL ||
15433 Cond.getOpcode() == X86ISD::UMUL)
15434 Cond = LowerXALUO(Cond, DAG);
15437 // Look pass (and (setcc_carry (cmp ...)), 1).
15438 if (Cond.getOpcode() == ISD::AND &&
15439 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY &&
15440 isOneConstant(Cond.getOperand(1)))
15441 Cond = Cond.getOperand(0);
15443 // If condition flag is set by a X86ISD::CMP, then use it as the condition
15444 // setting operand in place of the X86ISD::SETCC.
15445 unsigned CondOpcode = Cond.getOpcode();
15446 if (CondOpcode == X86ISD::SETCC ||
15447 CondOpcode == X86ISD::SETCC_CARRY) {
15448 CC = Cond.getOperand(0);
15450 SDValue Cmp = Cond.getOperand(1);
15451 unsigned Opc = Cmp.getOpcode();
15452 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
15453 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
15457 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
15461 // These can only come from an arithmetic instruction with overflow,
15462 // e.g. SADDO, UADDO.
15463 Cond = Cond.getNode()->getOperand(1);
15469 CondOpcode = Cond.getOpcode();
15470 if (CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
15471 CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
15472 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
15473 Cond.getOperand(0).getValueType() != MVT::i8)) {
15474 SDValue LHS = Cond.getOperand(0);
15475 SDValue RHS = Cond.getOperand(1);
15476 unsigned X86Opcode;
15479 // Keep this in sync with LowerXALUO, otherwise we might create redundant
15480 // instructions that can't be removed afterwards (i.e. X86ISD::ADD and
15482 switch (CondOpcode) {
15483 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
15485 if (isOneConstant(RHS)) {
15486 X86Opcode = X86ISD::INC; X86Cond = X86::COND_O;
15489 X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
15490 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
15492 if (isOneConstant(RHS)) {
15493 X86Opcode = X86ISD::DEC; X86Cond = X86::COND_O;
15496 X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
15497 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
15498 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
15499 default: llvm_unreachable("unexpected overflowing operator");
15502 X86Cond = X86::GetOppositeBranchCondition((X86::CondCode)X86Cond);
15503 if (CondOpcode == ISD::UMULO)
15504 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
15507 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
15509 SDValue X86Op = DAG.getNode(X86Opcode, dl, VTs, LHS, RHS);
15511 if (CondOpcode == ISD::UMULO)
15512 Cond = X86Op.getValue(2);
15514 Cond = X86Op.getValue(1);
15516 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
15520 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
15521 SDValue Cmp = Cond.getOperand(0).getOperand(1);
15522 if (CondOpc == ISD::OR) {
15523 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
15524 // two branches instead of an explicit OR instruction with a
15526 if (Cmp == Cond.getOperand(1).getOperand(1) &&
15527 isX86LogicalCmp(Cmp)) {
15528 CC = Cond.getOperand(0).getOperand(0);
15529 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15530 Chain, Dest, CC, Cmp);
15531 CC = Cond.getOperand(1).getOperand(0);
15535 } else { // ISD::AND
15536 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
15537 // two branches instead of an explicit AND instruction with a
15538 // separate test. However, we only do this if this block doesn't
15539 // have a fall-through edge, because this requires an explicit
15540 // jmp when the condition is false.
15541 if (Cmp == Cond.getOperand(1).getOperand(1) &&
15542 isX86LogicalCmp(Cmp) &&
15543 Op.getNode()->hasOneUse()) {
15544 X86::CondCode CCode =
15545 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
15546 CCode = X86::GetOppositeBranchCondition(CCode);
15547 CC = DAG.getConstant(CCode, dl, MVT::i8);
15548 SDNode *User = *Op.getNode()->use_begin();
15549 // Look for an unconditional branch following this conditional branch.
15550 // We need this because we need to reverse the successors in order
15551 // to implement FCMP_OEQ.
15552 if (User->getOpcode() == ISD::BR) {
15553 SDValue FalseBB = User->getOperand(1);
15555 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15556 assert(NewBR == User);
15560 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15561 Chain, Dest, CC, Cmp);
15562 X86::CondCode CCode =
15563 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
15564 CCode = X86::GetOppositeBranchCondition(CCode);
15565 CC = DAG.getConstant(CCode, dl, MVT::i8);
15571 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
15572 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
15573 // It should be transformed during dag combiner except when the condition
15574 // is set by a arithmetics with overflow node.
15575 X86::CondCode CCode =
15576 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
15577 CCode = X86::GetOppositeBranchCondition(CCode);
15578 CC = DAG.getConstant(CCode, dl, MVT::i8);
15579 Cond = Cond.getOperand(0).getOperand(1);
15581 } else if (Cond.getOpcode() == ISD::SETCC &&
15582 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETOEQ) {
15583 // For FCMP_OEQ, we can emit
15584 // two branches instead of an explicit AND instruction with a
15585 // separate test. However, we only do this if this block doesn't
15586 // have a fall-through edge, because this requires an explicit
15587 // jmp when the condition is false.
15588 if (Op.getNode()->hasOneUse()) {
15589 SDNode *User = *Op.getNode()->use_begin();
15590 // Look for an unconditional branch following this conditional branch.
15591 // We need this because we need to reverse the successors in order
15592 // to implement FCMP_OEQ.
15593 if (User->getOpcode() == ISD::BR) {
15594 SDValue FalseBB = User->getOperand(1);
15596 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15597 assert(NewBR == User);
15601 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
15602 Cond.getOperand(0), Cond.getOperand(1));
15603 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
15604 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
15605 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15606 Chain, Dest, CC, Cmp);
15607 CC = DAG.getConstant(X86::COND_P, dl, MVT::i8);
15612 } else if (Cond.getOpcode() == ISD::SETCC &&
15613 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETUNE) {
15614 // For FCMP_UNE, we can emit
15615 // two branches instead of an explicit AND instruction with a
15616 // separate test. However, we only do this if this block doesn't
15617 // have a fall-through edge, because this requires an explicit
15618 // jmp when the condition is false.
15619 if (Op.getNode()->hasOneUse()) {
15620 SDNode *User = *Op.getNode()->use_begin();
15621 // Look for an unconditional branch following this conditional branch.
15622 // We need this because we need to reverse the successors in order
15623 // to implement FCMP_UNE.
15624 if (User->getOpcode() == ISD::BR) {
15625 SDValue FalseBB = User->getOperand(1);
15627 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15628 assert(NewBR == User);
15631 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
15632 Cond.getOperand(0), Cond.getOperand(1));
15633 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
15634 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
15635 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15636 Chain, Dest, CC, Cmp);
15637 CC = DAG.getConstant(X86::COND_NP, dl, MVT::i8);
15647 // Look pass the truncate if the high bits are known zero.
15648 if (isTruncWithZeroHighBitsInput(Cond, DAG))
15649 Cond = Cond.getOperand(0);
15651 // We know the result of AND is compared against zero. Try to match
15653 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
15654 if (SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG)) {
15655 CC = NewSetCC.getOperand(0);
15656 Cond = NewSetCC.getOperand(1);
15663 X86::CondCode X86Cond = Inverted ? X86::COND_E : X86::COND_NE;
15664 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
15665 Cond = EmitTest(Cond, X86Cond, dl, DAG);
15667 Cond = ConvertCmpIfNecessary(Cond, DAG);
15668 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15669 Chain, Dest, CC, Cond);
15672 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
15673 // Calls to _alloca are needed to probe the stack when allocating more than 4k
15674 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
15675 // that the guard pages used by the OS virtual memory manager are allocated in
15676 // correct sequence.
15678 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
15679 SelectionDAG &DAG) const {
15680 MachineFunction &MF = DAG.getMachineFunction();
15681 bool SplitStack = MF.shouldSplitStack();
15682 bool Lower = (Subtarget->isOSWindows() && !Subtarget->isTargetMachO()) ||
15687 SDNode *Node = Op.getNode();
15688 SDValue Chain = Op.getOperand(0);
15689 SDValue Size = Op.getOperand(1);
15690 unsigned Align = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
15691 EVT VT = Node->getValueType(0);
15693 // Chain the dynamic stack allocation so that it doesn't modify the stack
15694 // pointer when other instructions are using the stack.
15695 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, dl, true), dl);
15697 bool Is64Bit = Subtarget->is64Bit();
15698 MVT SPTy = getPointerTy(DAG.getDataLayout());
15702 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15703 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
15704 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
15705 " not tell us which reg is the stack pointer!");
15706 EVT VT = Node->getValueType(0);
15707 SDValue Tmp3 = Node->getOperand(2);
15709 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
15710 Chain = SP.getValue(1);
15711 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
15712 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
15713 unsigned StackAlign = TFI.getStackAlignment();
15714 Result = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
15715 if (Align > StackAlign)
15716 Result = DAG.getNode(ISD::AND, dl, VT, Result,
15717 DAG.getConstant(-(uint64_t)Align, dl, VT));
15718 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Result); // Output chain
15719 } else if (SplitStack) {
15720 MachineRegisterInfo &MRI = MF.getRegInfo();
15723 // The 64 bit implementation of segmented stacks needs to clobber both r10
15724 // r11. This makes it impossible to use it along with nested parameters.
15725 const Function *F = MF.getFunction();
15727 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
15729 if (I->hasNestAttr())
15730 report_fatal_error("Cannot use segmented stacks with functions that "
15731 "have nested arguments.");
15734 const TargetRegisterClass *AddrRegClass = getRegClassFor(SPTy);
15735 unsigned Vreg = MRI.createVirtualRegister(AddrRegClass);
15736 Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size);
15737 Result = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain,
15738 DAG.getRegister(Vreg, SPTy));
15741 const unsigned Reg = (Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX);
15743 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
15744 Flag = Chain.getValue(1);
15745 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
15747 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
15749 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15750 unsigned SPReg = RegInfo->getStackRegister();
15751 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, SPTy);
15752 Chain = SP.getValue(1);
15755 SP = DAG.getNode(ISD::AND, dl, VT, SP.getValue(0),
15756 DAG.getConstant(-(uint64_t)Align, dl, VT));
15757 Chain = DAG.getCopyToReg(Chain, dl, SPReg, SP);
15763 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, dl, true),
15764 DAG.getIntPtrConstant(0, dl, true), SDValue(), dl);
15766 SDValue Ops[2] = {Result, Chain};
15767 return DAG.getMergeValues(Ops, dl);
15770 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
15771 MachineFunction &MF = DAG.getMachineFunction();
15772 auto PtrVT = getPointerTy(MF.getDataLayout());
15773 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
15775 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
15778 if (!Subtarget->is64Bit() ||
15779 Subtarget->isCallingConvWin64(MF.getFunction()->getCallingConv())) {
15780 // vastart just stores the address of the VarArgsFrameIndex slot into the
15781 // memory location argument.
15782 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
15783 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
15784 MachinePointerInfo(SV), false, false, 0);
15788 // gp_offset (0 - 6 * 8)
15789 // fp_offset (48 - 48 + 8 * 16)
15790 // overflow_arg_area (point to parameters coming in memory).
15792 SmallVector<SDValue, 8> MemOps;
15793 SDValue FIN = Op.getOperand(1);
15795 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
15796 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
15798 FIN, MachinePointerInfo(SV), false, false, 0);
15799 MemOps.push_back(Store);
15802 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL));
15803 Store = DAG.getStore(Op.getOperand(0), DL,
15804 DAG.getConstant(FuncInfo->getVarArgsFPOffset(), DL,
15806 FIN, MachinePointerInfo(SV, 4), false, false, 0);
15807 MemOps.push_back(Store);
15809 // Store ptr to overflow_arg_area
15810 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL));
15811 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
15812 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
15813 MachinePointerInfo(SV, 8),
15815 MemOps.push_back(Store);
15817 // Store ptr to reg_save_area.
15818 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(
15819 Subtarget->isTarget64BitLP64() ? 8 : 4, DL));
15820 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(), PtrVT);
15821 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN, MachinePointerInfo(
15822 SV, Subtarget->isTarget64BitLP64() ? 16 : 12), false, false, 0);
15823 MemOps.push_back(Store);
15824 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
15827 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
15828 assert(Subtarget->is64Bit() &&
15829 "LowerVAARG only handles 64-bit va_arg!");
15830 assert(Op.getNode()->getNumOperands() == 4);
15832 MachineFunction &MF = DAG.getMachineFunction();
15833 if (Subtarget->isCallingConvWin64(MF.getFunction()->getCallingConv()))
15834 // The Win64 ABI uses char* instead of a structure.
15835 return DAG.expandVAArg(Op.getNode());
15837 SDValue Chain = Op.getOperand(0);
15838 SDValue SrcPtr = Op.getOperand(1);
15839 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
15840 unsigned Align = Op.getConstantOperandVal(3);
15843 EVT ArgVT = Op.getNode()->getValueType(0);
15844 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
15845 uint32_t ArgSize = DAG.getDataLayout().getTypeAllocSize(ArgTy);
15848 // Decide which area this value should be read from.
15849 // TODO: Implement the AMD64 ABI in its entirety. This simple
15850 // selection mechanism works only for the basic types.
15851 if (ArgVT == MVT::f80) {
15852 llvm_unreachable("va_arg for f80 not yet implemented");
15853 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
15854 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
15855 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
15856 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
15858 llvm_unreachable("Unhandled argument type in LowerVAARG");
15861 if (ArgMode == 2) {
15862 // Sanity Check: Make sure using fp_offset makes sense.
15863 assert(!Subtarget->useSoftFloat() &&
15864 !(MF.getFunction()->hasFnAttribute(Attribute::NoImplicitFloat)) &&
15865 Subtarget->hasSSE1());
15868 // Insert VAARG_64 node into the DAG
15869 // VAARG_64 returns two values: Variable Argument Address, Chain
15870 SDValue InstOps[] = {Chain, SrcPtr, DAG.getConstant(ArgSize, dl, MVT::i32),
15871 DAG.getConstant(ArgMode, dl, MVT::i8),
15872 DAG.getConstant(Align, dl, MVT::i32)};
15873 SDVTList VTs = DAG.getVTList(getPointerTy(DAG.getDataLayout()), MVT::Other);
15874 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
15875 VTs, InstOps, MVT::i64,
15876 MachinePointerInfo(SV),
15878 /*Volatile=*/false,
15880 /*WriteMem=*/true);
15881 Chain = VAARG.getValue(1);
15883 // Load the next argument and return it
15884 return DAG.getLoad(ArgVT, dl,
15887 MachinePointerInfo(),
15888 false, false, false, 0);
15891 static SDValue LowerVACOPY(SDValue Op, const X86Subtarget *Subtarget,
15892 SelectionDAG &DAG) {
15893 // X86-64 va_list is a struct { i32, i32, i8*, i8* }, except on Windows,
15894 // where a va_list is still an i8*.
15895 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
15896 if (Subtarget->isCallingConvWin64(
15897 DAG.getMachineFunction().getFunction()->getCallingConv()))
15898 // Probably a Win64 va_copy.
15899 return DAG.expandVACopy(Op.getNode());
15901 SDValue Chain = Op.getOperand(0);
15902 SDValue DstPtr = Op.getOperand(1);
15903 SDValue SrcPtr = Op.getOperand(2);
15904 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
15905 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
15908 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
15909 DAG.getIntPtrConstant(24, DL), 8, /*isVolatile*/false,
15911 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
15914 // getTargetVShiftByConstNode - Handle vector element shifts where the shift
15915 // amount is a constant. Takes immediate version of shift as input.
15916 static SDValue getTargetVShiftByConstNode(unsigned Opc, SDLoc dl, MVT VT,
15917 SDValue SrcOp, uint64_t ShiftAmt,
15918 SelectionDAG &DAG) {
15919 MVT ElementType = VT.getVectorElementType();
15921 // Fold this packed shift into its first operand if ShiftAmt is 0.
15925 // Check for ShiftAmt >= element width
15926 if (ShiftAmt >= ElementType.getSizeInBits()) {
15927 if (Opc == X86ISD::VSRAI)
15928 ShiftAmt = ElementType.getSizeInBits() - 1;
15930 return DAG.getConstant(0, dl, VT);
15933 assert((Opc == X86ISD::VSHLI || Opc == X86ISD::VSRLI || Opc == X86ISD::VSRAI)
15934 && "Unknown target vector shift-by-constant node");
15936 // Fold this packed vector shift into a build vector if SrcOp is a
15937 // vector of Constants or UNDEFs, and SrcOp valuetype is the same as VT.
15938 if (VT == SrcOp.getSimpleValueType() &&
15939 ISD::isBuildVectorOfConstantSDNodes(SrcOp.getNode())) {
15940 SmallVector<SDValue, 8> Elts;
15941 unsigned NumElts = SrcOp->getNumOperands();
15942 ConstantSDNode *ND;
15945 default: llvm_unreachable(nullptr);
15946 case X86ISD::VSHLI:
15947 for (unsigned i=0; i!=NumElts; ++i) {
15948 SDValue CurrentOp = SrcOp->getOperand(i);
15949 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15950 Elts.push_back(CurrentOp);
15953 ND = cast<ConstantSDNode>(CurrentOp);
15954 const APInt &C = ND->getAPIntValue();
15955 Elts.push_back(DAG.getConstant(C.shl(ShiftAmt), dl, ElementType));
15958 case X86ISD::VSRLI:
15959 for (unsigned i=0; i!=NumElts; ++i) {
15960 SDValue CurrentOp = SrcOp->getOperand(i);
15961 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15962 Elts.push_back(CurrentOp);
15965 ND = cast<ConstantSDNode>(CurrentOp);
15966 const APInt &C = ND->getAPIntValue();
15967 Elts.push_back(DAG.getConstant(C.lshr(ShiftAmt), dl, ElementType));
15970 case X86ISD::VSRAI:
15971 for (unsigned i=0; i!=NumElts; ++i) {
15972 SDValue CurrentOp = SrcOp->getOperand(i);
15973 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15974 Elts.push_back(CurrentOp);
15977 ND = cast<ConstantSDNode>(CurrentOp);
15978 const APInt &C = ND->getAPIntValue();
15979 Elts.push_back(DAG.getConstant(C.ashr(ShiftAmt), dl, ElementType));
15984 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
15987 return DAG.getNode(Opc, dl, VT, SrcOp,
15988 DAG.getConstant(ShiftAmt, dl, MVT::i8));
15991 // getTargetVShiftNode - Handle vector element shifts where the shift amount
15992 // may or may not be a constant. Takes immediate version of shift as input.
15993 static SDValue getTargetVShiftNode(unsigned Opc, SDLoc dl, MVT VT,
15994 SDValue SrcOp, SDValue ShAmt,
15995 SelectionDAG &DAG) {
15996 MVT SVT = ShAmt.getSimpleValueType();
15997 assert((SVT == MVT::i32 || SVT == MVT::i64) && "Unexpected value type!");
15999 // Catch shift-by-constant.
16000 if (ConstantSDNode *CShAmt = dyn_cast<ConstantSDNode>(ShAmt))
16001 return getTargetVShiftByConstNode(Opc, dl, VT, SrcOp,
16002 CShAmt->getZExtValue(), DAG);
16004 // Change opcode to non-immediate version
16006 default: llvm_unreachable("Unknown target vector shift node");
16007 case X86ISD::VSHLI: Opc = X86ISD::VSHL; break;
16008 case X86ISD::VSRLI: Opc = X86ISD::VSRL; break;
16009 case X86ISD::VSRAI: Opc = X86ISD::VSRA; break;
16012 const X86Subtarget &Subtarget =
16013 static_cast<const X86Subtarget &>(DAG.getSubtarget());
16014 if (Subtarget.hasSSE41() && ShAmt.getOpcode() == ISD::ZERO_EXTEND &&
16015 ShAmt.getOperand(0).getSimpleValueType() == MVT::i16) {
16016 // Let the shuffle legalizer expand this shift amount node.
16017 SDValue Op0 = ShAmt.getOperand(0);
16018 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(Op0), MVT::v8i16, Op0);
16019 ShAmt = getShuffleVectorZeroOrUndef(Op0, 0, true, &Subtarget, DAG);
16021 // Need to build a vector containing shift amount.
16022 // SSE/AVX packed shifts only use the lower 64-bit of the shift count.
16023 SmallVector<SDValue, 4> ShOps;
16024 ShOps.push_back(ShAmt);
16025 if (SVT == MVT::i32) {
16026 ShOps.push_back(DAG.getConstant(0, dl, SVT));
16027 ShOps.push_back(DAG.getUNDEF(SVT));
16029 ShOps.push_back(DAG.getUNDEF(SVT));
16031 MVT BVT = SVT == MVT::i32 ? MVT::v4i32 : MVT::v2i64;
16032 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, BVT, ShOps);
16035 // The return type has to be a 128-bit type with the same element
16036 // type as the input type.
16037 MVT EltVT = VT.getVectorElementType();
16038 MVT ShVT = MVT::getVectorVT(EltVT, 128/EltVT.getSizeInBits());
16040 ShAmt = DAG.getBitcast(ShVT, ShAmt);
16041 return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt);
16044 /// \brief Return Mask with the necessary casting or extending
16045 /// for \p Mask according to \p MaskVT when lowering masking intrinsics
16046 static SDValue getMaskNode(SDValue Mask, MVT MaskVT,
16047 const X86Subtarget *Subtarget,
16048 SelectionDAG &DAG, SDLoc dl) {
16050 if (MaskVT.bitsGT(Mask.getSimpleValueType())) {
16051 // Mask should be extended
16052 Mask = DAG.getNode(ISD::ANY_EXTEND, dl,
16053 MVT::getIntegerVT(MaskVT.getSizeInBits()), Mask);
16056 if (Mask.getSimpleValueType() == MVT::i64 && Subtarget->is32Bit()) {
16057 if (MaskVT == MVT::v64i1) {
16058 assert(Subtarget->hasBWI() && "Expected AVX512BW target!");
16059 // In case 32bit mode, bitcast i64 is illegal, extend/split it.
16061 Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Mask,
16062 DAG.getConstant(0, dl, MVT::i32));
16063 Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Mask,
16064 DAG.getConstant(1, dl, MVT::i32));
16066 Lo = DAG.getBitcast(MVT::v32i1, Lo);
16067 Hi = DAG.getBitcast(MVT::v32i1, Hi);
16069 return DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v64i1, Lo, Hi);
16071 // MaskVT require < 64bit. Truncate mask (should succeed in any case),
16073 MVT TruncVT = MVT::getIntegerVT(MaskVT.getSizeInBits());
16074 return DAG.getBitcast(MaskVT,
16075 DAG.getNode(ISD::TRUNCATE, dl, TruncVT, Mask));
16079 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
16080 Mask.getSimpleValueType().getSizeInBits());
16081 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16082 // are extracted by EXTRACT_SUBVECTOR.
16083 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16084 DAG.getBitcast(BitcastVT, Mask),
16085 DAG.getIntPtrConstant(0, dl));
16089 /// \brief Return (and \p Op, \p Mask) for compare instructions or
16090 /// (vselect \p Mask, \p Op, \p PreservedSrc) for others along with the
16091 /// necessary casting or extending for \p Mask when lowering masking intrinsics
16092 static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask,
16093 SDValue PreservedSrc,
16094 const X86Subtarget *Subtarget,
16095 SelectionDAG &DAG) {
16096 MVT VT = Op.getSimpleValueType();
16097 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
16098 unsigned OpcodeSelect = ISD::VSELECT;
16101 if (isAllOnesConstant(Mask))
16104 SDValue VMask = getMaskNode(Mask, MaskVT, Subtarget, DAG, dl);
16106 switch (Op.getOpcode()) {
16108 case X86ISD::PCMPEQM:
16109 case X86ISD::PCMPGTM:
16111 case X86ISD::CMPMU:
16112 return DAG.getNode(ISD::AND, dl, VT, Op, VMask);
16113 case X86ISD::VFPCLASS:
16114 case X86ISD::VFPCLASSS:
16115 return DAG.getNode(ISD::OR, dl, VT, Op, VMask);
16116 case X86ISD::VTRUNC:
16117 case X86ISD::VTRUNCS:
16118 case X86ISD::VTRUNCUS:
16119 // We can't use ISD::VSELECT here because it is not always "Legal"
16120 // for the destination type. For example vpmovqb require only AVX512
16121 // and vselect that can operate on byte element type require BWI
16122 OpcodeSelect = X86ISD::SELECT;
16125 if (PreservedSrc.getOpcode() == ISD::UNDEF)
16126 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
16127 return DAG.getNode(OpcodeSelect, dl, VT, VMask, Op, PreservedSrc);
16130 /// \brief Creates an SDNode for a predicated scalar operation.
16131 /// \returns (X86vselect \p Mask, \p Op, \p PreservedSrc).
16132 /// The mask is coming as MVT::i8 and it should be truncated
16133 /// to MVT::i1 while lowering masking intrinsics.
16134 /// The main difference between ScalarMaskingNode and VectorMaskingNode is using
16135 /// "X86select" instead of "vselect". We just can't create the "vselect" node
16136 /// for a scalar instruction.
16137 static SDValue getScalarMaskingNode(SDValue Op, SDValue Mask,
16138 SDValue PreservedSrc,
16139 const X86Subtarget *Subtarget,
16140 SelectionDAG &DAG) {
16141 if (isAllOnesConstant(Mask))
16144 MVT VT = Op.getSimpleValueType();
16146 // The mask should be of type MVT::i1
16147 SDValue IMask = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Mask);
16149 if (Op.getOpcode() == X86ISD::FSETCC)
16150 return DAG.getNode(ISD::AND, dl, VT, Op, IMask);
16151 if (Op.getOpcode() == X86ISD::VFPCLASS ||
16152 Op.getOpcode() == X86ISD::VFPCLASSS)
16153 return DAG.getNode(ISD::OR, dl, VT, Op, IMask);
16155 if (PreservedSrc.getOpcode() == ISD::UNDEF)
16156 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
16157 return DAG.getNode(X86ISD::SELECT, dl, VT, IMask, Op, PreservedSrc);
16160 static int getSEHRegistrationNodeSize(const Function *Fn) {
16161 if (!Fn->hasPersonalityFn())
16162 report_fatal_error(
16163 "querying registration node size for function without personality");
16164 // The RegNodeSize is 6 32-bit words for SEH and 4 for C++ EH. See
16165 // WinEHStatePass for the full struct definition.
16166 switch (classifyEHPersonality(Fn->getPersonalityFn())) {
16167 case EHPersonality::MSVC_X86SEH: return 24;
16168 case EHPersonality::MSVC_CXX: return 16;
16171 report_fatal_error("can only recover FP for MSVC EH personality functions");
16174 /// When the 32-bit MSVC runtime transfers control to us, either to an outlined
16175 /// function or when returning to a parent frame after catching an exception, we
16176 /// recover the parent frame pointer by doing arithmetic on the incoming EBP.
16177 /// Here's the math:
16178 /// RegNodeBase = EntryEBP - RegNodeSize
16179 /// ParentFP = RegNodeBase - RegNodeFrameOffset
16180 /// Subtracting RegNodeSize takes us to the offset of the registration node, and
16181 /// subtracting the offset (negative on x86) takes us back to the parent FP.
16182 static SDValue recoverFramePointer(SelectionDAG &DAG, const Function *Fn,
16183 SDValue EntryEBP) {
16184 MachineFunction &MF = DAG.getMachineFunction();
16187 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16188 MVT PtrVT = TLI.getPointerTy(DAG.getDataLayout());
16190 // It's possible that the parent function no longer has a personality function
16191 // if the exceptional code was optimized away, in which case we just return
16192 // the incoming EBP.
16193 if (!Fn->hasPersonalityFn())
16196 int RegNodeSize = getSEHRegistrationNodeSize(Fn);
16198 // Get an MCSymbol that will ultimately resolve to the frame offset of the EH
16200 MCSymbol *OffsetSym =
16201 MF.getMMI().getContext().getOrCreateParentFrameOffsetSymbol(
16202 GlobalValue::getRealLinkageName(Fn->getName()));
16203 SDValue OffsetSymVal = DAG.getMCSymbol(OffsetSym, PtrVT);
16204 SDValue RegNodeFrameOffset =
16205 DAG.getNode(ISD::LOCAL_RECOVER, dl, PtrVT, OffsetSymVal);
16207 // RegNodeBase = EntryEBP - RegNodeSize
16208 // ParentFP = RegNodeBase - RegNodeFrameOffset
16209 SDValue RegNodeBase = DAG.getNode(ISD::SUB, dl, PtrVT, EntryEBP,
16210 DAG.getConstant(RegNodeSize, dl, PtrVT));
16211 return DAG.getNode(ISD::SUB, dl, PtrVT, RegNodeBase, RegNodeFrameOffset);
16214 static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
16215 SelectionDAG &DAG) {
16217 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
16218 MVT VT = Op.getSimpleValueType();
16219 const IntrinsicData* IntrData = getIntrinsicWithoutChain(IntNo);
16221 switch(IntrData->Type) {
16222 case INTR_TYPE_1OP:
16223 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1));
16224 case INTR_TYPE_2OP:
16225 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16227 case INTR_TYPE_2OP_IMM8:
16228 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16229 DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(2)));
16230 case INTR_TYPE_3OP:
16231 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16232 Op.getOperand(2), Op.getOperand(3));
16233 case INTR_TYPE_4OP:
16234 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16235 Op.getOperand(2), Op.getOperand(3), Op.getOperand(4));
16236 case INTR_TYPE_1OP_MASK_RM: {
16237 SDValue Src = Op.getOperand(1);
16238 SDValue PassThru = Op.getOperand(2);
16239 SDValue Mask = Op.getOperand(3);
16240 SDValue RoundingMode;
16241 // We allways add rounding mode to the Node.
16242 // If the rounding mode is not specified, we add the
16243 // "current direction" mode.
16244 if (Op.getNumOperands() == 4)
16246 DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
16248 RoundingMode = Op.getOperand(4);
16249 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16250 if (IntrWithRoundingModeOpcode != 0)
16251 if (cast<ConstantSDNode>(RoundingMode)->getZExtValue() !=
16252 X86::STATIC_ROUNDING::CUR_DIRECTION)
16253 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16254 dl, Op.getValueType(), Src, RoundingMode),
16255 Mask, PassThru, Subtarget, DAG);
16256 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src,
16258 Mask, PassThru, Subtarget, DAG);
16260 case INTR_TYPE_1OP_MASK: {
16261 SDValue Src = Op.getOperand(1);
16262 SDValue PassThru = Op.getOperand(2);
16263 SDValue Mask = Op.getOperand(3);
16264 // We add rounding mode to the Node when
16265 // - RM Opcode is specified and
16266 // - RM is not "current direction".
16267 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16268 if (IntrWithRoundingModeOpcode != 0) {
16269 SDValue Rnd = Op.getOperand(4);
16270 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
16271 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
16272 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16273 dl, Op.getValueType(),
16275 Mask, PassThru, Subtarget, DAG);
16278 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src),
16279 Mask, PassThru, Subtarget, DAG);
16281 case INTR_TYPE_SCALAR_MASK: {
16282 SDValue Src1 = Op.getOperand(1);
16283 SDValue Src2 = Op.getOperand(2);
16284 SDValue passThru = Op.getOperand(3);
16285 SDValue Mask = Op.getOperand(4);
16286 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2),
16287 Mask, passThru, Subtarget, DAG);
16289 case INTR_TYPE_SCALAR_MASK_RM: {
16290 SDValue Src1 = Op.getOperand(1);
16291 SDValue Src2 = Op.getOperand(2);
16292 SDValue Src0 = Op.getOperand(3);
16293 SDValue Mask = Op.getOperand(4);
16294 // There are 2 kinds of intrinsics in this group:
16295 // (1) With suppress-all-exceptions (sae) or rounding mode- 6 operands
16296 // (2) With rounding mode and sae - 7 operands.
16297 if (Op.getNumOperands() == 6) {
16298 SDValue Sae = Op.getOperand(5);
16299 unsigned Opc = IntrData->Opc1 ? IntrData->Opc1 : IntrData->Opc0;
16300 return getScalarMaskingNode(DAG.getNode(Opc, dl, VT, Src1, Src2,
16302 Mask, Src0, Subtarget, DAG);
16304 assert(Op.getNumOperands() == 7 && "Unexpected intrinsic form");
16305 SDValue RoundingMode = Op.getOperand(5);
16306 SDValue Sae = Op.getOperand(6);
16307 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2,
16308 RoundingMode, Sae),
16309 Mask, Src0, Subtarget, DAG);
16311 case INTR_TYPE_2OP_MASK:
16312 case INTR_TYPE_2OP_IMM8_MASK: {
16313 SDValue Src1 = Op.getOperand(1);
16314 SDValue Src2 = Op.getOperand(2);
16315 SDValue PassThru = Op.getOperand(3);
16316 SDValue Mask = Op.getOperand(4);
16318 if (IntrData->Type == INTR_TYPE_2OP_IMM8_MASK)
16319 Src2 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Src2);
16321 // We specify 2 possible opcodes for intrinsics with rounding modes.
16322 // First, we check if the intrinsic may have non-default rounding mode,
16323 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16324 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16325 if (IntrWithRoundingModeOpcode != 0) {
16326 SDValue Rnd = Op.getOperand(5);
16327 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
16328 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
16329 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16330 dl, Op.getValueType(),
16332 Mask, PassThru, Subtarget, DAG);
16335 // TODO: Intrinsics should have fast-math-flags to propagate.
16336 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,Src1,Src2),
16337 Mask, PassThru, Subtarget, DAG);
16339 case INTR_TYPE_2OP_MASK_RM: {
16340 SDValue Src1 = Op.getOperand(1);
16341 SDValue Src2 = Op.getOperand(2);
16342 SDValue PassThru = Op.getOperand(3);
16343 SDValue Mask = Op.getOperand(4);
16344 // We specify 2 possible modes for intrinsics, with/without rounding
16346 // First, we check if the intrinsic have rounding mode (6 operands),
16347 // if not, we set rounding mode to "current".
16349 if (Op.getNumOperands() == 6)
16350 Rnd = Op.getOperand(5);
16352 Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
16353 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16355 Mask, PassThru, Subtarget, DAG);
16357 case INTR_TYPE_3OP_SCALAR_MASK_RM: {
16358 SDValue Src1 = Op.getOperand(1);
16359 SDValue Src2 = Op.getOperand(2);
16360 SDValue Src3 = Op.getOperand(3);
16361 SDValue PassThru = Op.getOperand(4);
16362 SDValue Mask = Op.getOperand(5);
16363 SDValue Sae = Op.getOperand(6);
16365 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1,
16367 Mask, PassThru, Subtarget, DAG);
16369 case INTR_TYPE_3OP_MASK_RM: {
16370 SDValue Src1 = Op.getOperand(1);
16371 SDValue Src2 = Op.getOperand(2);
16372 SDValue Imm = Op.getOperand(3);
16373 SDValue PassThru = Op.getOperand(4);
16374 SDValue Mask = Op.getOperand(5);
16375 // We specify 2 possible modes for intrinsics, with/without rounding
16377 // First, we check if the intrinsic have rounding mode (7 operands),
16378 // if not, we set rounding mode to "current".
16380 if (Op.getNumOperands() == 7)
16381 Rnd = Op.getOperand(6);
16383 Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
16384 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16385 Src1, Src2, Imm, Rnd),
16386 Mask, PassThru, Subtarget, DAG);
16388 case INTR_TYPE_3OP_IMM8_MASK:
16389 case INTR_TYPE_3OP_MASK:
16390 case INSERT_SUBVEC: {
16391 SDValue Src1 = Op.getOperand(1);
16392 SDValue Src2 = Op.getOperand(2);
16393 SDValue Src3 = Op.getOperand(3);
16394 SDValue PassThru = Op.getOperand(4);
16395 SDValue Mask = Op.getOperand(5);
16397 if (IntrData->Type == INTR_TYPE_3OP_IMM8_MASK)
16398 Src3 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Src3);
16399 else if (IntrData->Type == INSERT_SUBVEC) {
16400 // imm should be adapted to ISD::INSERT_SUBVECTOR behavior
16401 assert(isa<ConstantSDNode>(Src3) && "Expected a ConstantSDNode here!");
16402 unsigned Imm = cast<ConstantSDNode>(Src3)->getZExtValue();
16403 Imm *= Src2.getSimpleValueType().getVectorNumElements();
16404 Src3 = DAG.getTargetConstant(Imm, dl, MVT::i32);
16407 // We specify 2 possible opcodes for intrinsics with rounding modes.
16408 // First, we check if the intrinsic may have non-default rounding mode,
16409 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16410 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16411 if (IntrWithRoundingModeOpcode != 0) {
16412 SDValue Rnd = Op.getOperand(6);
16413 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
16414 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
16415 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16416 dl, Op.getValueType(),
16417 Src1, Src2, Src3, Rnd),
16418 Mask, PassThru, Subtarget, DAG);
16421 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16423 Mask, PassThru, Subtarget, DAG);
16425 case VPERM_3OP_MASKZ:
16426 case VPERM_3OP_MASK:{
16427 // Src2 is the PassThru
16428 SDValue Src1 = Op.getOperand(1);
16429 SDValue Src2 = Op.getOperand(2);
16430 SDValue Src3 = Op.getOperand(3);
16431 SDValue Mask = Op.getOperand(4);
16432 MVT VT = Op.getSimpleValueType();
16433 SDValue PassThru = SDValue();
16435 // set PassThru element
16436 if (IntrData->Type == VPERM_3OP_MASKZ)
16437 PassThru = getZeroVector(VT, Subtarget, DAG, dl);
16439 PassThru = DAG.getBitcast(VT, Src2);
16441 // Swap Src1 and Src2 in the node creation
16442 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0,
16443 dl, Op.getValueType(),
16445 Mask, PassThru, Subtarget, DAG);
16449 case FMA_OP_MASK: {
16450 SDValue Src1 = Op.getOperand(1);
16451 SDValue Src2 = Op.getOperand(2);
16452 SDValue Src3 = Op.getOperand(3);
16453 SDValue Mask = Op.getOperand(4);
16454 MVT VT = Op.getSimpleValueType();
16455 SDValue PassThru = SDValue();
16457 // set PassThru element
16458 if (IntrData->Type == FMA_OP_MASKZ)
16459 PassThru = getZeroVector(VT, Subtarget, DAG, dl);
16460 else if (IntrData->Type == FMA_OP_MASK3)
16465 // We specify 2 possible opcodes for intrinsics with rounding modes.
16466 // First, we check if the intrinsic may have non-default rounding mode,
16467 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16468 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16469 if (IntrWithRoundingModeOpcode != 0) {
16470 SDValue Rnd = Op.getOperand(5);
16471 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16472 X86::STATIC_ROUNDING::CUR_DIRECTION)
16473 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16474 dl, Op.getValueType(),
16475 Src1, Src2, Src3, Rnd),
16476 Mask, PassThru, Subtarget, DAG);
16478 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0,
16479 dl, Op.getValueType(),
16481 Mask, PassThru, Subtarget, DAG);
16483 case TERLOG_OP_MASK:
16484 case TERLOG_OP_MASKZ: {
16485 SDValue Src1 = Op.getOperand(1);
16486 SDValue Src2 = Op.getOperand(2);
16487 SDValue Src3 = Op.getOperand(3);
16488 SDValue Src4 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(4));
16489 SDValue Mask = Op.getOperand(5);
16490 MVT VT = Op.getSimpleValueType();
16491 SDValue PassThru = Src1;
16492 // Set PassThru element.
16493 if (IntrData->Type == TERLOG_OP_MASKZ)
16494 PassThru = getZeroVector(VT, Subtarget, DAG, dl);
16496 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16497 Src1, Src2, Src3, Src4),
16498 Mask, PassThru, Subtarget, DAG);
16501 // FPclass intrinsics with mask
16502 SDValue Src1 = Op.getOperand(1);
16503 MVT VT = Src1.getSimpleValueType();
16504 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
16505 SDValue Imm = Op.getOperand(2);
16506 SDValue Mask = Op.getOperand(3);
16507 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
16508 Mask.getSimpleValueType().getSizeInBits());
16509 SDValue FPclass = DAG.getNode(IntrData->Opc0, dl, MaskVT, Src1, Imm);
16510 SDValue FPclassMask = getVectorMaskingNode(FPclass, Mask,
16511 DAG.getTargetConstant(0, dl, MaskVT),
16513 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
16514 DAG.getUNDEF(BitcastVT), FPclassMask,
16515 DAG.getIntPtrConstant(0, dl));
16516 return DAG.getBitcast(Op.getValueType(), Res);
16519 SDValue Src1 = Op.getOperand(1);
16520 SDValue Imm = Op.getOperand(2);
16521 SDValue Mask = Op.getOperand(3);
16522 SDValue FPclass = DAG.getNode(IntrData->Opc0, dl, MVT::i1, Src1, Imm);
16523 SDValue FPclassMask = getScalarMaskingNode(FPclass, Mask,
16524 DAG.getTargetConstant(0, dl, MVT::i1), Subtarget, DAG);
16525 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i8, FPclassMask);
16528 case CMP_MASK_CC: {
16529 // Comparison intrinsics with masks.
16530 // Example of transformation:
16531 // (i8 (int_x86_avx512_mask_pcmpeq_q_128
16532 // (v2i64 %a), (v2i64 %b), (i8 %mask))) ->
16534 // (v8i1 (insert_subvector undef,
16535 // (v2i1 (and (PCMPEQM %a, %b),
16536 // (extract_subvector
16537 // (v8i1 (bitcast %mask)), 0))), 0))))
16538 MVT VT = Op.getOperand(1).getSimpleValueType();
16539 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
16540 SDValue Mask = Op.getOperand((IntrData->Type == CMP_MASK_CC) ? 4 : 3);
16541 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
16542 Mask.getSimpleValueType().getSizeInBits());
16544 if (IntrData->Type == CMP_MASK_CC) {
16545 SDValue CC = Op.getOperand(3);
16546 CC = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, CC);
16547 // We specify 2 possible opcodes for intrinsics with rounding modes.
16548 // First, we check if the intrinsic may have non-default rounding mode,
16549 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16550 if (IntrData->Opc1 != 0) {
16551 SDValue Rnd = Op.getOperand(5);
16552 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16553 X86::STATIC_ROUNDING::CUR_DIRECTION)
16554 Cmp = DAG.getNode(IntrData->Opc1, dl, MaskVT, Op.getOperand(1),
16555 Op.getOperand(2), CC, Rnd);
16557 //default rounding mode
16559 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
16560 Op.getOperand(2), CC);
16563 assert(IntrData->Type == CMP_MASK && "Unexpected intrinsic type!");
16564 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
16567 SDValue CmpMask = getVectorMaskingNode(Cmp, Mask,
16568 DAG.getTargetConstant(0, dl,
16571 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
16572 DAG.getUNDEF(BitcastVT), CmpMask,
16573 DAG.getIntPtrConstant(0, dl));
16574 return DAG.getBitcast(Op.getValueType(), Res);
16576 case CMP_MASK_SCALAR_CC: {
16577 SDValue Src1 = Op.getOperand(1);
16578 SDValue Src2 = Op.getOperand(2);
16579 SDValue CC = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(3));
16580 SDValue Mask = Op.getOperand(4);
16583 if (IntrData->Opc1 != 0) {
16584 SDValue Rnd = Op.getOperand(5);
16585 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16586 X86::STATIC_ROUNDING::CUR_DIRECTION)
16587 Cmp = DAG.getNode(IntrData->Opc1, dl, MVT::i1, Src1, Src2, CC, Rnd);
16589 //default rounding mode
16591 Cmp = DAG.getNode(IntrData->Opc0, dl, MVT::i1, Src1, Src2, CC);
16593 SDValue CmpMask = getScalarMaskingNode(Cmp, Mask,
16594 DAG.getTargetConstant(0, dl,
16598 return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::i8,
16599 DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i8, CmpMask),
16600 DAG.getValueType(MVT::i1));
16602 case COMI: { // Comparison intrinsics
16603 ISD::CondCode CC = (ISD::CondCode)IntrData->Opc1;
16604 SDValue LHS = Op.getOperand(1);
16605 SDValue RHS = Op.getOperand(2);
16606 unsigned X86CC = TranslateX86CC(CC, dl, true, LHS, RHS, DAG);
16607 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
16608 SDValue Cond = DAG.getNode(IntrData->Opc0, dl, MVT::i32, LHS, RHS);
16609 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16610 DAG.getConstant(X86CC, dl, MVT::i8), Cond);
16611 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16613 case COMI_RM: { // Comparison intrinsics with Sae
16614 SDValue LHS = Op.getOperand(1);
16615 SDValue RHS = Op.getOperand(2);
16616 SDValue CC = Op.getOperand(3);
16617 SDValue Sae = Op.getOperand(4);
16618 auto ComiType = TranslateX86ConstCondToX86CC(CC);
16619 // choose between ordered and unordered (comi/ucomi)
16620 unsigned comiOp = std::get<0>(ComiType) ? IntrData->Opc0 : IntrData->Opc1;
16622 if (cast<ConstantSDNode>(Sae)->getZExtValue() !=
16623 X86::STATIC_ROUNDING::CUR_DIRECTION)
16624 Cond = DAG.getNode(comiOp, dl, MVT::i32, LHS, RHS, Sae);
16626 Cond = DAG.getNode(comiOp, dl, MVT::i32, LHS, RHS);
16627 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16628 DAG.getConstant(std::get<1>(ComiType), dl, MVT::i8), Cond);
16629 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16632 return getTargetVShiftNode(IntrData->Opc0, dl, Op.getSimpleValueType(),
16633 Op.getOperand(1), Op.getOperand(2), DAG);
16635 return getVectorMaskingNode(getTargetVShiftNode(IntrData->Opc0, dl,
16636 Op.getSimpleValueType(),
16638 Op.getOperand(2), DAG),
16639 Op.getOperand(4), Op.getOperand(3), Subtarget,
16641 case COMPRESS_EXPAND_IN_REG: {
16642 SDValue Mask = Op.getOperand(3);
16643 SDValue DataToCompress = Op.getOperand(1);
16644 SDValue PassThru = Op.getOperand(2);
16645 if (isAllOnesConstant(Mask)) // return data as is
16646 return Op.getOperand(1);
16648 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16650 Mask, PassThru, Subtarget, DAG);
16653 SDValue Mask = Op.getOperand(1);
16654 MVT MaskVT = MVT::getVectorVT(MVT::i1, Mask.getSimpleValueType().getSizeInBits());
16655 Mask = DAG.getBitcast(MaskVT, Mask);
16656 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Mask);
16659 SDValue Mask = Op.getOperand(3);
16660 MVT VT = Op.getSimpleValueType();
16661 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
16662 SDValue VMask = getMaskNode(Mask, MaskVT, Subtarget, DAG, dl);
16663 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, Op.getOperand(1),
16667 MVT VT = Op.getSimpleValueType();
16668 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getSizeInBits()/2);
16670 SDValue Src1 = getMaskNode(Op.getOperand(1), MaskVT, Subtarget, DAG, dl);
16671 SDValue Src2 = getMaskNode(Op.getOperand(2), MaskVT, Subtarget, DAG, dl);
16672 // Arguments should be swapped.
16673 SDValue Res = DAG.getNode(IntrData->Opc0, dl,
16674 MVT::getVectorVT(MVT::i1, VT.getSizeInBits()),
16676 return DAG.getBitcast(VT, Res);
16684 default: return SDValue(); // Don't custom lower most intrinsics.
16686 case Intrinsic::x86_avx2_permd:
16687 case Intrinsic::x86_avx2_permps:
16688 // Operands intentionally swapped. Mask is last operand to intrinsic,
16689 // but second operand for node/instruction.
16690 return DAG.getNode(X86ISD::VPERMV, dl, Op.getValueType(),
16691 Op.getOperand(2), Op.getOperand(1));
16693 // ptest and testp intrinsics. The intrinsic these come from are designed to
16694 // return an integer value, not just an instruction so lower it to the ptest
16695 // or testp pattern and a setcc for the result.
16696 case Intrinsic::x86_sse41_ptestz:
16697 case Intrinsic::x86_sse41_ptestc:
16698 case Intrinsic::x86_sse41_ptestnzc:
16699 case Intrinsic::x86_avx_ptestz_256:
16700 case Intrinsic::x86_avx_ptestc_256:
16701 case Intrinsic::x86_avx_ptestnzc_256:
16702 case Intrinsic::x86_avx_vtestz_ps:
16703 case Intrinsic::x86_avx_vtestc_ps:
16704 case Intrinsic::x86_avx_vtestnzc_ps:
16705 case Intrinsic::x86_avx_vtestz_pd:
16706 case Intrinsic::x86_avx_vtestc_pd:
16707 case Intrinsic::x86_avx_vtestnzc_pd:
16708 case Intrinsic::x86_avx_vtestz_ps_256:
16709 case Intrinsic::x86_avx_vtestc_ps_256:
16710 case Intrinsic::x86_avx_vtestnzc_ps_256:
16711 case Intrinsic::x86_avx_vtestz_pd_256:
16712 case Intrinsic::x86_avx_vtestc_pd_256:
16713 case Intrinsic::x86_avx_vtestnzc_pd_256: {
16714 bool IsTestPacked = false;
16717 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
16718 case Intrinsic::x86_avx_vtestz_ps:
16719 case Intrinsic::x86_avx_vtestz_pd:
16720 case Intrinsic::x86_avx_vtestz_ps_256:
16721 case Intrinsic::x86_avx_vtestz_pd_256:
16722 IsTestPacked = true; // Fallthrough
16723 case Intrinsic::x86_sse41_ptestz:
16724 case Intrinsic::x86_avx_ptestz_256:
16726 X86CC = X86::COND_E;
16728 case Intrinsic::x86_avx_vtestc_ps:
16729 case Intrinsic::x86_avx_vtestc_pd:
16730 case Intrinsic::x86_avx_vtestc_ps_256:
16731 case Intrinsic::x86_avx_vtestc_pd_256:
16732 IsTestPacked = true; // Fallthrough
16733 case Intrinsic::x86_sse41_ptestc:
16734 case Intrinsic::x86_avx_ptestc_256:
16736 X86CC = X86::COND_B;
16738 case Intrinsic::x86_avx_vtestnzc_ps:
16739 case Intrinsic::x86_avx_vtestnzc_pd:
16740 case Intrinsic::x86_avx_vtestnzc_ps_256:
16741 case Intrinsic::x86_avx_vtestnzc_pd_256:
16742 IsTestPacked = true; // Fallthrough
16743 case Intrinsic::x86_sse41_ptestnzc:
16744 case Intrinsic::x86_avx_ptestnzc_256:
16746 X86CC = X86::COND_A;
16750 SDValue LHS = Op.getOperand(1);
16751 SDValue RHS = Op.getOperand(2);
16752 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
16753 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
16754 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
16755 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
16756 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16758 case Intrinsic::x86_avx512_kortestz_w:
16759 case Intrinsic::x86_avx512_kortestc_w: {
16760 unsigned X86CC = (IntNo == Intrinsic::x86_avx512_kortestz_w)? X86::COND_E: X86::COND_B;
16761 SDValue LHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(1));
16762 SDValue RHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(2));
16763 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
16764 SDValue Test = DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, LHS, RHS);
16765 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i1, CC, Test);
16766 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16769 case Intrinsic::x86_sse42_pcmpistria128:
16770 case Intrinsic::x86_sse42_pcmpestria128:
16771 case Intrinsic::x86_sse42_pcmpistric128:
16772 case Intrinsic::x86_sse42_pcmpestric128:
16773 case Intrinsic::x86_sse42_pcmpistrio128:
16774 case Intrinsic::x86_sse42_pcmpestrio128:
16775 case Intrinsic::x86_sse42_pcmpistris128:
16776 case Intrinsic::x86_sse42_pcmpestris128:
16777 case Intrinsic::x86_sse42_pcmpistriz128:
16778 case Intrinsic::x86_sse42_pcmpestriz128: {
16782 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
16783 case Intrinsic::x86_sse42_pcmpistria128:
16784 Opcode = X86ISD::PCMPISTRI;
16785 X86CC = X86::COND_A;
16787 case Intrinsic::x86_sse42_pcmpestria128:
16788 Opcode = X86ISD::PCMPESTRI;
16789 X86CC = X86::COND_A;
16791 case Intrinsic::x86_sse42_pcmpistric128:
16792 Opcode = X86ISD::PCMPISTRI;
16793 X86CC = X86::COND_B;
16795 case Intrinsic::x86_sse42_pcmpestric128:
16796 Opcode = X86ISD::PCMPESTRI;
16797 X86CC = X86::COND_B;
16799 case Intrinsic::x86_sse42_pcmpistrio128:
16800 Opcode = X86ISD::PCMPISTRI;
16801 X86CC = X86::COND_O;
16803 case Intrinsic::x86_sse42_pcmpestrio128:
16804 Opcode = X86ISD::PCMPESTRI;
16805 X86CC = X86::COND_O;
16807 case Intrinsic::x86_sse42_pcmpistris128:
16808 Opcode = X86ISD::PCMPISTRI;
16809 X86CC = X86::COND_S;
16811 case Intrinsic::x86_sse42_pcmpestris128:
16812 Opcode = X86ISD::PCMPESTRI;
16813 X86CC = X86::COND_S;
16815 case Intrinsic::x86_sse42_pcmpistriz128:
16816 Opcode = X86ISD::PCMPISTRI;
16817 X86CC = X86::COND_E;
16819 case Intrinsic::x86_sse42_pcmpestriz128:
16820 Opcode = X86ISD::PCMPESTRI;
16821 X86CC = X86::COND_E;
16824 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
16825 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
16826 SDValue PCMP = DAG.getNode(Opcode, dl, VTs, NewOps);
16827 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16828 DAG.getConstant(X86CC, dl, MVT::i8),
16829 SDValue(PCMP.getNode(), 1));
16830 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16833 case Intrinsic::x86_sse42_pcmpistri128:
16834 case Intrinsic::x86_sse42_pcmpestri128: {
16836 if (IntNo == Intrinsic::x86_sse42_pcmpistri128)
16837 Opcode = X86ISD::PCMPISTRI;
16839 Opcode = X86ISD::PCMPESTRI;
16841 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
16842 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
16843 return DAG.getNode(Opcode, dl, VTs, NewOps);
16846 case Intrinsic::x86_seh_lsda: {
16847 // Compute the symbol for the LSDA. We know it'll get emitted later.
16848 MachineFunction &MF = DAG.getMachineFunction();
16849 SDValue Op1 = Op.getOperand(1);
16850 auto *Fn = cast<Function>(cast<GlobalAddressSDNode>(Op1)->getGlobal());
16851 MCSymbol *LSDASym = MF.getMMI().getContext().getOrCreateLSDASymbol(
16852 GlobalValue::getRealLinkageName(Fn->getName()));
16854 // Generate a simple absolute symbol reference. This intrinsic is only
16855 // supported on 32-bit Windows, which isn't PIC.
16856 SDValue Result = DAG.getMCSymbol(LSDASym, VT);
16857 return DAG.getNode(X86ISD::Wrapper, dl, VT, Result);
16860 case Intrinsic::x86_seh_recoverfp: {
16861 SDValue FnOp = Op.getOperand(1);
16862 SDValue IncomingFPOp = Op.getOperand(2);
16863 GlobalAddressSDNode *GSD = dyn_cast<GlobalAddressSDNode>(FnOp);
16864 auto *Fn = dyn_cast_or_null<Function>(GSD ? GSD->getGlobal() : nullptr);
16866 report_fatal_error(
16867 "llvm.x86.seh.recoverfp must take a function as the first argument");
16868 return recoverFramePointer(DAG, Fn, IncomingFPOp);
16871 case Intrinsic::localaddress: {
16872 // Returns one of the stack, base, or frame pointer registers, depending on
16873 // which is used to reference local variables.
16874 MachineFunction &MF = DAG.getMachineFunction();
16875 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16877 if (RegInfo->hasBasePointer(MF))
16878 Reg = RegInfo->getBaseRegister();
16879 else // This function handles the SP or FP case.
16880 Reg = RegInfo->getPtrSizedFrameRegister(MF);
16881 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
16886 static SDValue getGatherNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16887 SDValue Src, SDValue Mask, SDValue Base,
16888 SDValue Index, SDValue ScaleOp, SDValue Chain,
16889 const X86Subtarget * Subtarget) {
16891 auto *C = cast<ConstantSDNode>(ScaleOp);
16892 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16893 MVT MaskVT = MVT::getVectorVT(MVT::i1,
16894 Index.getSimpleValueType().getVectorNumElements());
16896 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16898 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16900 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
16901 Mask.getSimpleValueType().getSizeInBits());
16903 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16904 // are extracted by EXTRACT_SUBVECTOR.
16905 MaskInReg = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16906 DAG.getBitcast(BitcastVT, Mask),
16907 DAG.getIntPtrConstant(0, dl));
16909 SDVTList VTs = DAG.getVTList(Op.getValueType(), MaskVT, MVT::Other);
16910 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16911 SDValue Segment = DAG.getRegister(0, MVT::i32);
16912 if (Src.getOpcode() == ISD::UNDEF)
16913 Src = getZeroVector(Op.getSimpleValueType(), Subtarget, DAG, dl);
16914 SDValue Ops[] = {Src, MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
16915 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
16916 SDValue RetOps[] = { SDValue(Res, 0), SDValue(Res, 2) };
16917 return DAG.getMergeValues(RetOps, dl);
16920 static SDValue getScatterNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16921 SDValue Src, SDValue Mask, SDValue Base,
16922 SDValue Index, SDValue ScaleOp, SDValue Chain) {
16924 auto *C = cast<ConstantSDNode>(ScaleOp);
16925 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16926 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16927 SDValue Segment = DAG.getRegister(0, MVT::i32);
16928 MVT MaskVT = MVT::getVectorVT(MVT::i1,
16929 Index.getSimpleValueType().getVectorNumElements());
16931 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16933 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16935 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
16936 Mask.getSimpleValueType().getSizeInBits());
16938 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16939 // are extracted by EXTRACT_SUBVECTOR.
16940 MaskInReg = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16941 DAG.getBitcast(BitcastVT, Mask),
16942 DAG.getIntPtrConstant(0, dl));
16944 SDVTList VTs = DAG.getVTList(MaskVT, MVT::Other);
16945 SDValue Ops[] = {Base, Scale, Index, Disp, Segment, MaskInReg, Src, Chain};
16946 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
16947 return SDValue(Res, 1);
16950 static SDValue getPrefetchNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16951 SDValue Mask, SDValue Base, SDValue Index,
16952 SDValue ScaleOp, SDValue Chain) {
16954 auto *C = cast<ConstantSDNode>(ScaleOp);
16955 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16956 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16957 SDValue Segment = DAG.getRegister(0, MVT::i32);
16959 MVT::getVectorVT(MVT::i1, Index.getSimpleValueType().getVectorNumElements());
16961 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16963 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16965 MaskInReg = DAG.getBitcast(MaskVT, Mask);
16966 //SDVTList VTs = DAG.getVTList(MVT::Other);
16967 SDValue Ops[] = {MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
16968 SDNode *Res = DAG.getMachineNode(Opc, dl, MVT::Other, Ops);
16969 return SDValue(Res, 0);
16972 // getReadPerformanceCounter - Handles the lowering of builtin intrinsics that
16973 // read performance monitor counters (x86_rdpmc).
16974 static void getReadPerformanceCounter(SDNode *N, SDLoc DL,
16975 SelectionDAG &DAG, const X86Subtarget *Subtarget,
16976 SmallVectorImpl<SDValue> &Results) {
16977 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
16978 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
16981 // The ECX register is used to select the index of the performance counter
16983 SDValue Chain = DAG.getCopyToReg(N->getOperand(0), DL, X86::ECX,
16985 SDValue rd = DAG.getNode(X86ISD::RDPMC_DAG, DL, Tys, Chain);
16987 // Reads the content of a 64-bit performance counter and returns it in the
16988 // registers EDX:EAX.
16989 if (Subtarget->is64Bit()) {
16990 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
16991 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
16994 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
16995 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
16998 Chain = HI.getValue(1);
17000 if (Subtarget->is64Bit()) {
17001 // The EAX register is loaded with the low-order 32 bits. The EDX register
17002 // is loaded with the supported high-order bits of the counter.
17003 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
17004 DAG.getConstant(32, DL, MVT::i8));
17005 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
17006 Results.push_back(Chain);
17010 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
17011 SDValue Ops[] = { LO, HI };
17012 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
17013 Results.push_back(Pair);
17014 Results.push_back(Chain);
17017 // getReadTimeStampCounter - Handles the lowering of builtin intrinsics that
17018 // read the time stamp counter (x86_rdtsc and x86_rdtscp). This function is
17019 // also used to custom lower READCYCLECOUNTER nodes.
17020 static void getReadTimeStampCounter(SDNode *N, SDLoc DL, unsigned Opcode,
17021 SelectionDAG &DAG, const X86Subtarget *Subtarget,
17022 SmallVectorImpl<SDValue> &Results) {
17023 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
17024 SDValue rd = DAG.getNode(Opcode, DL, Tys, N->getOperand(0));
17027 // The processor's time-stamp counter (a 64-bit MSR) is stored into the
17028 // EDX:EAX registers. EDX is loaded with the high-order 32 bits of the MSR
17029 // and the EAX register is loaded with the low-order 32 bits.
17030 if (Subtarget->is64Bit()) {
17031 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
17032 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
17035 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
17036 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
17039 SDValue Chain = HI.getValue(1);
17041 if (Opcode == X86ISD::RDTSCP_DAG) {
17042 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
17044 // Instruction RDTSCP loads the IA32:TSC_AUX_MSR (address C000_0103H) into
17045 // the ECX register. Add 'ecx' explicitly to the chain.
17046 SDValue ecx = DAG.getCopyFromReg(Chain, DL, X86::ECX, MVT::i32,
17048 // Explicitly store the content of ECX at the location passed in input
17049 // to the 'rdtscp' intrinsic.
17050 Chain = DAG.getStore(ecx.getValue(1), DL, ecx, N->getOperand(2),
17051 MachinePointerInfo(), false, false, 0);
17054 if (Subtarget->is64Bit()) {
17055 // The EDX register is loaded with the high-order 32 bits of the MSR, and
17056 // the EAX register is loaded with the low-order 32 bits.
17057 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
17058 DAG.getConstant(32, DL, MVT::i8));
17059 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
17060 Results.push_back(Chain);
17064 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
17065 SDValue Ops[] = { LO, HI };
17066 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
17067 Results.push_back(Pair);
17068 Results.push_back(Chain);
17071 static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget *Subtarget,
17072 SelectionDAG &DAG) {
17073 SmallVector<SDValue, 2> Results;
17075 getReadTimeStampCounter(Op.getNode(), DL, X86ISD::RDTSC_DAG, DAG, Subtarget,
17077 return DAG.getMergeValues(Results, DL);
17080 static SDValue LowerSEHRESTOREFRAME(SDValue Op, const X86Subtarget *Subtarget,
17081 SelectionDAG &DAG) {
17082 MachineFunction &MF = DAG.getMachineFunction();
17083 const Function *Fn = MF.getFunction();
17085 SDValue Chain = Op.getOperand(0);
17087 assert(Subtarget->getFrameLowering()->hasFP(MF) &&
17088 "using llvm.x86.seh.restoreframe requires a frame pointer");
17090 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
17091 MVT VT = TLI.getPointerTy(DAG.getDataLayout());
17093 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17094 unsigned FrameReg =
17095 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
17096 unsigned SPReg = RegInfo->getStackRegister();
17097 unsigned SlotSize = RegInfo->getSlotSize();
17099 // Get incoming EBP.
17100 SDValue IncomingEBP =
17101 DAG.getCopyFromReg(Chain, dl, FrameReg, VT);
17103 // SP is saved in the first field of every registration node, so load
17104 // [EBP-RegNodeSize] into SP.
17105 int RegNodeSize = getSEHRegistrationNodeSize(Fn);
17106 SDValue SPAddr = DAG.getNode(ISD::ADD, dl, VT, IncomingEBP,
17107 DAG.getConstant(-RegNodeSize, dl, VT));
17109 DAG.getLoad(VT, dl, Chain, SPAddr, MachinePointerInfo(), false, false,
17110 false, VT.getScalarSizeInBits() / 8);
17111 Chain = DAG.getCopyToReg(Chain, dl, SPReg, NewSP);
17113 if (!RegInfo->needsStackRealignment(MF)) {
17114 // Adjust EBP to point back to the original frame position.
17115 SDValue NewFP = recoverFramePointer(DAG, Fn, IncomingEBP);
17116 Chain = DAG.getCopyToReg(Chain, dl, FrameReg, NewFP);
17118 assert(RegInfo->hasBasePointer(MF) &&
17119 "functions with Win32 EH must use frame or base pointer register");
17121 // Reload the base pointer (ESI) with the adjusted incoming EBP.
17122 SDValue NewBP = recoverFramePointer(DAG, Fn, IncomingEBP);
17123 Chain = DAG.getCopyToReg(Chain, dl, RegInfo->getBaseRegister(), NewBP);
17125 // Reload the spilled EBP value, now that the stack and base pointers are
17127 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
17128 X86FI->setHasSEHFramePtrSave(true);
17129 int FI = MF.getFrameInfo()->CreateSpillStackObject(SlotSize, SlotSize);
17130 X86FI->setSEHFramePtrSaveIndex(FI);
17131 SDValue NewFP = DAG.getLoad(VT, dl, Chain, DAG.getFrameIndex(FI, VT),
17132 MachinePointerInfo(), false, false, false,
17133 VT.getScalarSizeInBits() / 8);
17134 Chain = DAG.getCopyToReg(NewFP, dl, FrameReg, NewFP);
17140 static SDValue MarkEHRegistrationNode(SDValue Op, SelectionDAG &DAG) {
17141 MachineFunction &MF = DAG.getMachineFunction();
17142 SDValue Chain = Op.getOperand(0);
17143 SDValue RegNode = Op.getOperand(2);
17144 WinEHFuncInfo *EHInfo = MF.getWinEHFuncInfo();
17146 report_fatal_error("EH registrations only live in functions using WinEH");
17148 // Cast the operand to an alloca, and remember the frame index.
17149 auto *FINode = dyn_cast<FrameIndexSDNode>(RegNode);
17151 report_fatal_error("llvm.x86.seh.ehregnode expects a static alloca");
17152 EHInfo->EHRegNodeFrameIndex = FINode->getIndex();
17154 // Return the chain operand without making any DAG nodes.
17158 /// \brief Lower intrinsics for TRUNCATE_TO_MEM case
17159 /// return truncate Store/MaskedStore Node
17160 static SDValue LowerINTRINSIC_TRUNCATE_TO_MEM(const SDValue & Op,
17164 SDValue Mask = Op.getOperand(4);
17165 SDValue DataToTruncate = Op.getOperand(3);
17166 SDValue Addr = Op.getOperand(2);
17167 SDValue Chain = Op.getOperand(0);
17169 MVT VT = DataToTruncate.getSimpleValueType();
17170 MVT SVT = MVT::getVectorVT(ElementType, VT.getVectorNumElements());
17172 if (isAllOnesConstant(Mask)) // return just a truncate store
17173 return DAG.getTruncStore(Chain, dl, DataToTruncate, Addr,
17174 MachinePointerInfo(), SVT, false, false,
17175 SVT.getScalarSizeInBits()/8);
17177 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
17178 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
17179 Mask.getSimpleValueType().getSizeInBits());
17180 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
17181 // are extracted by EXTRACT_SUBVECTOR.
17182 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
17183 DAG.getBitcast(BitcastVT, Mask),
17184 DAG.getIntPtrConstant(0, dl));
17186 MachineMemOperand *MMO = DAG.getMachineFunction().
17187 getMachineMemOperand(MachinePointerInfo(),
17188 MachineMemOperand::MOStore, SVT.getStoreSize(),
17189 SVT.getScalarSizeInBits()/8);
17191 return DAG.getMaskedStore(Chain, dl, DataToTruncate, Addr,
17192 VMask, SVT, MMO, true);
17195 static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
17196 SelectionDAG &DAG) {
17197 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
17199 const IntrinsicData* IntrData = getIntrinsicWithChain(IntNo);
17201 if (IntNo == llvm::Intrinsic::x86_seh_restoreframe)
17202 return LowerSEHRESTOREFRAME(Op, Subtarget, DAG);
17203 else if (IntNo == llvm::Intrinsic::x86_seh_ehregnode)
17204 return MarkEHRegistrationNode(Op, DAG);
17209 switch(IntrData->Type) {
17210 default: llvm_unreachable("Unknown Intrinsic Type");
17213 // Emit the node with the right value type.
17214 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other);
17215 SDValue Result = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
17217 // If the value returned by RDRAND/RDSEED was valid (CF=1), return 1.
17218 // Otherwise return the value from Rand, which is always 0, casted to i32.
17219 SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)),
17220 DAG.getConstant(1, dl, Op->getValueType(1)),
17221 DAG.getConstant(X86::COND_B, dl, MVT::i32),
17222 SDValue(Result.getNode(), 1) };
17223 SDValue isValid = DAG.getNode(X86ISD::CMOV, dl,
17224 DAG.getVTList(Op->getValueType(1), MVT::Glue),
17227 // Return { result, isValid, chain }.
17228 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid,
17229 SDValue(Result.getNode(), 2));
17232 //gather(v1, mask, index, base, scale);
17233 SDValue Chain = Op.getOperand(0);
17234 SDValue Src = Op.getOperand(2);
17235 SDValue Base = Op.getOperand(3);
17236 SDValue Index = Op.getOperand(4);
17237 SDValue Mask = Op.getOperand(5);
17238 SDValue Scale = Op.getOperand(6);
17239 return getGatherNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale,
17243 //scatter(base, mask, index, v1, scale);
17244 SDValue Chain = Op.getOperand(0);
17245 SDValue Base = Op.getOperand(2);
17246 SDValue Mask = Op.getOperand(3);
17247 SDValue Index = Op.getOperand(4);
17248 SDValue Src = Op.getOperand(5);
17249 SDValue Scale = Op.getOperand(6);
17250 return getScatterNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index,
17254 SDValue Hint = Op.getOperand(6);
17255 unsigned HintVal = cast<ConstantSDNode>(Hint)->getZExtValue();
17256 assert(HintVal < 2 && "Wrong prefetch hint in intrinsic: should be 0 or 1");
17257 unsigned Opcode = (HintVal ? IntrData->Opc1 : IntrData->Opc0);
17258 SDValue Chain = Op.getOperand(0);
17259 SDValue Mask = Op.getOperand(2);
17260 SDValue Index = Op.getOperand(3);
17261 SDValue Base = Op.getOperand(4);
17262 SDValue Scale = Op.getOperand(5);
17263 return getPrefetchNode(Opcode, Op, DAG, Mask, Base, Index, Scale, Chain);
17265 // Read Time Stamp Counter (RDTSC) and Processor ID (RDTSCP).
17267 SmallVector<SDValue, 2> Results;
17268 getReadTimeStampCounter(Op.getNode(), dl, IntrData->Opc0, DAG, Subtarget,
17270 return DAG.getMergeValues(Results, dl);
17272 // Read Performance Monitoring Counters.
17274 SmallVector<SDValue, 2> Results;
17275 getReadPerformanceCounter(Op.getNode(), dl, DAG, Subtarget, Results);
17276 return DAG.getMergeValues(Results, dl);
17278 // XTEST intrinsics.
17280 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
17281 SDValue InTrans = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
17282 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17283 DAG.getConstant(X86::COND_NE, dl, MVT::i8),
17285 SDValue Ret = DAG.getNode(ISD::ZERO_EXTEND, dl, Op->getValueType(0), SetCC);
17286 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(),
17287 Ret, SDValue(InTrans.getNode(), 1));
17291 SmallVector<SDValue, 2> Results;
17292 SDVTList CFVTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
17293 SDVTList VTs = DAG.getVTList(Op.getOperand(3)->getValueType(0), MVT::Other);
17294 SDValue GenCF = DAG.getNode(X86ISD::ADD, dl, CFVTs, Op.getOperand(2),
17295 DAG.getConstant(-1, dl, MVT::i8));
17296 SDValue Res = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(3),
17297 Op.getOperand(4), GenCF.getValue(1));
17298 SDValue Store = DAG.getStore(Op.getOperand(0), dl, Res.getValue(0),
17299 Op.getOperand(5), MachinePointerInfo(),
17301 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17302 DAG.getConstant(X86::COND_B, dl, MVT::i8),
17304 Results.push_back(SetCC);
17305 Results.push_back(Store);
17306 return DAG.getMergeValues(Results, dl);
17308 case COMPRESS_TO_MEM: {
17310 SDValue Mask = Op.getOperand(4);
17311 SDValue DataToCompress = Op.getOperand(3);
17312 SDValue Addr = Op.getOperand(2);
17313 SDValue Chain = Op.getOperand(0);
17315 MVT VT = DataToCompress.getSimpleValueType();
17316 if (isAllOnesConstant(Mask)) // return just a store
17317 return DAG.getStore(Chain, dl, DataToCompress, Addr,
17318 MachinePointerInfo(), false, false,
17319 VT.getScalarSizeInBits()/8);
17321 SDValue Compressed =
17322 getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, DataToCompress),
17323 Mask, DAG.getUNDEF(VT), Subtarget, DAG);
17324 return DAG.getStore(Chain, dl, Compressed, Addr,
17325 MachinePointerInfo(), false, false,
17326 VT.getScalarSizeInBits()/8);
17328 case TRUNCATE_TO_MEM_VI8:
17329 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i8);
17330 case TRUNCATE_TO_MEM_VI16:
17331 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i16);
17332 case TRUNCATE_TO_MEM_VI32:
17333 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i32);
17334 case EXPAND_FROM_MEM: {
17336 SDValue Mask = Op.getOperand(4);
17337 SDValue PassThru = Op.getOperand(3);
17338 SDValue Addr = Op.getOperand(2);
17339 SDValue Chain = Op.getOperand(0);
17340 MVT VT = Op.getSimpleValueType();
17342 if (isAllOnesConstant(Mask)) // return just a load
17343 return DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(), false, false,
17344 false, VT.getScalarSizeInBits()/8);
17346 SDValue DataToExpand = DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(),
17347 false, false, false,
17348 VT.getScalarSizeInBits()/8);
17350 SDValue Results[] = {
17351 getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, DataToExpand),
17352 Mask, PassThru, Subtarget, DAG), Chain};
17353 return DAG.getMergeValues(Results, dl);
17358 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
17359 SelectionDAG &DAG) const {
17360 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
17361 MFI->setReturnAddressIsTaken(true);
17363 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
17366 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
17368 EVT PtrVT = getPointerTy(DAG.getDataLayout());
17371 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
17372 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17373 SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), dl, PtrVT);
17374 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
17375 DAG.getNode(ISD::ADD, dl, PtrVT,
17376 FrameAddr, Offset),
17377 MachinePointerInfo(), false, false, false, 0);
17380 // Just load the return address.
17381 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
17382 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
17383 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
17386 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
17387 MachineFunction &MF = DAG.getMachineFunction();
17388 MachineFrameInfo *MFI = MF.getFrameInfo();
17389 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
17390 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17391 EVT VT = Op.getValueType();
17393 MFI->setFrameAddressIsTaken(true);
17395 if (MF.getTarget().getMCAsmInfo()->usesWindowsCFI()) {
17396 // Depth > 0 makes no sense on targets which use Windows unwind codes. It
17397 // is not possible to crawl up the stack without looking at the unwind codes
17399 int FrameAddrIndex = FuncInfo->getFAIndex();
17400 if (!FrameAddrIndex) {
17401 // Set up a frame object for the return address.
17402 unsigned SlotSize = RegInfo->getSlotSize();
17403 FrameAddrIndex = MF.getFrameInfo()->CreateFixedObject(
17404 SlotSize, /*Offset=*/0, /*IsImmutable=*/false);
17405 FuncInfo->setFAIndex(FrameAddrIndex);
17407 return DAG.getFrameIndex(FrameAddrIndex, VT);
17410 unsigned FrameReg =
17411 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
17412 SDLoc dl(Op); // FIXME probably not meaningful
17413 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
17414 assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
17415 (FrameReg == X86::EBP && VT == MVT::i32)) &&
17416 "Invalid Frame Register!");
17417 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
17419 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
17420 MachinePointerInfo(),
17421 false, false, false, 0);
17425 // FIXME? Maybe this could be a TableGen attribute on some registers and
17426 // this table could be generated automatically from RegInfo.
17427 unsigned X86TargetLowering::getRegisterByName(const char* RegName, EVT VT,
17428 SelectionDAG &DAG) const {
17429 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
17430 const MachineFunction &MF = DAG.getMachineFunction();
17432 unsigned Reg = StringSwitch<unsigned>(RegName)
17433 .Case("esp", X86::ESP)
17434 .Case("rsp", X86::RSP)
17435 .Case("ebp", X86::EBP)
17436 .Case("rbp", X86::RBP)
17439 if (Reg == X86::EBP || Reg == X86::RBP) {
17440 if (!TFI.hasFP(MF))
17441 report_fatal_error("register " + StringRef(RegName) +
17442 " is allocatable: function has no frame pointer");
17445 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17446 unsigned FrameReg =
17447 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
17448 assert((FrameReg == X86::EBP || FrameReg == X86::RBP) &&
17449 "Invalid Frame Register!");
17457 report_fatal_error("Invalid register name global variable");
17460 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
17461 SelectionDAG &DAG) const {
17462 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17463 return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize(), SDLoc(Op));
17466 unsigned X86TargetLowering::getExceptionPointerRegister(
17467 const Constant *PersonalityFn) const {
17468 if (classifyEHPersonality(PersonalityFn) == EHPersonality::CoreCLR)
17469 return Subtarget->isTarget64BitLP64() ? X86::RDX : X86::EDX;
17471 return Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX;
17474 unsigned X86TargetLowering::getExceptionSelectorRegister(
17475 const Constant *PersonalityFn) const {
17476 // Funclet personalities don't use selectors (the runtime does the selection).
17477 assert(!isFuncletEHPersonality(classifyEHPersonality(PersonalityFn)));
17478 return Subtarget->isTarget64BitLP64() ? X86::RDX : X86::EDX;
17481 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
17482 SDValue Chain = Op.getOperand(0);
17483 SDValue Offset = Op.getOperand(1);
17484 SDValue Handler = Op.getOperand(2);
17487 EVT PtrVT = getPointerTy(DAG.getDataLayout());
17488 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17489 unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
17490 assert(((FrameReg == X86::RBP && PtrVT == MVT::i64) ||
17491 (FrameReg == X86::EBP && PtrVT == MVT::i32)) &&
17492 "Invalid Frame Register!");
17493 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, PtrVT);
17494 unsigned StoreAddrReg = (PtrVT == MVT::i64) ? X86::RCX : X86::ECX;
17496 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Frame,
17497 DAG.getIntPtrConstant(RegInfo->getSlotSize(),
17499 StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, StoreAddr, Offset);
17500 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
17502 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
17504 return DAG.getNode(X86ISD::EH_RETURN, dl, MVT::Other, Chain,
17505 DAG.getRegister(StoreAddrReg, PtrVT));
17508 SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
17509 SelectionDAG &DAG) const {
17511 return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL,
17512 DAG.getVTList(MVT::i32, MVT::Other),
17513 Op.getOperand(0), Op.getOperand(1));
17516 SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
17517 SelectionDAG &DAG) const {
17519 return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
17520 Op.getOperand(0), Op.getOperand(1));
17523 static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
17524 return Op.getOperand(0);
17527 SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
17528 SelectionDAG &DAG) const {
17529 SDValue Root = Op.getOperand(0);
17530 SDValue Trmp = Op.getOperand(1); // trampoline
17531 SDValue FPtr = Op.getOperand(2); // nested function
17532 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
17535 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
17536 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
17538 if (Subtarget->is64Bit()) {
17539 SDValue OutChains[6];
17541 // Large code-model.
17542 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
17543 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
17545 const unsigned char N86R10 = TRI->getEncodingValue(X86::R10) & 0x7;
17546 const unsigned char N86R11 = TRI->getEncodingValue(X86::R11) & 0x7;
17548 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
17550 // Load the pointer to the nested function into R11.
17551 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
17552 SDValue Addr = Trmp;
17553 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17554 Addr, MachinePointerInfo(TrmpAddr),
17557 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17558 DAG.getConstant(2, dl, MVT::i64));
17559 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
17560 MachinePointerInfo(TrmpAddr, 2),
17563 // Load the 'nest' parameter value into R10.
17564 // R10 is specified in X86CallingConv.td
17565 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
17566 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17567 DAG.getConstant(10, dl, MVT::i64));
17568 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17569 Addr, MachinePointerInfo(TrmpAddr, 10),
17572 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17573 DAG.getConstant(12, dl, MVT::i64));
17574 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
17575 MachinePointerInfo(TrmpAddr, 12),
17578 // Jump to the nested function.
17579 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
17580 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17581 DAG.getConstant(20, dl, MVT::i64));
17582 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17583 Addr, MachinePointerInfo(TrmpAddr, 20),
17586 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
17587 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17588 DAG.getConstant(22, dl, MVT::i64));
17589 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, dl, MVT::i8),
17590 Addr, MachinePointerInfo(TrmpAddr, 22),
17593 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
17595 const Function *Func =
17596 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
17597 CallingConv::ID CC = Func->getCallingConv();
17602 llvm_unreachable("Unsupported calling convention");
17603 case CallingConv::C:
17604 case CallingConv::X86_StdCall: {
17605 // Pass 'nest' parameter in ECX.
17606 // Must be kept in sync with X86CallingConv.td
17607 NestReg = X86::ECX;
17609 // Check that ECX wasn't needed by an 'inreg' parameter.
17610 FunctionType *FTy = Func->getFunctionType();
17611 const AttributeSet &Attrs = Func->getAttributes();
17613 if (!Attrs.isEmpty() && !Func->isVarArg()) {
17614 unsigned InRegCount = 0;
17617 for (FunctionType::param_iterator I = FTy->param_begin(),
17618 E = FTy->param_end(); I != E; ++I, ++Idx)
17619 if (Attrs.hasAttribute(Idx, Attribute::InReg)) {
17620 auto &DL = DAG.getDataLayout();
17621 // FIXME: should only count parameters that are lowered to integers.
17622 InRegCount += (DL.getTypeSizeInBits(*I) + 31) / 32;
17625 if (InRegCount > 2) {
17626 report_fatal_error("Nest register in use - reduce number of inreg"
17632 case CallingConv::X86_FastCall:
17633 case CallingConv::X86_ThisCall:
17634 case CallingConv::Fast:
17635 // Pass 'nest' parameter in EAX.
17636 // Must be kept in sync with X86CallingConv.td
17637 NestReg = X86::EAX;
17641 SDValue OutChains[4];
17642 SDValue Addr, Disp;
17644 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17645 DAG.getConstant(10, dl, MVT::i32));
17646 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
17648 // This is storing the opcode for MOV32ri.
17649 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
17650 const unsigned char N86Reg = TRI->getEncodingValue(NestReg) & 0x7;
17651 OutChains[0] = DAG.getStore(Root, dl,
17652 DAG.getConstant(MOV32ri|N86Reg, dl, MVT::i8),
17653 Trmp, MachinePointerInfo(TrmpAddr),
17656 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17657 DAG.getConstant(1, dl, MVT::i32));
17658 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
17659 MachinePointerInfo(TrmpAddr, 1),
17662 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
17663 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17664 DAG.getConstant(5, dl, MVT::i32));
17665 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, dl, MVT::i8),
17666 Addr, MachinePointerInfo(TrmpAddr, 5),
17669 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17670 DAG.getConstant(6, dl, MVT::i32));
17671 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
17672 MachinePointerInfo(TrmpAddr, 6),
17675 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
17679 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
17680 SelectionDAG &DAG) const {
17682 The rounding mode is in bits 11:10 of FPSR, and has the following
17684 00 Round to nearest
17689 FLT_ROUNDS, on the other hand, expects the following:
17696 To perform the conversion, we do:
17697 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
17700 MachineFunction &MF = DAG.getMachineFunction();
17701 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
17702 unsigned StackAlignment = TFI.getStackAlignment();
17703 MVT VT = Op.getSimpleValueType();
17706 // Save FP Control Word to stack slot
17707 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
17708 SDValue StackSlot =
17709 DAG.getFrameIndex(SSFI, getPointerTy(DAG.getDataLayout()));
17711 MachineMemOperand *MMO =
17712 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
17713 MachineMemOperand::MOStore, 2, 2);
17715 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
17716 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
17717 DAG.getVTList(MVT::Other),
17718 Ops, MVT::i16, MMO);
17720 // Load FP Control Word from stack slot
17721 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
17722 MachinePointerInfo(), false, false, false, 0);
17724 // Transform as necessary
17726 DAG.getNode(ISD::SRL, DL, MVT::i16,
17727 DAG.getNode(ISD::AND, DL, MVT::i16,
17728 CWD, DAG.getConstant(0x800, DL, MVT::i16)),
17729 DAG.getConstant(11, DL, MVT::i8));
17731 DAG.getNode(ISD::SRL, DL, MVT::i16,
17732 DAG.getNode(ISD::AND, DL, MVT::i16,
17733 CWD, DAG.getConstant(0x400, DL, MVT::i16)),
17734 DAG.getConstant(9, DL, MVT::i8));
17737 DAG.getNode(ISD::AND, DL, MVT::i16,
17738 DAG.getNode(ISD::ADD, DL, MVT::i16,
17739 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
17740 DAG.getConstant(1, DL, MVT::i16)),
17741 DAG.getConstant(3, DL, MVT::i16));
17743 return DAG.getNode((VT.getSizeInBits() < 16 ?
17744 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
17747 /// \brief Lower a vector CTLZ using native supported vector CTLZ instruction.
17749 // 1. i32/i64 128/256-bit vector (native support require VLX) are expended
17750 // to 512-bit vector.
17751 // 2. i8/i16 vector implemented using dword LZCNT vector instruction
17752 // ( sub(trunc(lzcnt(zext32(x)))) ). In case zext32(x) is illegal,
17753 // split the vector, perform operation on it's Lo a Hi part and
17754 // concatenate the results.
17755 static SDValue LowerVectorCTLZ_AVX512(SDValue Op, SelectionDAG &DAG) {
17757 MVT VT = Op.getSimpleValueType();
17758 MVT EltVT = VT.getVectorElementType();
17759 unsigned NumElems = VT.getVectorNumElements();
17761 if (EltVT == MVT::i64 || EltVT == MVT::i32) {
17762 // Extend to 512 bit vector.
17763 assert((VT.is256BitVector() || VT.is128BitVector()) &&
17764 "Unsupported value type for operation");
17766 MVT NewVT = MVT::getVectorVT(EltVT, 512 / VT.getScalarSizeInBits());
17767 SDValue Vec512 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, NewVT,
17768 DAG.getUNDEF(NewVT),
17770 DAG.getIntPtrConstant(0, dl));
17771 SDValue CtlzNode = DAG.getNode(ISD::CTLZ, dl, NewVT, Vec512);
17773 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, CtlzNode,
17774 DAG.getIntPtrConstant(0, dl));
17777 assert((EltVT == MVT::i8 || EltVT == MVT::i16) &&
17778 "Unsupported element type");
17780 if (16 < NumElems) {
17781 // Split vector, it's Lo and Hi parts will be handled in next iteration.
17783 std::tie(Lo, Hi) = DAG.SplitVector(Op.getOperand(0), dl);
17784 MVT OutVT = MVT::getVectorVT(EltVT, NumElems/2);
17786 Lo = DAG.getNode(Op.getOpcode(), dl, OutVT, Lo);
17787 Hi = DAG.getNode(Op.getOpcode(), dl, OutVT, Hi);
17789 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Lo, Hi);
17792 MVT NewVT = MVT::getVectorVT(MVT::i32, NumElems);
17794 assert((NewVT.is256BitVector() || NewVT.is512BitVector()) &&
17795 "Unsupported value type for operation");
17797 // Use native supported vector instruction vplzcntd.
17798 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, NewVT, Op.getOperand(0));
17799 SDValue CtlzNode = DAG.getNode(ISD::CTLZ, dl, NewVT, Op);
17800 SDValue TruncNode = DAG.getNode(ISD::TRUNCATE, dl, VT, CtlzNode);
17801 SDValue Delta = DAG.getConstant(32 - EltVT.getSizeInBits(), dl, VT);
17803 return DAG.getNode(ISD::SUB, dl, VT, TruncNode, Delta);
17806 static SDValue LowerCTLZ(SDValue Op, const X86Subtarget *Subtarget,
17807 SelectionDAG &DAG) {
17808 MVT VT = Op.getSimpleValueType();
17810 unsigned NumBits = VT.getSizeInBits();
17813 if (VT.isVector() && Subtarget->hasAVX512())
17814 return LowerVectorCTLZ_AVX512(Op, DAG);
17816 Op = Op.getOperand(0);
17817 if (VT == MVT::i8) {
17818 // Zero extend to i32 since there is not an i8 bsr.
17820 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
17823 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
17824 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
17825 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
17827 // If src is zero (i.e. bsr sets ZF), returns NumBits.
17830 DAG.getConstant(NumBits + NumBits - 1, dl, OpVT),
17831 DAG.getConstant(X86::COND_E, dl, MVT::i8),
17834 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops);
17836 // Finally xor with NumBits-1.
17837 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
17838 DAG.getConstant(NumBits - 1, dl, OpVT));
17841 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
17845 static SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, const X86Subtarget *Subtarget,
17846 SelectionDAG &DAG) {
17847 MVT VT = Op.getSimpleValueType();
17849 unsigned NumBits = VT.getSizeInBits();
17852 if (VT.isVector() && Subtarget->hasAVX512())
17853 return LowerVectorCTLZ_AVX512(Op, DAG);
17855 Op = Op.getOperand(0);
17856 if (VT == MVT::i8) {
17857 // Zero extend to i32 since there is not an i8 bsr.
17859 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
17862 // Issue a bsr (scan bits in reverse).
17863 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
17864 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
17866 // And xor with NumBits-1.
17867 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
17868 DAG.getConstant(NumBits - 1, dl, OpVT));
17871 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
17875 static SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
17876 MVT VT = Op.getSimpleValueType();
17877 unsigned NumBits = VT.getScalarSizeInBits();
17880 if (VT.isVector()) {
17881 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
17883 SDValue N0 = Op.getOperand(0);
17884 SDValue Zero = DAG.getConstant(0, dl, VT);
17886 // lsb(x) = (x & -x)
17887 SDValue LSB = DAG.getNode(ISD::AND, dl, VT, N0,
17888 DAG.getNode(ISD::SUB, dl, VT, Zero, N0));
17890 // cttz_undef(x) = (width - 1) - ctlz(lsb)
17891 if (Op.getOpcode() == ISD::CTTZ_ZERO_UNDEF &&
17892 TLI.isOperationLegal(ISD::CTLZ, VT)) {
17893 SDValue WidthMinusOne = DAG.getConstant(NumBits - 1, dl, VT);
17894 return DAG.getNode(ISD::SUB, dl, VT, WidthMinusOne,
17895 DAG.getNode(ISD::CTLZ, dl, VT, LSB));
17898 // cttz(x) = ctpop(lsb - 1)
17899 SDValue One = DAG.getConstant(1, dl, VT);
17900 return DAG.getNode(ISD::CTPOP, dl, VT,
17901 DAG.getNode(ISD::SUB, dl, VT, LSB, One));
17904 assert(Op.getOpcode() == ISD::CTTZ &&
17905 "Only scalar CTTZ requires custom lowering");
17907 // Issue a bsf (scan bits forward) which also sets EFLAGS.
17908 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
17909 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op.getOperand(0));
17911 // If src is zero (i.e. bsf sets ZF), returns NumBits.
17914 DAG.getConstant(NumBits, dl, VT),
17915 DAG.getConstant(X86::COND_E, dl, MVT::i8),
17918 return DAG.getNode(X86ISD::CMOV, dl, VT, Ops);
17921 // Lower256IntArith - Break a 256-bit integer operation into two new 128-bit
17922 // ones, and then concatenate the result back.
17923 static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) {
17924 MVT VT = Op.getSimpleValueType();
17926 assert(VT.is256BitVector() && VT.isInteger() &&
17927 "Unsupported value type for operation");
17929 unsigned NumElems = VT.getVectorNumElements();
17932 // Extract the LHS vectors
17933 SDValue LHS = Op.getOperand(0);
17934 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
17935 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
17937 // Extract the RHS vectors
17938 SDValue RHS = Op.getOperand(1);
17939 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
17940 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
17942 MVT EltVT = VT.getVectorElementType();
17943 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
17945 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
17946 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1),
17947 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
17950 static SDValue LowerADD(SDValue Op, SelectionDAG &DAG) {
17951 if (Op.getValueType() == MVT::i1)
17952 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
17953 Op.getOperand(0), Op.getOperand(1));
17954 assert(Op.getSimpleValueType().is256BitVector() &&
17955 Op.getSimpleValueType().isInteger() &&
17956 "Only handle AVX 256-bit vector integer operation");
17957 return Lower256IntArith(Op, DAG);
17960 static SDValue LowerSUB(SDValue Op, SelectionDAG &DAG) {
17961 if (Op.getValueType() == MVT::i1)
17962 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
17963 Op.getOperand(0), Op.getOperand(1));
17964 assert(Op.getSimpleValueType().is256BitVector() &&
17965 Op.getSimpleValueType().isInteger() &&
17966 "Only handle AVX 256-bit vector integer operation");
17967 return Lower256IntArith(Op, DAG);
17970 static SDValue LowerMINMAX(SDValue Op, SelectionDAG &DAG) {
17971 assert(Op.getSimpleValueType().is256BitVector() &&
17972 Op.getSimpleValueType().isInteger() &&
17973 "Only handle AVX 256-bit vector integer operation");
17974 return Lower256IntArith(Op, DAG);
17977 static SDValue LowerMUL(SDValue Op, const X86Subtarget *Subtarget,
17978 SelectionDAG &DAG) {
17980 MVT VT = Op.getSimpleValueType();
17983 return DAG.getNode(ISD::AND, dl, VT, Op.getOperand(0), Op.getOperand(1));
17985 // Decompose 256-bit ops into smaller 128-bit ops.
17986 if (VT.is256BitVector() && !Subtarget->hasInt256())
17987 return Lower256IntArith(Op, DAG);
17989 SDValue A = Op.getOperand(0);
17990 SDValue B = Op.getOperand(1);
17992 // Lower v16i8/v32i8 mul as promotion to v8i16/v16i16 vector
17993 // pairs, multiply and truncate.
17994 if (VT == MVT::v16i8 || VT == MVT::v32i8) {
17995 if (Subtarget->hasInt256()) {
17996 if (VT == MVT::v32i8) {
17997 MVT SubVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() / 2);
17998 SDValue Lo = DAG.getIntPtrConstant(0, dl);
17999 SDValue Hi = DAG.getIntPtrConstant(VT.getVectorNumElements() / 2, dl);
18000 SDValue ALo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Lo);
18001 SDValue BLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Lo);
18002 SDValue AHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Hi);
18003 SDValue BHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Hi);
18004 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
18005 DAG.getNode(ISD::MUL, dl, SubVT, ALo, BLo),
18006 DAG.getNode(ISD::MUL, dl, SubVT, AHi, BHi));
18009 MVT ExVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements());
18010 return DAG.getNode(
18011 ISD::TRUNCATE, dl, VT,
18012 DAG.getNode(ISD::MUL, dl, ExVT,
18013 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, A),
18014 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, B)));
18017 assert(VT == MVT::v16i8 &&
18018 "Pre-AVX2 support only supports v16i8 multiplication");
18019 MVT ExVT = MVT::v8i16;
18021 // Extract the lo parts and sign extend to i16
18023 if (Subtarget->hasSSE41()) {
18024 ALo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, A);
18025 BLo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, B);
18027 const int ShufMask[] = {-1, 0, -1, 1, -1, 2, -1, 3,
18028 -1, 4, -1, 5, -1, 6, -1, 7};
18029 ALo = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
18030 BLo = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
18031 ALo = DAG.getBitcast(ExVT, ALo);
18032 BLo = DAG.getBitcast(ExVT, BLo);
18033 ALo = DAG.getNode(ISD::SRA, dl, ExVT, ALo, DAG.getConstant(8, dl, ExVT));
18034 BLo = DAG.getNode(ISD::SRA, dl, ExVT, BLo, DAG.getConstant(8, dl, ExVT));
18037 // Extract the hi parts and sign extend to i16
18039 if (Subtarget->hasSSE41()) {
18040 const int ShufMask[] = {8, 9, 10, 11, 12, 13, 14, 15,
18041 -1, -1, -1, -1, -1, -1, -1, -1};
18042 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
18043 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
18044 AHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, AHi);
18045 BHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, BHi);
18047 const int ShufMask[] = {-1, 8, -1, 9, -1, 10, -1, 11,
18048 -1, 12, -1, 13, -1, 14, -1, 15};
18049 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
18050 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
18051 AHi = DAG.getBitcast(ExVT, AHi);
18052 BHi = DAG.getBitcast(ExVT, BHi);
18053 AHi = DAG.getNode(ISD::SRA, dl, ExVT, AHi, DAG.getConstant(8, dl, ExVT));
18054 BHi = DAG.getNode(ISD::SRA, dl, ExVT, BHi, DAG.getConstant(8, dl, ExVT));
18057 // Multiply, mask the lower 8bits of the lo/hi results and pack
18058 SDValue RLo = DAG.getNode(ISD::MUL, dl, ExVT, ALo, BLo);
18059 SDValue RHi = DAG.getNode(ISD::MUL, dl, ExVT, AHi, BHi);
18060 RLo = DAG.getNode(ISD::AND, dl, ExVT, RLo, DAG.getConstant(255, dl, ExVT));
18061 RHi = DAG.getNode(ISD::AND, dl, ExVT, RHi, DAG.getConstant(255, dl, ExVT));
18062 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
18065 // Lower v4i32 mul as 2x shuffle, 2x pmuludq, 2x shuffle.
18066 if (VT == MVT::v4i32) {
18067 assert(Subtarget->hasSSE2() && !Subtarget->hasSSE41() &&
18068 "Should not custom lower when pmuldq is available!");
18070 // Extract the odd parts.
18071 static const int UnpackMask[] = { 1, -1, 3, -1 };
18072 SDValue Aodds = DAG.getVectorShuffle(VT, dl, A, A, UnpackMask);
18073 SDValue Bodds = DAG.getVectorShuffle(VT, dl, B, B, UnpackMask);
18075 // Multiply the even parts.
18076 SDValue Evens = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, A, B);
18077 // Now multiply odd parts.
18078 SDValue Odds = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, Aodds, Bodds);
18080 Evens = DAG.getBitcast(VT, Evens);
18081 Odds = DAG.getBitcast(VT, Odds);
18083 // Merge the two vectors back together with a shuffle. This expands into 2
18085 static const int ShufMask[] = { 0, 4, 2, 6 };
18086 return DAG.getVectorShuffle(VT, dl, Evens, Odds, ShufMask);
18089 assert((VT == MVT::v2i64 || VT == MVT::v4i64 || VT == MVT::v8i64) &&
18090 "Only know how to lower V2I64/V4I64/V8I64 multiply");
18092 // Ahi = psrlqi(a, 32);
18093 // Bhi = psrlqi(b, 32);
18095 // AloBlo = pmuludq(a, b);
18096 // AloBhi = pmuludq(a, Bhi);
18097 // AhiBlo = pmuludq(Ahi, b);
18099 // AloBhi = psllqi(AloBhi, 32);
18100 // AhiBlo = psllqi(AhiBlo, 32);
18101 // return AloBlo + AloBhi + AhiBlo;
18103 SDValue Ahi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, A, 32, DAG);
18104 SDValue Bhi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, B, 32, DAG);
18106 SDValue AhiBlo = Ahi;
18107 SDValue AloBhi = Bhi;
18108 // Bit cast to 32-bit vectors for MULUDQ
18109 MVT MulVT = (VT == MVT::v2i64) ? MVT::v4i32 :
18110 (VT == MVT::v4i64) ? MVT::v8i32 : MVT::v16i32;
18111 A = DAG.getBitcast(MulVT, A);
18112 B = DAG.getBitcast(MulVT, B);
18113 Ahi = DAG.getBitcast(MulVT, Ahi);
18114 Bhi = DAG.getBitcast(MulVT, Bhi);
18116 SDValue AloBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, B);
18117 // After shifting right const values the result may be all-zero.
18118 if (!ISD::isBuildVectorAllZeros(Ahi.getNode())) {
18119 AhiBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Ahi, B);
18120 AhiBlo = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AhiBlo, 32, DAG);
18122 if (!ISD::isBuildVectorAllZeros(Bhi.getNode())) {
18123 AloBhi = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, Bhi);
18124 AloBhi = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AloBhi, 32, DAG);
18127 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
18128 return DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
18131 SDValue X86TargetLowering::LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const {
18132 assert(Subtarget->isTargetWin64() && "Unexpected target");
18133 EVT VT = Op.getValueType();
18134 assert(VT.isInteger() && VT.getSizeInBits() == 128 &&
18135 "Unexpected return type for lowering");
18139 switch (Op->getOpcode()) {
18140 default: llvm_unreachable("Unexpected request for libcall!");
18141 case ISD::SDIV: isSigned = true; LC = RTLIB::SDIV_I128; break;
18142 case ISD::UDIV: isSigned = false; LC = RTLIB::UDIV_I128; break;
18143 case ISD::SREM: isSigned = true; LC = RTLIB::SREM_I128; break;
18144 case ISD::UREM: isSigned = false; LC = RTLIB::UREM_I128; break;
18145 case ISD::SDIVREM: isSigned = true; LC = RTLIB::SDIVREM_I128; break;
18146 case ISD::UDIVREM: isSigned = false; LC = RTLIB::UDIVREM_I128; break;
18150 SDValue InChain = DAG.getEntryNode();
18152 TargetLowering::ArgListTy Args;
18153 TargetLowering::ArgListEntry Entry;
18154 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
18155 EVT ArgVT = Op->getOperand(i).getValueType();
18156 assert(ArgVT.isInteger() && ArgVT.getSizeInBits() == 128 &&
18157 "Unexpected argument type for lowering");
18158 SDValue StackPtr = DAG.CreateStackTemporary(ArgVT, 16);
18159 Entry.Node = StackPtr;
18160 InChain = DAG.getStore(InChain, dl, Op->getOperand(i), StackPtr, MachinePointerInfo(),
18162 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
18163 Entry.Ty = PointerType::get(ArgTy,0);
18164 Entry.isSExt = false;
18165 Entry.isZExt = false;
18166 Args.push_back(Entry);
18169 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
18170 getPointerTy(DAG.getDataLayout()));
18172 TargetLowering::CallLoweringInfo CLI(DAG);
18173 CLI.setDebugLoc(dl).setChain(InChain)
18174 .setCallee(getLibcallCallingConv(LC),
18175 static_cast<EVT>(MVT::v2i64).getTypeForEVT(*DAG.getContext()),
18176 Callee, std::move(Args), 0)
18177 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
18179 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
18180 return DAG.getBitcast(VT, CallInfo.first);
18183 static SDValue LowerMUL_LOHI(SDValue Op, const X86Subtarget *Subtarget,
18184 SelectionDAG &DAG) {
18185 SDValue Op0 = Op.getOperand(0), Op1 = Op.getOperand(1);
18186 MVT VT = Op0.getSimpleValueType();
18189 assert((VT == MVT::v4i32 && Subtarget->hasSSE2()) ||
18190 (VT == MVT::v8i32 && Subtarget->hasInt256()));
18192 // PMULxD operations multiply each even value (starting at 0) of LHS with
18193 // the related value of RHS and produce a widen result.
18194 // E.g., PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
18195 // => <2 x i64> <ae|cg>
18197 // In other word, to have all the results, we need to perform two PMULxD:
18198 // 1. one with the even values.
18199 // 2. one with the odd values.
18200 // To achieve #2, with need to place the odd values at an even position.
18202 // Place the odd value at an even position (basically, shift all values 1
18203 // step to the left):
18204 const int Mask[] = {1, -1, 3, -1, 5, -1, 7, -1};
18205 // <a|b|c|d> => <b|undef|d|undef>
18206 SDValue Odd0 = DAG.getVectorShuffle(VT, dl, Op0, Op0, Mask);
18207 // <e|f|g|h> => <f|undef|h|undef>
18208 SDValue Odd1 = DAG.getVectorShuffle(VT, dl, Op1, Op1, Mask);
18210 // Emit two multiplies, one for the lower 2 ints and one for the higher 2
18212 MVT MulVT = VT == MVT::v4i32 ? MVT::v2i64 : MVT::v4i64;
18213 bool IsSigned = Op->getOpcode() == ISD::SMUL_LOHI;
18215 (!IsSigned || !Subtarget->hasSSE41()) ? X86ISD::PMULUDQ : X86ISD::PMULDQ;
18216 // PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
18217 // => <2 x i64> <ae|cg>
18218 SDValue Mul1 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Op0, Op1));
18219 // PMULUDQ <4 x i32> <b|undef|d|undef>, <4 x i32> <f|undef|h|undef>
18220 // => <2 x i64> <bf|dh>
18221 SDValue Mul2 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Odd0, Odd1));
18223 // Shuffle it back into the right order.
18224 SDValue Highs, Lows;
18225 if (VT == MVT::v8i32) {
18226 const int HighMask[] = {1, 9, 3, 11, 5, 13, 7, 15};
18227 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
18228 const int LowMask[] = {0, 8, 2, 10, 4, 12, 6, 14};
18229 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
18231 const int HighMask[] = {1, 5, 3, 7};
18232 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
18233 const int LowMask[] = {0, 4, 2, 6};
18234 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
18237 // If we have a signed multiply but no PMULDQ fix up the high parts of a
18238 // unsigned multiply.
18239 if (IsSigned && !Subtarget->hasSSE41()) {
18240 SDValue ShAmt = DAG.getConstant(
18242 DAG.getTargetLoweringInfo().getShiftAmountTy(VT, DAG.getDataLayout()));
18243 SDValue T1 = DAG.getNode(ISD::AND, dl, VT,
18244 DAG.getNode(ISD::SRA, dl, VT, Op0, ShAmt), Op1);
18245 SDValue T2 = DAG.getNode(ISD::AND, dl, VT,
18246 DAG.getNode(ISD::SRA, dl, VT, Op1, ShAmt), Op0);
18248 SDValue Fixup = DAG.getNode(ISD::ADD, dl, VT, T1, T2);
18249 Highs = DAG.getNode(ISD::SUB, dl, VT, Highs, Fixup);
18252 // The first result of MUL_LOHI is actually the low value, followed by the
18254 SDValue Ops[] = {Lows, Highs};
18255 return DAG.getMergeValues(Ops, dl);
18258 // Return true if the required (according to Opcode) shift-imm form is natively
18259 // supported by the Subtarget
18260 static bool SupportedVectorShiftWithImm(MVT VT, const X86Subtarget *Subtarget,
18262 if (VT.getScalarSizeInBits() < 16)
18265 if (VT.is512BitVector() &&
18266 (VT.getScalarSizeInBits() > 16 || Subtarget->hasBWI()))
18269 bool LShift = VT.is128BitVector() ||
18270 (VT.is256BitVector() && Subtarget->hasInt256());
18272 bool AShift = LShift && (Subtarget->hasVLX() ||
18273 (VT != MVT::v2i64 && VT != MVT::v4i64));
18274 return (Opcode == ISD::SRA) ? AShift : LShift;
18277 // The shift amount is a variable, but it is the same for all vector lanes.
18278 // These instructions are defined together with shift-immediate.
18280 bool SupportedVectorShiftWithBaseAmnt(MVT VT, const X86Subtarget *Subtarget,
18282 return SupportedVectorShiftWithImm(VT, Subtarget, Opcode);
18285 // Return true if the required (according to Opcode) variable-shift form is
18286 // natively supported by the Subtarget
18287 static bool SupportedVectorVarShift(MVT VT, const X86Subtarget *Subtarget,
18290 if (!Subtarget->hasInt256() || VT.getScalarSizeInBits() < 16)
18293 // vXi16 supported only on AVX-512, BWI
18294 if (VT.getScalarSizeInBits() == 16 && !Subtarget->hasBWI())
18297 if (VT.is512BitVector() || Subtarget->hasVLX())
18300 bool LShift = VT.is128BitVector() || VT.is256BitVector();
18301 bool AShift = LShift && VT != MVT::v2i64 && VT != MVT::v4i64;
18302 return (Opcode == ISD::SRA) ? AShift : LShift;
18305 static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
18306 const X86Subtarget *Subtarget) {
18307 MVT VT = Op.getSimpleValueType();
18309 SDValue R = Op.getOperand(0);
18310 SDValue Amt = Op.getOperand(1);
18312 unsigned X86Opc = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
18313 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
18315 auto ArithmeticShiftRight64 = [&](uint64_t ShiftAmt) {
18316 assert((VT == MVT::v2i64 || VT == MVT::v4i64) && "Unexpected SRA type");
18317 MVT ExVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() * 2);
18318 SDValue Ex = DAG.getBitcast(ExVT, R);
18320 if (ShiftAmt >= 32) {
18321 // Splat sign to upper i32 dst, and SRA upper i32 src to lower i32.
18323 getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex, 31, DAG);
18324 SDValue Lower = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex,
18325 ShiftAmt - 32, DAG);
18326 if (VT == MVT::v2i64)
18327 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {5, 1, 7, 3});
18328 if (VT == MVT::v4i64)
18329 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower,
18330 {9, 1, 11, 3, 13, 5, 15, 7});
18332 // SRA upper i32, SHL whole i64 and select lower i32.
18333 SDValue Upper = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex,
18336 getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt, DAG);
18337 Lower = DAG.getBitcast(ExVT, Lower);
18338 if (VT == MVT::v2i64)
18339 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {4, 1, 6, 3});
18340 if (VT == MVT::v4i64)
18341 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower,
18342 {8, 1, 10, 3, 12, 5, 14, 7});
18344 return DAG.getBitcast(VT, Ex);
18347 // Optimize shl/srl/sra with constant shift amount.
18348 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
18349 if (auto *ShiftConst = BVAmt->getConstantSplatNode()) {
18350 uint64_t ShiftAmt = ShiftConst->getZExtValue();
18352 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
18353 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
18355 // i64 SRA needs to be performed as partial shifts.
18356 if ((VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
18357 Op.getOpcode() == ISD::SRA && !Subtarget->hasXOP())
18358 return ArithmeticShiftRight64(ShiftAmt);
18360 if (VT == MVT::v16i8 || (Subtarget->hasInt256() && VT == MVT::v32i8)) {
18361 unsigned NumElts = VT.getVectorNumElements();
18362 MVT ShiftVT = MVT::getVectorVT(MVT::i16, NumElts / 2);
18364 // Simple i8 add case
18365 if (Op.getOpcode() == ISD::SHL && ShiftAmt == 1)
18366 return DAG.getNode(ISD::ADD, dl, VT, R, R);
18368 // ashr(R, 7) === cmp_slt(R, 0)
18369 if (Op.getOpcode() == ISD::SRA && ShiftAmt == 7) {
18370 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
18371 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
18374 // XOP can shift v16i8 directly instead of as shift v8i16 + mask.
18375 if (VT == MVT::v16i8 && Subtarget->hasXOP())
18378 if (Op.getOpcode() == ISD::SHL) {
18379 // Make a large shift.
18380 SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, ShiftVT,
18382 SHL = DAG.getBitcast(VT, SHL);
18383 // Zero out the rightmost bits.
18384 SmallVector<SDValue, 32> V(
18385 NumElts, DAG.getConstant(uint8_t(-1U << ShiftAmt), dl, MVT::i8));
18386 return DAG.getNode(ISD::AND, dl, VT, SHL,
18387 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
18389 if (Op.getOpcode() == ISD::SRL) {
18390 // Make a large shift.
18391 SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, ShiftVT,
18393 SRL = DAG.getBitcast(VT, SRL);
18394 // Zero out the leftmost bits.
18395 SmallVector<SDValue, 32> V(
18396 NumElts, DAG.getConstant(uint8_t(-1U) >> ShiftAmt, dl, MVT::i8));
18397 return DAG.getNode(ISD::AND, dl, VT, SRL,
18398 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
18400 if (Op.getOpcode() == ISD::SRA) {
18401 // ashr(R, Amt) === sub(xor(lshr(R, Amt), Mask), Mask)
18402 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
18403 SmallVector<SDValue, 32> V(NumElts,
18404 DAG.getConstant(128 >> ShiftAmt, dl,
18406 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V);
18407 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
18408 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
18411 llvm_unreachable("Unknown shift opcode.");
18416 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
18417 if (!Subtarget->is64Bit() && !Subtarget->hasXOP() &&
18418 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64))) {
18420 // Peek through any splat that was introduced for i64 shift vectorization.
18421 int SplatIndex = -1;
18422 if (ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt.getNode()))
18423 if (SVN->isSplat()) {
18424 SplatIndex = SVN->getSplatIndex();
18425 Amt = Amt.getOperand(0);
18426 assert(SplatIndex < (int)VT.getVectorNumElements() &&
18427 "Splat shuffle referencing second operand");
18430 if (Amt.getOpcode() != ISD::BITCAST ||
18431 Amt.getOperand(0).getOpcode() != ISD::BUILD_VECTOR)
18434 Amt = Amt.getOperand(0);
18435 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
18436 VT.getVectorNumElements();
18437 unsigned RatioInLog2 = Log2_32_Ceil(Ratio);
18438 uint64_t ShiftAmt = 0;
18439 unsigned BaseOp = (SplatIndex < 0 ? 0 : SplatIndex * Ratio);
18440 for (unsigned i = 0; i != Ratio; ++i) {
18441 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i + BaseOp));
18445 ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2)));
18448 // Check remaining shift amounts (if not a splat).
18449 if (SplatIndex < 0) {
18450 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
18451 uint64_t ShAmt = 0;
18452 for (unsigned j = 0; j != Ratio; ++j) {
18453 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i + j));
18457 ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2)));
18459 if (ShAmt != ShiftAmt)
18464 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
18465 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
18467 if (Op.getOpcode() == ISD::SRA)
18468 return ArithmeticShiftRight64(ShiftAmt);
18474 static SDValue LowerScalarVariableShift(SDValue Op, SelectionDAG &DAG,
18475 const X86Subtarget* Subtarget) {
18476 MVT VT = Op.getSimpleValueType();
18478 SDValue R = Op.getOperand(0);
18479 SDValue Amt = Op.getOperand(1);
18481 unsigned X86OpcI = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
18482 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
18484 unsigned X86OpcV = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHL :
18485 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRL : X86ISD::VSRA;
18487 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode())) {
18489 MVT EltVT = VT.getVectorElementType();
18491 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Amt)) {
18492 // Check if this build_vector node is doing a splat.
18493 // If so, then set BaseShAmt equal to the splat value.
18494 BaseShAmt = BV->getSplatValue();
18495 if (BaseShAmt && BaseShAmt.getOpcode() == ISD::UNDEF)
18496 BaseShAmt = SDValue();
18498 if (Amt.getOpcode() == ISD::EXTRACT_SUBVECTOR)
18499 Amt = Amt.getOperand(0);
18501 ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt);
18502 if (SVN && SVN->isSplat()) {
18503 unsigned SplatIdx = (unsigned)SVN->getSplatIndex();
18504 SDValue InVec = Amt.getOperand(0);
18505 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
18506 assert((SplatIdx < InVec.getSimpleValueType().getVectorNumElements()) &&
18507 "Unexpected shuffle index found!");
18508 BaseShAmt = InVec.getOperand(SplatIdx);
18509 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
18510 if (ConstantSDNode *C =
18511 dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
18512 if (C->getZExtValue() == SplatIdx)
18513 BaseShAmt = InVec.getOperand(1);
18518 // Avoid introducing an extract element from a shuffle.
18519 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InVec,
18520 DAG.getIntPtrConstant(SplatIdx, dl));
18524 if (BaseShAmt.getNode()) {
18525 assert(EltVT.bitsLE(MVT::i64) && "Unexpected element type!");
18526 if (EltVT != MVT::i64 && EltVT.bitsGT(MVT::i32))
18527 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, BaseShAmt);
18528 else if (EltVT.bitsLT(MVT::i32))
18529 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt);
18531 return getTargetVShiftNode(X86OpcI, dl, VT, R, BaseShAmt, DAG);
18535 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
18536 if (!Subtarget->is64Bit() && VT == MVT::v2i64 &&
18537 Amt.getOpcode() == ISD::BITCAST &&
18538 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
18539 Amt = Amt.getOperand(0);
18540 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
18541 VT.getVectorNumElements();
18542 std::vector<SDValue> Vals(Ratio);
18543 for (unsigned i = 0; i != Ratio; ++i)
18544 Vals[i] = Amt.getOperand(i);
18545 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
18546 for (unsigned j = 0; j != Ratio; ++j)
18547 if (Vals[j] != Amt.getOperand(i + j))
18551 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode()))
18552 return DAG.getNode(X86OpcV, dl, VT, R, Op.getOperand(1));
18557 static SDValue LowerShift(SDValue Op, const X86Subtarget* Subtarget,
18558 SelectionDAG &DAG) {
18559 MVT VT = Op.getSimpleValueType();
18561 SDValue R = Op.getOperand(0);
18562 SDValue Amt = Op.getOperand(1);
18564 assert(VT.isVector() && "Custom lowering only for vector shifts!");
18565 assert(Subtarget->hasSSE2() && "Only custom lower when we have SSE2!");
18567 if (SDValue V = LowerScalarImmediateShift(Op, DAG, Subtarget))
18570 if (SDValue V = LowerScalarVariableShift(Op, DAG, Subtarget))
18573 if (SupportedVectorVarShift(VT, Subtarget, Op.getOpcode()))
18576 // XOP has 128-bit variable logical/arithmetic shifts.
18577 // +ve/-ve Amt = shift left/right.
18578 if (Subtarget->hasXOP() &&
18579 (VT == MVT::v2i64 || VT == MVT::v4i32 ||
18580 VT == MVT::v8i16 || VT == MVT::v16i8)) {
18581 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SRA) {
18582 SDValue Zero = getZeroVector(VT, Subtarget, DAG, dl);
18583 Amt = DAG.getNode(ISD::SUB, dl, VT, Zero, Amt);
18585 if (Op.getOpcode() == ISD::SHL || Op.getOpcode() == ISD::SRL)
18586 return DAG.getNode(X86ISD::VPSHL, dl, VT, R, Amt);
18587 if (Op.getOpcode() == ISD::SRA)
18588 return DAG.getNode(X86ISD::VPSHA, dl, VT, R, Amt);
18591 // 2i64 vector logical shifts can efficiently avoid scalarization - do the
18592 // shifts per-lane and then shuffle the partial results back together.
18593 if (VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) {
18594 // Splat the shift amounts so the scalar shifts above will catch it.
18595 SDValue Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {0, 0});
18596 SDValue Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {1, 1});
18597 SDValue R0 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt0);
18598 SDValue R1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt1);
18599 return DAG.getVectorShuffle(VT, dl, R0, R1, {0, 3});
18602 // i64 vector arithmetic shift can be emulated with the transform:
18603 // M = lshr(SIGN_BIT, Amt)
18604 // ashr(R, Amt) === sub(xor(lshr(R, Amt), M), M)
18605 if ((VT == MVT::v2i64 || (VT == MVT::v4i64 && Subtarget->hasInt256())) &&
18606 Op.getOpcode() == ISD::SRA) {
18607 SDValue S = DAG.getConstant(APInt::getSignBit(64), dl, VT);
18608 SDValue M = DAG.getNode(ISD::SRL, dl, VT, S, Amt);
18609 R = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
18610 R = DAG.getNode(ISD::XOR, dl, VT, R, M);
18611 R = DAG.getNode(ISD::SUB, dl, VT, R, M);
18615 // If possible, lower this packed shift into a vector multiply instead of
18616 // expanding it into a sequence of scalar shifts.
18617 // Do this only if the vector shift count is a constant build_vector.
18618 if (Op.getOpcode() == ISD::SHL &&
18619 (VT == MVT::v8i16 || VT == MVT::v4i32 ||
18620 (Subtarget->hasInt256() && VT == MVT::v16i16)) &&
18621 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18622 SmallVector<SDValue, 8> Elts;
18623 MVT SVT = VT.getVectorElementType();
18624 unsigned SVTBits = SVT.getSizeInBits();
18625 APInt One(SVTBits, 1);
18626 unsigned NumElems = VT.getVectorNumElements();
18628 for (unsigned i=0; i !=NumElems; ++i) {
18629 SDValue Op = Amt->getOperand(i);
18630 if (Op->getOpcode() == ISD::UNDEF) {
18631 Elts.push_back(Op);
18635 ConstantSDNode *ND = cast<ConstantSDNode>(Op);
18636 APInt C(SVTBits, ND->getAPIntValue().getZExtValue());
18637 uint64_t ShAmt = C.getZExtValue();
18638 if (ShAmt >= SVTBits) {
18639 Elts.push_back(DAG.getUNDEF(SVT));
18642 Elts.push_back(DAG.getConstant(One.shl(ShAmt), dl, SVT));
18644 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
18645 return DAG.getNode(ISD::MUL, dl, VT, R, BV);
18648 // Lower SHL with variable shift amount.
18649 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
18650 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, dl, VT));
18652 Op = DAG.getNode(ISD::ADD, dl, VT, Op,
18653 DAG.getConstant(0x3f800000U, dl, VT));
18654 Op = DAG.getBitcast(MVT::v4f32, Op);
18655 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
18656 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
18659 // If possible, lower this shift as a sequence of two shifts by
18660 // constant plus a MOVSS/MOVSD instead of scalarizing it.
18662 // (v4i32 (srl A, (build_vector < X, Y, Y, Y>)))
18664 // Could be rewritten as:
18665 // (v4i32 (MOVSS (srl A, <Y,Y,Y,Y>), (srl A, <X,X,X,X>)))
18667 // The advantage is that the two shifts from the example would be
18668 // lowered as X86ISD::VSRLI nodes. This would be cheaper than scalarizing
18669 // the vector shift into four scalar shifts plus four pairs of vector
18671 if ((VT == MVT::v8i16 || VT == MVT::v4i32) &&
18672 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18673 unsigned TargetOpcode = X86ISD::MOVSS;
18674 bool CanBeSimplified;
18675 // The splat value for the first packed shift (the 'X' from the example).
18676 SDValue Amt1 = Amt->getOperand(0);
18677 // The splat value for the second packed shift (the 'Y' from the example).
18678 SDValue Amt2 = (VT == MVT::v4i32) ? Amt->getOperand(1) :
18679 Amt->getOperand(2);
18681 // See if it is possible to replace this node with a sequence of
18682 // two shifts followed by a MOVSS/MOVSD
18683 if (VT == MVT::v4i32) {
18684 // Check if it is legal to use a MOVSS.
18685 CanBeSimplified = Amt2 == Amt->getOperand(2) &&
18686 Amt2 == Amt->getOperand(3);
18687 if (!CanBeSimplified) {
18688 // Otherwise, check if we can still simplify this node using a MOVSD.
18689 CanBeSimplified = Amt1 == Amt->getOperand(1) &&
18690 Amt->getOperand(2) == Amt->getOperand(3);
18691 TargetOpcode = X86ISD::MOVSD;
18692 Amt2 = Amt->getOperand(2);
18695 // Do similar checks for the case where the machine value type
18697 CanBeSimplified = Amt1 == Amt->getOperand(1);
18698 for (unsigned i=3; i != 8 && CanBeSimplified; ++i)
18699 CanBeSimplified = Amt2 == Amt->getOperand(i);
18701 if (!CanBeSimplified) {
18702 TargetOpcode = X86ISD::MOVSD;
18703 CanBeSimplified = true;
18704 Amt2 = Amt->getOperand(4);
18705 for (unsigned i=0; i != 4 && CanBeSimplified; ++i)
18706 CanBeSimplified = Amt1 == Amt->getOperand(i);
18707 for (unsigned j=4; j != 8 && CanBeSimplified; ++j)
18708 CanBeSimplified = Amt2 == Amt->getOperand(j);
18712 if (CanBeSimplified && isa<ConstantSDNode>(Amt1) &&
18713 isa<ConstantSDNode>(Amt2)) {
18714 // Replace this node with two shifts followed by a MOVSS/MOVSD.
18715 MVT CastVT = MVT::v4i32;
18717 DAG.getConstant(cast<ConstantSDNode>(Amt1)->getAPIntValue(), dl, VT);
18718 SDValue Shift1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat1);
18720 DAG.getConstant(cast<ConstantSDNode>(Amt2)->getAPIntValue(), dl, VT);
18721 SDValue Shift2 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat2);
18722 if (TargetOpcode == X86ISD::MOVSD)
18723 CastVT = MVT::v2i64;
18724 SDValue BitCast1 = DAG.getBitcast(CastVT, Shift1);
18725 SDValue BitCast2 = DAG.getBitcast(CastVT, Shift2);
18726 SDValue Result = getTargetShuffleNode(TargetOpcode, dl, CastVT, BitCast2,
18728 return DAG.getBitcast(VT, Result);
18732 // v4i32 Non Uniform Shifts.
18733 // If the shift amount is constant we can shift each lane using the SSE2
18734 // immediate shifts, else we need to zero-extend each lane to the lower i64
18735 // and shift using the SSE2 variable shifts.
18736 // The separate results can then be blended together.
18737 if (VT == MVT::v4i32) {
18738 unsigned Opc = Op.getOpcode();
18739 SDValue Amt0, Amt1, Amt2, Amt3;
18740 if (ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18741 Amt0 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {0, 0, 0, 0});
18742 Amt1 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {1, 1, 1, 1});
18743 Amt2 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {2, 2, 2, 2});
18744 Amt3 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {3, 3, 3, 3});
18746 // ISD::SHL is handled above but we include it here for completeness.
18749 llvm_unreachable("Unknown target vector shift node");
18751 Opc = X86ISD::VSHL;
18754 Opc = X86ISD::VSRL;
18757 Opc = X86ISD::VSRA;
18760 // The SSE2 shifts use the lower i64 as the same shift amount for
18761 // all lanes and the upper i64 is ignored. These shuffle masks
18762 // optimally zero-extend each lanes on SSE2/SSE41/AVX targets.
18763 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
18764 Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Z, {0, 4, -1, -1});
18765 Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Z, {1, 5, -1, -1});
18766 Amt2 = DAG.getVectorShuffle(VT, dl, Amt, Z, {2, 6, -1, -1});
18767 Amt3 = DAG.getVectorShuffle(VT, dl, Amt, Z, {3, 7, -1, -1});
18770 SDValue R0 = DAG.getNode(Opc, dl, VT, R, Amt0);
18771 SDValue R1 = DAG.getNode(Opc, dl, VT, R, Amt1);
18772 SDValue R2 = DAG.getNode(Opc, dl, VT, R, Amt2);
18773 SDValue R3 = DAG.getNode(Opc, dl, VT, R, Amt3);
18774 SDValue R02 = DAG.getVectorShuffle(VT, dl, R0, R2, {0, -1, 6, -1});
18775 SDValue R13 = DAG.getVectorShuffle(VT, dl, R1, R3, {-1, 1, -1, 7});
18776 return DAG.getVectorShuffle(VT, dl, R02, R13, {0, 5, 2, 7});
18779 if (VT == MVT::v16i8 ||
18780 (VT == MVT::v32i8 && Subtarget->hasInt256() && !Subtarget->hasXOP())) {
18781 MVT ExtVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements() / 2);
18782 unsigned ShiftOpcode = Op->getOpcode();
18784 auto SignBitSelect = [&](MVT SelVT, SDValue Sel, SDValue V0, SDValue V1) {
18785 // On SSE41 targets we make use of the fact that VSELECT lowers
18786 // to PBLENDVB which selects bytes based just on the sign bit.
18787 if (Subtarget->hasSSE41()) {
18788 V0 = DAG.getBitcast(VT, V0);
18789 V1 = DAG.getBitcast(VT, V1);
18790 Sel = DAG.getBitcast(VT, Sel);
18791 return DAG.getBitcast(SelVT,
18792 DAG.getNode(ISD::VSELECT, dl, VT, Sel, V0, V1));
18794 // On pre-SSE41 targets we test for the sign bit by comparing to
18795 // zero - a negative value will set all bits of the lanes to true
18796 // and VSELECT uses that in its OR(AND(V0,C),AND(V1,~C)) lowering.
18797 SDValue Z = getZeroVector(SelVT, Subtarget, DAG, dl);
18798 SDValue C = DAG.getNode(X86ISD::PCMPGT, dl, SelVT, Z, Sel);
18799 return DAG.getNode(ISD::VSELECT, dl, SelVT, C, V0, V1);
18802 // Turn 'a' into a mask suitable for VSELECT: a = a << 5;
18803 // We can safely do this using i16 shifts as we're only interested in
18804 // the 3 lower bits of each byte.
18805 Amt = DAG.getBitcast(ExtVT, Amt);
18806 Amt = DAG.getNode(ISD::SHL, dl, ExtVT, Amt, DAG.getConstant(5, dl, ExtVT));
18807 Amt = DAG.getBitcast(VT, Amt);
18809 if (Op->getOpcode() == ISD::SHL || Op->getOpcode() == ISD::SRL) {
18810 // r = VSELECT(r, shift(r, 4), a);
18812 DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT));
18813 R = SignBitSelect(VT, Amt, M, R);
18816 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18818 // r = VSELECT(r, shift(r, 2), a);
18819 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT));
18820 R = SignBitSelect(VT, Amt, M, R);
18823 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18825 // return VSELECT(r, shift(r, 1), a);
18826 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT));
18827 R = SignBitSelect(VT, Amt, M, R);
18831 if (Op->getOpcode() == ISD::SRA) {
18832 // For SRA we need to unpack each byte to the higher byte of a i16 vector
18833 // so we can correctly sign extend. We don't care what happens to the
18835 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), Amt);
18836 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), Amt);
18837 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), R);
18838 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), R);
18839 ALo = DAG.getBitcast(ExtVT, ALo);
18840 AHi = DAG.getBitcast(ExtVT, AHi);
18841 RLo = DAG.getBitcast(ExtVT, RLo);
18842 RHi = DAG.getBitcast(ExtVT, RHi);
18844 // r = VSELECT(r, shift(r, 4), a);
18845 SDValue MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
18846 DAG.getConstant(4, dl, ExtVT));
18847 SDValue MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
18848 DAG.getConstant(4, dl, ExtVT));
18849 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
18850 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
18853 ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo);
18854 AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi);
18856 // r = VSELECT(r, shift(r, 2), a);
18857 MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
18858 DAG.getConstant(2, dl, ExtVT));
18859 MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
18860 DAG.getConstant(2, dl, ExtVT));
18861 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
18862 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
18865 ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo);
18866 AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi);
18868 // r = VSELECT(r, shift(r, 1), a);
18869 MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
18870 DAG.getConstant(1, dl, ExtVT));
18871 MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
18872 DAG.getConstant(1, dl, ExtVT));
18873 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
18874 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
18876 // Logical shift the result back to the lower byte, leaving a zero upper
18878 // meaning that we can safely pack with PACKUSWB.
18880 DAG.getNode(ISD::SRL, dl, ExtVT, RLo, DAG.getConstant(8, dl, ExtVT));
18882 DAG.getNode(ISD::SRL, dl, ExtVT, RHi, DAG.getConstant(8, dl, ExtVT));
18883 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
18887 // It's worth extending once and using the v8i32 shifts for 16-bit types, but
18888 // the extra overheads to get from v16i8 to v8i32 make the existing SSE
18889 // solution better.
18890 if (Subtarget->hasInt256() && VT == MVT::v8i16) {
18891 MVT ExtVT = MVT::v8i32;
18893 Op.getOpcode() == ISD::SRA ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
18894 R = DAG.getNode(ExtOpc, dl, ExtVT, R);
18895 Amt = DAG.getNode(ISD::ANY_EXTEND, dl, ExtVT, Amt);
18896 return DAG.getNode(ISD::TRUNCATE, dl, VT,
18897 DAG.getNode(Op.getOpcode(), dl, ExtVT, R, Amt));
18900 if (Subtarget->hasInt256() && !Subtarget->hasXOP() && VT == MVT::v16i16) {
18901 MVT ExtVT = MVT::v8i32;
18902 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
18903 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, Amt, Z);
18904 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, Amt, Z);
18905 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, R, R);
18906 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, R, R);
18907 ALo = DAG.getBitcast(ExtVT, ALo);
18908 AHi = DAG.getBitcast(ExtVT, AHi);
18909 RLo = DAG.getBitcast(ExtVT, RLo);
18910 RHi = DAG.getBitcast(ExtVT, RHi);
18911 SDValue Lo = DAG.getNode(Op.getOpcode(), dl, ExtVT, RLo, ALo);
18912 SDValue Hi = DAG.getNode(Op.getOpcode(), dl, ExtVT, RHi, AHi);
18913 Lo = DAG.getNode(ISD::SRL, dl, ExtVT, Lo, DAG.getConstant(16, dl, ExtVT));
18914 Hi = DAG.getNode(ISD::SRL, dl, ExtVT, Hi, DAG.getConstant(16, dl, ExtVT));
18915 return DAG.getNode(X86ISD::PACKUS, dl, VT, Lo, Hi);
18918 if (VT == MVT::v8i16) {
18919 unsigned ShiftOpcode = Op->getOpcode();
18921 auto SignBitSelect = [&](SDValue Sel, SDValue V0, SDValue V1) {
18922 // On SSE41 targets we make use of the fact that VSELECT lowers
18923 // to PBLENDVB which selects bytes based just on the sign bit.
18924 if (Subtarget->hasSSE41()) {
18925 MVT ExtVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() * 2);
18926 V0 = DAG.getBitcast(ExtVT, V0);
18927 V1 = DAG.getBitcast(ExtVT, V1);
18928 Sel = DAG.getBitcast(ExtVT, Sel);
18929 return DAG.getBitcast(
18930 VT, DAG.getNode(ISD::VSELECT, dl, ExtVT, Sel, V0, V1));
18932 // On pre-SSE41 targets we splat the sign bit - a negative value will
18933 // set all bits of the lanes to true and VSELECT uses that in
18934 // its OR(AND(V0,C),AND(V1,~C)) lowering.
18936 DAG.getNode(ISD::SRA, dl, VT, Sel, DAG.getConstant(15, dl, VT));
18937 return DAG.getNode(ISD::VSELECT, dl, VT, C, V0, V1);
18940 // Turn 'a' into a mask suitable for VSELECT: a = a << 12;
18941 if (Subtarget->hasSSE41()) {
18942 // On SSE41 targets we need to replicate the shift mask in both
18943 // bytes for PBLENDVB.
18946 DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(4, dl, VT)),
18947 DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT)));
18949 Amt = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT));
18952 // r = VSELECT(r, shift(r, 8), a);
18953 SDValue M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(8, dl, VT));
18954 R = SignBitSelect(Amt, M, R);
18957 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18959 // r = VSELECT(r, shift(r, 4), a);
18960 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT));
18961 R = SignBitSelect(Amt, M, R);
18964 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18966 // r = VSELECT(r, shift(r, 2), a);
18967 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT));
18968 R = SignBitSelect(Amt, M, R);
18971 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18973 // return VSELECT(r, shift(r, 1), a);
18974 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT));
18975 R = SignBitSelect(Amt, M, R);
18979 // Decompose 256-bit shifts into smaller 128-bit shifts.
18980 if (VT.is256BitVector()) {
18981 unsigned NumElems = VT.getVectorNumElements();
18982 MVT EltVT = VT.getVectorElementType();
18983 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
18985 // Extract the two vectors
18986 SDValue V1 = Extract128BitVector(R, 0, DAG, dl);
18987 SDValue V2 = Extract128BitVector(R, NumElems/2, DAG, dl);
18989 // Recreate the shift amount vectors
18990 SDValue Amt1, Amt2;
18991 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
18992 // Constant shift amount
18993 SmallVector<SDValue, 8> Ops(Amt->op_begin(), Amt->op_begin() + NumElems);
18994 ArrayRef<SDValue> Amt1Csts = makeArrayRef(Ops).slice(0, NumElems / 2);
18995 ArrayRef<SDValue> Amt2Csts = makeArrayRef(Ops).slice(NumElems / 2);
18997 Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt1Csts);
18998 Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt2Csts);
19000 // Variable shift amount
19001 Amt1 = Extract128BitVector(Amt, 0, DAG, dl);
19002 Amt2 = Extract128BitVector(Amt, NumElems/2, DAG, dl);
19005 // Issue new vector shifts for the smaller types
19006 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
19007 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
19009 // Concatenate the result back
19010 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
19016 static SDValue LowerRotate(SDValue Op, const X86Subtarget *Subtarget,
19017 SelectionDAG &DAG) {
19018 MVT VT = Op.getSimpleValueType();
19020 SDValue R = Op.getOperand(0);
19021 SDValue Amt = Op.getOperand(1);
19023 assert(VT.isVector() && "Custom lowering only for vector rotates!");
19024 assert(Subtarget->hasXOP() && "XOP support required for vector rotates!");
19025 assert((Op.getOpcode() == ISD::ROTL) && "Only ROTL supported");
19027 // XOP has 128-bit vector variable + immediate rotates.
19028 // +ve/-ve Amt = rotate left/right.
19030 // Split 256-bit integers.
19031 if (VT.is256BitVector())
19032 return Lower256IntArith(Op, DAG);
19034 assert(VT.is128BitVector() && "Only rotate 128-bit vectors!");
19036 // Attempt to rotate by immediate.
19037 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
19038 if (auto *RotateConst = BVAmt->getConstantSplatNode()) {
19039 uint64_t RotateAmt = RotateConst->getAPIntValue().getZExtValue();
19040 assert(RotateAmt < VT.getScalarSizeInBits() && "Rotation out of range");
19041 return DAG.getNode(X86ISD::VPROTI, DL, VT, R,
19042 DAG.getConstant(RotateAmt, DL, MVT::i8));
19046 // Use general rotate by variable (per-element).
19047 return DAG.getNode(X86ISD::VPROT, DL, VT, R, Amt);
19050 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
19051 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
19052 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
19053 // looks for this combo and may remove the "setcc" instruction if the "setcc"
19054 // has only one use.
19055 SDNode *N = Op.getNode();
19056 SDValue LHS = N->getOperand(0);
19057 SDValue RHS = N->getOperand(1);
19058 unsigned BaseOp = 0;
19061 switch (Op.getOpcode()) {
19062 default: llvm_unreachable("Unknown ovf instruction!");
19064 // A subtract of one will be selected as a INC. Note that INC doesn't
19065 // set CF, so we can't do this for UADDO.
19066 if (isOneConstant(RHS)) {
19067 BaseOp = X86ISD::INC;
19068 Cond = X86::COND_O;
19071 BaseOp = X86ISD::ADD;
19072 Cond = X86::COND_O;
19075 BaseOp = X86ISD::ADD;
19076 Cond = X86::COND_B;
19079 // A subtract of one will be selected as a DEC. Note that DEC doesn't
19080 // set CF, so we can't do this for USUBO.
19081 if (isOneConstant(RHS)) {
19082 BaseOp = X86ISD::DEC;
19083 Cond = X86::COND_O;
19086 BaseOp = X86ISD::SUB;
19087 Cond = X86::COND_O;
19090 BaseOp = X86ISD::SUB;
19091 Cond = X86::COND_B;
19094 BaseOp = N->getValueType(0) == MVT::i8 ? X86ISD::SMUL8 : X86ISD::SMUL;
19095 Cond = X86::COND_O;
19097 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
19098 if (N->getValueType(0) == MVT::i8) {
19099 BaseOp = X86ISD::UMUL8;
19100 Cond = X86::COND_O;
19103 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
19105 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
19108 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
19109 DAG.getConstant(X86::COND_O, DL, MVT::i32),
19110 SDValue(Sum.getNode(), 2));
19112 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
19116 // Also sets EFLAGS.
19117 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
19118 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
19121 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
19122 DAG.getConstant(Cond, DL, MVT::i32),
19123 SDValue(Sum.getNode(), 1));
19125 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
19128 /// Returns true if the operand type is exactly twice the native width, and
19129 /// the corresponding cmpxchg8b or cmpxchg16b instruction is available.
19130 /// Used to know whether to use cmpxchg8/16b when expanding atomic operations
19131 /// (otherwise we leave them alone to become __sync_fetch_and_... calls).
19132 bool X86TargetLowering::needsCmpXchgNb(Type *MemType) const {
19133 unsigned OpWidth = MemType->getPrimitiveSizeInBits();
19136 return !Subtarget->is64Bit(); // FIXME this should be Subtarget.hasCmpxchg8b
19137 else if (OpWidth == 128)
19138 return Subtarget->hasCmpxchg16b();
19143 bool X86TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
19144 return needsCmpXchgNb(SI->getValueOperand()->getType());
19147 // Note: this turns large loads into lock cmpxchg8b/16b.
19148 // FIXME: On 32 bits x86, fild/movq might be faster than lock cmpxchg8b.
19149 TargetLowering::AtomicExpansionKind
19150 X86TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
19151 auto PTy = cast<PointerType>(LI->getPointerOperand()->getType());
19152 return needsCmpXchgNb(PTy->getElementType()) ? AtomicExpansionKind::CmpXChg
19153 : AtomicExpansionKind::None;
19156 TargetLowering::AtomicExpansionKind
19157 X86TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
19158 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
19159 Type *MemType = AI->getType();
19161 // If the operand is too big, we must see if cmpxchg8/16b is available
19162 // and default to library calls otherwise.
19163 if (MemType->getPrimitiveSizeInBits() > NativeWidth) {
19164 return needsCmpXchgNb(MemType) ? AtomicExpansionKind::CmpXChg
19165 : AtomicExpansionKind::None;
19168 AtomicRMWInst::BinOp Op = AI->getOperation();
19171 llvm_unreachable("Unknown atomic operation");
19172 case AtomicRMWInst::Xchg:
19173 case AtomicRMWInst::Add:
19174 case AtomicRMWInst::Sub:
19175 // It's better to use xadd, xsub or xchg for these in all cases.
19176 return AtomicExpansionKind::None;
19177 case AtomicRMWInst::Or:
19178 case AtomicRMWInst::And:
19179 case AtomicRMWInst::Xor:
19180 // If the atomicrmw's result isn't actually used, we can just add a "lock"
19181 // prefix to a normal instruction for these operations.
19182 return !AI->use_empty() ? AtomicExpansionKind::CmpXChg
19183 : AtomicExpansionKind::None;
19184 case AtomicRMWInst::Nand:
19185 case AtomicRMWInst::Max:
19186 case AtomicRMWInst::Min:
19187 case AtomicRMWInst::UMax:
19188 case AtomicRMWInst::UMin:
19189 // These always require a non-trivial set of data operations on x86. We must
19190 // use a cmpxchg loop.
19191 return AtomicExpansionKind::CmpXChg;
19195 static bool hasMFENCE(const X86Subtarget& Subtarget) {
19196 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
19197 // no-sse2). There isn't any reason to disable it if the target processor
19199 return Subtarget.hasSSE2() || Subtarget.is64Bit();
19203 X86TargetLowering::lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const {
19204 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
19205 Type *MemType = AI->getType();
19206 // Accesses larger than the native width are turned into cmpxchg/libcalls, so
19207 // there is no benefit in turning such RMWs into loads, and it is actually
19208 // harmful as it introduces a mfence.
19209 if (MemType->getPrimitiveSizeInBits() > NativeWidth)
19212 auto Builder = IRBuilder<>(AI);
19213 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
19214 auto SynchScope = AI->getSynchScope();
19215 // We must restrict the ordering to avoid generating loads with Release or
19216 // ReleaseAcquire orderings.
19217 auto Order = AtomicCmpXchgInst::getStrongestFailureOrdering(AI->getOrdering());
19218 auto Ptr = AI->getPointerOperand();
19220 // Before the load we need a fence. Here is an example lifted from
19221 // http://www.hpl.hp.com/techreports/2012/HPL-2012-68.pdf showing why a fence
19224 // x.store(1, relaxed);
19225 // r1 = y.fetch_add(0, release);
19227 // y.fetch_add(42, acquire);
19228 // r2 = x.load(relaxed);
19229 // r1 = r2 = 0 is impossible, but becomes possible if the idempotent rmw is
19230 // lowered to just a load without a fence. A mfence flushes the store buffer,
19231 // making the optimization clearly correct.
19232 // FIXME: it is required if isAtLeastRelease(Order) but it is not clear
19233 // otherwise, we might be able to be more aggressive on relaxed idempotent
19234 // rmw. In practice, they do not look useful, so we don't try to be
19235 // especially clever.
19236 if (SynchScope == SingleThread)
19237 // FIXME: we could just insert an X86ISD::MEMBARRIER here, except we are at
19238 // the IR level, so we must wrap it in an intrinsic.
19241 if (!hasMFENCE(*Subtarget))
19242 // FIXME: it might make sense to use a locked operation here but on a
19243 // different cache-line to prevent cache-line bouncing. In practice it
19244 // is probably a small win, and x86 processors without mfence are rare
19245 // enough that we do not bother.
19249 llvm::Intrinsic::getDeclaration(M, Intrinsic::x86_sse2_mfence);
19250 Builder.CreateCall(MFence, {});
19252 // Finally we can emit the atomic load.
19253 LoadInst *Loaded = Builder.CreateAlignedLoad(Ptr,
19254 AI->getType()->getPrimitiveSizeInBits());
19255 Loaded->setAtomic(Order, SynchScope);
19256 AI->replaceAllUsesWith(Loaded);
19257 AI->eraseFromParent();
19261 static SDValue LowerATOMIC_FENCE(SDValue Op, const X86Subtarget *Subtarget,
19262 SelectionDAG &DAG) {
19264 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
19265 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
19266 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
19267 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
19269 // The only fence that needs an instruction is a sequentially-consistent
19270 // cross-thread fence.
19271 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
19272 if (hasMFENCE(*Subtarget))
19273 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
19275 SDValue Chain = Op.getOperand(0);
19276 SDValue Zero = DAG.getConstant(0, dl, MVT::i32);
19278 DAG.getRegister(X86::ESP, MVT::i32), // Base
19279 DAG.getTargetConstant(1, dl, MVT::i8), // Scale
19280 DAG.getRegister(0, MVT::i32), // Index
19281 DAG.getTargetConstant(0, dl, MVT::i32), // Disp
19282 DAG.getRegister(0, MVT::i32), // Segment.
19286 SDNode *Res = DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops);
19287 return SDValue(Res, 0);
19290 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
19291 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
19294 static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget *Subtarget,
19295 SelectionDAG &DAG) {
19296 MVT T = Op.getSimpleValueType();
19300 switch(T.SimpleTy) {
19301 default: llvm_unreachable("Invalid value type!");
19302 case MVT::i8: Reg = X86::AL; size = 1; break;
19303 case MVT::i16: Reg = X86::AX; size = 2; break;
19304 case MVT::i32: Reg = X86::EAX; size = 4; break;
19306 assert(Subtarget->is64Bit() && "Node not type legal!");
19307 Reg = X86::RAX; size = 8;
19310 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
19311 Op.getOperand(2), SDValue());
19312 SDValue Ops[] = { cpIn.getValue(0),
19315 DAG.getTargetConstant(size, DL, MVT::i8),
19316 cpIn.getValue(1) };
19317 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
19318 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
19319 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
19323 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
19324 SDValue EFLAGS = DAG.getCopyFromReg(cpOut.getValue(1), DL, X86::EFLAGS,
19325 MVT::i32, cpOut.getValue(2));
19326 SDValue Success = DAG.getNode(X86ISD::SETCC, DL, Op->getValueType(1),
19327 DAG.getConstant(X86::COND_E, DL, MVT::i8),
19330 DAG.ReplaceAllUsesOfValueWith(Op.getValue(0), cpOut);
19331 DAG.ReplaceAllUsesOfValueWith(Op.getValue(1), Success);
19332 DAG.ReplaceAllUsesOfValueWith(Op.getValue(2), EFLAGS.getValue(1));
19336 static SDValue LowerBITCAST(SDValue Op, const X86Subtarget *Subtarget,
19337 SelectionDAG &DAG) {
19338 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
19339 MVT DstVT = Op.getSimpleValueType();
19341 if (SrcVT == MVT::v2i32 || SrcVT == MVT::v4i16 || SrcVT == MVT::v8i8) {
19342 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19343 if (DstVT != MVT::f64)
19344 // This conversion needs to be expanded.
19347 SDValue InVec = Op->getOperand(0);
19349 unsigned NumElts = SrcVT.getVectorNumElements();
19350 MVT SVT = SrcVT.getVectorElementType();
19352 // Widen the vector in input in the case of MVT::v2i32.
19353 // Example: from MVT::v2i32 to MVT::v4i32.
19354 SmallVector<SDValue, 16> Elts;
19355 for (unsigned i = 0, e = NumElts; i != e; ++i)
19356 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT, InVec,
19357 DAG.getIntPtrConstant(i, dl)));
19359 // Explicitly mark the extra elements as Undef.
19360 Elts.append(NumElts, DAG.getUNDEF(SVT));
19362 EVT NewVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
19363 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Elts);
19364 SDValue ToV2F64 = DAG.getBitcast(MVT::v2f64, BV);
19365 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, ToV2F64,
19366 DAG.getIntPtrConstant(0, dl));
19369 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
19370 Subtarget->hasMMX() && "Unexpected custom BITCAST");
19371 assert((DstVT == MVT::i64 ||
19372 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
19373 "Unexpected custom BITCAST");
19374 // i64 <=> MMX conversions are Legal.
19375 if (SrcVT==MVT::i64 && DstVT.isVector())
19377 if (DstVT==MVT::i64 && SrcVT.isVector())
19379 // MMX <=> MMX conversions are Legal.
19380 if (SrcVT.isVector() && DstVT.isVector())
19382 // All other conversions need to be expanded.
19386 /// Compute the horizontal sum of bytes in V for the elements of VT.
19388 /// Requires V to be a byte vector and VT to be an integer vector type with
19389 /// wider elements than V's type. The width of the elements of VT determines
19390 /// how many bytes of V are summed horizontally to produce each element of the
19392 static SDValue LowerHorizontalByteSum(SDValue V, MVT VT,
19393 const X86Subtarget *Subtarget,
19394 SelectionDAG &DAG) {
19396 MVT ByteVecVT = V.getSimpleValueType();
19397 MVT EltVT = VT.getVectorElementType();
19398 int NumElts = VT.getVectorNumElements();
19399 assert(ByteVecVT.getVectorElementType() == MVT::i8 &&
19400 "Expected value to have byte element type.");
19401 assert(EltVT != MVT::i8 &&
19402 "Horizontal byte sum only makes sense for wider elements!");
19403 unsigned VecSize = VT.getSizeInBits();
19404 assert(ByteVecVT.getSizeInBits() == VecSize && "Cannot change vector size!");
19406 // PSADBW instruction horizontally add all bytes and leave the result in i64
19407 // chunks, thus directly computes the pop count for v2i64 and v4i64.
19408 if (EltVT == MVT::i64) {
19409 SDValue Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
19410 MVT SadVecVT = MVT::getVectorVT(MVT::i64, VecSize / 64);
19411 V = DAG.getNode(X86ISD::PSADBW, DL, SadVecVT, V, Zeros);
19412 return DAG.getBitcast(VT, V);
19415 if (EltVT == MVT::i32) {
19416 // We unpack the low half and high half into i32s interleaved with zeros so
19417 // that we can use PSADBW to horizontally sum them. The most useful part of
19418 // this is that it lines up the results of two PSADBW instructions to be
19419 // two v2i64 vectors which concatenated are the 4 population counts. We can
19420 // then use PACKUSWB to shrink and concatenate them into a v4i32 again.
19421 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, DL);
19422 SDValue Low = DAG.getNode(X86ISD::UNPCKL, DL, VT, V, Zeros);
19423 SDValue High = DAG.getNode(X86ISD::UNPCKH, DL, VT, V, Zeros);
19425 // Do the horizontal sums into two v2i64s.
19426 Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
19427 MVT SadVecVT = MVT::getVectorVT(MVT::i64, VecSize / 64);
19428 Low = DAG.getNode(X86ISD::PSADBW, DL, SadVecVT,
19429 DAG.getBitcast(ByteVecVT, Low), Zeros);
19430 High = DAG.getNode(X86ISD::PSADBW, DL, SadVecVT,
19431 DAG.getBitcast(ByteVecVT, High), Zeros);
19433 // Merge them together.
19434 MVT ShortVecVT = MVT::getVectorVT(MVT::i16, VecSize / 16);
19435 V = DAG.getNode(X86ISD::PACKUS, DL, ByteVecVT,
19436 DAG.getBitcast(ShortVecVT, Low),
19437 DAG.getBitcast(ShortVecVT, High));
19439 return DAG.getBitcast(VT, V);
19442 // The only element type left is i16.
19443 assert(EltVT == MVT::i16 && "Unknown how to handle type");
19445 // To obtain pop count for each i16 element starting from the pop count for
19446 // i8 elements, shift the i16s left by 8, sum as i8s, and then shift as i16s
19447 // right by 8. It is important to shift as i16s as i8 vector shift isn't
19448 // directly supported.
19449 SmallVector<SDValue, 16> Shifters(NumElts, DAG.getConstant(8, DL, EltVT));
19450 SDValue Shifter = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters);
19451 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, DAG.getBitcast(VT, V), Shifter);
19452 V = DAG.getNode(ISD::ADD, DL, ByteVecVT, DAG.getBitcast(ByteVecVT, Shl),
19453 DAG.getBitcast(ByteVecVT, V));
19454 return DAG.getNode(ISD::SRL, DL, VT, DAG.getBitcast(VT, V), Shifter);
19457 static SDValue LowerVectorCTPOPInRegLUT(SDValue Op, SDLoc DL,
19458 const X86Subtarget *Subtarget,
19459 SelectionDAG &DAG) {
19460 MVT VT = Op.getSimpleValueType();
19461 MVT EltVT = VT.getVectorElementType();
19462 unsigned VecSize = VT.getSizeInBits();
19464 // Implement a lookup table in register by using an algorithm based on:
19465 // http://wm.ite.pl/articles/sse-popcount.html
19467 // The general idea is that every lower byte nibble in the input vector is an
19468 // index into a in-register pre-computed pop count table. We then split up the
19469 // input vector in two new ones: (1) a vector with only the shifted-right
19470 // higher nibbles for each byte and (2) a vector with the lower nibbles (and
19471 // masked out higher ones) for each byte. PSHUB is used separately with both
19472 // to index the in-register table. Next, both are added and the result is a
19473 // i8 vector where each element contains the pop count for input byte.
19475 // To obtain the pop count for elements != i8, we follow up with the same
19476 // approach and use additional tricks as described below.
19478 const int LUT[16] = {/* 0 */ 0, /* 1 */ 1, /* 2 */ 1, /* 3 */ 2,
19479 /* 4 */ 1, /* 5 */ 2, /* 6 */ 2, /* 7 */ 3,
19480 /* 8 */ 1, /* 9 */ 2, /* a */ 2, /* b */ 3,
19481 /* c */ 2, /* d */ 3, /* e */ 3, /* f */ 4};
19483 int NumByteElts = VecSize / 8;
19484 MVT ByteVecVT = MVT::getVectorVT(MVT::i8, NumByteElts);
19485 SDValue In = DAG.getBitcast(ByteVecVT, Op);
19486 SmallVector<SDValue, 16> LUTVec;
19487 for (int i = 0; i < NumByteElts; ++i)
19488 LUTVec.push_back(DAG.getConstant(LUT[i % 16], DL, MVT::i8));
19489 SDValue InRegLUT = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, LUTVec);
19490 SmallVector<SDValue, 16> Mask0F(NumByteElts,
19491 DAG.getConstant(0x0F, DL, MVT::i8));
19492 SDValue M0F = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Mask0F);
19495 SmallVector<SDValue, 16> Four(NumByteElts, DAG.getConstant(4, DL, MVT::i8));
19496 SDValue FourV = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Four);
19497 SDValue HighNibbles = DAG.getNode(ISD::SRL, DL, ByteVecVT, In, FourV);
19500 SDValue LowNibbles = DAG.getNode(ISD::AND, DL, ByteVecVT, In, M0F);
19502 // The input vector is used as the shuffle mask that index elements into the
19503 // LUT. After counting low and high nibbles, add the vector to obtain the
19504 // final pop count per i8 element.
19505 SDValue HighPopCnt =
19506 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, HighNibbles);
19507 SDValue LowPopCnt =
19508 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, LowNibbles);
19509 SDValue PopCnt = DAG.getNode(ISD::ADD, DL, ByteVecVT, HighPopCnt, LowPopCnt);
19511 if (EltVT == MVT::i8)
19514 return LowerHorizontalByteSum(PopCnt, VT, Subtarget, DAG);
19517 static SDValue LowerVectorCTPOPBitmath(SDValue Op, SDLoc DL,
19518 const X86Subtarget *Subtarget,
19519 SelectionDAG &DAG) {
19520 MVT VT = Op.getSimpleValueType();
19521 assert(VT.is128BitVector() &&
19522 "Only 128-bit vector bitmath lowering supported.");
19524 int VecSize = VT.getSizeInBits();
19525 MVT EltVT = VT.getVectorElementType();
19526 int Len = EltVT.getSizeInBits();
19528 // This is the vectorized version of the "best" algorithm from
19529 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
19530 // with a minor tweak to use a series of adds + shifts instead of vector
19531 // multiplications. Implemented for all integer vector types. We only use
19532 // this when we don't have SSSE3 which allows a LUT-based lowering that is
19533 // much faster, even faster than using native popcnt instructions.
19535 auto GetShift = [&](unsigned OpCode, SDValue V, int Shifter) {
19536 MVT VT = V.getSimpleValueType();
19537 SmallVector<SDValue, 32> Shifters(
19538 VT.getVectorNumElements(),
19539 DAG.getConstant(Shifter, DL, VT.getVectorElementType()));
19540 return DAG.getNode(OpCode, DL, VT, V,
19541 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters));
19543 auto GetMask = [&](SDValue V, APInt Mask) {
19544 MVT VT = V.getSimpleValueType();
19545 SmallVector<SDValue, 32> Masks(
19546 VT.getVectorNumElements(),
19547 DAG.getConstant(Mask, DL, VT.getVectorElementType()));
19548 return DAG.getNode(ISD::AND, DL, VT, V,
19549 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Masks));
19552 // We don't want to incur the implicit masks required to SRL vNi8 vectors on
19553 // x86, so set the SRL type to have elements at least i16 wide. This is
19554 // correct because all of our SRLs are followed immediately by a mask anyways
19555 // that handles any bits that sneak into the high bits of the byte elements.
19556 MVT SrlVT = Len > 8 ? VT : MVT::getVectorVT(MVT::i16, VecSize / 16);
19560 // v = v - ((v >> 1) & 0x55555555...)
19562 DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 1));
19563 SDValue And = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x55)));
19564 V = DAG.getNode(ISD::SUB, DL, VT, V, And);
19566 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
19567 SDValue AndLHS = GetMask(V, APInt::getSplat(Len, APInt(8, 0x33)));
19568 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 2));
19569 SDValue AndRHS = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x33)));
19570 V = DAG.getNode(ISD::ADD, DL, VT, AndLHS, AndRHS);
19572 // v = (v + (v >> 4)) & 0x0F0F0F0F...
19573 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 4));
19574 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, V, Srl);
19575 V = GetMask(Add, APInt::getSplat(Len, APInt(8, 0x0F)));
19577 // At this point, V contains the byte-wise population count, and we are
19578 // merely doing a horizontal sum if necessary to get the wider element
19580 if (EltVT == MVT::i8)
19583 return LowerHorizontalByteSum(
19584 DAG.getBitcast(MVT::getVectorVT(MVT::i8, VecSize / 8), V), VT, Subtarget,
19588 static SDValue LowerVectorCTPOP(SDValue Op, const X86Subtarget *Subtarget,
19589 SelectionDAG &DAG) {
19590 MVT VT = Op.getSimpleValueType();
19591 // FIXME: Need to add AVX-512 support here!
19592 assert((VT.is256BitVector() || VT.is128BitVector()) &&
19593 "Unknown CTPOP type to handle");
19594 SDLoc DL(Op.getNode());
19595 SDValue Op0 = Op.getOperand(0);
19597 if (!Subtarget->hasSSSE3()) {
19598 // We can't use the fast LUT approach, so fall back on vectorized bitmath.
19599 assert(VT.is128BitVector() && "Only 128-bit vectors supported in SSE!");
19600 return LowerVectorCTPOPBitmath(Op0, DL, Subtarget, DAG);
19603 if (VT.is256BitVector() && !Subtarget->hasInt256()) {
19604 unsigned NumElems = VT.getVectorNumElements();
19606 // Extract each 128-bit vector, compute pop count and concat the result.
19607 SDValue LHS = Extract128BitVector(Op0, 0, DAG, DL);
19608 SDValue RHS = Extract128BitVector(Op0, NumElems/2, DAG, DL);
19610 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT,
19611 LowerVectorCTPOPInRegLUT(LHS, DL, Subtarget, DAG),
19612 LowerVectorCTPOPInRegLUT(RHS, DL, Subtarget, DAG));
19615 return LowerVectorCTPOPInRegLUT(Op0, DL, Subtarget, DAG);
19618 static SDValue LowerCTPOP(SDValue Op, const X86Subtarget *Subtarget,
19619 SelectionDAG &DAG) {
19620 assert(Op.getSimpleValueType().isVector() &&
19621 "We only do custom lowering for vector population count.");
19622 return LowerVectorCTPOP(Op, Subtarget, DAG);
19625 static SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
19626 SDNode *Node = Op.getNode();
19628 EVT T = Node->getValueType(0);
19629 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
19630 DAG.getConstant(0, dl, T), Node->getOperand(2));
19631 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
19632 cast<AtomicSDNode>(Node)->getMemoryVT(),
19633 Node->getOperand(0),
19634 Node->getOperand(1), negOp,
19635 cast<AtomicSDNode>(Node)->getMemOperand(),
19636 cast<AtomicSDNode>(Node)->getOrdering(),
19637 cast<AtomicSDNode>(Node)->getSynchScope());
19640 static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) {
19641 SDNode *Node = Op.getNode();
19643 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
19645 // Convert seq_cst store -> xchg
19646 // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b)
19647 // FIXME: On 32-bit, store -> fist or movq would be more efficient
19648 // (The only way to get a 16-byte store is cmpxchg16b)
19649 // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment.
19650 if (cast<AtomicSDNode>(Node)->getOrdering() == SequentiallyConsistent ||
19651 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
19652 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
19653 cast<AtomicSDNode>(Node)->getMemoryVT(),
19654 Node->getOperand(0),
19655 Node->getOperand(1), Node->getOperand(2),
19656 cast<AtomicSDNode>(Node)->getMemOperand(),
19657 cast<AtomicSDNode>(Node)->getOrdering(),
19658 cast<AtomicSDNode>(Node)->getSynchScope());
19659 return Swap.getValue(1);
19661 // Other atomic stores have a simple pattern.
19665 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
19666 MVT VT = Op.getNode()->getSimpleValueType(0);
19668 // Let legalize expand this if it isn't a legal type yet.
19669 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
19672 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
19675 bool ExtraOp = false;
19676 switch (Op.getOpcode()) {
19677 default: llvm_unreachable("Invalid code");
19678 case ISD::ADDC: Opc = X86ISD::ADD; break;
19679 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
19680 case ISD::SUBC: Opc = X86ISD::SUB; break;
19681 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
19685 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
19687 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
19688 Op.getOperand(1), Op.getOperand(2));
19691 static SDValue LowerFSINCOS(SDValue Op, const X86Subtarget *Subtarget,
19692 SelectionDAG &DAG) {
19693 assert(Subtarget->isTargetDarwin() && Subtarget->is64Bit());
19695 // For MacOSX, we want to call an alternative entry point: __sincos_stret,
19696 // which returns the values as { float, float } (in XMM0) or
19697 // { double, double } (which is returned in XMM0, XMM1).
19699 SDValue Arg = Op.getOperand(0);
19700 EVT ArgVT = Arg.getValueType();
19701 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
19703 TargetLowering::ArgListTy Args;
19704 TargetLowering::ArgListEntry Entry;
19708 Entry.isSExt = false;
19709 Entry.isZExt = false;
19710 Args.push_back(Entry);
19712 bool isF64 = ArgVT == MVT::f64;
19713 // Only optimize x86_64 for now. i386 is a bit messy. For f32,
19714 // the small struct {f32, f32} is returned in (eax, edx). For f64,
19715 // the results are returned via SRet in memory.
19716 const char *LibcallName = isF64 ? "__sincos_stret" : "__sincosf_stret";
19717 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
19719 DAG.getExternalSymbol(LibcallName, TLI.getPointerTy(DAG.getDataLayout()));
19721 Type *RetTy = isF64
19722 ? (Type*)StructType::get(ArgTy, ArgTy, nullptr)
19723 : (Type*)VectorType::get(ArgTy, 4);
19725 TargetLowering::CallLoweringInfo CLI(DAG);
19726 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
19727 .setCallee(CallingConv::C, RetTy, Callee, std::move(Args), 0);
19729 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
19732 // Returned in xmm0 and xmm1.
19733 return CallResult.first;
19735 // Returned in bits 0:31 and 32:64 xmm0.
19736 SDValue SinVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
19737 CallResult.first, DAG.getIntPtrConstant(0, dl));
19738 SDValue CosVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
19739 CallResult.first, DAG.getIntPtrConstant(1, dl));
19740 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
19741 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys, SinVal, CosVal);
19744 /// Widen a vector input to a vector of NVT. The
19745 /// input vector must have the same element type as NVT.
19746 static SDValue ExtendToType(SDValue InOp, MVT NVT, SelectionDAG &DAG,
19747 bool FillWithZeroes = false) {
19748 // Check if InOp already has the right width.
19749 MVT InVT = InOp.getSimpleValueType();
19753 if (InOp.isUndef())
19754 return DAG.getUNDEF(NVT);
19756 assert(InVT.getVectorElementType() == NVT.getVectorElementType() &&
19757 "input and widen element type must match");
19759 unsigned InNumElts = InVT.getVectorNumElements();
19760 unsigned WidenNumElts = NVT.getVectorNumElements();
19761 assert(WidenNumElts > InNumElts && WidenNumElts % InNumElts == 0 &&
19762 "Unexpected request for vector widening");
19764 EVT EltVT = NVT.getVectorElementType();
19767 if (InOp.getOpcode() == ISD::CONCAT_VECTORS &&
19768 InOp.getNumOperands() == 2) {
19769 SDValue N1 = InOp.getOperand(1);
19770 if ((ISD::isBuildVectorAllZeros(N1.getNode()) && FillWithZeroes) ||
19772 InOp = InOp.getOperand(0);
19773 InVT = InOp.getSimpleValueType();
19774 InNumElts = InVT.getVectorNumElements();
19777 if (ISD::isBuildVectorOfConstantSDNodes(InOp.getNode()) ||
19778 ISD::isBuildVectorOfConstantFPSDNodes(InOp.getNode())) {
19779 SmallVector<SDValue, 16> Ops;
19780 for (unsigned i = 0; i < InNumElts; ++i)
19781 Ops.push_back(InOp.getOperand(i));
19783 SDValue FillVal = FillWithZeroes ? DAG.getConstant(0, dl, EltVT) :
19784 DAG.getUNDEF(EltVT);
19785 for (unsigned i = 0; i < WidenNumElts - InNumElts; ++i)
19786 Ops.push_back(FillVal);
19787 return DAG.getNode(ISD::BUILD_VECTOR, dl, NVT, Ops);
19789 SDValue FillVal = FillWithZeroes ? DAG.getConstant(0, dl, NVT) :
19791 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, NVT, FillVal,
19792 InOp, DAG.getIntPtrConstant(0, dl));
19795 static SDValue LowerMSCATTER(SDValue Op, const X86Subtarget *Subtarget,
19796 SelectionDAG &DAG) {
19797 assert(Subtarget->hasAVX512() &&
19798 "MGATHER/MSCATTER are supported on AVX-512 arch only");
19800 // X86 scatter kills mask register, so its type should be added to
19801 // the list of return values.
19802 // If the "scatter" has 2 return values, it is already handled.
19803 if (Op.getNode()->getNumValues() == 2)
19806 MaskedScatterSDNode *N = cast<MaskedScatterSDNode>(Op.getNode());
19807 SDValue Src = N->getValue();
19808 MVT VT = Src.getSimpleValueType();
19809 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported scatter op");
19812 SDValue NewScatter;
19813 SDValue Index = N->getIndex();
19814 SDValue Mask = N->getMask();
19815 SDValue Chain = N->getChain();
19816 SDValue BasePtr = N->getBasePtr();
19817 MVT MemVT = N->getMemoryVT().getSimpleVT();
19818 MVT IndexVT = Index.getSimpleValueType();
19819 MVT MaskVT = Mask.getSimpleValueType();
19821 if (MemVT.getScalarSizeInBits() < VT.getScalarSizeInBits()) {
19822 // The v2i32 value was promoted to v2i64.
19823 // Now we "redo" the type legalizer's work and widen the original
19824 // v2i32 value to v4i32. The original v2i32 is retrieved from v2i64
19826 assert((MemVT == MVT::v2i32 && VT == MVT::v2i64) &&
19827 "Unexpected memory type");
19828 int ShuffleMask[] = {0, 2, -1, -1};
19829 Src = DAG.getVectorShuffle(MVT::v4i32, dl, DAG.getBitcast(MVT::v4i32, Src),
19830 DAG.getUNDEF(MVT::v4i32), ShuffleMask);
19831 // Now we have 4 elements instead of 2.
19832 // Expand the index.
19833 MVT NewIndexVT = MVT::getVectorVT(IndexVT.getScalarType(), 4);
19834 Index = ExtendToType(Index, NewIndexVT, DAG);
19836 // Expand the mask with zeroes
19837 // Mask may be <2 x i64> or <2 x i1> at this moment
19838 assert((MaskVT == MVT::v2i1 || MaskVT == MVT::v2i64) &&
19839 "Unexpected mask type");
19840 MVT ExtMaskVT = MVT::getVectorVT(MaskVT.getScalarType(), 4);
19841 Mask = ExtendToType(Mask, ExtMaskVT, DAG, true);
19845 unsigned NumElts = VT.getVectorNumElements();
19846 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
19847 !Index.getSimpleValueType().is512BitVector()) {
19848 // AVX512F supports only 512-bit vectors. Or data or index should
19849 // be 512 bit wide. If now the both index and data are 256-bit, but
19850 // the vector contains 8 elements, we just sign-extend the index
19851 if (IndexVT == MVT::v8i32)
19852 // Just extend index
19853 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
19855 // The minimal number of elts in scatter is 8
19858 MVT NewIndexVT = MVT::getVectorVT(IndexVT.getScalarType(), NumElts);
19859 // Use original index here, do not modify the index twice
19860 Index = ExtendToType(N->getIndex(), NewIndexVT, DAG);
19861 if (IndexVT.getScalarType() == MVT::i32)
19862 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
19865 // At this point we have promoted mask operand
19866 assert(MaskVT.getScalarSizeInBits() >= 32 && "unexpected mask type");
19867 MVT ExtMaskVT = MVT::getVectorVT(MaskVT.getScalarType(), NumElts);
19868 // Use the original mask here, do not modify the mask twice
19869 Mask = ExtendToType(N->getMask(), ExtMaskVT, DAG, true);
19871 // The value that should be stored
19872 MVT NewVT = MVT::getVectorVT(VT.getScalarType(), NumElts);
19873 Src = ExtendToType(Src, NewVT, DAG);
19876 // If the mask is "wide" at this point - truncate it to i1 vector
19877 MVT BitMaskVT = MVT::getVectorVT(MVT::i1, NumElts);
19878 Mask = DAG.getNode(ISD::TRUNCATE, dl, BitMaskVT, Mask);
19880 // The mask is killed by scatter, add it to the values
19881 SDVTList VTs = DAG.getVTList(BitMaskVT, MVT::Other);
19882 SDValue Ops[] = {Chain, Src, Mask, BasePtr, Index};
19883 NewScatter = DAG.getMaskedScatter(VTs, N->getMemoryVT(), dl, Ops,
19884 N->getMemOperand());
19885 DAG.ReplaceAllUsesWith(Op, SDValue(NewScatter.getNode(), 1));
19886 return SDValue(NewScatter.getNode(), 0);
19889 static SDValue LowerMLOAD(SDValue Op, const X86Subtarget *Subtarget,
19890 SelectionDAG &DAG) {
19892 MaskedLoadSDNode *N = cast<MaskedLoadSDNode>(Op.getNode());
19893 MVT VT = Op.getSimpleValueType();
19894 SDValue Mask = N->getMask();
19897 if (Subtarget->hasAVX512() && !Subtarget->hasVLX() &&
19898 !VT.is512BitVector() && Mask.getValueType() == MVT::v8i1) {
19899 // This operation is legal for targets with VLX, but without
19900 // VLX the vector should be widened to 512 bit
19901 unsigned NumEltsInWideVec = 512/VT.getScalarSizeInBits();
19902 MVT WideDataVT = MVT::getVectorVT(VT.getScalarType(), NumEltsInWideVec);
19903 MVT WideMaskVT = MVT::getVectorVT(MVT::i1, NumEltsInWideVec);
19904 SDValue Src0 = N->getSrc0();
19905 Src0 = ExtendToType(Src0, WideDataVT, DAG);
19906 Mask = ExtendToType(Mask, WideMaskVT, DAG, true);
19907 SDValue NewLoad = DAG.getMaskedLoad(WideDataVT, dl, N->getChain(),
19908 N->getBasePtr(), Mask, Src0,
19909 N->getMemoryVT(), N->getMemOperand(),
19910 N->getExtensionType());
19912 SDValue Exract = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
19913 NewLoad.getValue(0),
19914 DAG.getIntPtrConstant(0, dl));
19915 SDValue RetOps[] = {Exract, NewLoad.getValue(1)};
19916 return DAG.getMergeValues(RetOps, dl);
19921 static SDValue LowerMSTORE(SDValue Op, const X86Subtarget *Subtarget,
19922 SelectionDAG &DAG) {
19923 MaskedStoreSDNode *N = cast<MaskedStoreSDNode>(Op.getNode());
19924 SDValue DataToStore = N->getValue();
19925 MVT VT = DataToStore.getSimpleValueType();
19926 SDValue Mask = N->getMask();
19929 if (Subtarget->hasAVX512() && !Subtarget->hasVLX() &&
19930 !VT.is512BitVector() && Mask.getValueType() == MVT::v8i1) {
19931 // This operation is legal for targets with VLX, but without
19932 // VLX the vector should be widened to 512 bit
19933 unsigned NumEltsInWideVec = 512/VT.getScalarSizeInBits();
19934 MVT WideDataVT = MVT::getVectorVT(VT.getScalarType(), NumEltsInWideVec);
19935 MVT WideMaskVT = MVT::getVectorVT(MVT::i1, NumEltsInWideVec);
19936 DataToStore = ExtendToType(DataToStore, WideDataVT, DAG);
19937 Mask = ExtendToType(Mask, WideMaskVT, DAG, true);
19938 return DAG.getMaskedStore(N->getChain(), dl, DataToStore, N->getBasePtr(),
19939 Mask, N->getMemoryVT(), N->getMemOperand(),
19940 N->isTruncatingStore());
19945 static SDValue LowerMGATHER(SDValue Op, const X86Subtarget *Subtarget,
19946 SelectionDAG &DAG) {
19947 assert(Subtarget->hasAVX512() &&
19948 "MGATHER/MSCATTER are supported on AVX-512 arch only");
19950 MaskedGatherSDNode *N = cast<MaskedGatherSDNode>(Op.getNode());
19952 MVT VT = Op.getSimpleValueType();
19953 SDValue Index = N->getIndex();
19954 SDValue Mask = N->getMask();
19955 SDValue Src0 = N->getValue();
19956 MVT IndexVT = Index.getSimpleValueType();
19957 MVT MaskVT = Mask.getSimpleValueType();
19959 unsigned NumElts = VT.getVectorNumElements();
19960 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported gather op");
19962 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
19963 !Index.getSimpleValueType().is512BitVector()) {
19964 // AVX512F supports only 512-bit vectors. Or data or index should
19965 // be 512 bit wide. If now the both index and data are 256-bit, but
19966 // the vector contains 8 elements, we just sign-extend the index
19967 if (NumElts == 8) {
19968 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
19969 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
19970 N->getOperand(3), Index };
19971 DAG.UpdateNodeOperands(N, Ops);
19975 // Minimal number of elements in Gather
19978 MVT NewIndexVT = MVT::getVectorVT(IndexVT.getScalarType(), NumElts);
19979 Index = ExtendToType(Index, NewIndexVT, DAG);
19980 if (IndexVT.getScalarType() == MVT::i32)
19981 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
19984 MVT MaskBitVT = MVT::getVectorVT(MVT::i1, NumElts);
19985 // At this point we have promoted mask operand
19986 assert(MaskVT.getScalarSizeInBits() >= 32 && "unexpected mask type");
19987 MVT ExtMaskVT = MVT::getVectorVT(MaskVT.getScalarType(), NumElts);
19988 Mask = ExtendToType(Mask, ExtMaskVT, DAG, true);
19989 Mask = DAG.getNode(ISD::TRUNCATE, dl, MaskBitVT, Mask);
19991 // The pass-thru value
19992 MVT NewVT = MVT::getVectorVT(VT.getScalarType(), NumElts);
19993 Src0 = ExtendToType(Src0, NewVT, DAG);
19995 SDValue Ops[] = { N->getChain(), Src0, Mask, N->getBasePtr(), Index };
19996 SDValue NewGather = DAG.getMaskedGather(DAG.getVTList(NewVT, MVT::Other),
19997 N->getMemoryVT(), dl, Ops,
19998 N->getMemOperand());
19999 SDValue Exract = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
20000 NewGather.getValue(0),
20001 DAG.getIntPtrConstant(0, dl));
20002 SDValue RetOps[] = {Exract, NewGather.getValue(1)};
20003 return DAG.getMergeValues(RetOps, dl);
20008 SDValue X86TargetLowering::LowerGC_TRANSITION_START(SDValue Op,
20009 SelectionDAG &DAG) const {
20010 // TODO: Eventually, the lowering of these nodes should be informed by or
20011 // deferred to the GC strategy for the function in which they appear. For
20012 // now, however, they must be lowered to something. Since they are logically
20013 // no-ops in the case of a null GC strategy (or a GC strategy which does not
20014 // require special handling for these nodes), lower them as literal NOOPs for
20016 SmallVector<SDValue, 2> Ops;
20018 Ops.push_back(Op.getOperand(0));
20019 if (Op->getGluedNode())
20020 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
20023 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
20024 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
20029 SDValue X86TargetLowering::LowerGC_TRANSITION_END(SDValue Op,
20030 SelectionDAG &DAG) const {
20031 // TODO: Eventually, the lowering of these nodes should be informed by or
20032 // deferred to the GC strategy for the function in which they appear. For
20033 // now, however, they must be lowered to something. Since they are logically
20034 // no-ops in the case of a null GC strategy (or a GC strategy which does not
20035 // require special handling for these nodes), lower them as literal NOOPs for
20037 SmallVector<SDValue, 2> Ops;
20039 Ops.push_back(Op.getOperand(0));
20040 if (Op->getGluedNode())
20041 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
20044 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
20045 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
20050 /// LowerOperation - Provide custom lowering hooks for some operations.
20052 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
20053 switch (Op.getOpcode()) {
20054 default: llvm_unreachable("Should not custom lower this!");
20055 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, Subtarget, DAG);
20056 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
20057 return LowerCMP_SWAP(Op, Subtarget, DAG);
20058 case ISD::CTPOP: return LowerCTPOP(Op, Subtarget, DAG);
20059 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
20060 case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG);
20061 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
20062 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, Subtarget, DAG);
20063 case ISD::VECTOR_SHUFFLE: return lowerVectorShuffle(Op, Subtarget, DAG);
20064 case ISD::VSELECT: return LowerVSELECT(Op, DAG);
20065 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
20066 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
20067 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op,Subtarget,DAG);
20068 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, Subtarget,DAG);
20069 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
20070 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
20071 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
20072 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
20073 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
20074 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
20075 case ISD::SHL_PARTS:
20076 case ISD::SRA_PARTS:
20077 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
20078 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
20079 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
20080 case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG);
20081 case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, Subtarget, DAG);
20082 case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, Subtarget, DAG);
20083 case ISD::ANY_EXTEND: return LowerANY_EXTEND(Op, Subtarget, DAG);
20084 case ISD::SIGN_EXTEND_VECTOR_INREG:
20085 return LowerSIGN_EXTEND_VECTOR_INREG(Op, Subtarget, DAG);
20086 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
20087 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
20088 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
20089 case ISD::LOAD: return LowerExtendedLoad(Op, Subtarget, DAG);
20091 case ISD::FNEG: return LowerFABSorFNEG(Op, DAG);
20092 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
20093 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
20094 case ISD::SETCC: return LowerSETCC(Op, DAG);
20095 case ISD::SETCCE: return LowerSETCCE(Op, DAG);
20096 case ISD::SELECT: return LowerSELECT(Op, DAG);
20097 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
20098 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
20099 case ISD::VASTART: return LowerVASTART(Op, DAG);
20100 case ISD::VAARG: return LowerVAARG(Op, DAG);
20101 case ISD::VACOPY: return LowerVACOPY(Op, Subtarget, DAG);
20102 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, Subtarget, DAG);
20103 case ISD::INTRINSIC_VOID:
20104 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, Subtarget, DAG);
20105 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
20106 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
20107 case ISD::FRAME_TO_ARGS_OFFSET:
20108 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
20109 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
20110 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
20111 case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
20112 case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
20113 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
20114 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
20115 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
20116 case ISD::CTLZ: return LowerCTLZ(Op, Subtarget, DAG);
20117 case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, Subtarget, DAG);
20119 case ISD::CTTZ_ZERO_UNDEF: return LowerCTTZ(Op, DAG);
20120 case ISD::MUL: return LowerMUL(Op, Subtarget, DAG);
20121 case ISD::UMUL_LOHI:
20122 case ISD::SMUL_LOHI: return LowerMUL_LOHI(Op, Subtarget, DAG);
20123 case ISD::ROTL: return LowerRotate(Op, Subtarget, DAG);
20126 case ISD::SHL: return LowerShift(Op, Subtarget, DAG);
20132 case ISD::UMULO: return LowerXALUO(Op, DAG);
20133 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, Subtarget,DAG);
20134 case ISD::BITCAST: return LowerBITCAST(Op, Subtarget, DAG);
20138 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
20139 case ISD::ADD: return LowerADD(Op, DAG);
20140 case ISD::SUB: return LowerSUB(Op, DAG);
20144 case ISD::UMIN: return LowerMINMAX(Op, DAG);
20145 case ISD::FSINCOS: return LowerFSINCOS(Op, Subtarget, DAG);
20146 case ISD::MLOAD: return LowerMLOAD(Op, Subtarget, DAG);
20147 case ISD::MSTORE: return LowerMSTORE(Op, Subtarget, DAG);
20148 case ISD::MGATHER: return LowerMGATHER(Op, Subtarget, DAG);
20149 case ISD::MSCATTER: return LowerMSCATTER(Op, Subtarget, DAG);
20150 case ISD::GC_TRANSITION_START:
20151 return LowerGC_TRANSITION_START(Op, DAG);
20152 case ISD::GC_TRANSITION_END: return LowerGC_TRANSITION_END(Op, DAG);
20156 /// ReplaceNodeResults - Replace a node with an illegal result type
20157 /// with a new node built out of custom code.
20158 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
20159 SmallVectorImpl<SDValue>&Results,
20160 SelectionDAG &DAG) const {
20162 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
20163 switch (N->getOpcode()) {
20165 llvm_unreachable("Do not know how to custom type legalize this operation!");
20166 case X86ISD::AVG: {
20167 // Legalize types for X86ISD::AVG by expanding vectors.
20168 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
20170 auto InVT = N->getValueType(0);
20171 auto InVTSize = InVT.getSizeInBits();
20172 const unsigned RegSize =
20173 (InVTSize > 128) ? ((InVTSize > 256) ? 512 : 256) : 128;
20174 assert((!Subtarget->hasAVX512() || RegSize < 512) &&
20175 "512-bit vector requires AVX512");
20176 assert((!Subtarget->hasAVX2() || RegSize < 256) &&
20177 "256-bit vector requires AVX2");
20179 auto ElemVT = InVT.getVectorElementType();
20180 auto RegVT = EVT::getVectorVT(*DAG.getContext(), ElemVT,
20181 RegSize / ElemVT.getSizeInBits());
20182 assert(RegSize % InVT.getSizeInBits() == 0);
20183 unsigned NumConcat = RegSize / InVT.getSizeInBits();
20185 SmallVector<SDValue, 16> Ops(NumConcat, DAG.getUNDEF(InVT));
20186 Ops[0] = N->getOperand(0);
20187 SDValue InVec0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, RegVT, Ops);
20188 Ops[0] = N->getOperand(1);
20189 SDValue InVec1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, RegVT, Ops);
20191 SDValue Res = DAG.getNode(X86ISD::AVG, dl, RegVT, InVec0, InVec1);
20192 Results.push_back(DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, InVT, Res,
20193 DAG.getIntPtrConstant(0, dl)));
20196 // We might have generated v2f32 FMIN/FMAX operations. Widen them to v4f32.
20197 case X86ISD::FMINC:
20199 case X86ISD::FMAXC:
20200 case X86ISD::FMAX: {
20201 EVT VT = N->getValueType(0);
20202 assert(VT == MVT::v2f32 && "Unexpected type (!= v2f32) on FMIN/FMAX.");
20203 SDValue UNDEF = DAG.getUNDEF(VT);
20204 SDValue LHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
20205 N->getOperand(0), UNDEF);
20206 SDValue RHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
20207 N->getOperand(1), UNDEF);
20208 Results.push_back(DAG.getNode(N->getOpcode(), dl, MVT::v4f32, LHS, RHS));
20211 case ISD::SIGN_EXTEND_INREG:
20216 // We don't want to expand or promote these.
20223 case ISD::UDIVREM: {
20224 SDValue V = LowerWin64_i128OP(SDValue(N,0), DAG);
20225 Results.push_back(V);
20228 case ISD::FP_TO_SINT:
20229 case ISD::FP_TO_UINT: {
20230 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
20232 std::pair<SDValue,SDValue> Vals =
20233 FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true);
20234 SDValue FIST = Vals.first, StackSlot = Vals.second;
20235 if (FIST.getNode()) {
20236 EVT VT = N->getValueType(0);
20237 // Return a load from the stack slot.
20238 if (StackSlot.getNode())
20239 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
20240 MachinePointerInfo(),
20241 false, false, false, 0));
20243 Results.push_back(FIST);
20247 case ISD::UINT_TO_FP: {
20248 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
20249 if (N->getOperand(0).getValueType() != MVT::v2i32 ||
20250 N->getValueType(0) != MVT::v2f32)
20252 SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64,
20254 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
20256 SDValue VBias = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2f64, Bias, Bias);
20257 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn,
20258 DAG.getBitcast(MVT::v2i64, VBias));
20259 Or = DAG.getBitcast(MVT::v2f64, Or);
20260 // TODO: Are there any fast-math-flags to propagate here?
20261 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias);
20262 Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub));
20265 case ISD::FP_ROUND: {
20266 if (!TLI.isTypeLegal(N->getOperand(0).getValueType()))
20268 SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0));
20269 Results.push_back(V);
20272 case ISD::FP_EXTEND: {
20273 // Right now, only MVT::v2f32 has OperationAction for FP_EXTEND.
20274 // No other ValueType for FP_EXTEND should reach this point.
20275 assert(N->getValueType(0) == MVT::v2f32 &&
20276 "Do not know how to legalize this Node");
20279 case ISD::INTRINSIC_W_CHAIN: {
20280 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
20282 default : llvm_unreachable("Do not know how to custom type "
20283 "legalize this intrinsic operation!");
20284 case Intrinsic::x86_rdtsc:
20285 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
20287 case Intrinsic::x86_rdtscp:
20288 return getReadTimeStampCounter(N, dl, X86ISD::RDTSCP_DAG, DAG, Subtarget,
20290 case Intrinsic::x86_rdpmc:
20291 return getReadPerformanceCounter(N, dl, DAG, Subtarget, Results);
20294 case ISD::INTRINSIC_WO_CHAIN: {
20295 if (SDValue V = LowerINTRINSIC_WO_CHAIN(SDValue(N, 0), Subtarget, DAG))
20296 Results.push_back(V);
20299 case ISD::READCYCLECOUNTER: {
20300 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
20303 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
20304 EVT T = N->getValueType(0);
20305 assert((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair");
20306 bool Regs64bit = T == MVT::i128;
20307 MVT HalfT = Regs64bit ? MVT::i64 : MVT::i32;
20308 SDValue cpInL, cpInH;
20309 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
20310 DAG.getConstant(0, dl, HalfT));
20311 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
20312 DAG.getConstant(1, dl, HalfT));
20313 cpInL = DAG.getCopyToReg(N->getOperand(0), dl,
20314 Regs64bit ? X86::RAX : X86::EAX,
20316 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl,
20317 Regs64bit ? X86::RDX : X86::EDX,
20318 cpInH, cpInL.getValue(1));
20319 SDValue swapInL, swapInH;
20320 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
20321 DAG.getConstant(0, dl, HalfT));
20322 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
20323 DAG.getConstant(1, dl, HalfT));
20324 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl,
20325 Regs64bit ? X86::RBX : X86::EBX,
20326 swapInL, cpInH.getValue(1));
20327 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl,
20328 Regs64bit ? X86::RCX : X86::ECX,
20329 swapInH, swapInL.getValue(1));
20330 SDValue Ops[] = { swapInH.getValue(0),
20332 swapInH.getValue(1) };
20333 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
20334 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
20335 unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_DAG :
20336 X86ISD::LCMPXCHG8_DAG;
20337 SDValue Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys, Ops, T, MMO);
20338 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl,
20339 Regs64bit ? X86::RAX : X86::EAX,
20340 HalfT, Result.getValue(1));
20341 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl,
20342 Regs64bit ? X86::RDX : X86::EDX,
20343 HalfT, cpOutL.getValue(2));
20344 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
20346 SDValue EFLAGS = DAG.getCopyFromReg(cpOutH.getValue(1), dl, X86::EFLAGS,
20347 MVT::i32, cpOutH.getValue(2));
20349 DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
20350 DAG.getConstant(X86::COND_E, dl, MVT::i8), EFLAGS);
20351 Success = DAG.getZExtOrTrunc(Success, dl, N->getValueType(1));
20353 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF));
20354 Results.push_back(Success);
20355 Results.push_back(EFLAGS.getValue(1));
20358 case ISD::ATOMIC_SWAP:
20359 case ISD::ATOMIC_LOAD_ADD:
20360 case ISD::ATOMIC_LOAD_SUB:
20361 case ISD::ATOMIC_LOAD_AND:
20362 case ISD::ATOMIC_LOAD_OR:
20363 case ISD::ATOMIC_LOAD_XOR:
20364 case ISD::ATOMIC_LOAD_NAND:
20365 case ISD::ATOMIC_LOAD_MIN:
20366 case ISD::ATOMIC_LOAD_MAX:
20367 case ISD::ATOMIC_LOAD_UMIN:
20368 case ISD::ATOMIC_LOAD_UMAX:
20369 case ISD::ATOMIC_LOAD: {
20370 // Delegate to generic TypeLegalization. Situations we can really handle
20371 // should have already been dealt with by AtomicExpandPass.cpp.
20374 case ISD::BITCAST: {
20375 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
20376 EVT DstVT = N->getValueType(0);
20377 EVT SrcVT = N->getOperand(0)->getValueType(0);
20379 if (SrcVT != MVT::f64 ||
20380 (DstVT != MVT::v2i32 && DstVT != MVT::v4i16 && DstVT != MVT::v8i8))
20383 unsigned NumElts = DstVT.getVectorNumElements();
20384 EVT SVT = DstVT.getVectorElementType();
20385 EVT WiderVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
20386 SDValue Expanded = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
20387 MVT::v2f64, N->getOperand(0));
20388 SDValue ToVecInt = DAG.getBitcast(WiderVT, Expanded);
20390 if (ExperimentalVectorWideningLegalization) {
20391 // If we are legalizing vectors by widening, we already have the desired
20392 // legal vector type, just return it.
20393 Results.push_back(ToVecInt);
20397 SmallVector<SDValue, 8> Elts;
20398 for (unsigned i = 0, e = NumElts; i != e; ++i)
20399 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT,
20400 ToVecInt, DAG.getIntPtrConstant(i, dl)));
20402 Results.push_back(DAG.getNode(ISD::BUILD_VECTOR, dl, DstVT, Elts));
20407 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
20408 switch ((X86ISD::NodeType)Opcode) {
20409 case X86ISD::FIRST_NUMBER: break;
20410 case X86ISD::BSF: return "X86ISD::BSF";
20411 case X86ISD::BSR: return "X86ISD::BSR";
20412 case X86ISD::SHLD: return "X86ISD::SHLD";
20413 case X86ISD::SHRD: return "X86ISD::SHRD";
20414 case X86ISD::FAND: return "X86ISD::FAND";
20415 case X86ISD::FANDN: return "X86ISD::FANDN";
20416 case X86ISD::FOR: return "X86ISD::FOR";
20417 case X86ISD::FXOR: return "X86ISD::FXOR";
20418 case X86ISD::FILD: return "X86ISD::FILD";
20419 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
20420 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
20421 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
20422 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
20423 case X86ISD::FLD: return "X86ISD::FLD";
20424 case X86ISD::FST: return "X86ISD::FST";
20425 case X86ISD::CALL: return "X86ISD::CALL";
20426 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
20427 case X86ISD::RDTSCP_DAG: return "X86ISD::RDTSCP_DAG";
20428 case X86ISD::RDPMC_DAG: return "X86ISD::RDPMC_DAG";
20429 case X86ISD::BT: return "X86ISD::BT";
20430 case X86ISD::CMP: return "X86ISD::CMP";
20431 case X86ISD::COMI: return "X86ISD::COMI";
20432 case X86ISD::UCOMI: return "X86ISD::UCOMI";
20433 case X86ISD::CMPM: return "X86ISD::CMPM";
20434 case X86ISD::CMPMU: return "X86ISD::CMPMU";
20435 case X86ISD::CMPM_RND: return "X86ISD::CMPM_RND";
20436 case X86ISD::SETCC: return "X86ISD::SETCC";
20437 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
20438 case X86ISD::FSETCC: return "X86ISD::FSETCC";
20439 case X86ISD::FGETSIGNx86: return "X86ISD::FGETSIGNx86";
20440 case X86ISD::CMOV: return "X86ISD::CMOV";
20441 case X86ISD::BRCOND: return "X86ISD::BRCOND";
20442 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
20443 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
20444 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
20445 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
20446 case X86ISD::Wrapper: return "X86ISD::Wrapper";
20447 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
20448 case X86ISD::MOVDQ2Q: return "X86ISD::MOVDQ2Q";
20449 case X86ISD::MMX_MOVD2W: return "X86ISD::MMX_MOVD2W";
20450 case X86ISD::MMX_MOVW2D: return "X86ISD::MMX_MOVW2D";
20451 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
20452 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
20453 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
20454 case X86ISD::PINSRB: return "X86ISD::PINSRB";
20455 case X86ISD::PINSRW: return "X86ISD::PINSRW";
20456 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
20457 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
20458 case X86ISD::ANDNP: return "X86ISD::ANDNP";
20459 case X86ISD::PSIGN: return "X86ISD::PSIGN";
20460 case X86ISD::BLENDI: return "X86ISD::BLENDI";
20461 case X86ISD::SHRUNKBLEND: return "X86ISD::SHRUNKBLEND";
20462 case X86ISD::ADDUS: return "X86ISD::ADDUS";
20463 case X86ISD::SUBUS: return "X86ISD::SUBUS";
20464 case X86ISD::HADD: return "X86ISD::HADD";
20465 case X86ISD::HSUB: return "X86ISD::HSUB";
20466 case X86ISD::FHADD: return "X86ISD::FHADD";
20467 case X86ISD::FHSUB: return "X86ISD::FHSUB";
20468 case X86ISD::ABS: return "X86ISD::ABS";
20469 case X86ISD::CONFLICT: return "X86ISD::CONFLICT";
20470 case X86ISD::FMAX: return "X86ISD::FMAX";
20471 case X86ISD::FMAX_RND: return "X86ISD::FMAX_RND";
20472 case X86ISD::FMIN: return "X86ISD::FMIN";
20473 case X86ISD::FMIN_RND: return "X86ISD::FMIN_RND";
20474 case X86ISD::FMAXC: return "X86ISD::FMAXC";
20475 case X86ISD::FMINC: return "X86ISD::FMINC";
20476 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
20477 case X86ISD::FRCP: return "X86ISD::FRCP";
20478 case X86ISD::EXTRQI: return "X86ISD::EXTRQI";
20479 case X86ISD::INSERTQI: return "X86ISD::INSERTQI";
20480 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
20481 case X86ISD::TLSBASEADDR: return "X86ISD::TLSBASEADDR";
20482 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
20483 case X86ISD::EH_SJLJ_SETJMP: return "X86ISD::EH_SJLJ_SETJMP";
20484 case X86ISD::EH_SJLJ_LONGJMP: return "X86ISD::EH_SJLJ_LONGJMP";
20485 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
20486 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
20487 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
20488 case X86ISD::FNSTSW16r: return "X86ISD::FNSTSW16r";
20489 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
20490 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
20491 case X86ISD::LCMPXCHG16_DAG: return "X86ISD::LCMPXCHG16_DAG";
20492 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
20493 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
20494 case X86ISD::VZEXT: return "X86ISD::VZEXT";
20495 case X86ISD::VSEXT: return "X86ISD::VSEXT";
20496 case X86ISD::VTRUNC: return "X86ISD::VTRUNC";
20497 case X86ISD::VTRUNCS: return "X86ISD::VTRUNCS";
20498 case X86ISD::VTRUNCUS: return "X86ISD::VTRUNCUS";
20499 case X86ISD::VINSERT: return "X86ISD::VINSERT";
20500 case X86ISD::VFPEXT: return "X86ISD::VFPEXT";
20501 case X86ISD::VFPROUND: return "X86ISD::VFPROUND";
20502 case X86ISD::CVTDQ2PD: return "X86ISD::CVTDQ2PD";
20503 case X86ISD::CVTUDQ2PD: return "X86ISD::CVTUDQ2PD";
20504 case X86ISD::VSHLDQ: return "X86ISD::VSHLDQ";
20505 case X86ISD::VSRLDQ: return "X86ISD::VSRLDQ";
20506 case X86ISD::VSHL: return "X86ISD::VSHL";
20507 case X86ISD::VSRL: return "X86ISD::VSRL";
20508 case X86ISD::VSRA: return "X86ISD::VSRA";
20509 case X86ISD::VSHLI: return "X86ISD::VSHLI";
20510 case X86ISD::VSRLI: return "X86ISD::VSRLI";
20511 case X86ISD::VSRAI: return "X86ISD::VSRAI";
20512 case X86ISD::CMPP: return "X86ISD::CMPP";
20513 case X86ISD::PCMPEQ: return "X86ISD::PCMPEQ";
20514 case X86ISD::PCMPGT: return "X86ISD::PCMPGT";
20515 case X86ISD::PCMPEQM: return "X86ISD::PCMPEQM";
20516 case X86ISD::PCMPGTM: return "X86ISD::PCMPGTM";
20517 case X86ISD::ADD: return "X86ISD::ADD";
20518 case X86ISD::SUB: return "X86ISD::SUB";
20519 case X86ISD::ADC: return "X86ISD::ADC";
20520 case X86ISD::SBB: return "X86ISD::SBB";
20521 case X86ISD::SMUL: return "X86ISD::SMUL";
20522 case X86ISD::UMUL: return "X86ISD::UMUL";
20523 case X86ISD::SMUL8: return "X86ISD::SMUL8";
20524 case X86ISD::UMUL8: return "X86ISD::UMUL8";
20525 case X86ISD::SDIVREM8_SEXT_HREG: return "X86ISD::SDIVREM8_SEXT_HREG";
20526 case X86ISD::UDIVREM8_ZEXT_HREG: return "X86ISD::UDIVREM8_ZEXT_HREG";
20527 case X86ISD::INC: return "X86ISD::INC";
20528 case X86ISD::DEC: return "X86ISD::DEC";
20529 case X86ISD::OR: return "X86ISD::OR";
20530 case X86ISD::XOR: return "X86ISD::XOR";
20531 case X86ISD::AND: return "X86ISD::AND";
20532 case X86ISD::BEXTR: return "X86ISD::BEXTR";
20533 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
20534 case X86ISD::PTEST: return "X86ISD::PTEST";
20535 case X86ISD::TESTP: return "X86ISD::TESTP";
20536 case X86ISD::TESTM: return "X86ISD::TESTM";
20537 case X86ISD::TESTNM: return "X86ISD::TESTNM";
20538 case X86ISD::KORTEST: return "X86ISD::KORTEST";
20539 case X86ISD::KTEST: return "X86ISD::KTEST";
20540 case X86ISD::PACKSS: return "X86ISD::PACKSS";
20541 case X86ISD::PACKUS: return "X86ISD::PACKUS";
20542 case X86ISD::PALIGNR: return "X86ISD::PALIGNR";
20543 case X86ISD::VALIGN: return "X86ISD::VALIGN";
20544 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
20545 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
20546 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
20547 case X86ISD::SHUFP: return "X86ISD::SHUFP";
20548 case X86ISD::SHUF128: return "X86ISD::SHUF128";
20549 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
20550 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
20551 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
20552 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
20553 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
20554 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
20555 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
20556 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
20557 case X86ISD::MOVSD: return "X86ISD::MOVSD";
20558 case X86ISD::MOVSS: return "X86ISD::MOVSS";
20559 case X86ISD::UNPCKL: return "X86ISD::UNPCKL";
20560 case X86ISD::UNPCKH: return "X86ISD::UNPCKH";
20561 case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
20562 case X86ISD::VBROADCASTM: return "X86ISD::VBROADCASTM";
20563 case X86ISD::SUBV_BROADCAST: return "X86ISD::SUBV_BROADCAST";
20564 case X86ISD::VEXTRACT: return "X86ISD::VEXTRACT";
20565 case X86ISD::VPERMILPV: return "X86ISD::VPERMILPV";
20566 case X86ISD::VPERMILPI: return "X86ISD::VPERMILPI";
20567 case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128";
20568 case X86ISD::VPERMV: return "X86ISD::VPERMV";
20569 case X86ISD::VPERMV3: return "X86ISD::VPERMV3";
20570 case X86ISD::VPERMIV3: return "X86ISD::VPERMIV3";
20571 case X86ISD::VPERMI: return "X86ISD::VPERMI";
20572 case X86ISD::VPTERNLOG: return "X86ISD::VPTERNLOG";
20573 case X86ISD::VFIXUPIMM: return "X86ISD::VFIXUPIMM";
20574 case X86ISD::VRANGE: return "X86ISD::VRANGE";
20575 case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
20576 case X86ISD::PMULDQ: return "X86ISD::PMULDQ";
20577 case X86ISD::PSADBW: return "X86ISD::PSADBW";
20578 case X86ISD::DBPSADBW: return "X86ISD::DBPSADBW";
20579 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
20580 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
20581 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
20582 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
20583 case X86ISD::MFENCE: return "X86ISD::MFENCE";
20584 case X86ISD::SFENCE: return "X86ISD::SFENCE";
20585 case X86ISD::LFENCE: return "X86ISD::LFENCE";
20586 case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA";
20587 case X86ISD::SAHF: return "X86ISD::SAHF";
20588 case X86ISD::RDRAND: return "X86ISD::RDRAND";
20589 case X86ISD::RDSEED: return "X86ISD::RDSEED";
20590 case X86ISD::VPMADDUBSW: return "X86ISD::VPMADDUBSW";
20591 case X86ISD::VPMADDWD: return "X86ISD::VPMADDWD";
20592 case X86ISD::VPROT: return "X86ISD::VPROT";
20593 case X86ISD::VPROTI: return "X86ISD::VPROTI";
20594 case X86ISD::VPSHA: return "X86ISD::VPSHA";
20595 case X86ISD::VPSHL: return "X86ISD::VPSHL";
20596 case X86ISD::VPCOM: return "X86ISD::VPCOM";
20597 case X86ISD::VPCOMU: return "X86ISD::VPCOMU";
20598 case X86ISD::FMADD: return "X86ISD::FMADD";
20599 case X86ISD::FMSUB: return "X86ISD::FMSUB";
20600 case X86ISD::FNMADD: return "X86ISD::FNMADD";
20601 case X86ISD::FNMSUB: return "X86ISD::FNMSUB";
20602 case X86ISD::FMADDSUB: return "X86ISD::FMADDSUB";
20603 case X86ISD::FMSUBADD: return "X86ISD::FMSUBADD";
20604 case X86ISD::FMADD_RND: return "X86ISD::FMADD_RND";
20605 case X86ISD::FNMADD_RND: return "X86ISD::FNMADD_RND";
20606 case X86ISD::FMSUB_RND: return "X86ISD::FMSUB_RND";
20607 case X86ISD::FNMSUB_RND: return "X86ISD::FNMSUB_RND";
20608 case X86ISD::FMADDSUB_RND: return "X86ISD::FMADDSUB_RND";
20609 case X86ISD::FMSUBADD_RND: return "X86ISD::FMSUBADD_RND";
20610 case X86ISD::VRNDSCALE: return "X86ISD::VRNDSCALE";
20611 case X86ISD::VREDUCE: return "X86ISD::VREDUCE";
20612 case X86ISD::VGETMANT: return "X86ISD::VGETMANT";
20613 case X86ISD::PCMPESTRI: return "X86ISD::PCMPESTRI";
20614 case X86ISD::PCMPISTRI: return "X86ISD::PCMPISTRI";
20615 case X86ISD::XTEST: return "X86ISD::XTEST";
20616 case X86ISD::COMPRESS: return "X86ISD::COMPRESS";
20617 case X86ISD::EXPAND: return "X86ISD::EXPAND";
20618 case X86ISD::SELECT: return "X86ISD::SELECT";
20619 case X86ISD::ADDSUB: return "X86ISD::ADDSUB";
20620 case X86ISD::RCP28: return "X86ISD::RCP28";
20621 case X86ISD::EXP2: return "X86ISD::EXP2";
20622 case X86ISD::RSQRT28: return "X86ISD::RSQRT28";
20623 case X86ISD::FADD_RND: return "X86ISD::FADD_RND";
20624 case X86ISD::FSUB_RND: return "X86ISD::FSUB_RND";
20625 case X86ISD::FMUL_RND: return "X86ISD::FMUL_RND";
20626 case X86ISD::FDIV_RND: return "X86ISD::FDIV_RND";
20627 case X86ISD::FSQRT_RND: return "X86ISD::FSQRT_RND";
20628 case X86ISD::FGETEXP_RND: return "X86ISD::FGETEXP_RND";
20629 case X86ISD::SCALEF: return "X86ISD::SCALEF";
20630 case X86ISD::ADDS: return "X86ISD::ADDS";
20631 case X86ISD::SUBS: return "X86ISD::SUBS";
20632 case X86ISD::AVG: return "X86ISD::AVG";
20633 case X86ISD::MULHRS: return "X86ISD::MULHRS";
20634 case X86ISD::SINT_TO_FP_RND: return "X86ISD::SINT_TO_FP_RND";
20635 case X86ISD::UINT_TO_FP_RND: return "X86ISD::UINT_TO_FP_RND";
20636 case X86ISD::FP_TO_SINT_RND: return "X86ISD::FP_TO_SINT_RND";
20637 case X86ISD::FP_TO_UINT_RND: return "X86ISD::FP_TO_UINT_RND";
20638 case X86ISD::VFPCLASS: return "X86ISD::VFPCLASS";
20639 case X86ISD::VFPCLASSS: return "X86ISD::VFPCLASSS";
20644 // isLegalAddressingMode - Return true if the addressing mode represented
20645 // by AM is legal for this target, for a load/store of the specified type.
20646 bool X86TargetLowering::isLegalAddressingMode(const DataLayout &DL,
20647 const AddrMode &AM, Type *Ty,
20648 unsigned AS) const {
20649 // X86 supports extremely general addressing modes.
20650 CodeModel::Model M = getTargetMachine().getCodeModel();
20651 Reloc::Model R = getTargetMachine().getRelocationModel();
20653 // X86 allows a sign-extended 32-bit immediate field as a displacement.
20654 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != nullptr))
20659 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
20661 // If a reference to this global requires an extra load, we can't fold it.
20662 if (isGlobalStubReference(GVFlags))
20665 // If BaseGV requires a register for the PIC base, we cannot also have a
20666 // BaseReg specified.
20667 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
20670 // If lower 4G is not available, then we must use rip-relative addressing.
20671 if ((M != CodeModel::Small || R != Reloc::Static) &&
20672 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
20676 switch (AM.Scale) {
20682 // These scales always work.
20687 // These scales are formed with basereg+scalereg. Only accept if there is
20692 default: // Other stuff never works.
20699 bool X86TargetLowering::isVectorShiftByScalarCheap(Type *Ty) const {
20700 unsigned Bits = Ty->getScalarSizeInBits();
20702 // 8-bit shifts are always expensive, but versions with a scalar amount aren't
20703 // particularly cheaper than those without.
20707 // On AVX2 there are new vpsllv[dq] instructions (and other shifts), that make
20708 // variable shifts just as cheap as scalar ones.
20709 if (Subtarget->hasInt256() && (Bits == 32 || Bits == 64))
20712 // Otherwise, it's significantly cheaper to shift by a scalar amount than by a
20713 // fully general vector.
20717 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
20718 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
20720 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
20721 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
20722 return NumBits1 > NumBits2;
20725 bool X86TargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
20726 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
20729 if (!isTypeLegal(EVT::getEVT(Ty1)))
20732 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
20734 // Assuming the caller doesn't have a zeroext or signext return parameter,
20735 // truncation all the way down to i1 is valid.
20739 bool X86TargetLowering::isLegalICmpImmediate(int64_t Imm) const {
20740 return isInt<32>(Imm);
20743 bool X86TargetLowering::isLegalAddImmediate(int64_t Imm) const {
20744 // Can also use sub to handle negated immediates.
20745 return isInt<32>(Imm);
20748 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
20749 if (!VT1.isInteger() || !VT2.isInteger())
20751 unsigned NumBits1 = VT1.getSizeInBits();
20752 unsigned NumBits2 = VT2.getSizeInBits();
20753 return NumBits1 > NumBits2;
20756 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
20757 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
20758 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
20761 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
20762 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
20763 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
20766 bool X86TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
20767 EVT VT1 = Val.getValueType();
20768 if (isZExtFree(VT1, VT2))
20771 if (Val.getOpcode() != ISD::LOAD)
20774 if (!VT1.isSimple() || !VT1.isInteger() ||
20775 !VT2.isSimple() || !VT2.isInteger())
20778 switch (VT1.getSimpleVT().SimpleTy) {
20783 // X86 has 8, 16, and 32-bit zero-extending loads.
20790 bool X86TargetLowering::isVectorLoadExtDesirable(SDValue) const { return true; }
20793 X86TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
20794 if (!Subtarget->hasAnyFMA())
20797 VT = VT.getScalarType();
20799 if (!VT.isSimple())
20802 switch (VT.getSimpleVT().SimpleTy) {
20813 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
20814 // i16 instructions are longer (0x66 prefix) and potentially slower.
20815 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
20818 /// isShuffleMaskLegal - Targets can use this to indicate that they only
20819 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
20820 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
20821 /// are assumed to be legal.
20823 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
20825 if (!VT.isSimple())
20828 // Not for i1 vectors
20829 if (VT.getSimpleVT().getScalarType() == MVT::i1)
20832 // Very little shuffling can be done for 64-bit vectors right now.
20833 if (VT.getSimpleVT().getSizeInBits() == 64)
20836 // We only care that the types being shuffled are legal. The lowering can
20837 // handle any possible shuffle mask that results.
20838 return isTypeLegal(VT.getSimpleVT());
20842 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
20844 // Just delegate to the generic legality, clear masks aren't special.
20845 return isShuffleMaskLegal(Mask, VT);
20848 //===----------------------------------------------------------------------===//
20849 // X86 Scheduler Hooks
20850 //===----------------------------------------------------------------------===//
20852 /// Utility function to emit xbegin specifying the start of an RTM region.
20853 static MachineBasicBlock *EmitXBegin(MachineInstr *MI, MachineBasicBlock *MBB,
20854 const TargetInstrInfo *TII) {
20855 DebugLoc DL = MI->getDebugLoc();
20857 const BasicBlock *BB = MBB->getBasicBlock();
20858 MachineFunction::iterator I = ++MBB->getIterator();
20860 // For the v = xbegin(), we generate
20871 MachineBasicBlock *thisMBB = MBB;
20872 MachineFunction *MF = MBB->getParent();
20873 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
20874 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
20875 MF->insert(I, mainMBB);
20876 MF->insert(I, sinkMBB);
20878 // Transfer the remainder of BB and its successor edges to sinkMBB.
20879 sinkMBB->splice(sinkMBB->begin(), MBB,
20880 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
20881 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
20885 // # fallthrough to mainMBB
20886 // # abortion to sinkMBB
20887 BuildMI(thisMBB, DL, TII->get(X86::XBEGIN_4)).addMBB(sinkMBB);
20888 thisMBB->addSuccessor(mainMBB);
20889 thisMBB->addSuccessor(sinkMBB);
20893 BuildMI(mainMBB, DL, TII->get(X86::MOV32ri), X86::EAX).addImm(-1);
20894 mainMBB->addSuccessor(sinkMBB);
20897 // EAX is live into the sinkMBB
20898 sinkMBB->addLiveIn(X86::EAX);
20899 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
20900 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20903 MI->eraseFromParent();
20907 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
20908 // or XMM0_V32I8 in AVX all of this code can be replaced with that
20909 // in the .td file.
20910 static MachineBasicBlock *EmitPCMPSTRM(MachineInstr *MI, MachineBasicBlock *BB,
20911 const TargetInstrInfo *TII) {
20913 switch (MI->getOpcode()) {
20914 default: llvm_unreachable("illegal opcode!");
20915 case X86::PCMPISTRM128REG: Opc = X86::PCMPISTRM128rr; break;
20916 case X86::VPCMPISTRM128REG: Opc = X86::VPCMPISTRM128rr; break;
20917 case X86::PCMPISTRM128MEM: Opc = X86::PCMPISTRM128rm; break;
20918 case X86::VPCMPISTRM128MEM: Opc = X86::VPCMPISTRM128rm; break;
20919 case X86::PCMPESTRM128REG: Opc = X86::PCMPESTRM128rr; break;
20920 case X86::VPCMPESTRM128REG: Opc = X86::VPCMPESTRM128rr; break;
20921 case X86::PCMPESTRM128MEM: Opc = X86::PCMPESTRM128rm; break;
20922 case X86::VPCMPESTRM128MEM: Opc = X86::VPCMPESTRM128rm; break;
20925 DebugLoc dl = MI->getDebugLoc();
20926 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
20928 unsigned NumArgs = MI->getNumOperands();
20929 for (unsigned i = 1; i < NumArgs; ++i) {
20930 MachineOperand &Op = MI->getOperand(i);
20931 if (!(Op.isReg() && Op.isImplicit()))
20932 MIB.addOperand(Op);
20934 if (MI->hasOneMemOperand())
20935 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
20937 BuildMI(*BB, MI, dl,
20938 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20939 .addReg(X86::XMM0);
20941 MI->eraseFromParent();
20945 // FIXME: Custom handling because TableGen doesn't support multiple implicit
20946 // defs in an instruction pattern
20947 static MachineBasicBlock *EmitPCMPSTRI(MachineInstr *MI, MachineBasicBlock *BB,
20948 const TargetInstrInfo *TII) {
20950 switch (MI->getOpcode()) {
20951 default: llvm_unreachable("illegal opcode!");
20952 case X86::PCMPISTRIREG: Opc = X86::PCMPISTRIrr; break;
20953 case X86::VPCMPISTRIREG: Opc = X86::VPCMPISTRIrr; break;
20954 case X86::PCMPISTRIMEM: Opc = X86::PCMPISTRIrm; break;
20955 case X86::VPCMPISTRIMEM: Opc = X86::VPCMPISTRIrm; break;
20956 case X86::PCMPESTRIREG: Opc = X86::PCMPESTRIrr; break;
20957 case X86::VPCMPESTRIREG: Opc = X86::VPCMPESTRIrr; break;
20958 case X86::PCMPESTRIMEM: Opc = X86::PCMPESTRIrm; break;
20959 case X86::VPCMPESTRIMEM: Opc = X86::VPCMPESTRIrm; break;
20962 DebugLoc dl = MI->getDebugLoc();
20963 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
20965 unsigned NumArgs = MI->getNumOperands(); // remove the results
20966 for (unsigned i = 1; i < NumArgs; ++i) {
20967 MachineOperand &Op = MI->getOperand(i);
20968 if (!(Op.isReg() && Op.isImplicit()))
20969 MIB.addOperand(Op);
20971 if (MI->hasOneMemOperand())
20972 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
20974 BuildMI(*BB, MI, dl,
20975 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20978 MI->eraseFromParent();
20982 static MachineBasicBlock *EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB,
20983 const X86Subtarget *Subtarget) {
20984 DebugLoc dl = MI->getDebugLoc();
20985 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20986 // Address into RAX/EAX, other two args into ECX, EDX.
20987 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
20988 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
20989 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
20990 for (int i = 0; i < X86::AddrNumOperands; ++i)
20991 MIB.addOperand(MI->getOperand(i));
20993 unsigned ValOps = X86::AddrNumOperands;
20994 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
20995 .addReg(MI->getOperand(ValOps).getReg());
20996 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
20997 .addReg(MI->getOperand(ValOps+1).getReg());
20999 // The instruction doesn't actually take any operands though.
21000 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
21002 MI->eraseFromParent(); // The pseudo is gone now.
21006 MachineBasicBlock *
21007 X86TargetLowering::EmitVAARG64WithCustomInserter(MachineInstr *MI,
21008 MachineBasicBlock *MBB) const {
21009 // Emit va_arg instruction on X86-64.
21011 // Operands to this pseudo-instruction:
21012 // 0 ) Output : destination address (reg)
21013 // 1-5) Input : va_list address (addr, i64mem)
21014 // 6 ) ArgSize : Size (in bytes) of vararg type
21015 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
21016 // 8 ) Align : Alignment of type
21017 // 9 ) EFLAGS (implicit-def)
21019 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
21020 static_assert(X86::AddrNumOperands == 5,
21021 "VAARG_64 assumes 5 address operands");
21023 unsigned DestReg = MI->getOperand(0).getReg();
21024 MachineOperand &Base = MI->getOperand(1);
21025 MachineOperand &Scale = MI->getOperand(2);
21026 MachineOperand &Index = MI->getOperand(3);
21027 MachineOperand &Disp = MI->getOperand(4);
21028 MachineOperand &Segment = MI->getOperand(5);
21029 unsigned ArgSize = MI->getOperand(6).getImm();
21030 unsigned ArgMode = MI->getOperand(7).getImm();
21031 unsigned Align = MI->getOperand(8).getImm();
21033 // Memory Reference
21034 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
21035 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
21036 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
21038 // Machine Information
21039 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21040 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
21041 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
21042 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
21043 DebugLoc DL = MI->getDebugLoc();
21045 // struct va_list {
21048 // i64 overflow_area (address)
21049 // i64 reg_save_area (address)
21051 // sizeof(va_list) = 24
21052 // alignment(va_list) = 8
21054 unsigned TotalNumIntRegs = 6;
21055 unsigned TotalNumXMMRegs = 8;
21056 bool UseGPOffset = (ArgMode == 1);
21057 bool UseFPOffset = (ArgMode == 2);
21058 unsigned MaxOffset = TotalNumIntRegs * 8 +
21059 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
21061 /* Align ArgSize to a multiple of 8 */
21062 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
21063 bool NeedsAlign = (Align > 8);
21065 MachineBasicBlock *thisMBB = MBB;
21066 MachineBasicBlock *overflowMBB;
21067 MachineBasicBlock *offsetMBB;
21068 MachineBasicBlock *endMBB;
21070 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
21071 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
21072 unsigned OffsetReg = 0;
21074 if (!UseGPOffset && !UseFPOffset) {
21075 // If we only pull from the overflow region, we don't create a branch.
21076 // We don't need to alter control flow.
21077 OffsetDestReg = 0; // unused
21078 OverflowDestReg = DestReg;
21080 offsetMBB = nullptr;
21081 overflowMBB = thisMBB;
21084 // First emit code to check if gp_offset (or fp_offset) is below the bound.
21085 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
21086 // If not, pull from overflow_area. (branch to overflowMBB)
21091 // offsetMBB overflowMBB
21096 // Registers for the PHI in endMBB
21097 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
21098 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
21100 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
21101 MachineFunction *MF = MBB->getParent();
21102 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21103 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21104 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21106 MachineFunction::iterator MBBIter = ++MBB->getIterator();
21108 // Insert the new basic blocks
21109 MF->insert(MBBIter, offsetMBB);
21110 MF->insert(MBBIter, overflowMBB);
21111 MF->insert(MBBIter, endMBB);
21113 // Transfer the remainder of MBB and its successor edges to endMBB.
21114 endMBB->splice(endMBB->begin(), thisMBB,
21115 std::next(MachineBasicBlock::iterator(MI)), thisMBB->end());
21116 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
21118 // Make offsetMBB and overflowMBB successors of thisMBB
21119 thisMBB->addSuccessor(offsetMBB);
21120 thisMBB->addSuccessor(overflowMBB);
21122 // endMBB is a successor of both offsetMBB and overflowMBB
21123 offsetMBB->addSuccessor(endMBB);
21124 overflowMBB->addSuccessor(endMBB);
21126 // Load the offset value into a register
21127 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
21128 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
21132 .addDisp(Disp, UseFPOffset ? 4 : 0)
21133 .addOperand(Segment)
21134 .setMemRefs(MMOBegin, MMOEnd);
21136 // Check if there is enough room left to pull this argument.
21137 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
21139 .addImm(MaxOffset + 8 - ArgSizeA8);
21141 // Branch to "overflowMBB" if offset >= max
21142 // Fall through to "offsetMBB" otherwise
21143 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
21144 .addMBB(overflowMBB);
21147 // In offsetMBB, emit code to use the reg_save_area.
21149 assert(OffsetReg != 0);
21151 // Read the reg_save_area address.
21152 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
21153 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
21158 .addOperand(Segment)
21159 .setMemRefs(MMOBegin, MMOEnd);
21161 // Zero-extend the offset
21162 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
21163 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
21166 .addImm(X86::sub_32bit);
21168 // Add the offset to the reg_save_area to get the final address.
21169 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
21170 .addReg(OffsetReg64)
21171 .addReg(RegSaveReg);
21173 // Compute the offset for the next argument
21174 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
21175 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
21177 .addImm(UseFPOffset ? 16 : 8);
21179 // Store it back into the va_list.
21180 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
21184 .addDisp(Disp, UseFPOffset ? 4 : 0)
21185 .addOperand(Segment)
21186 .addReg(NextOffsetReg)
21187 .setMemRefs(MMOBegin, MMOEnd);
21190 BuildMI(offsetMBB, DL, TII->get(X86::JMP_1))
21195 // Emit code to use overflow area
21198 // Load the overflow_area address into a register.
21199 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
21200 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
21205 .addOperand(Segment)
21206 .setMemRefs(MMOBegin, MMOEnd);
21208 // If we need to align it, do so. Otherwise, just copy the address
21209 // to OverflowDestReg.
21211 // Align the overflow address
21212 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
21213 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
21215 // aligned_addr = (addr + (align-1)) & ~(align-1)
21216 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
21217 .addReg(OverflowAddrReg)
21220 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
21222 .addImm(~(uint64_t)(Align-1));
21224 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
21225 .addReg(OverflowAddrReg);
21228 // Compute the next overflow address after this argument.
21229 // (the overflow address should be kept 8-byte aligned)
21230 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
21231 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
21232 .addReg(OverflowDestReg)
21233 .addImm(ArgSizeA8);
21235 // Store the new overflow address.
21236 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
21241 .addOperand(Segment)
21242 .addReg(NextAddrReg)
21243 .setMemRefs(MMOBegin, MMOEnd);
21245 // If we branched, emit the PHI to the front of endMBB.
21247 BuildMI(*endMBB, endMBB->begin(), DL,
21248 TII->get(X86::PHI), DestReg)
21249 .addReg(OffsetDestReg).addMBB(offsetMBB)
21250 .addReg(OverflowDestReg).addMBB(overflowMBB);
21253 // Erase the pseudo instruction
21254 MI->eraseFromParent();
21259 MachineBasicBlock *
21260 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
21262 MachineBasicBlock *MBB) const {
21263 // Emit code to save XMM registers to the stack. The ABI says that the
21264 // number of registers to save is given in %al, so it's theoretically
21265 // possible to do an indirect jump trick to avoid saving all of them,
21266 // however this code takes a simpler approach and just executes all
21267 // of the stores if %al is non-zero. It's less code, and it's probably
21268 // easier on the hardware branch predictor, and stores aren't all that
21269 // expensive anyway.
21271 // Create the new basic blocks. One block contains all the XMM stores,
21272 // and one block is the final destination regardless of whether any
21273 // stores were performed.
21274 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
21275 MachineFunction *F = MBB->getParent();
21276 MachineFunction::iterator MBBIter = ++MBB->getIterator();
21277 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
21278 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
21279 F->insert(MBBIter, XMMSaveMBB);
21280 F->insert(MBBIter, EndMBB);
21282 // Transfer the remainder of MBB and its successor edges to EndMBB.
21283 EndMBB->splice(EndMBB->begin(), MBB,
21284 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
21285 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
21287 // The original block will now fall through to the XMM save block.
21288 MBB->addSuccessor(XMMSaveMBB);
21289 // The XMMSaveMBB will fall through to the end block.
21290 XMMSaveMBB->addSuccessor(EndMBB);
21292 // Now add the instructions.
21293 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21294 DebugLoc DL = MI->getDebugLoc();
21296 unsigned CountReg = MI->getOperand(0).getReg();
21297 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
21298 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
21300 if (!Subtarget->isCallingConvWin64(F->getFunction()->getCallingConv())) {
21301 // If %al is 0, branch around the XMM save block.
21302 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
21303 BuildMI(MBB, DL, TII->get(X86::JE_1)).addMBB(EndMBB);
21304 MBB->addSuccessor(EndMBB);
21307 // Make sure the last operand is EFLAGS, which gets clobbered by the branch
21308 // that was just emitted, but clearly shouldn't be "saved".
21309 assert((MI->getNumOperands() <= 3 ||
21310 !MI->getOperand(MI->getNumOperands() - 1).isReg() ||
21311 MI->getOperand(MI->getNumOperands() - 1).getReg() == X86::EFLAGS)
21312 && "Expected last argument to be EFLAGS");
21313 unsigned MOVOpc = Subtarget->hasFp256() ? X86::VMOVAPSmr : X86::MOVAPSmr;
21314 // In the XMM save block, save all the XMM argument registers.
21315 for (int i = 3, e = MI->getNumOperands() - 1; i != e; ++i) {
21316 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
21317 MachineMemOperand *MMO = F->getMachineMemOperand(
21318 MachinePointerInfo::getFixedStack(*F, RegSaveFrameIndex, Offset),
21319 MachineMemOperand::MOStore,
21320 /*Size=*/16, /*Align=*/16);
21321 BuildMI(XMMSaveMBB, DL, TII->get(MOVOpc))
21322 .addFrameIndex(RegSaveFrameIndex)
21323 .addImm(/*Scale=*/1)
21324 .addReg(/*IndexReg=*/0)
21325 .addImm(/*Disp=*/Offset)
21326 .addReg(/*Segment=*/0)
21327 .addReg(MI->getOperand(i).getReg())
21328 .addMemOperand(MMO);
21331 MI->eraseFromParent(); // The pseudo instruction is gone now.
21336 // The EFLAGS operand of SelectItr might be missing a kill marker
21337 // because there were multiple uses of EFLAGS, and ISel didn't know
21338 // which to mark. Figure out whether SelectItr should have had a
21339 // kill marker, and set it if it should. Returns the correct kill
21341 static bool checkAndUpdateEFLAGSKill(MachineBasicBlock::iterator SelectItr,
21342 MachineBasicBlock* BB,
21343 const TargetRegisterInfo* TRI) {
21344 // Scan forward through BB for a use/def of EFLAGS.
21345 MachineBasicBlock::iterator miI(std::next(SelectItr));
21346 for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) {
21347 const MachineInstr& mi = *miI;
21348 if (mi.readsRegister(X86::EFLAGS))
21350 if (mi.definesRegister(X86::EFLAGS))
21351 break; // Should have kill-flag - update below.
21354 // If we hit the end of the block, check whether EFLAGS is live into a
21356 if (miI == BB->end()) {
21357 for (MachineBasicBlock::succ_iterator sItr = BB->succ_begin(),
21358 sEnd = BB->succ_end();
21359 sItr != sEnd; ++sItr) {
21360 MachineBasicBlock* succ = *sItr;
21361 if (succ->isLiveIn(X86::EFLAGS))
21366 // We found a def, or hit the end of the basic block and EFLAGS wasn't live
21367 // out. SelectMI should have a kill flag on EFLAGS.
21368 SelectItr->addRegisterKilled(X86::EFLAGS, TRI);
21372 // Return true if it is OK for this CMOV pseudo-opcode to be cascaded
21373 // together with other CMOV pseudo-opcodes into a single basic-block with
21374 // conditional jump around it.
21375 static bool isCMOVPseudo(MachineInstr *MI) {
21376 switch (MI->getOpcode()) {
21377 case X86::CMOV_FR32:
21378 case X86::CMOV_FR64:
21379 case X86::CMOV_GR8:
21380 case X86::CMOV_GR16:
21381 case X86::CMOV_GR32:
21382 case X86::CMOV_RFP32:
21383 case X86::CMOV_RFP64:
21384 case X86::CMOV_RFP80:
21385 case X86::CMOV_V2F64:
21386 case X86::CMOV_V2I64:
21387 case X86::CMOV_V4F32:
21388 case X86::CMOV_V4F64:
21389 case X86::CMOV_V4I64:
21390 case X86::CMOV_V16F32:
21391 case X86::CMOV_V8F32:
21392 case X86::CMOV_V8F64:
21393 case X86::CMOV_V8I64:
21394 case X86::CMOV_V8I1:
21395 case X86::CMOV_V16I1:
21396 case X86::CMOV_V32I1:
21397 case X86::CMOV_V64I1:
21405 MachineBasicBlock *
21406 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
21407 MachineBasicBlock *BB) const {
21408 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21409 DebugLoc DL = MI->getDebugLoc();
21411 // To "insert" a SELECT_CC instruction, we actually have to insert the
21412 // diamond control-flow pattern. The incoming instruction knows the
21413 // destination vreg to set, the condition code register to branch on, the
21414 // true/false values to select between, and a branch opcode to use.
21415 const BasicBlock *LLVM_BB = BB->getBasicBlock();
21416 MachineFunction::iterator It = ++BB->getIterator();
21421 // cmpTY ccX, r1, r2
21423 // fallthrough --> copy0MBB
21424 MachineBasicBlock *thisMBB = BB;
21425 MachineFunction *F = BB->getParent();
21427 // This code lowers all pseudo-CMOV instructions. Generally it lowers these
21428 // as described above, by inserting a BB, and then making a PHI at the join
21429 // point to select the true and false operands of the CMOV in the PHI.
21431 // The code also handles two different cases of multiple CMOV opcodes
21435 // In this case, there are multiple CMOVs in a row, all which are based on
21436 // the same condition setting (or the exact opposite condition setting).
21437 // In this case we can lower all the CMOVs using a single inserted BB, and
21438 // then make a number of PHIs at the join point to model the CMOVs. The only
21439 // trickiness here, is that in a case like:
21441 // t2 = CMOV cond1 t1, f1
21442 // t3 = CMOV cond1 t2, f2
21444 // when rewriting this into PHIs, we have to perform some renaming on the
21445 // temps since you cannot have a PHI operand refer to a PHI result earlier
21446 // in the same block. The "simple" but wrong lowering would be:
21448 // t2 = PHI t1(BB1), f1(BB2)
21449 // t3 = PHI t2(BB1), f2(BB2)
21451 // but clearly t2 is not defined in BB1, so that is incorrect. The proper
21452 // renaming is to note that on the path through BB1, t2 is really just a
21453 // copy of t1, and do that renaming, properly generating:
21455 // t2 = PHI t1(BB1), f1(BB2)
21456 // t3 = PHI t1(BB1), f2(BB2)
21458 // Case 2, we lower cascaded CMOVs such as
21460 // (CMOV (CMOV F, T, cc1), T, cc2)
21462 // to two successives branches. For that, we look for another CMOV as the
21463 // following instruction.
21465 // Without this, we would add a PHI between the two jumps, which ends up
21466 // creating a few copies all around. For instance, for
21468 // (sitofp (zext (fcmp une)))
21470 // we would generate:
21472 // ucomiss %xmm1, %xmm0
21473 // movss <1.0f>, %xmm0
21474 // movaps %xmm0, %xmm1
21476 // xorps %xmm1, %xmm1
21479 // movaps %xmm1, %xmm0
21483 // because this custom-inserter would have generated:
21495 // A: X = ...; Y = ...
21497 // C: Z = PHI [X, A], [Y, B]
21499 // E: PHI [X, C], [Z, D]
21501 // If we lower both CMOVs in a single step, we can instead generate:
21513 // A: X = ...; Y = ...
21515 // E: PHI [X, A], [X, C], [Y, D]
21517 // Which, in our sitofp/fcmp example, gives us something like:
21519 // ucomiss %xmm1, %xmm0
21520 // movss <1.0f>, %xmm0
21523 // xorps %xmm0, %xmm0
21527 MachineInstr *CascadedCMOV = nullptr;
21528 MachineInstr *LastCMOV = MI;
21529 X86::CondCode CC = X86::CondCode(MI->getOperand(3).getImm());
21530 X86::CondCode OppCC = X86::GetOppositeBranchCondition(CC);
21531 MachineBasicBlock::iterator NextMIIt =
21532 std::next(MachineBasicBlock::iterator(MI));
21534 // Check for case 1, where there are multiple CMOVs with the same condition
21535 // first. Of the two cases of multiple CMOV lowerings, case 1 reduces the
21536 // number of jumps the most.
21538 if (isCMOVPseudo(MI)) {
21539 // See if we have a string of CMOVS with the same condition.
21540 while (NextMIIt != BB->end() &&
21541 isCMOVPseudo(NextMIIt) &&
21542 (NextMIIt->getOperand(3).getImm() == CC ||
21543 NextMIIt->getOperand(3).getImm() == OppCC)) {
21544 LastCMOV = &*NextMIIt;
21549 // This checks for case 2, but only do this if we didn't already find
21550 // case 1, as indicated by LastCMOV == MI.
21551 if (LastCMOV == MI &&
21552 NextMIIt != BB->end() && NextMIIt->getOpcode() == MI->getOpcode() &&
21553 NextMIIt->getOperand(2).getReg() == MI->getOperand(2).getReg() &&
21554 NextMIIt->getOperand(1).getReg() == MI->getOperand(0).getReg()) {
21555 CascadedCMOV = &*NextMIIt;
21558 MachineBasicBlock *jcc1MBB = nullptr;
21560 // If we have a cascaded CMOV, we lower it to two successive branches to
21561 // the same block. EFLAGS is used by both, so mark it as live in the second.
21562 if (CascadedCMOV) {
21563 jcc1MBB = F->CreateMachineBasicBlock(LLVM_BB);
21564 F->insert(It, jcc1MBB);
21565 jcc1MBB->addLiveIn(X86::EFLAGS);
21568 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
21569 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
21570 F->insert(It, copy0MBB);
21571 F->insert(It, sinkMBB);
21573 // If the EFLAGS register isn't dead in the terminator, then claim that it's
21574 // live into the sink and copy blocks.
21575 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
21577 MachineInstr *LastEFLAGSUser = CascadedCMOV ? CascadedCMOV : LastCMOV;
21578 if (!LastEFLAGSUser->killsRegister(X86::EFLAGS) &&
21579 !checkAndUpdateEFLAGSKill(LastEFLAGSUser, BB, TRI)) {
21580 copy0MBB->addLiveIn(X86::EFLAGS);
21581 sinkMBB->addLiveIn(X86::EFLAGS);
21584 // Transfer the remainder of BB and its successor edges to sinkMBB.
21585 sinkMBB->splice(sinkMBB->begin(), BB,
21586 std::next(MachineBasicBlock::iterator(LastCMOV)), BB->end());
21587 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
21589 // Add the true and fallthrough blocks as its successors.
21590 if (CascadedCMOV) {
21591 // The fallthrough block may be jcc1MBB, if we have a cascaded CMOV.
21592 BB->addSuccessor(jcc1MBB);
21594 // In that case, jcc1MBB will itself fallthrough the copy0MBB, and
21595 // jump to the sinkMBB.
21596 jcc1MBB->addSuccessor(copy0MBB);
21597 jcc1MBB->addSuccessor(sinkMBB);
21599 BB->addSuccessor(copy0MBB);
21602 // The true block target of the first (or only) branch is always sinkMBB.
21603 BB->addSuccessor(sinkMBB);
21605 // Create the conditional branch instruction.
21606 unsigned Opc = X86::GetCondBranchFromCond(CC);
21607 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
21609 if (CascadedCMOV) {
21610 unsigned Opc2 = X86::GetCondBranchFromCond(
21611 (X86::CondCode)CascadedCMOV->getOperand(3).getImm());
21612 BuildMI(jcc1MBB, DL, TII->get(Opc2)).addMBB(sinkMBB);
21616 // %FalseValue = ...
21617 // # fallthrough to sinkMBB
21618 copy0MBB->addSuccessor(sinkMBB);
21621 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
21623 MachineBasicBlock::iterator MIItBegin = MachineBasicBlock::iterator(MI);
21624 MachineBasicBlock::iterator MIItEnd =
21625 std::next(MachineBasicBlock::iterator(LastCMOV));
21626 MachineBasicBlock::iterator SinkInsertionPoint = sinkMBB->begin();
21627 DenseMap<unsigned, std::pair<unsigned, unsigned>> RegRewriteTable;
21628 MachineInstrBuilder MIB;
21630 // As we are creating the PHIs, we have to be careful if there is more than
21631 // one. Later CMOVs may reference the results of earlier CMOVs, but later
21632 // PHIs have to reference the individual true/false inputs from earlier PHIs.
21633 // That also means that PHI construction must work forward from earlier to
21634 // later, and that the code must maintain a mapping from earlier PHI's
21635 // destination registers, and the registers that went into the PHI.
21637 for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd; ++MIIt) {
21638 unsigned DestReg = MIIt->getOperand(0).getReg();
21639 unsigned Op1Reg = MIIt->getOperand(1).getReg();
21640 unsigned Op2Reg = MIIt->getOperand(2).getReg();
21642 // If this CMOV we are generating is the opposite condition from
21643 // the jump we generated, then we have to swap the operands for the
21644 // PHI that is going to be generated.
21645 if (MIIt->getOperand(3).getImm() == OppCC)
21646 std::swap(Op1Reg, Op2Reg);
21648 if (RegRewriteTable.find(Op1Reg) != RegRewriteTable.end())
21649 Op1Reg = RegRewriteTable[Op1Reg].first;
21651 if (RegRewriteTable.find(Op2Reg) != RegRewriteTable.end())
21652 Op2Reg = RegRewriteTable[Op2Reg].second;
21654 MIB = BuildMI(*sinkMBB, SinkInsertionPoint, DL,
21655 TII->get(X86::PHI), DestReg)
21656 .addReg(Op1Reg).addMBB(copy0MBB)
21657 .addReg(Op2Reg).addMBB(thisMBB);
21659 // Add this PHI to the rewrite table.
21660 RegRewriteTable[DestReg] = std::make_pair(Op1Reg, Op2Reg);
21663 // If we have a cascaded CMOV, the second Jcc provides the same incoming
21664 // value as the first Jcc (the True operand of the SELECT_CC/CMOV nodes).
21665 if (CascadedCMOV) {
21666 MIB.addReg(MI->getOperand(2).getReg()).addMBB(jcc1MBB);
21667 // Copy the PHI result to the register defined by the second CMOV.
21668 BuildMI(*sinkMBB, std::next(MachineBasicBlock::iterator(MIB.getInstr())),
21669 DL, TII->get(TargetOpcode::COPY),
21670 CascadedCMOV->getOperand(0).getReg())
21671 .addReg(MI->getOperand(0).getReg());
21672 CascadedCMOV->eraseFromParent();
21675 // Now remove the CMOV(s).
21676 for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd; )
21677 (MIIt++)->eraseFromParent();
21682 MachineBasicBlock *
21683 X86TargetLowering::EmitLoweredAtomicFP(MachineInstr *MI,
21684 MachineBasicBlock *BB) const {
21685 // Combine the following atomic floating-point modification pattern:
21686 // a.store(reg OP a.load(acquire), release)
21687 // Transform them into:
21688 // OPss (%gpr), %xmm
21689 // movss %xmm, (%gpr)
21690 // Or sd equivalent for 64-bit operations.
21692 switch (MI->getOpcode()) {
21693 default: llvm_unreachable("unexpected instr type for EmitLoweredAtomicFP");
21694 case X86::RELEASE_FADD32mr: MOp = X86::MOVSSmr; FOp = X86::ADDSSrm; break;
21695 case X86::RELEASE_FADD64mr: MOp = X86::MOVSDmr; FOp = X86::ADDSDrm; break;
21697 const X86InstrInfo *TII = Subtarget->getInstrInfo();
21698 DebugLoc DL = MI->getDebugLoc();
21699 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
21700 MachineOperand MSrc = MI->getOperand(0);
21701 unsigned VSrc = MI->getOperand(5).getReg();
21702 const MachineOperand &Disp = MI->getOperand(3);
21703 MachineOperand ZeroDisp = MachineOperand::CreateImm(0);
21704 bool hasDisp = Disp.isGlobal() || Disp.isImm();
21705 if (hasDisp && MSrc.isReg())
21706 MSrc.setIsKill(false);
21707 MachineInstrBuilder MIM = BuildMI(*BB, MI, DL, TII->get(MOp))
21708 .addOperand(/*Base=*/MSrc)
21709 .addImm(/*Scale=*/1)
21710 .addReg(/*Index=*/0)
21711 .addDisp(hasDisp ? Disp : ZeroDisp, /*off=*/0)
21713 MachineInstr *MIO = BuildMI(*BB, (MachineInstr *)MIM, DL, TII->get(FOp),
21714 MRI.createVirtualRegister(MRI.getRegClass(VSrc)))
21716 .addOperand(/*Base=*/MSrc)
21717 .addImm(/*Scale=*/1)
21718 .addReg(/*Index=*/0)
21719 .addDisp(hasDisp ? Disp : ZeroDisp, /*off=*/0)
21720 .addReg(/*Segment=*/0);
21721 MIM.addReg(MIO->getOperand(0).getReg(), RegState::Kill);
21722 MI->eraseFromParent(); // The pseudo instruction is gone now.
21726 MachineBasicBlock *
21727 X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI,
21728 MachineBasicBlock *BB) const {
21729 MachineFunction *MF = BB->getParent();
21730 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21731 DebugLoc DL = MI->getDebugLoc();
21732 const BasicBlock *LLVM_BB = BB->getBasicBlock();
21734 assert(MF->shouldSplitStack());
21736 const bool Is64Bit = Subtarget->is64Bit();
21737 const bool IsLP64 = Subtarget->isTarget64BitLP64();
21739 const unsigned TlsReg = Is64Bit ? X86::FS : X86::GS;
21740 const unsigned TlsOffset = IsLP64 ? 0x70 : Is64Bit ? 0x40 : 0x30;
21743 // ... [Till the alloca]
21744 // If stacklet is not large enough, jump to mallocMBB
21747 // Allocate by subtracting from RSP
21748 // Jump to continueMBB
21751 // Allocate by call to runtime
21755 // [rest of original BB]
21758 MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21759 MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21760 MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21762 MachineRegisterInfo &MRI = MF->getRegInfo();
21763 const TargetRegisterClass *AddrRegClass =
21764 getRegClassFor(getPointerTy(MF->getDataLayout()));
21766 unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass),
21767 bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass),
21768 tmpSPVReg = MRI.createVirtualRegister(AddrRegClass),
21769 SPLimitVReg = MRI.createVirtualRegister(AddrRegClass),
21770 sizeVReg = MI->getOperand(1).getReg(),
21771 physSPReg = IsLP64 || Subtarget->isTargetNaCl64() ? X86::RSP : X86::ESP;
21773 MachineFunction::iterator MBBIter = ++BB->getIterator();
21775 MF->insert(MBBIter, bumpMBB);
21776 MF->insert(MBBIter, mallocMBB);
21777 MF->insert(MBBIter, continueMBB);
21779 continueMBB->splice(continueMBB->begin(), BB,
21780 std::next(MachineBasicBlock::iterator(MI)), BB->end());
21781 continueMBB->transferSuccessorsAndUpdatePHIs(BB);
21783 // Add code to the main basic block to check if the stack limit has been hit,
21784 // and if so, jump to mallocMBB otherwise to bumpMBB.
21785 BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg);
21786 BuildMI(BB, DL, TII->get(IsLP64 ? X86::SUB64rr:X86::SUB32rr), SPLimitVReg)
21787 .addReg(tmpSPVReg).addReg(sizeVReg);
21788 BuildMI(BB, DL, TII->get(IsLP64 ? X86::CMP64mr:X86::CMP32mr))
21789 .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg)
21790 .addReg(SPLimitVReg);
21791 BuildMI(BB, DL, TII->get(X86::JG_1)).addMBB(mallocMBB);
21793 // bumpMBB simply decreases the stack pointer, since we know the current
21794 // stacklet has enough space.
21795 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg)
21796 .addReg(SPLimitVReg);
21797 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg)
21798 .addReg(SPLimitVReg);
21799 BuildMI(bumpMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
21801 // Calls into a routine in libgcc to allocate more space from the heap.
21802 const uint32_t *RegMask =
21803 Subtarget->getRegisterInfo()->getCallPreservedMask(*MF, CallingConv::C);
21805 BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
21807 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
21808 .addExternalSymbol("__morestack_allocate_stack_space")
21809 .addRegMask(RegMask)
21810 .addReg(X86::RDI, RegState::Implicit)
21811 .addReg(X86::RAX, RegState::ImplicitDefine);
21812 } else if (Is64Bit) {
21813 BuildMI(mallocMBB, DL, TII->get(X86::MOV32rr), X86::EDI)
21815 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
21816 .addExternalSymbol("__morestack_allocate_stack_space")
21817 .addRegMask(RegMask)
21818 .addReg(X86::EDI, RegState::Implicit)
21819 .addReg(X86::EAX, RegState::ImplicitDefine);
21821 BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg)
21823 BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg);
21824 BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32))
21825 .addExternalSymbol("__morestack_allocate_stack_space")
21826 .addRegMask(RegMask)
21827 .addReg(X86::EAX, RegState::ImplicitDefine);
21831 BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg)
21834 BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg)
21835 .addReg(IsLP64 ? X86::RAX : X86::EAX);
21836 BuildMI(mallocMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
21838 // Set up the CFG correctly.
21839 BB->addSuccessor(bumpMBB);
21840 BB->addSuccessor(mallocMBB);
21841 mallocMBB->addSuccessor(continueMBB);
21842 bumpMBB->addSuccessor(continueMBB);
21844 // Take care of the PHI nodes.
21845 BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI),
21846 MI->getOperand(0).getReg())
21847 .addReg(mallocPtrVReg).addMBB(mallocMBB)
21848 .addReg(bumpSPPtrVReg).addMBB(bumpMBB);
21850 // Delete the original pseudo instruction.
21851 MI->eraseFromParent();
21854 return continueMBB;
21857 MachineBasicBlock *
21858 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
21859 MachineBasicBlock *BB) const {
21860 assert(!Subtarget->isTargetMachO());
21861 DebugLoc DL = MI->getDebugLoc();
21862 MachineInstr *ResumeMI = Subtarget->getFrameLowering()->emitStackProbe(
21863 *BB->getParent(), *BB, MI, DL, false);
21864 MachineBasicBlock *ResumeBB = ResumeMI->getParent();
21865 MI->eraseFromParent(); // The pseudo instruction is gone now.
21869 MachineBasicBlock *
21870 X86TargetLowering::EmitLoweredCatchRet(MachineInstr *MI,
21871 MachineBasicBlock *BB) const {
21872 MachineFunction *MF = BB->getParent();
21873 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
21874 MachineBasicBlock *TargetMBB = MI->getOperand(0).getMBB();
21875 DebugLoc DL = MI->getDebugLoc();
21877 assert(!isAsynchronousEHPersonality(
21878 classifyEHPersonality(MF->getFunction()->getPersonalityFn())) &&
21879 "SEH does not use catchret!");
21881 // Only 32-bit EH needs to worry about manually restoring stack pointers.
21882 if (!Subtarget->is32Bit())
21885 // C++ EH creates a new target block to hold the restore code, and wires up
21886 // the new block to the return destination with a normal JMP_4.
21887 MachineBasicBlock *RestoreMBB =
21888 MF->CreateMachineBasicBlock(BB->getBasicBlock());
21889 assert(BB->succ_size() == 1);
21890 MF->insert(std::next(BB->getIterator()), RestoreMBB);
21891 RestoreMBB->transferSuccessorsAndUpdatePHIs(BB);
21892 BB->addSuccessor(RestoreMBB);
21893 MI->getOperand(0).setMBB(RestoreMBB);
21895 auto RestoreMBBI = RestoreMBB->begin();
21896 BuildMI(*RestoreMBB, RestoreMBBI, DL, TII.get(X86::EH_RESTORE));
21897 BuildMI(*RestoreMBB, RestoreMBBI, DL, TII.get(X86::JMP_4)).addMBB(TargetMBB);
21901 MachineBasicBlock *
21902 X86TargetLowering::EmitLoweredCatchPad(MachineInstr *MI,
21903 MachineBasicBlock *BB) const {
21904 MachineFunction *MF = BB->getParent();
21905 const Constant *PerFn = MF->getFunction()->getPersonalityFn();
21906 bool IsSEH = isAsynchronousEHPersonality(classifyEHPersonality(PerFn));
21907 // Only 32-bit SEH requires special handling for catchpad.
21908 if (IsSEH && Subtarget->is32Bit()) {
21909 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
21910 DebugLoc DL = MI->getDebugLoc();
21911 BuildMI(*BB, MI, DL, TII.get(X86::EH_RESTORE));
21913 MI->eraseFromParent();
21917 MachineBasicBlock *
21918 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
21919 MachineBasicBlock *BB) const {
21920 // This is pretty easy. We're taking the value that we received from
21921 // our load from the relocation, sticking it in either RDI (x86-64)
21922 // or EAX and doing an indirect call. The return value will then
21923 // be in the normal return register.
21924 MachineFunction *F = BB->getParent();
21925 const X86InstrInfo *TII = Subtarget->getInstrInfo();
21926 DebugLoc DL = MI->getDebugLoc();
21928 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
21929 assert(MI->getOperand(3).isGlobal() && "This should be a global");
21931 // Get a register mask for the lowered call.
21932 // FIXME: The 32-bit calls have non-standard calling conventions. Use a
21933 // proper register mask.
21934 const uint32_t *RegMask =
21935 Subtarget->is64Bit() ?
21936 Subtarget->getRegisterInfo()->getDarwinTLSCallPreservedMask() :
21937 Subtarget->getRegisterInfo()->getCallPreservedMask(*F, CallingConv::C);
21938 if (Subtarget->is64Bit()) {
21939 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
21940 TII->get(X86::MOV64rm), X86::RDI)
21942 .addImm(0).addReg(0)
21943 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
21944 MI->getOperand(3).getTargetFlags())
21946 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
21947 addDirectMem(MIB, X86::RDI);
21948 MIB.addReg(X86::RAX, RegState::ImplicitDefine).addRegMask(RegMask);
21949 } else if (F->getTarget().getRelocationModel() != Reloc::PIC_) {
21950 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
21951 TII->get(X86::MOV32rm), X86::EAX)
21953 .addImm(0).addReg(0)
21954 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
21955 MI->getOperand(3).getTargetFlags())
21957 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
21958 addDirectMem(MIB, X86::EAX);
21959 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
21961 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
21962 TII->get(X86::MOV32rm), X86::EAX)
21963 .addReg(TII->getGlobalBaseReg(F))
21964 .addImm(0).addReg(0)
21965 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
21966 MI->getOperand(3).getTargetFlags())
21968 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
21969 addDirectMem(MIB, X86::EAX);
21970 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
21973 MI->eraseFromParent(); // The pseudo instruction is gone now.
21977 MachineBasicBlock *
21978 X86TargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
21979 MachineBasicBlock *MBB) const {
21980 DebugLoc DL = MI->getDebugLoc();
21981 MachineFunction *MF = MBB->getParent();
21982 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21983 MachineRegisterInfo &MRI = MF->getRegInfo();
21985 const BasicBlock *BB = MBB->getBasicBlock();
21986 MachineFunction::iterator I = ++MBB->getIterator();
21988 // Memory Reference
21989 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
21990 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
21993 unsigned MemOpndSlot = 0;
21995 unsigned CurOp = 0;
21997 DstReg = MI->getOperand(CurOp++).getReg();
21998 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
21999 assert(RC->hasType(MVT::i32) && "Invalid destination!");
22000 unsigned mainDstReg = MRI.createVirtualRegister(RC);
22001 unsigned restoreDstReg = MRI.createVirtualRegister(RC);
22003 MemOpndSlot = CurOp;
22005 MVT PVT = getPointerTy(MF->getDataLayout());
22006 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
22007 "Invalid Pointer Size!");
22009 // For v = setjmp(buf), we generate
22012 // buf[LabelOffset] = restoreMBB <-- takes address of restoreMBB
22013 // SjLjSetup restoreMBB
22019 // v = phi(main, restore)
22022 // if base pointer being used, load it from frame
22025 MachineBasicBlock *thisMBB = MBB;
22026 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
22027 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
22028 MachineBasicBlock *restoreMBB = MF->CreateMachineBasicBlock(BB);
22029 MF->insert(I, mainMBB);
22030 MF->insert(I, sinkMBB);
22031 MF->push_back(restoreMBB);
22032 restoreMBB->setHasAddressTaken();
22034 MachineInstrBuilder MIB;
22036 // Transfer the remainder of BB and its successor edges to sinkMBB.
22037 sinkMBB->splice(sinkMBB->begin(), MBB,
22038 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
22039 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
22042 unsigned PtrStoreOpc = 0;
22043 unsigned LabelReg = 0;
22044 const int64_t LabelOffset = 1 * PVT.getStoreSize();
22045 Reloc::Model RM = MF->getTarget().getRelocationModel();
22046 bool UseImmLabel = (MF->getTarget().getCodeModel() == CodeModel::Small) &&
22047 (RM == Reloc::Static || RM == Reloc::DynamicNoPIC);
22049 // Prepare IP either in reg or imm.
22050 if (!UseImmLabel) {
22051 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr;
22052 const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
22053 LabelReg = MRI.createVirtualRegister(PtrRC);
22054 if (Subtarget->is64Bit()) {
22055 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA64r), LabelReg)
22059 .addMBB(restoreMBB)
22062 const X86InstrInfo *XII = static_cast<const X86InstrInfo*>(TII);
22063 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA32r), LabelReg)
22064 .addReg(XII->getGlobalBaseReg(MF))
22067 .addMBB(restoreMBB, Subtarget->ClassifyBlockAddressReference())
22071 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi;
22073 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PtrStoreOpc));
22074 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
22075 if (i == X86::AddrDisp)
22076 MIB.addDisp(MI->getOperand(MemOpndSlot + i), LabelOffset);
22078 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
22081 MIB.addReg(LabelReg);
22083 MIB.addMBB(restoreMBB);
22084 MIB.setMemRefs(MMOBegin, MMOEnd);
22086 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
22087 .addMBB(restoreMBB);
22089 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
22090 MIB.addRegMask(RegInfo->getNoPreservedMask());
22091 thisMBB->addSuccessor(mainMBB);
22092 thisMBB->addSuccessor(restoreMBB);
22096 BuildMI(mainMBB, DL, TII->get(X86::MOV32r0), mainDstReg);
22097 mainMBB->addSuccessor(sinkMBB);
22100 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
22101 TII->get(X86::PHI), DstReg)
22102 .addReg(mainDstReg).addMBB(mainMBB)
22103 .addReg(restoreDstReg).addMBB(restoreMBB);
22106 if (RegInfo->hasBasePointer(*MF)) {
22107 const bool Uses64BitFramePtr =
22108 Subtarget->isTarget64BitLP64() || Subtarget->isTargetNaCl64();
22109 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
22110 X86FI->setRestoreBasePointer(MF);
22111 unsigned FramePtr = RegInfo->getFrameRegister(*MF);
22112 unsigned BasePtr = RegInfo->getBaseRegister();
22113 unsigned Opm = Uses64BitFramePtr ? X86::MOV64rm : X86::MOV32rm;
22114 addRegOffset(BuildMI(restoreMBB, DL, TII->get(Opm), BasePtr),
22115 FramePtr, true, X86FI->getRestoreBasePointerOffset())
22116 .setMIFlag(MachineInstr::FrameSetup);
22118 BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1);
22119 BuildMI(restoreMBB, DL, TII->get(X86::JMP_1)).addMBB(sinkMBB);
22120 restoreMBB->addSuccessor(sinkMBB);
22122 MI->eraseFromParent();
22126 MachineBasicBlock *
22127 X86TargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
22128 MachineBasicBlock *MBB) const {
22129 DebugLoc DL = MI->getDebugLoc();
22130 MachineFunction *MF = MBB->getParent();
22131 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
22132 MachineRegisterInfo &MRI = MF->getRegInfo();
22134 // Memory Reference
22135 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
22136 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
22138 MVT PVT = getPointerTy(MF->getDataLayout());
22139 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
22140 "Invalid Pointer Size!");
22142 const TargetRegisterClass *RC =
22143 (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass;
22144 unsigned Tmp = MRI.createVirtualRegister(RC);
22145 // Since FP is only updated here but NOT referenced, it's treated as GPR.
22146 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
22147 unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
22148 unsigned SP = RegInfo->getStackRegister();
22150 MachineInstrBuilder MIB;
22152 const int64_t LabelOffset = 1 * PVT.getStoreSize();
22153 const int64_t SPOffset = 2 * PVT.getStoreSize();
22155 unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm;
22156 unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r;
22159 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP);
22160 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
22161 MIB.addOperand(MI->getOperand(i));
22162 MIB.setMemRefs(MMOBegin, MMOEnd);
22164 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp);
22165 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
22166 if (i == X86::AddrDisp)
22167 MIB.addDisp(MI->getOperand(i), LabelOffset);
22169 MIB.addOperand(MI->getOperand(i));
22171 MIB.setMemRefs(MMOBegin, MMOEnd);
22173 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP);
22174 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
22175 if (i == X86::AddrDisp)
22176 MIB.addDisp(MI->getOperand(i), SPOffset);
22178 MIB.addOperand(MI->getOperand(i));
22180 MIB.setMemRefs(MMOBegin, MMOEnd);
22182 BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp);
22184 MI->eraseFromParent();
22188 // Replace 213-type (isel default) FMA3 instructions with 231-type for
22189 // accumulator loops. Writing back to the accumulator allows the coalescer
22190 // to remove extra copies in the loop.
22191 // FIXME: Do this on AVX512. We don't support 231 variants yet (PR23937).
22192 MachineBasicBlock *
22193 X86TargetLowering::emitFMA3Instr(MachineInstr *MI,
22194 MachineBasicBlock *MBB) const {
22195 MachineOperand &AddendOp = MI->getOperand(3);
22197 // Bail out early if the addend isn't a register - we can't switch these.
22198 if (!AddendOp.isReg())
22201 MachineFunction &MF = *MBB->getParent();
22202 MachineRegisterInfo &MRI = MF.getRegInfo();
22204 // Check whether the addend is defined by a PHI:
22205 assert(MRI.hasOneDef(AddendOp.getReg()) && "Multiple defs in SSA?");
22206 MachineInstr &AddendDef = *MRI.def_instr_begin(AddendOp.getReg());
22207 if (!AddendDef.isPHI())
22210 // Look for the following pattern:
22212 // %addend = phi [%entry, 0], [%loop, %result]
22214 // %result<tied1> = FMA213 %m2<tied0>, %m1, %addend
22218 // %addend = phi [%entry, 0], [%loop, %result]
22220 // %result<tied1> = FMA231 %addend<tied0>, %m1, %m2
22222 for (unsigned i = 1, e = AddendDef.getNumOperands(); i < e; i += 2) {
22223 assert(AddendDef.getOperand(i).isReg());
22224 MachineOperand PHISrcOp = AddendDef.getOperand(i);
22225 MachineInstr &PHISrcInst = *MRI.def_instr_begin(PHISrcOp.getReg());
22226 if (&PHISrcInst == MI) {
22227 // Found a matching instruction.
22228 unsigned NewFMAOpc = 0;
22229 switch (MI->getOpcode()) {
22230 case X86::VFMADDPDr213r: NewFMAOpc = X86::VFMADDPDr231r; break;
22231 case X86::VFMADDPSr213r: NewFMAOpc = X86::VFMADDPSr231r; break;
22232 case X86::VFMADDSDr213r: NewFMAOpc = X86::VFMADDSDr231r; break;
22233 case X86::VFMADDSSr213r: NewFMAOpc = X86::VFMADDSSr231r; break;
22234 case X86::VFMSUBPDr213r: NewFMAOpc = X86::VFMSUBPDr231r; break;
22235 case X86::VFMSUBPSr213r: NewFMAOpc = X86::VFMSUBPSr231r; break;
22236 case X86::VFMSUBSDr213r: NewFMAOpc = X86::VFMSUBSDr231r; break;
22237 case X86::VFMSUBSSr213r: NewFMAOpc = X86::VFMSUBSSr231r; break;
22238 case X86::VFNMADDPDr213r: NewFMAOpc = X86::VFNMADDPDr231r; break;
22239 case X86::VFNMADDPSr213r: NewFMAOpc = X86::VFNMADDPSr231r; break;
22240 case X86::VFNMADDSDr213r: NewFMAOpc = X86::VFNMADDSDr231r; break;
22241 case X86::VFNMADDSSr213r: NewFMAOpc = X86::VFNMADDSSr231r; break;
22242 case X86::VFNMSUBPDr213r: NewFMAOpc = X86::VFNMSUBPDr231r; break;
22243 case X86::VFNMSUBPSr213r: NewFMAOpc = X86::VFNMSUBPSr231r; break;
22244 case X86::VFNMSUBSDr213r: NewFMAOpc = X86::VFNMSUBSDr231r; break;
22245 case X86::VFNMSUBSSr213r: NewFMAOpc = X86::VFNMSUBSSr231r; break;
22246 case X86::VFMADDSUBPDr213r: NewFMAOpc = X86::VFMADDSUBPDr231r; break;
22247 case X86::VFMADDSUBPSr213r: NewFMAOpc = X86::VFMADDSUBPSr231r; break;
22248 case X86::VFMSUBADDPDr213r: NewFMAOpc = X86::VFMSUBADDPDr231r; break;
22249 case X86::VFMSUBADDPSr213r: NewFMAOpc = X86::VFMSUBADDPSr231r; break;
22251 case X86::VFMADDPDr213rY: NewFMAOpc = X86::VFMADDPDr231rY; break;
22252 case X86::VFMADDPSr213rY: NewFMAOpc = X86::VFMADDPSr231rY; break;
22253 case X86::VFMSUBPDr213rY: NewFMAOpc = X86::VFMSUBPDr231rY; break;
22254 case X86::VFMSUBPSr213rY: NewFMAOpc = X86::VFMSUBPSr231rY; break;
22255 case X86::VFNMADDPDr213rY: NewFMAOpc = X86::VFNMADDPDr231rY; break;
22256 case X86::VFNMADDPSr213rY: NewFMAOpc = X86::VFNMADDPSr231rY; break;
22257 case X86::VFNMSUBPDr213rY: NewFMAOpc = X86::VFNMSUBPDr231rY; break;
22258 case X86::VFNMSUBPSr213rY: NewFMAOpc = X86::VFNMSUBPSr231rY; break;
22259 case X86::VFMADDSUBPDr213rY: NewFMAOpc = X86::VFMADDSUBPDr231rY; break;
22260 case X86::VFMADDSUBPSr213rY: NewFMAOpc = X86::VFMADDSUBPSr231rY; break;
22261 case X86::VFMSUBADDPDr213rY: NewFMAOpc = X86::VFMSUBADDPDr231rY; break;
22262 case X86::VFMSUBADDPSr213rY: NewFMAOpc = X86::VFMSUBADDPSr231rY; break;
22263 default: llvm_unreachable("Unrecognized FMA variant.");
22266 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
22267 MachineInstrBuilder MIB =
22268 BuildMI(MF, MI->getDebugLoc(), TII.get(NewFMAOpc))
22269 .addOperand(MI->getOperand(0))
22270 .addOperand(MI->getOperand(3))
22271 .addOperand(MI->getOperand(2))
22272 .addOperand(MI->getOperand(1));
22273 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
22274 MI->eraseFromParent();
22281 MachineBasicBlock *
22282 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
22283 MachineBasicBlock *BB) const {
22284 switch (MI->getOpcode()) {
22285 default: llvm_unreachable("Unexpected instr type to insert");
22286 case X86::TAILJMPd64:
22287 case X86::TAILJMPr64:
22288 case X86::TAILJMPm64:
22289 case X86::TAILJMPd64_REX:
22290 case X86::TAILJMPr64_REX:
22291 case X86::TAILJMPm64_REX:
22292 llvm_unreachable("TAILJMP64 would not be touched here.");
22293 case X86::TCRETURNdi64:
22294 case X86::TCRETURNri64:
22295 case X86::TCRETURNmi64:
22297 case X86::WIN_ALLOCA:
22298 return EmitLoweredWinAlloca(MI, BB);
22299 case X86::CATCHRET:
22300 return EmitLoweredCatchRet(MI, BB);
22301 case X86::CATCHPAD:
22302 return EmitLoweredCatchPad(MI, BB);
22303 case X86::SEG_ALLOCA_32:
22304 case X86::SEG_ALLOCA_64:
22305 return EmitLoweredSegAlloca(MI, BB);
22306 case X86::TLSCall_32:
22307 case X86::TLSCall_64:
22308 return EmitLoweredTLSCall(MI, BB);
22309 case X86::CMOV_FR32:
22310 case X86::CMOV_FR64:
22311 case X86::CMOV_FR128:
22312 case X86::CMOV_GR8:
22313 case X86::CMOV_GR16:
22314 case X86::CMOV_GR32:
22315 case X86::CMOV_RFP32:
22316 case X86::CMOV_RFP64:
22317 case X86::CMOV_RFP80:
22318 case X86::CMOV_V2F64:
22319 case X86::CMOV_V2I64:
22320 case X86::CMOV_V4F32:
22321 case X86::CMOV_V4F64:
22322 case X86::CMOV_V4I64:
22323 case X86::CMOV_V16F32:
22324 case X86::CMOV_V8F32:
22325 case X86::CMOV_V8F64:
22326 case X86::CMOV_V8I64:
22327 case X86::CMOV_V8I1:
22328 case X86::CMOV_V16I1:
22329 case X86::CMOV_V32I1:
22330 case X86::CMOV_V64I1:
22331 return EmitLoweredSelect(MI, BB);
22333 case X86::RELEASE_FADD32mr:
22334 case X86::RELEASE_FADD64mr:
22335 return EmitLoweredAtomicFP(MI, BB);
22337 case X86::FP32_TO_INT16_IN_MEM:
22338 case X86::FP32_TO_INT32_IN_MEM:
22339 case X86::FP32_TO_INT64_IN_MEM:
22340 case X86::FP64_TO_INT16_IN_MEM:
22341 case X86::FP64_TO_INT32_IN_MEM:
22342 case X86::FP64_TO_INT64_IN_MEM:
22343 case X86::FP80_TO_INT16_IN_MEM:
22344 case X86::FP80_TO_INT32_IN_MEM:
22345 case X86::FP80_TO_INT64_IN_MEM: {
22346 MachineFunction *F = BB->getParent();
22347 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
22348 DebugLoc DL = MI->getDebugLoc();
22350 // Change the floating point control register to use "round towards zero"
22351 // mode when truncating to an integer value.
22352 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
22353 addFrameReference(BuildMI(*BB, MI, DL,
22354 TII->get(X86::FNSTCW16m)), CWFrameIdx);
22356 // Load the old value of the high byte of the control word...
22358 F->getRegInfo().createVirtualRegister(&X86::GR16RegClass);
22359 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
22362 // Set the high part to be round to zero...
22363 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
22366 // Reload the modified control word now...
22367 addFrameReference(BuildMI(*BB, MI, DL,
22368 TII->get(X86::FLDCW16m)), CWFrameIdx);
22370 // Restore the memory image of control word to original value
22371 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
22374 // Get the X86 opcode to use.
22376 switch (MI->getOpcode()) {
22377 default: llvm_unreachable("illegal opcode!");
22378 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
22379 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
22380 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
22381 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
22382 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
22383 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
22384 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
22385 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
22386 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
22390 MachineOperand &Op = MI->getOperand(0);
22392 AM.BaseType = X86AddressMode::RegBase;
22393 AM.Base.Reg = Op.getReg();
22395 AM.BaseType = X86AddressMode::FrameIndexBase;
22396 AM.Base.FrameIndex = Op.getIndex();
22398 Op = MI->getOperand(1);
22400 AM.Scale = Op.getImm();
22401 Op = MI->getOperand(2);
22403 AM.IndexReg = Op.getImm();
22404 Op = MI->getOperand(3);
22405 if (Op.isGlobal()) {
22406 AM.GV = Op.getGlobal();
22408 AM.Disp = Op.getImm();
22410 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
22411 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
22413 // Reload the original control word now.
22414 addFrameReference(BuildMI(*BB, MI, DL,
22415 TII->get(X86::FLDCW16m)), CWFrameIdx);
22417 MI->eraseFromParent(); // The pseudo instruction is gone now.
22420 // String/text processing lowering.
22421 case X86::PCMPISTRM128REG:
22422 case X86::VPCMPISTRM128REG:
22423 case X86::PCMPISTRM128MEM:
22424 case X86::VPCMPISTRM128MEM:
22425 case X86::PCMPESTRM128REG:
22426 case X86::VPCMPESTRM128REG:
22427 case X86::PCMPESTRM128MEM:
22428 case X86::VPCMPESTRM128MEM:
22429 assert(Subtarget->hasSSE42() &&
22430 "Target must have SSE4.2 or AVX features enabled");
22431 return EmitPCMPSTRM(MI, BB, Subtarget->getInstrInfo());
22433 // String/text processing lowering.
22434 case X86::PCMPISTRIREG:
22435 case X86::VPCMPISTRIREG:
22436 case X86::PCMPISTRIMEM:
22437 case X86::VPCMPISTRIMEM:
22438 case X86::PCMPESTRIREG:
22439 case X86::VPCMPESTRIREG:
22440 case X86::PCMPESTRIMEM:
22441 case X86::VPCMPESTRIMEM:
22442 assert(Subtarget->hasSSE42() &&
22443 "Target must have SSE4.2 or AVX features enabled");
22444 return EmitPCMPSTRI(MI, BB, Subtarget->getInstrInfo());
22446 // Thread synchronization.
22448 return EmitMonitor(MI, BB, Subtarget);
22452 return EmitXBegin(MI, BB, Subtarget->getInstrInfo());
22454 case X86::VASTART_SAVE_XMM_REGS:
22455 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
22457 case X86::VAARG_64:
22458 return EmitVAARG64WithCustomInserter(MI, BB);
22460 case X86::EH_SjLj_SetJmp32:
22461 case X86::EH_SjLj_SetJmp64:
22462 return emitEHSjLjSetJmp(MI, BB);
22464 case X86::EH_SjLj_LongJmp32:
22465 case X86::EH_SjLj_LongJmp64:
22466 return emitEHSjLjLongJmp(MI, BB);
22468 case TargetOpcode::STATEPOINT:
22469 // As an implementation detail, STATEPOINT shares the STACKMAP format at
22470 // this point in the process. We diverge later.
22471 return emitPatchPoint(MI, BB);
22473 case TargetOpcode::STACKMAP:
22474 case TargetOpcode::PATCHPOINT:
22475 return emitPatchPoint(MI, BB);
22477 case X86::VFMADDPDr213r:
22478 case X86::VFMADDPSr213r:
22479 case X86::VFMADDSDr213r:
22480 case X86::VFMADDSSr213r:
22481 case X86::VFMSUBPDr213r:
22482 case X86::VFMSUBPSr213r:
22483 case X86::VFMSUBSDr213r:
22484 case X86::VFMSUBSSr213r:
22485 case X86::VFNMADDPDr213r:
22486 case X86::VFNMADDPSr213r:
22487 case X86::VFNMADDSDr213r:
22488 case X86::VFNMADDSSr213r:
22489 case X86::VFNMSUBPDr213r:
22490 case X86::VFNMSUBPSr213r:
22491 case X86::VFNMSUBSDr213r:
22492 case X86::VFNMSUBSSr213r:
22493 case X86::VFMADDSUBPDr213r:
22494 case X86::VFMADDSUBPSr213r:
22495 case X86::VFMSUBADDPDr213r:
22496 case X86::VFMSUBADDPSr213r:
22497 case X86::VFMADDPDr213rY:
22498 case X86::VFMADDPSr213rY:
22499 case X86::VFMSUBPDr213rY:
22500 case X86::VFMSUBPSr213rY:
22501 case X86::VFNMADDPDr213rY:
22502 case X86::VFNMADDPSr213rY:
22503 case X86::VFNMSUBPDr213rY:
22504 case X86::VFNMSUBPSr213rY:
22505 case X86::VFMADDSUBPDr213rY:
22506 case X86::VFMADDSUBPSr213rY:
22507 case X86::VFMSUBADDPDr213rY:
22508 case X86::VFMSUBADDPSr213rY:
22509 return emitFMA3Instr(MI, BB);
22513 //===----------------------------------------------------------------------===//
22514 // X86 Optimization Hooks
22515 //===----------------------------------------------------------------------===//
22517 void X86TargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
22520 const SelectionDAG &DAG,
22521 unsigned Depth) const {
22522 unsigned BitWidth = KnownZero.getBitWidth();
22523 unsigned Opc = Op.getOpcode();
22524 assert((Opc >= ISD::BUILTIN_OP_END ||
22525 Opc == ISD::INTRINSIC_WO_CHAIN ||
22526 Opc == ISD::INTRINSIC_W_CHAIN ||
22527 Opc == ISD::INTRINSIC_VOID) &&
22528 "Should use MaskedValueIsZero if you don't know whether Op"
22529 " is a target node!");
22531 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
22545 // These nodes' second result is a boolean.
22546 if (Op.getResNo() == 0)
22549 case X86ISD::SETCC:
22550 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
22552 case ISD::INTRINSIC_WO_CHAIN: {
22553 unsigned IntId = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
22554 unsigned NumLoBits = 0;
22557 case Intrinsic::x86_sse_movmsk_ps:
22558 case Intrinsic::x86_avx_movmsk_ps_256:
22559 case Intrinsic::x86_sse2_movmsk_pd:
22560 case Intrinsic::x86_avx_movmsk_pd_256:
22561 case Intrinsic::x86_mmx_pmovmskb:
22562 case Intrinsic::x86_sse2_pmovmskb_128:
22563 case Intrinsic::x86_avx2_pmovmskb: {
22564 // High bits of movmskp{s|d}, pmovmskb are known zero.
22566 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
22567 case Intrinsic::x86_sse_movmsk_ps: NumLoBits = 4; break;
22568 case Intrinsic::x86_avx_movmsk_ps_256: NumLoBits = 8; break;
22569 case Intrinsic::x86_sse2_movmsk_pd: NumLoBits = 2; break;
22570 case Intrinsic::x86_avx_movmsk_pd_256: NumLoBits = 4; break;
22571 case Intrinsic::x86_mmx_pmovmskb: NumLoBits = 8; break;
22572 case Intrinsic::x86_sse2_pmovmskb_128: NumLoBits = 16; break;
22573 case Intrinsic::x86_avx2_pmovmskb: NumLoBits = 32; break;
22575 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - NumLoBits);
22584 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(
22586 const SelectionDAG &,
22587 unsigned Depth) const {
22588 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
22589 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
22590 return Op.getValueType().getScalarSizeInBits();
22596 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
22597 /// node is a GlobalAddress + offset.
22598 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
22599 const GlobalValue* &GA,
22600 int64_t &Offset) const {
22601 if (N->getOpcode() == X86ISD::Wrapper) {
22602 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
22603 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
22604 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
22608 return TargetLowering::isGAPlusOffset(N, GA, Offset);
22611 /// isShuffleHigh128VectorInsertLow - Checks whether the shuffle node is the
22612 /// same as extracting the high 128-bit part of 256-bit vector and then
22613 /// inserting the result into the low part of a new 256-bit vector
22614 static bool isShuffleHigh128VectorInsertLow(ShuffleVectorSDNode *SVOp) {
22615 EVT VT = SVOp->getValueType(0);
22616 unsigned NumElems = VT.getVectorNumElements();
22618 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
22619 for (unsigned i = 0, j = NumElems/2; i != NumElems/2; ++i, ++j)
22620 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
22621 SVOp->getMaskElt(j) >= 0)
22627 /// isShuffleLow128VectorInsertHigh - Checks whether the shuffle node is the
22628 /// same as extracting the low 128-bit part of 256-bit vector and then
22629 /// inserting the result into the high part of a new 256-bit vector
22630 static bool isShuffleLow128VectorInsertHigh(ShuffleVectorSDNode *SVOp) {
22631 EVT VT = SVOp->getValueType(0);
22632 unsigned NumElems = VT.getVectorNumElements();
22634 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
22635 for (unsigned i = NumElems/2, j = 0; i != NumElems; ++i, ++j)
22636 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
22637 SVOp->getMaskElt(j) >= 0)
22643 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
22644 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
22645 TargetLowering::DAGCombinerInfo &DCI,
22646 const X86Subtarget* Subtarget) {
22648 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
22649 SDValue V1 = SVOp->getOperand(0);
22650 SDValue V2 = SVOp->getOperand(1);
22651 MVT VT = SVOp->getSimpleValueType(0);
22652 unsigned NumElems = VT.getVectorNumElements();
22654 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
22655 V2.getOpcode() == ISD::CONCAT_VECTORS) {
22659 // V UNDEF BUILD_VECTOR UNDEF
22661 // CONCAT_VECTOR CONCAT_VECTOR
22664 // RESULT: V + zero extended
22666 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
22667 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
22668 V1.getOperand(1).getOpcode() != ISD::UNDEF)
22671 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
22674 // To match the shuffle mask, the first half of the mask should
22675 // be exactly the first vector, and all the rest a splat with the
22676 // first element of the second one.
22677 for (unsigned i = 0; i != NumElems/2; ++i)
22678 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
22679 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
22682 // If V1 is coming from a vector load then just fold to a VZEXT_LOAD.
22683 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(V1.getOperand(0))) {
22684 if (Ld->hasNUsesOfValue(1, 0)) {
22685 SDVTList Tys = DAG.getVTList(MVT::v4i64, MVT::Other);
22686 SDValue Ops[] = { Ld->getChain(), Ld->getBasePtr() };
22688 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops,
22690 Ld->getPointerInfo(),
22691 Ld->getAlignment(),
22692 false/*isVolatile*/, true/*ReadMem*/,
22693 false/*WriteMem*/);
22695 // Make sure the newly-created LOAD is in the same position as Ld in
22696 // terms of dependency. We create a TokenFactor for Ld and ResNode,
22697 // and update uses of Ld's output chain to use the TokenFactor.
22698 if (Ld->hasAnyUseOfValue(1)) {
22699 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
22700 SDValue(Ld, 1), SDValue(ResNode.getNode(), 1));
22701 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), NewChain);
22702 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(Ld, 1),
22703 SDValue(ResNode.getNode(), 1));
22706 return DAG.getBitcast(VT, ResNode);
22710 // Emit a zeroed vector and insert the desired subvector on its
22712 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
22713 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0), 0, DAG, dl);
22714 return DCI.CombineTo(N, InsV);
22717 //===--------------------------------------------------------------------===//
22718 // Combine some shuffles into subvector extracts and inserts:
22721 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
22722 if (isShuffleHigh128VectorInsertLow(SVOp)) {
22723 SDValue V = Extract128BitVector(V1, NumElems/2, DAG, dl);
22724 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, 0, DAG, dl);
22725 return DCI.CombineTo(N, InsV);
22728 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
22729 if (isShuffleLow128VectorInsertHigh(SVOp)) {
22730 SDValue V = Extract128BitVector(V1, 0, DAG, dl);
22731 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, NumElems/2, DAG, dl);
22732 return DCI.CombineTo(N, InsV);
22738 /// \brief Combine an arbitrary chain of shuffles into a single instruction if
22741 /// This is the leaf of the recursive combinine below. When we have found some
22742 /// chain of single-use x86 shuffle instructions and accumulated the combined
22743 /// shuffle mask represented by them, this will try to pattern match that mask
22744 /// into either a single instruction if there is a special purpose instruction
22745 /// for this operation, or into a PSHUFB instruction which is a fully general
22746 /// instruction but should only be used to replace chains over a certain depth.
22747 static bool combineX86ShuffleChain(SDValue Op, SDValue Root, ArrayRef<int> Mask,
22748 int Depth, bool HasPSHUFB, SelectionDAG &DAG,
22749 TargetLowering::DAGCombinerInfo &DCI,
22750 const X86Subtarget *Subtarget) {
22751 assert(!Mask.empty() && "Cannot combine an empty shuffle mask!");
22753 // Find the operand that enters the chain. Note that multiple uses are OK
22754 // here, we're not going to remove the operand we find.
22755 SDValue Input = Op.getOperand(0);
22756 while (Input.getOpcode() == ISD::BITCAST)
22757 Input = Input.getOperand(0);
22759 MVT VT = Input.getSimpleValueType();
22760 MVT RootVT = Root.getSimpleValueType();
22763 if (Mask.size() == 1) {
22764 int Index = Mask[0];
22765 assert((Index >= 0 || Index == SM_SentinelUndef ||
22766 Index == SM_SentinelZero) &&
22767 "Invalid shuffle index found!");
22769 // We may end up with an accumulated mask of size 1 as a result of
22770 // widening of shuffle operands (see function canWidenShuffleElements).
22771 // If the only shuffle index is equal to SM_SentinelZero then propagate
22772 // a zero vector. Otherwise, the combine shuffle mask is a no-op shuffle
22773 // mask, and therefore the entire chain of shuffles can be folded away.
22774 if (Index == SM_SentinelZero)
22775 DCI.CombineTo(Root.getNode(), getZeroVector(RootVT, Subtarget, DAG, DL));
22777 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Input),
22782 // Use the float domain if the operand type is a floating point type.
22783 bool FloatDomain = VT.isFloatingPoint();
22785 // For floating point shuffles, we don't have free copies in the shuffle
22786 // instructions or the ability to load as part of the instruction, so
22787 // canonicalize their shuffles to UNPCK or MOV variants.
22789 // Note that even with AVX we prefer the PSHUFD form of shuffle for integer
22790 // vectors because it can have a load folded into it that UNPCK cannot. This
22791 // doesn't preclude something switching to the shorter encoding post-RA.
22793 // FIXME: Should teach these routines about AVX vector widths.
22794 if (FloatDomain && VT.is128BitVector()) {
22795 if (Mask.equals({0, 0}) || Mask.equals({1, 1})) {
22796 bool Lo = Mask.equals({0, 0});
22799 // Check if we have SSE3 which will let us use MOVDDUP. That instruction
22800 // is no slower than UNPCKLPD but has the option to fold the input operand
22801 // into even an unaligned memory load.
22802 if (Lo && Subtarget->hasSSE3()) {
22803 Shuffle = X86ISD::MOVDDUP;
22804 ShuffleVT = MVT::v2f64;
22806 // We have MOVLHPS and MOVHLPS throughout SSE and they encode smaller
22807 // than the UNPCK variants.
22808 Shuffle = Lo ? X86ISD::MOVLHPS : X86ISD::MOVHLPS;
22809 ShuffleVT = MVT::v4f32;
22811 if (Depth == 1 && Root->getOpcode() == Shuffle)
22812 return false; // Nothing to do!
22813 Op = DAG.getBitcast(ShuffleVT, Input);
22814 DCI.AddToWorklist(Op.getNode());
22815 if (Shuffle == X86ISD::MOVDDUP)
22816 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
22818 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
22819 DCI.AddToWorklist(Op.getNode());
22820 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22824 if (Subtarget->hasSSE3() &&
22825 (Mask.equals({0, 0, 2, 2}) || Mask.equals({1, 1, 3, 3}))) {
22826 bool Lo = Mask.equals({0, 0, 2, 2});
22827 unsigned Shuffle = Lo ? X86ISD::MOVSLDUP : X86ISD::MOVSHDUP;
22828 MVT ShuffleVT = MVT::v4f32;
22829 if (Depth == 1 && Root->getOpcode() == Shuffle)
22830 return false; // Nothing to do!
22831 Op = DAG.getBitcast(ShuffleVT, Input);
22832 DCI.AddToWorklist(Op.getNode());
22833 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
22834 DCI.AddToWorklist(Op.getNode());
22835 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22839 if (Mask.equals({0, 0, 1, 1}) || Mask.equals({2, 2, 3, 3})) {
22840 bool Lo = Mask.equals({0, 0, 1, 1});
22841 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
22842 MVT ShuffleVT = MVT::v4f32;
22843 if (Depth == 1 && Root->getOpcode() == Shuffle)
22844 return false; // Nothing to do!
22845 Op = DAG.getBitcast(ShuffleVT, Input);
22846 DCI.AddToWorklist(Op.getNode());
22847 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
22848 DCI.AddToWorklist(Op.getNode());
22849 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22855 // We always canonicalize the 8 x i16 and 16 x i8 shuffles into their UNPCK
22856 // variants as none of these have single-instruction variants that are
22857 // superior to the UNPCK formulation.
22858 if (!FloatDomain && VT.is128BitVector() &&
22859 (Mask.equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
22860 Mask.equals({4, 4, 5, 5, 6, 6, 7, 7}) ||
22861 Mask.equals({0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7}) ||
22863 {8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14, 14, 15, 15}))) {
22864 bool Lo = Mask[0] == 0;
22865 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
22866 if (Depth == 1 && Root->getOpcode() == Shuffle)
22867 return false; // Nothing to do!
22869 switch (Mask.size()) {
22871 ShuffleVT = MVT::v8i16;
22874 ShuffleVT = MVT::v16i8;
22877 llvm_unreachable("Impossible mask size!");
22879 Op = DAG.getBitcast(ShuffleVT, Input);
22880 DCI.AddToWorklist(Op.getNode());
22881 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
22882 DCI.AddToWorklist(Op.getNode());
22883 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22888 // Don't try to re-form single instruction chains under any circumstances now
22889 // that we've done encoding canonicalization for them.
22893 // If we have 3 or more shuffle instructions or a chain involving PSHUFB, we
22894 // can replace them with a single PSHUFB instruction profitably. Intel's
22895 // manuals suggest only using PSHUFB if doing so replacing 5 instructions, but
22896 // in practice PSHUFB tends to be *very* fast so we're more aggressive.
22897 if ((Depth >= 3 || HasPSHUFB) && Subtarget->hasSSSE3()) {
22898 SmallVector<SDValue, 16> PSHUFBMask;
22899 int NumBytes = VT.getSizeInBits() / 8;
22900 int Ratio = NumBytes / Mask.size();
22901 for (int i = 0; i < NumBytes; ++i) {
22902 if (Mask[i / Ratio] == SM_SentinelUndef) {
22903 PSHUFBMask.push_back(DAG.getUNDEF(MVT::i8));
22906 int M = Mask[i / Ratio] != SM_SentinelZero
22907 ? Ratio * Mask[i / Ratio] + i % Ratio
22909 PSHUFBMask.push_back(DAG.getConstant(M, DL, MVT::i8));
22911 MVT ByteVT = MVT::getVectorVT(MVT::i8, NumBytes);
22912 Op = DAG.getBitcast(ByteVT, Input);
22913 DCI.AddToWorklist(Op.getNode());
22914 SDValue PSHUFBMaskOp =
22915 DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVT, PSHUFBMask);
22916 DCI.AddToWorklist(PSHUFBMaskOp.getNode());
22917 Op = DAG.getNode(X86ISD::PSHUFB, DL, ByteVT, Op, PSHUFBMaskOp);
22918 DCI.AddToWorklist(Op.getNode());
22919 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22924 // Failed to find any combines.
22928 /// \brief Fully generic combining of x86 shuffle instructions.
22930 /// This should be the last combine run over the x86 shuffle instructions. Once
22931 /// they have been fully optimized, this will recursively consider all chains
22932 /// of single-use shuffle instructions, build a generic model of the cumulative
22933 /// shuffle operation, and check for simpler instructions which implement this
22934 /// operation. We use this primarily for two purposes:
22936 /// 1) Collapse generic shuffles to specialized single instructions when
22937 /// equivalent. In most cases, this is just an encoding size win, but
22938 /// sometimes we will collapse multiple generic shuffles into a single
22939 /// special-purpose shuffle.
22940 /// 2) Look for sequences of shuffle instructions with 3 or more total
22941 /// instructions, and replace them with the slightly more expensive SSSE3
22942 /// PSHUFB instruction if available. We do this as the last combining step
22943 /// to ensure we avoid using PSHUFB if we can implement the shuffle with
22944 /// a suitable short sequence of other instructions. The PHUFB will either
22945 /// use a register or have to read from memory and so is slightly (but only
22946 /// slightly) more expensive than the other shuffle instructions.
22948 /// Because this is inherently a quadratic operation (for each shuffle in
22949 /// a chain, we recurse up the chain), the depth is limited to 8 instructions.
22950 /// This should never be an issue in practice as the shuffle lowering doesn't
22951 /// produce sequences of more than 8 instructions.
22953 /// FIXME: We will currently miss some cases where the redundant shuffling
22954 /// would simplify under the threshold for PSHUFB formation because of
22955 /// combine-ordering. To fix this, we should do the redundant instruction
22956 /// combining in this recursive walk.
22957 static bool combineX86ShufflesRecursively(SDValue Op, SDValue Root,
22958 ArrayRef<int> RootMask,
22959 int Depth, bool HasPSHUFB,
22961 TargetLowering::DAGCombinerInfo &DCI,
22962 const X86Subtarget *Subtarget) {
22963 // Bound the depth of our recursive combine because this is ultimately
22964 // quadratic in nature.
22968 // Directly rip through bitcasts to find the underlying operand.
22969 while (Op.getOpcode() == ISD::BITCAST && Op.getOperand(0).hasOneUse())
22970 Op = Op.getOperand(0);
22972 MVT VT = Op.getSimpleValueType();
22973 if (!VT.isVector())
22974 return false; // Bail if we hit a non-vector.
22976 assert(Root.getSimpleValueType().isVector() &&
22977 "Shuffles operate on vector types!");
22978 assert(VT.getSizeInBits() == Root.getSimpleValueType().getSizeInBits() &&
22979 "Can only combine shuffles of the same vector register size.");
22981 if (!isTargetShuffle(Op.getOpcode()))
22983 SmallVector<int, 16> OpMask;
22985 bool HaveMask = getTargetShuffleMask(Op.getNode(), VT, OpMask, IsUnary);
22986 // We only can combine unary shuffles which we can decode the mask for.
22987 if (!HaveMask || !IsUnary)
22990 assert(VT.getVectorNumElements() == OpMask.size() &&
22991 "Different mask size from vector size!");
22992 assert(((RootMask.size() > OpMask.size() &&
22993 RootMask.size() % OpMask.size() == 0) ||
22994 (OpMask.size() > RootMask.size() &&
22995 OpMask.size() % RootMask.size() == 0) ||
22996 OpMask.size() == RootMask.size()) &&
22997 "The smaller number of elements must divide the larger.");
22998 int RootRatio = std::max<int>(1, OpMask.size() / RootMask.size());
22999 int OpRatio = std::max<int>(1, RootMask.size() / OpMask.size());
23000 assert(((RootRatio == 1 && OpRatio == 1) ||
23001 (RootRatio == 1) != (OpRatio == 1)) &&
23002 "Must not have a ratio for both incoming and op masks!");
23004 SmallVector<int, 16> Mask;
23005 Mask.reserve(std::max(OpMask.size(), RootMask.size()));
23007 // Merge this shuffle operation's mask into our accumulated mask. Note that
23008 // this shuffle's mask will be the first applied to the input, followed by the
23009 // root mask to get us all the way to the root value arrangement. The reason
23010 // for this order is that we are recursing up the operation chain.
23011 for (int i = 0, e = std::max(OpMask.size(), RootMask.size()); i < e; ++i) {
23012 int RootIdx = i / RootRatio;
23013 if (RootMask[RootIdx] < 0) {
23014 // This is a zero or undef lane, we're done.
23015 Mask.push_back(RootMask[RootIdx]);
23019 int RootMaskedIdx = RootMask[RootIdx] * RootRatio + i % RootRatio;
23020 int OpIdx = RootMaskedIdx / OpRatio;
23021 if (OpMask[OpIdx] < 0) {
23022 // The incoming lanes are zero or undef, it doesn't matter which ones we
23024 Mask.push_back(OpMask[OpIdx]);
23028 // Ok, we have non-zero lanes, map them through.
23029 Mask.push_back(OpMask[OpIdx] * OpRatio +
23030 RootMaskedIdx % OpRatio);
23033 // See if we can recurse into the operand to combine more things.
23034 switch (Op.getOpcode()) {
23035 case X86ISD::PSHUFB:
23037 case X86ISD::PSHUFD:
23038 case X86ISD::PSHUFHW:
23039 case X86ISD::PSHUFLW:
23040 if (Op.getOperand(0).hasOneUse() &&
23041 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
23042 HasPSHUFB, DAG, DCI, Subtarget))
23046 case X86ISD::UNPCKL:
23047 case X86ISD::UNPCKH:
23048 assert(Op.getOperand(0) == Op.getOperand(1) &&
23049 "We only combine unary shuffles!");
23050 // We can't check for single use, we have to check that this shuffle is the
23052 if (Op->isOnlyUserOf(Op.getOperand(0).getNode()) &&
23053 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
23054 HasPSHUFB, DAG, DCI, Subtarget))
23059 // Minor canonicalization of the accumulated shuffle mask to make it easier
23060 // to match below. All this does is detect masks with squential pairs of
23061 // elements, and shrink them to the half-width mask. It does this in a loop
23062 // so it will reduce the size of the mask to the minimal width mask which
23063 // performs an equivalent shuffle.
23064 SmallVector<int, 16> WidenedMask;
23065 while (Mask.size() > 1 && canWidenShuffleElements(Mask, WidenedMask)) {
23066 Mask = std::move(WidenedMask);
23067 WidenedMask.clear();
23070 return combineX86ShuffleChain(Op, Root, Mask, Depth, HasPSHUFB, DAG, DCI,
23074 /// \brief Get the PSHUF-style mask from PSHUF node.
23076 /// This is a very minor wrapper around getTargetShuffleMask to easy forming v4
23077 /// PSHUF-style masks that can be reused with such instructions.
23078 static SmallVector<int, 4> getPSHUFShuffleMask(SDValue N) {
23079 MVT VT = N.getSimpleValueType();
23080 SmallVector<int, 4> Mask;
23082 bool HaveMask = getTargetShuffleMask(N.getNode(), VT, Mask, IsUnary);
23086 // If we have more than 128-bits, only the low 128-bits of shuffle mask
23087 // matter. Check that the upper masks are repeats and remove them.
23088 if (VT.getSizeInBits() > 128) {
23089 int LaneElts = 128 / VT.getScalarSizeInBits();
23091 for (int i = 1, NumLanes = VT.getSizeInBits() / 128; i < NumLanes; ++i)
23092 for (int j = 0; j < LaneElts; ++j)
23093 assert(Mask[j] == Mask[i * LaneElts + j] - (LaneElts * i) &&
23094 "Mask doesn't repeat in high 128-bit lanes!");
23096 Mask.resize(LaneElts);
23099 switch (N.getOpcode()) {
23100 case X86ISD::PSHUFD:
23102 case X86ISD::PSHUFLW:
23105 case X86ISD::PSHUFHW:
23106 Mask.erase(Mask.begin(), Mask.begin() + 4);
23107 for (int &M : Mask)
23111 llvm_unreachable("No valid shuffle instruction found!");
23115 /// \brief Search for a combinable shuffle across a chain ending in pshufd.
23117 /// We walk up the chain and look for a combinable shuffle, skipping over
23118 /// shuffles that we could hoist this shuffle's transformation past without
23119 /// altering anything.
23121 combineRedundantDWordShuffle(SDValue N, MutableArrayRef<int> Mask,
23123 TargetLowering::DAGCombinerInfo &DCI) {
23124 assert(N.getOpcode() == X86ISD::PSHUFD &&
23125 "Called with something other than an x86 128-bit half shuffle!");
23128 // Walk up a single-use chain looking for a combinable shuffle. Keep a stack
23129 // of the shuffles in the chain so that we can form a fresh chain to replace
23131 SmallVector<SDValue, 8> Chain;
23132 SDValue V = N.getOperand(0);
23133 for (; V.hasOneUse(); V = V.getOperand(0)) {
23134 switch (V.getOpcode()) {
23136 return SDValue(); // Nothing combined!
23139 // Skip bitcasts as we always know the type for the target specific
23143 case X86ISD::PSHUFD:
23144 // Found another dword shuffle.
23147 case X86ISD::PSHUFLW:
23148 // Check that the low words (being shuffled) are the identity in the
23149 // dword shuffle, and the high words are self-contained.
23150 if (Mask[0] != 0 || Mask[1] != 1 ||
23151 !(Mask[2] >= 2 && Mask[2] < 4 && Mask[3] >= 2 && Mask[3] < 4))
23154 Chain.push_back(V);
23157 case X86ISD::PSHUFHW:
23158 // Check that the high words (being shuffled) are the identity in the
23159 // dword shuffle, and the low words are self-contained.
23160 if (Mask[2] != 2 || Mask[3] != 3 ||
23161 !(Mask[0] >= 0 && Mask[0] < 2 && Mask[1] >= 0 && Mask[1] < 2))
23164 Chain.push_back(V);
23167 case X86ISD::UNPCKL:
23168 case X86ISD::UNPCKH:
23169 // For either i8 -> i16 or i16 -> i32 unpacks, we can combine a dword
23170 // shuffle into a preceding word shuffle.
23171 if (V.getSimpleValueType().getVectorElementType() != MVT::i8 &&
23172 V.getSimpleValueType().getVectorElementType() != MVT::i16)
23175 // Search for a half-shuffle which we can combine with.
23176 unsigned CombineOp =
23177 V.getOpcode() == X86ISD::UNPCKL ? X86ISD::PSHUFLW : X86ISD::PSHUFHW;
23178 if (V.getOperand(0) != V.getOperand(1) ||
23179 !V->isOnlyUserOf(V.getOperand(0).getNode()))
23181 Chain.push_back(V);
23182 V = V.getOperand(0);
23184 switch (V.getOpcode()) {
23186 return SDValue(); // Nothing to combine.
23188 case X86ISD::PSHUFLW:
23189 case X86ISD::PSHUFHW:
23190 if (V.getOpcode() == CombineOp)
23193 Chain.push_back(V);
23197 V = V.getOperand(0);
23201 } while (V.hasOneUse());
23204 // Break out of the loop if we break out of the switch.
23208 if (!V.hasOneUse())
23209 // We fell out of the loop without finding a viable combining instruction.
23212 // Merge this node's mask and our incoming mask.
23213 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
23214 for (int &M : Mask)
23216 V = DAG.getNode(V.getOpcode(), DL, V.getValueType(), V.getOperand(0),
23217 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
23219 // Rebuild the chain around this new shuffle.
23220 while (!Chain.empty()) {
23221 SDValue W = Chain.pop_back_val();
23223 if (V.getValueType() != W.getOperand(0).getValueType())
23224 V = DAG.getBitcast(W.getOperand(0).getValueType(), V);
23226 switch (W.getOpcode()) {
23228 llvm_unreachable("Only PSHUF and UNPCK instructions get here!");
23230 case X86ISD::UNPCKL:
23231 case X86ISD::UNPCKH:
23232 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, V);
23235 case X86ISD::PSHUFD:
23236 case X86ISD::PSHUFLW:
23237 case X86ISD::PSHUFHW:
23238 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, W.getOperand(1));
23242 if (V.getValueType() != N.getValueType())
23243 V = DAG.getBitcast(N.getValueType(), V);
23245 // Return the new chain to replace N.
23249 /// \brief Search for a combinable shuffle across a chain ending in pshuflw or
23252 /// We walk up the chain, skipping shuffles of the other half and looking
23253 /// through shuffles which switch halves trying to find a shuffle of the same
23254 /// pair of dwords.
23255 static bool combineRedundantHalfShuffle(SDValue N, MutableArrayRef<int> Mask,
23257 TargetLowering::DAGCombinerInfo &DCI) {
23259 (N.getOpcode() == X86ISD::PSHUFLW || N.getOpcode() == X86ISD::PSHUFHW) &&
23260 "Called with something other than an x86 128-bit half shuffle!");
23262 unsigned CombineOpcode = N.getOpcode();
23264 // Walk up a single-use chain looking for a combinable shuffle.
23265 SDValue V = N.getOperand(0);
23266 for (; V.hasOneUse(); V = V.getOperand(0)) {
23267 switch (V.getOpcode()) {
23269 return false; // Nothing combined!
23272 // Skip bitcasts as we always know the type for the target specific
23276 case X86ISD::PSHUFLW:
23277 case X86ISD::PSHUFHW:
23278 if (V.getOpcode() == CombineOpcode)
23281 // Other-half shuffles are no-ops.
23284 // Break out of the loop if we break out of the switch.
23288 if (!V.hasOneUse())
23289 // We fell out of the loop without finding a viable combining instruction.
23292 // Combine away the bottom node as its shuffle will be accumulated into
23293 // a preceding shuffle.
23294 DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
23296 // Record the old value.
23299 // Merge this node's mask and our incoming mask (adjusted to account for all
23300 // the pshufd instructions encountered).
23301 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
23302 for (int &M : Mask)
23304 V = DAG.getNode(V.getOpcode(), DL, MVT::v8i16, V.getOperand(0),
23305 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
23307 // Check that the shuffles didn't cancel each other out. If not, we need to
23308 // combine to the new one.
23310 // Replace the combinable shuffle with the combined one, updating all users
23311 // so that we re-evaluate the chain here.
23312 DCI.CombineTo(Old.getNode(), V, /*AddTo*/ true);
23317 /// \brief Try to combine x86 target specific shuffles.
23318 static SDValue PerformTargetShuffleCombine(SDValue N, SelectionDAG &DAG,
23319 TargetLowering::DAGCombinerInfo &DCI,
23320 const X86Subtarget *Subtarget) {
23322 MVT VT = N.getSimpleValueType();
23323 SmallVector<int, 4> Mask;
23325 switch (N.getOpcode()) {
23326 case X86ISD::PSHUFD:
23327 case X86ISD::PSHUFLW:
23328 case X86ISD::PSHUFHW:
23329 Mask = getPSHUFShuffleMask(N);
23330 assert(Mask.size() == 4);
23332 case X86ISD::UNPCKL: {
23333 // Combine X86ISD::UNPCKL and ISD::VECTOR_SHUFFLE into X86ISD::UNPCKH, in
23334 // which X86ISD::UNPCKL has a ISD::UNDEF operand, and ISD::VECTOR_SHUFFLE
23335 // moves upper half elements into the lower half part. For example:
23337 // t2: v16i8 = vector_shuffle<8,9,10,11,12,13,14,15,u,u,u,u,u,u,u,u> t1,
23339 // t3: v16i8 = X86ISD::UNPCKL undef:v16i8, t2
23341 // will be combined to:
23343 // t3: v16i8 = X86ISD::UNPCKH undef:v16i8, t1
23345 // This is only for 128-bit vectors. From SSE4.1 onward this combine may not
23346 // happen due to advanced instructions.
23347 if (!VT.is128BitVector())
23350 auto Op0 = N.getOperand(0);
23351 auto Op1 = N.getOperand(1);
23352 if (Op0.getOpcode() == ISD::UNDEF &&
23353 Op1.getNode()->getOpcode() == ISD::VECTOR_SHUFFLE) {
23354 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Op1.getNode())->getMask();
23356 unsigned NumElts = VT.getVectorNumElements();
23357 SmallVector<int, 8> ExpectedMask(NumElts, -1);
23358 std::iota(ExpectedMask.begin(), ExpectedMask.begin() + NumElts / 2,
23361 auto ShufOp = Op1.getOperand(0);
23362 if (isShuffleEquivalent(Op1, ShufOp, Mask, ExpectedMask))
23363 return DAG.getNode(X86ISD::UNPCKH, DL, VT, N.getOperand(0), ShufOp);
23371 // Nuke no-op shuffles that show up after combining.
23372 if (isNoopShuffleMask(Mask))
23373 return DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
23375 // Look for simplifications involving one or two shuffle instructions.
23376 SDValue V = N.getOperand(0);
23377 switch (N.getOpcode()) {
23380 case X86ISD::PSHUFLW:
23381 case X86ISD::PSHUFHW:
23382 assert(VT.getVectorElementType() == MVT::i16 && "Bad word shuffle type!");
23384 if (combineRedundantHalfShuffle(N, Mask, DAG, DCI))
23385 return SDValue(); // We combined away this shuffle, so we're done.
23387 // See if this reduces to a PSHUFD which is no more expensive and can
23388 // combine with more operations. Note that it has to at least flip the
23389 // dwords as otherwise it would have been removed as a no-op.
23390 if (makeArrayRef(Mask).equals({2, 3, 0, 1})) {
23391 int DMask[] = {0, 1, 2, 3};
23392 int DOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 2;
23393 DMask[DOffset + 0] = DOffset + 1;
23394 DMask[DOffset + 1] = DOffset + 0;
23395 MVT DVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
23396 V = DAG.getBitcast(DVT, V);
23397 DCI.AddToWorklist(V.getNode());
23398 V = DAG.getNode(X86ISD::PSHUFD, DL, DVT, V,
23399 getV4X86ShuffleImm8ForMask(DMask, DL, DAG));
23400 DCI.AddToWorklist(V.getNode());
23401 return DAG.getBitcast(VT, V);
23404 // Look for shuffle patterns which can be implemented as a single unpack.
23405 // FIXME: This doesn't handle the location of the PSHUFD generically, and
23406 // only works when we have a PSHUFD followed by two half-shuffles.
23407 if (Mask[0] == Mask[1] && Mask[2] == Mask[3] &&
23408 (V.getOpcode() == X86ISD::PSHUFLW ||
23409 V.getOpcode() == X86ISD::PSHUFHW) &&
23410 V.getOpcode() != N.getOpcode() &&
23412 SDValue D = V.getOperand(0);
23413 while (D.getOpcode() == ISD::BITCAST && D.hasOneUse())
23414 D = D.getOperand(0);
23415 if (D.getOpcode() == X86ISD::PSHUFD && D.hasOneUse()) {
23416 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
23417 SmallVector<int, 4> DMask = getPSHUFShuffleMask(D);
23418 int NOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
23419 int VOffset = V.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
23421 for (int i = 0; i < 4; ++i) {
23422 WordMask[i + NOffset] = Mask[i] + NOffset;
23423 WordMask[i + VOffset] = VMask[i] + VOffset;
23425 // Map the word mask through the DWord mask.
23427 for (int i = 0; i < 8; ++i)
23428 MappedMask[i] = 2 * DMask[WordMask[i] / 2] + WordMask[i] % 2;
23429 if (makeArrayRef(MappedMask).equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
23430 makeArrayRef(MappedMask).equals({4, 4, 5, 5, 6, 6, 7, 7})) {
23431 // We can replace all three shuffles with an unpack.
23432 V = DAG.getBitcast(VT, D.getOperand(0));
23433 DCI.AddToWorklist(V.getNode());
23434 return DAG.getNode(MappedMask[0] == 0 ? X86ISD::UNPCKL
23443 case X86ISD::PSHUFD:
23444 if (SDValue NewN = combineRedundantDWordShuffle(N, Mask, DAG, DCI))
23453 /// \brief Try to combine a shuffle into a target-specific add-sub node.
23455 /// We combine this directly on the abstract vector shuffle nodes so it is
23456 /// easier to generically match. We also insert dummy vector shuffle nodes for
23457 /// the operands which explicitly discard the lanes which are unused by this
23458 /// operation to try to flow through the rest of the combiner the fact that
23459 /// they're unused.
23460 static SDValue combineShuffleToAddSub(SDNode *N, SelectionDAG &DAG) {
23462 EVT VT = N->getValueType(0);
23464 // We only handle target-independent shuffles.
23465 // FIXME: It would be easy and harmless to use the target shuffle mask
23466 // extraction tool to support more.
23467 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
23470 auto *SVN = cast<ShuffleVectorSDNode>(N);
23471 SmallVector<int, 8> Mask;
23472 for (int M : SVN->getMask())
23475 SDValue V1 = N->getOperand(0);
23476 SDValue V2 = N->getOperand(1);
23478 // We require the first shuffle operand to be the FSUB node, and the second to
23479 // be the FADD node.
23480 if (V1.getOpcode() == ISD::FADD && V2.getOpcode() == ISD::FSUB) {
23481 ShuffleVectorSDNode::commuteMask(Mask);
23483 } else if (V1.getOpcode() != ISD::FSUB || V2.getOpcode() != ISD::FADD)
23486 // If there are other uses of these operations we can't fold them.
23487 if (!V1->hasOneUse() || !V2->hasOneUse())
23490 // Ensure that both operations have the same operands. Note that we can
23491 // commute the FADD operands.
23492 SDValue LHS = V1->getOperand(0), RHS = V1->getOperand(1);
23493 if ((V2->getOperand(0) != LHS || V2->getOperand(1) != RHS) &&
23494 (V2->getOperand(0) != RHS || V2->getOperand(1) != LHS))
23497 // We're looking for blends between FADD and FSUB nodes. We insist on these
23498 // nodes being lined up in a specific expected pattern.
23499 if (!(isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
23500 isShuffleEquivalent(V1, V2, Mask, {0, 5, 2, 7}) ||
23501 isShuffleEquivalent(V1, V2, Mask, {0, 9, 2, 11, 4, 13, 6, 15})))
23504 // Only specific types are legal at this point, assert so we notice if and
23505 // when these change.
23506 assert((VT == MVT::v4f32 || VT == MVT::v2f64 || VT == MVT::v8f32 ||
23507 VT == MVT::v4f64) &&
23508 "Unknown vector type encountered!");
23510 return DAG.getNode(X86ISD::ADDSUB, DL, VT, LHS, RHS);
23513 /// PerformShuffleCombine - Performs several different shuffle combines.
23514 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
23515 TargetLowering::DAGCombinerInfo &DCI,
23516 const X86Subtarget *Subtarget) {
23518 SDValue N0 = N->getOperand(0);
23519 SDValue N1 = N->getOperand(1);
23520 EVT VT = N->getValueType(0);
23522 // Don't create instructions with illegal types after legalize types has run.
23523 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23524 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
23527 // If we have legalized the vector types, look for blends of FADD and FSUB
23528 // nodes that we can fuse into an ADDSUB node.
23529 if (TLI.isTypeLegal(VT) && Subtarget->hasSSE3())
23530 if (SDValue AddSub = combineShuffleToAddSub(N, DAG))
23533 // Combine 256-bit vector shuffles. This is only profitable when in AVX mode
23534 if (TLI.isTypeLegal(VT) && Subtarget->hasFp256() && VT.is256BitVector() &&
23535 N->getOpcode() == ISD::VECTOR_SHUFFLE)
23536 return PerformShuffleCombine256(N, DAG, DCI, Subtarget);
23538 // During Type Legalization, when promoting illegal vector types,
23539 // the backend might introduce new shuffle dag nodes and bitcasts.
23541 // This code performs the following transformation:
23542 // fold: (shuffle (bitcast (BINOP A, B)), Undef, <Mask>) ->
23543 // (shuffle (BINOP (bitcast A), (bitcast B)), Undef, <Mask>)
23545 // We do this only if both the bitcast and the BINOP dag nodes have
23546 // one use. Also, perform this transformation only if the new binary
23547 // operation is legal. This is to avoid introducing dag nodes that
23548 // potentially need to be further expanded (or custom lowered) into a
23549 // less optimal sequence of dag nodes.
23550 if (!DCI.isBeforeLegalize() && DCI.isBeforeLegalizeOps() &&
23551 N1.getOpcode() == ISD::UNDEF && N0.hasOneUse() &&
23552 N0.getOpcode() == ISD::BITCAST) {
23553 SDValue BC0 = N0.getOperand(0);
23554 EVT SVT = BC0.getValueType();
23555 unsigned Opcode = BC0.getOpcode();
23556 unsigned NumElts = VT.getVectorNumElements();
23558 if (BC0.hasOneUse() && SVT.isVector() &&
23559 SVT.getVectorNumElements() * 2 == NumElts &&
23560 TLI.isOperationLegal(Opcode, VT)) {
23561 bool CanFold = false;
23573 unsigned SVTNumElts = SVT.getVectorNumElements();
23574 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
23575 for (unsigned i = 0, e = SVTNumElts; i != e && CanFold; ++i)
23576 CanFold = SVOp->getMaskElt(i) == (int)(i * 2);
23577 for (unsigned i = SVTNumElts, e = NumElts; i != e && CanFold; ++i)
23578 CanFold = SVOp->getMaskElt(i) < 0;
23581 SDValue BC00 = DAG.getBitcast(VT, BC0.getOperand(0));
23582 SDValue BC01 = DAG.getBitcast(VT, BC0.getOperand(1));
23583 SDValue NewBinOp = DAG.getNode(BC0.getOpcode(), dl, VT, BC00, BC01);
23584 return DAG.getVectorShuffle(VT, dl, NewBinOp, N1, &SVOp->getMask()[0]);
23589 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
23590 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
23591 // consecutive, non-overlapping, and in the right order.
23592 SmallVector<SDValue, 16> Elts;
23593 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
23594 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
23596 if (SDValue LD = EltsFromConsecutiveLoads(VT, Elts, dl, DAG, true))
23599 if (isTargetShuffle(N->getOpcode())) {
23601 PerformTargetShuffleCombine(SDValue(N, 0), DAG, DCI, Subtarget);
23602 if (Shuffle.getNode())
23605 // Try recursively combining arbitrary sequences of x86 shuffle
23606 // instructions into higher-order shuffles. We do this after combining
23607 // specific PSHUF instruction sequences into their minimal form so that we
23608 // can evaluate how many specialized shuffle instructions are involved in
23609 // a particular chain.
23610 SmallVector<int, 1> NonceMask; // Just a placeholder.
23611 NonceMask.push_back(0);
23612 if (combineX86ShufflesRecursively(SDValue(N, 0), SDValue(N, 0), NonceMask,
23613 /*Depth*/ 1, /*HasPSHUFB*/ false, DAG,
23615 return SDValue(); // This routine will use CombineTo to replace N.
23621 /// XFormVExtractWithShuffleIntoLoad - Check if a vector extract from a target
23622 /// specific shuffle of a load can be folded into a single element load.
23623 /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
23624 /// shuffles have been custom lowered so we need to handle those here.
23625 static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
23626 TargetLowering::DAGCombinerInfo &DCI) {
23627 if (DCI.isBeforeLegalizeOps())
23630 SDValue InVec = N->getOperand(0);
23631 SDValue EltNo = N->getOperand(1);
23633 if (!isa<ConstantSDNode>(EltNo))
23636 EVT OriginalVT = InVec.getValueType();
23638 if (InVec.getOpcode() == ISD::BITCAST) {
23639 // Don't duplicate a load with other uses.
23640 if (!InVec.hasOneUse())
23642 EVT BCVT = InVec.getOperand(0).getValueType();
23643 if (!BCVT.isVector() ||
23644 BCVT.getVectorNumElements() != OriginalVT.getVectorNumElements())
23646 InVec = InVec.getOperand(0);
23649 EVT CurrentVT = InVec.getValueType();
23651 if (!isTargetShuffle(InVec.getOpcode()))
23654 // Don't duplicate a load with other uses.
23655 if (!InVec.hasOneUse())
23658 SmallVector<int, 16> ShuffleMask;
23660 if (!getTargetShuffleMask(InVec.getNode(), CurrentVT.getSimpleVT(),
23661 ShuffleMask, UnaryShuffle))
23664 // Select the input vector, guarding against out of range extract vector.
23665 unsigned NumElems = CurrentVT.getVectorNumElements();
23666 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
23667 int Idx = (Elt > (int)NumElems) ? -1 : ShuffleMask[Elt];
23668 SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0)
23669 : InVec.getOperand(1);
23671 // If inputs to shuffle are the same for both ops, then allow 2 uses
23672 unsigned AllowedUses = InVec.getNumOperands() > 1 &&
23673 InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
23675 if (LdNode.getOpcode() == ISD::BITCAST) {
23676 // Don't duplicate a load with other uses.
23677 if (!LdNode.getNode()->hasNUsesOfValue(AllowedUses, 0))
23680 AllowedUses = 1; // only allow 1 load use if we have a bitcast
23681 LdNode = LdNode.getOperand(0);
23684 if (!ISD::isNormalLoad(LdNode.getNode()))
23687 LoadSDNode *LN0 = cast<LoadSDNode>(LdNode);
23689 if (!LN0 ||!LN0->hasNUsesOfValue(AllowedUses, 0) || LN0->isVolatile())
23692 EVT EltVT = N->getValueType(0);
23693 // If there's a bitcast before the shuffle, check if the load type and
23694 // alignment is valid.
23695 unsigned Align = LN0->getAlignment();
23696 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23697 unsigned NewAlign = DAG.getDataLayout().getABITypeAlignment(
23698 EltVT.getTypeForEVT(*DAG.getContext()));
23700 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, EltVT))
23703 // All checks match so transform back to vector_shuffle so that DAG combiner
23704 // can finish the job
23707 // Create shuffle node taking into account the case that its a unary shuffle
23708 SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(CurrentVT)
23709 : InVec.getOperand(1);
23710 Shuffle = DAG.getVectorShuffle(CurrentVT, dl,
23711 InVec.getOperand(0), Shuffle,
23713 Shuffle = DAG.getBitcast(OriginalVT, Shuffle);
23714 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle,
23718 static SDValue PerformBITCASTCombine(SDNode *N, SelectionDAG &DAG,
23719 const X86Subtarget *Subtarget) {
23720 SDValue N0 = N->getOperand(0);
23721 EVT VT = N->getValueType(0);
23723 // Detect bitcasts between i32 to x86mmx low word. Since MMX types are
23724 // special and don't usually play with other vector types, it's better to
23725 // handle them early to be sure we emit efficient code by avoiding
23726 // store-load conversions.
23727 if (VT == MVT::x86mmx && N0.getOpcode() == ISD::BUILD_VECTOR &&
23728 N0.getValueType() == MVT::v2i32 &&
23729 isNullConstant(N0.getOperand(1))) {
23730 SDValue N00 = N0->getOperand(0);
23731 if (N00.getValueType() == MVT::i32)
23732 return DAG.getNode(X86ISD::MMX_MOVW2D, SDLoc(N00), VT, N00);
23735 // Convert a bitcasted integer logic operation that has one bitcasted
23736 // floating-point operand and one constant operand into a floating-point
23737 // logic operation. This may create a load of the constant, but that is
23738 // cheaper than materializing the constant in an integer register and
23739 // transferring it to an SSE register or transferring the SSE operand to
23740 // integer register and back.
23742 switch (N0.getOpcode()) {
23743 case ISD::AND: FPOpcode = X86ISD::FAND; break;
23744 case ISD::OR: FPOpcode = X86ISD::FOR; break;
23745 case ISD::XOR: FPOpcode = X86ISD::FXOR; break;
23746 default: return SDValue();
23748 if (((Subtarget->hasSSE1() && VT == MVT::f32) ||
23749 (Subtarget->hasSSE2() && VT == MVT::f64)) &&
23750 isa<ConstantSDNode>(N0.getOperand(1)) &&
23751 N0.getOperand(0).getOpcode() == ISD::BITCAST &&
23752 N0.getOperand(0).getOperand(0).getValueType() == VT) {
23753 SDValue N000 = N0.getOperand(0).getOperand(0);
23754 SDValue FPConst = DAG.getBitcast(VT, N0.getOperand(1));
23755 return DAG.getNode(FPOpcode, SDLoc(N0), VT, N000, FPConst);
23761 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
23762 /// generation and convert it from being a bunch of shuffles and extracts
23763 /// into a somewhat faster sequence. For i686, the best sequence is apparently
23764 /// storing the value and loading scalars back, while for x64 we should
23765 /// use 64-bit extracts and shifts.
23766 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
23767 TargetLowering::DAGCombinerInfo &DCI) {
23768 if (SDValue NewOp = XFormVExtractWithShuffleIntoLoad(N, DAG, DCI))
23771 SDValue InputVector = N->getOperand(0);
23772 SDLoc dl(InputVector);
23773 // Detect mmx to i32 conversion through a v2i32 elt extract.
23774 if (InputVector.getOpcode() == ISD::BITCAST && InputVector.hasOneUse() &&
23775 N->getValueType(0) == MVT::i32 &&
23776 InputVector.getValueType() == MVT::v2i32) {
23778 // The bitcast source is a direct mmx result.
23779 SDValue MMXSrc = InputVector.getNode()->getOperand(0);
23780 if (MMXSrc.getValueType() == MVT::x86mmx)
23781 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
23782 N->getValueType(0),
23783 InputVector.getNode()->getOperand(0));
23785 // The mmx is indirect: (i64 extract_elt (v1i64 bitcast (x86mmx ...))).
23786 if (MMXSrc.getOpcode() == ISD::EXTRACT_VECTOR_ELT && MMXSrc.hasOneUse() &&
23787 MMXSrc.getValueType() == MVT::i64) {
23788 SDValue MMXSrcOp = MMXSrc.getOperand(0);
23789 if (MMXSrcOp.hasOneUse() && MMXSrcOp.getOpcode() == ISD::BITCAST &&
23790 MMXSrcOp.getValueType() == MVT::v1i64 &&
23791 MMXSrcOp.getOperand(0).getValueType() == MVT::x86mmx)
23792 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
23793 N->getValueType(0), MMXSrcOp.getOperand(0));
23797 EVT VT = N->getValueType(0);
23799 if (VT == MVT::i1 && isa<ConstantSDNode>(N->getOperand(1)) &&
23800 InputVector.getOpcode() == ISD::BITCAST &&
23801 isa<ConstantSDNode>(InputVector.getOperand(0))) {
23802 uint64_t ExtractedElt =
23803 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
23804 uint64_t InputValue =
23805 cast<ConstantSDNode>(InputVector.getOperand(0))->getZExtValue();
23806 uint64_t Res = (InputValue >> ExtractedElt) & 1;
23807 return DAG.getConstant(Res, dl, MVT::i1);
23809 // Only operate on vectors of 4 elements, where the alternative shuffling
23810 // gets to be more expensive.
23811 if (InputVector.getValueType() != MVT::v4i32)
23814 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
23815 // single use which is a sign-extend or zero-extend, and all elements are
23817 SmallVector<SDNode *, 4> Uses;
23818 unsigned ExtractedElements = 0;
23819 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
23820 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
23821 if (UI.getUse().getResNo() != InputVector.getResNo())
23824 SDNode *Extract = *UI;
23825 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
23828 if (Extract->getValueType(0) != MVT::i32)
23830 if (!Extract->hasOneUse())
23832 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
23833 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
23835 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
23838 // Record which element was extracted.
23839 ExtractedElements |=
23840 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
23842 Uses.push_back(Extract);
23845 // If not all the elements were used, this may not be worthwhile.
23846 if (ExtractedElements != 15)
23849 // Ok, we've now decided to do the transformation.
23850 // If 64-bit shifts are legal, use the extract-shift sequence,
23851 // otherwise bounce the vector off the cache.
23852 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23855 if (TLI.isOperationLegal(ISD::SRA, MVT::i64)) {
23856 SDValue Cst = DAG.getBitcast(MVT::v2i64, InputVector);
23857 auto &DL = DAG.getDataLayout();
23858 EVT VecIdxTy = DAG.getTargetLoweringInfo().getVectorIdxTy(DL);
23859 SDValue BottomHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
23860 DAG.getConstant(0, dl, VecIdxTy));
23861 SDValue TopHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
23862 DAG.getConstant(1, dl, VecIdxTy));
23864 SDValue ShAmt = DAG.getConstant(
23865 32, dl, DAG.getTargetLoweringInfo().getShiftAmountTy(MVT::i64, DL));
23866 Vals[0] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BottomHalf);
23867 Vals[1] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
23868 DAG.getNode(ISD::SRA, dl, MVT::i64, BottomHalf, ShAmt));
23869 Vals[2] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, TopHalf);
23870 Vals[3] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
23871 DAG.getNode(ISD::SRA, dl, MVT::i64, TopHalf, ShAmt));
23873 // Store the value to a temporary stack slot.
23874 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
23875 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
23876 MachinePointerInfo(), false, false, 0);
23878 EVT ElementType = InputVector.getValueType().getVectorElementType();
23879 unsigned EltSize = ElementType.getSizeInBits() / 8;
23881 // Replace each use (extract) with a load of the appropriate element.
23882 for (unsigned i = 0; i < 4; ++i) {
23883 uint64_t Offset = EltSize * i;
23884 auto PtrVT = TLI.getPointerTy(DAG.getDataLayout());
23885 SDValue OffsetVal = DAG.getConstant(Offset, dl, PtrVT);
23887 SDValue ScalarAddr =
23888 DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, OffsetVal);
23890 // Load the scalar.
23891 Vals[i] = DAG.getLoad(ElementType, dl, Ch,
23892 ScalarAddr, MachinePointerInfo(),
23893 false, false, false, 0);
23898 // Replace the extracts
23899 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
23900 UE = Uses.end(); UI != UE; ++UI) {
23901 SDNode *Extract = *UI;
23903 SDValue Idx = Extract->getOperand(1);
23904 uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
23905 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), Vals[IdxVal]);
23908 // The replacement was made in place; don't return anything.
23913 transformVSELECTtoBlendVECTOR_SHUFFLE(SDNode *N, SelectionDAG &DAG,
23914 const X86Subtarget *Subtarget) {
23916 SDValue Cond = N->getOperand(0);
23917 SDValue LHS = N->getOperand(1);
23918 SDValue RHS = N->getOperand(2);
23920 if (Cond.getOpcode() == ISD::SIGN_EXTEND) {
23921 SDValue CondSrc = Cond->getOperand(0);
23922 if (CondSrc->getOpcode() == ISD::SIGN_EXTEND_INREG)
23923 Cond = CondSrc->getOperand(0);
23926 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
23929 // A vselect where all conditions and data are constants can be optimized into
23930 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
23931 if (ISD::isBuildVectorOfConstantSDNodes(LHS.getNode()) &&
23932 ISD::isBuildVectorOfConstantSDNodes(RHS.getNode()))
23935 unsigned MaskValue = 0;
23936 if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue))
23939 MVT VT = N->getSimpleValueType(0);
23940 unsigned NumElems = VT.getVectorNumElements();
23941 SmallVector<int, 8> ShuffleMask(NumElems, -1);
23942 for (unsigned i = 0; i < NumElems; ++i) {
23943 // Be sure we emit undef where we can.
23944 if (Cond.getOperand(i)->getOpcode() == ISD::UNDEF)
23945 ShuffleMask[i] = -1;
23947 ShuffleMask[i] = i + NumElems * ((MaskValue >> i) & 1);
23950 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23951 if (!TLI.isShuffleMaskLegal(ShuffleMask, VT))
23953 return DAG.getVectorShuffle(VT, dl, LHS, RHS, &ShuffleMask[0]);
23956 /// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT
23958 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
23959 TargetLowering::DAGCombinerInfo &DCI,
23960 const X86Subtarget *Subtarget) {
23962 SDValue Cond = N->getOperand(0);
23963 // Get the LHS/RHS of the select.
23964 SDValue LHS = N->getOperand(1);
23965 SDValue RHS = N->getOperand(2);
23966 EVT VT = LHS.getValueType();
23967 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23969 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
23970 // instructions match the semantics of the common C idiom x<y?x:y but not
23971 // x<=y?x:y, because of how they handle negative zero (which can be
23972 // ignored in unsafe-math mode).
23973 // We also try to create v2f32 min/max nodes, which we later widen to v4f32.
23974 if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() &&
23975 VT != MVT::f80 && VT != MVT::f128 &&
23976 (TLI.isTypeLegal(VT) || VT == MVT::v2f32) &&
23977 (Subtarget->hasSSE2() ||
23978 (Subtarget->hasSSE1() && VT.getScalarType() == MVT::f32))) {
23979 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23981 unsigned Opcode = 0;
23982 // Check for x CC y ? x : y.
23983 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
23984 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
23988 // Converting this to a min would handle NaNs incorrectly, and swapping
23989 // the operands would cause it to handle comparisons between positive
23990 // and negative zero incorrectly.
23991 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
23992 if (!DAG.getTarget().Options.UnsafeFPMath &&
23993 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
23995 std::swap(LHS, RHS);
23997 Opcode = X86ISD::FMIN;
24000 // Converting this to a min would handle comparisons between positive
24001 // and negative zero incorrectly.
24002 if (!DAG.getTarget().Options.UnsafeFPMath &&
24003 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
24005 Opcode = X86ISD::FMIN;
24008 // Converting this to a min would handle both negative zeros and NaNs
24009 // incorrectly, but we can swap the operands to fix both.
24010 std::swap(LHS, RHS);
24014 Opcode = X86ISD::FMIN;
24018 // Converting this to a max would handle comparisons between positive
24019 // and negative zero incorrectly.
24020 if (!DAG.getTarget().Options.UnsafeFPMath &&
24021 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
24023 Opcode = X86ISD::FMAX;
24026 // Converting this to a max would handle NaNs incorrectly, and swapping
24027 // the operands would cause it to handle comparisons between positive
24028 // and negative zero incorrectly.
24029 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
24030 if (!DAG.getTarget().Options.UnsafeFPMath &&
24031 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
24033 std::swap(LHS, RHS);
24035 Opcode = X86ISD::FMAX;
24038 // Converting this to a max would handle both negative zeros and NaNs
24039 // incorrectly, but we can swap the operands to fix both.
24040 std::swap(LHS, RHS);
24044 Opcode = X86ISD::FMAX;
24047 // Check for x CC y ? y : x -- a min/max with reversed arms.
24048 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
24049 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
24053 // Converting this to a min would handle comparisons between positive
24054 // and negative zero incorrectly, and swapping the operands would
24055 // cause it to handle NaNs incorrectly.
24056 if (!DAG.getTarget().Options.UnsafeFPMath &&
24057 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
24058 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
24060 std::swap(LHS, RHS);
24062 Opcode = X86ISD::FMIN;
24065 // Converting this to a min would handle NaNs incorrectly.
24066 if (!DAG.getTarget().Options.UnsafeFPMath &&
24067 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
24069 Opcode = X86ISD::FMIN;
24072 // Converting this to a min would handle both negative zeros and NaNs
24073 // incorrectly, but we can swap the operands to fix both.
24074 std::swap(LHS, RHS);
24078 Opcode = X86ISD::FMIN;
24082 // Converting this to a max would handle NaNs incorrectly.
24083 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
24085 Opcode = X86ISD::FMAX;
24088 // Converting this to a max would handle comparisons between positive
24089 // and negative zero incorrectly, and swapping the operands would
24090 // cause it to handle NaNs incorrectly.
24091 if (!DAG.getTarget().Options.UnsafeFPMath &&
24092 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
24093 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
24095 std::swap(LHS, RHS);
24097 Opcode = X86ISD::FMAX;
24100 // Converting this to a max would handle both negative zeros and NaNs
24101 // incorrectly, but we can swap the operands to fix both.
24102 std::swap(LHS, RHS);
24106 Opcode = X86ISD::FMAX;
24112 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
24115 EVT CondVT = Cond.getValueType();
24116 if (Subtarget->hasAVX512() && VT.isVector() && CondVT.isVector() &&
24117 CondVT.getVectorElementType() == MVT::i1) {
24118 // v16i8 (select v16i1, v16i8, v16i8) does not have a proper
24119 // lowering on KNL. In this case we convert it to
24120 // v16i8 (select v16i8, v16i8, v16i8) and use AVX instruction.
24121 // The same situation for all 128 and 256-bit vectors of i8 and i16.
24122 // Since SKX these selects have a proper lowering.
24123 EVT OpVT = LHS.getValueType();
24124 if ((OpVT.is128BitVector() || OpVT.is256BitVector()) &&
24125 (OpVT.getVectorElementType() == MVT::i8 ||
24126 OpVT.getVectorElementType() == MVT::i16) &&
24127 !(Subtarget->hasBWI() && Subtarget->hasVLX())) {
24128 Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, OpVT, Cond);
24129 DCI.AddToWorklist(Cond.getNode());
24130 return DAG.getNode(N->getOpcode(), DL, OpVT, Cond, LHS, RHS);
24133 // If this is a select between two integer constants, try to do some
24135 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
24136 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
24137 // Don't do this for crazy integer types.
24138 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
24139 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
24140 // so that TrueC (the true value) is larger than FalseC.
24141 bool NeedsCondInvert = false;
24143 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
24144 // Efficiently invertible.
24145 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
24146 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
24147 isa<ConstantSDNode>(Cond.getOperand(1))))) {
24148 NeedsCondInvert = true;
24149 std::swap(TrueC, FalseC);
24152 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
24153 if (FalseC->getAPIntValue() == 0 &&
24154 TrueC->getAPIntValue().isPowerOf2()) {
24155 if (NeedsCondInvert) // Invert the condition if needed.
24156 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
24157 DAG.getConstant(1, DL, Cond.getValueType()));
24159 // Zero extend the condition if needed.
24160 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
24162 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
24163 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
24164 DAG.getConstant(ShAmt, DL, MVT::i8));
24167 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
24168 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
24169 if (NeedsCondInvert) // Invert the condition if needed.
24170 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
24171 DAG.getConstant(1, DL, Cond.getValueType()));
24173 // Zero extend the condition if needed.
24174 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
24175 FalseC->getValueType(0), Cond);
24176 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
24177 SDValue(FalseC, 0));
24180 // Optimize cases that will turn into an LEA instruction. This requires
24181 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
24182 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
24183 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
24184 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
24186 bool isFastMultiplier = false;
24188 switch ((unsigned char)Diff) {
24190 case 1: // result = add base, cond
24191 case 2: // result = lea base( , cond*2)
24192 case 3: // result = lea base(cond, cond*2)
24193 case 4: // result = lea base( , cond*4)
24194 case 5: // result = lea base(cond, cond*4)
24195 case 8: // result = lea base( , cond*8)
24196 case 9: // result = lea base(cond, cond*8)
24197 isFastMultiplier = true;
24202 if (isFastMultiplier) {
24203 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
24204 if (NeedsCondInvert) // Invert the condition if needed.
24205 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
24206 DAG.getConstant(1, DL, Cond.getValueType()));
24208 // Zero extend the condition if needed.
24209 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
24211 // Scale the condition by the difference.
24213 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
24214 DAG.getConstant(Diff, DL,
24215 Cond.getValueType()));
24217 // Add the base if non-zero.
24218 if (FalseC->getAPIntValue() != 0)
24219 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
24220 SDValue(FalseC, 0));
24227 // Canonicalize max and min:
24228 // (x > y) ? x : y -> (x >= y) ? x : y
24229 // (x < y) ? x : y -> (x <= y) ? x : y
24230 // This allows use of COND_S / COND_NS (see TranslateX86CC) which eliminates
24231 // the need for an extra compare
24232 // against zero. e.g.
24233 // (x - y) > 0 : (x - y) ? 0 -> (x - y) >= 0 : (x - y) ? 0
24235 // testl %edi, %edi
24237 // cmovgl %edi, %eax
24241 // cmovsl %eax, %edi
24242 if (N->getOpcode() == ISD::SELECT && Cond.getOpcode() == ISD::SETCC &&
24243 DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
24244 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
24245 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
24250 ISD::CondCode NewCC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGE;
24251 Cond = DAG.getSetCC(SDLoc(Cond), Cond.getValueType(),
24252 Cond.getOperand(0), Cond.getOperand(1), NewCC);
24253 return DAG.getNode(ISD::SELECT, DL, VT, Cond, LHS, RHS);
24258 // Early exit check
24259 if (!TLI.isTypeLegal(VT))
24262 // Match VSELECTs into subs with unsigned saturation.
24263 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
24264 // psubus is available in SSE2 and AVX2 for i8 and i16 vectors.
24265 ((Subtarget->hasSSE2() && (VT == MVT::v16i8 || VT == MVT::v8i16)) ||
24266 (Subtarget->hasAVX2() && (VT == MVT::v32i8 || VT == MVT::v16i16)))) {
24267 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
24269 // Check if one of the arms of the VSELECT is a zero vector. If it's on the
24270 // left side invert the predicate to simplify logic below.
24272 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
24274 CC = ISD::getSetCCInverse(CC, true);
24275 } else if (ISD::isBuildVectorAllZeros(RHS.getNode())) {
24279 if (Other.getNode() && Other->getNumOperands() == 2 &&
24280 DAG.isEqualTo(Other->getOperand(0), Cond.getOperand(0))) {
24281 SDValue OpLHS = Other->getOperand(0), OpRHS = Other->getOperand(1);
24282 SDValue CondRHS = Cond->getOperand(1);
24284 // Look for a general sub with unsigned saturation first.
24285 // x >= y ? x-y : 0 --> subus x, y
24286 // x > y ? x-y : 0 --> subus x, y
24287 if ((CC == ISD::SETUGE || CC == ISD::SETUGT) &&
24288 Other->getOpcode() == ISD::SUB && DAG.isEqualTo(OpRHS, CondRHS))
24289 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
24291 if (auto *OpRHSBV = dyn_cast<BuildVectorSDNode>(OpRHS))
24292 if (auto *OpRHSConst = OpRHSBV->getConstantSplatNode()) {
24293 if (auto *CondRHSBV = dyn_cast<BuildVectorSDNode>(CondRHS))
24294 if (auto *CondRHSConst = CondRHSBV->getConstantSplatNode())
24295 // If the RHS is a constant we have to reverse the const
24296 // canonicalization.
24297 // x > C-1 ? x+-C : 0 --> subus x, C
24298 if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD &&
24299 CondRHSConst->getAPIntValue() ==
24300 (-OpRHSConst->getAPIntValue() - 1))
24301 return DAG.getNode(
24302 X86ISD::SUBUS, DL, VT, OpLHS,
24303 DAG.getConstant(-OpRHSConst->getAPIntValue(), DL, VT));
24305 // Another special case: If C was a sign bit, the sub has been
24306 // canonicalized into a xor.
24307 // FIXME: Would it be better to use computeKnownBits to determine
24308 // whether it's safe to decanonicalize the xor?
24309 // x s< 0 ? x^C : 0 --> subus x, C
24310 if (CC == ISD::SETLT && Other->getOpcode() == ISD::XOR &&
24311 ISD::isBuildVectorAllZeros(CondRHS.getNode()) &&
24312 OpRHSConst->getAPIntValue().isSignBit())
24313 // Note that we have to rebuild the RHS constant here to ensure we
24314 // don't rely on particular values of undef lanes.
24315 return DAG.getNode(
24316 X86ISD::SUBUS, DL, VT, OpLHS,
24317 DAG.getConstant(OpRHSConst->getAPIntValue(), DL, VT));
24322 // Simplify vector selection if condition value type matches vselect
24324 if (N->getOpcode() == ISD::VSELECT && CondVT == VT) {
24325 assert(Cond.getValueType().isVector() &&
24326 "vector select expects a vector selector!");
24328 bool TValIsAllOnes = ISD::isBuildVectorAllOnes(LHS.getNode());
24329 bool FValIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
24331 // Try invert the condition if true value is not all 1s and false value
24333 if (!TValIsAllOnes && !FValIsAllZeros &&
24334 // Check if the selector will be produced by CMPP*/PCMP*
24335 Cond.getOpcode() == ISD::SETCC &&
24336 // Check if SETCC has already been promoted
24337 TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT) ==
24339 bool TValIsAllZeros = ISD::isBuildVectorAllZeros(LHS.getNode());
24340 bool FValIsAllOnes = ISD::isBuildVectorAllOnes(RHS.getNode());
24342 if (TValIsAllZeros || FValIsAllOnes) {
24343 SDValue CC = Cond.getOperand(2);
24344 ISD::CondCode NewCC =
24345 ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
24346 Cond.getOperand(0).getValueType().isInteger());
24347 Cond = DAG.getSetCC(DL, CondVT, Cond.getOperand(0), Cond.getOperand(1), NewCC);
24348 std::swap(LHS, RHS);
24349 TValIsAllOnes = FValIsAllOnes;
24350 FValIsAllZeros = TValIsAllZeros;
24354 if (TValIsAllOnes || FValIsAllZeros) {
24357 if (TValIsAllOnes && FValIsAllZeros)
24359 else if (TValIsAllOnes)
24361 DAG.getNode(ISD::OR, DL, CondVT, Cond, DAG.getBitcast(CondVT, RHS));
24362 else if (FValIsAllZeros)
24363 Ret = DAG.getNode(ISD::AND, DL, CondVT, Cond,
24364 DAG.getBitcast(CondVT, LHS));
24366 return DAG.getBitcast(VT, Ret);
24370 // We should generate an X86ISD::BLENDI from a vselect if its argument
24371 // is a sign_extend_inreg of an any_extend of a BUILD_VECTOR of
24372 // constants. This specific pattern gets generated when we split a
24373 // selector for a 512 bit vector in a machine without AVX512 (but with
24374 // 256-bit vectors), during legalization:
24376 // (vselect (sign_extend (any_extend (BUILD_VECTOR)) i1) LHS RHS)
24378 // Iff we find this pattern and the build_vectors are built from
24379 // constants, we translate the vselect into a shuffle_vector that we
24380 // know will be matched by LowerVECTOR_SHUFFLEtoBlend.
24381 if ((N->getOpcode() == ISD::VSELECT ||
24382 N->getOpcode() == X86ISD::SHRUNKBLEND) &&
24383 !DCI.isBeforeLegalize() && !VT.is512BitVector()) {
24384 SDValue Shuffle = transformVSELECTtoBlendVECTOR_SHUFFLE(N, DAG, Subtarget);
24385 if (Shuffle.getNode())
24389 // If this is a *dynamic* select (non-constant condition) and we can match
24390 // this node with one of the variable blend instructions, restructure the
24391 // condition so that the blends can use the high bit of each element and use
24392 // SimplifyDemandedBits to simplify the condition operand.
24393 if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() &&
24394 !DCI.isBeforeLegalize() &&
24395 !ISD::isBuildVectorOfConstantSDNodes(Cond.getNode())) {
24396 unsigned BitWidth = Cond.getValueType().getScalarSizeInBits();
24398 // Don't optimize vector selects that map to mask-registers.
24402 // We can only handle the cases where VSELECT is directly legal on the
24403 // subtarget. We custom lower VSELECT nodes with constant conditions and
24404 // this makes it hard to see whether a dynamic VSELECT will correctly
24405 // lower, so we both check the operation's status and explicitly handle the
24406 // cases where a *dynamic* blend will fail even though a constant-condition
24407 // blend could be custom lowered.
24408 // FIXME: We should find a better way to handle this class of problems.
24409 // Potentially, we should combine constant-condition vselect nodes
24410 // pre-legalization into shuffles and not mark as many types as custom
24412 if (!TLI.isOperationLegalOrCustom(ISD::VSELECT, VT))
24414 // FIXME: We don't support i16-element blends currently. We could and
24415 // should support them by making *all* the bits in the condition be set
24416 // rather than just the high bit and using an i8-element blend.
24417 if (VT.getVectorElementType() == MVT::i16)
24419 // Dynamic blending was only available from SSE4.1 onward.
24420 if (VT.is128BitVector() && !Subtarget->hasSSE41())
24422 // Byte blends are only available in AVX2
24423 if (VT == MVT::v32i8 && !Subtarget->hasAVX2())
24426 assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size");
24427 APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);
24429 APInt KnownZero, KnownOne;
24430 TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(),
24431 DCI.isBeforeLegalizeOps());
24432 if (TLO.ShrinkDemandedConstant(Cond, DemandedMask) ||
24433 TLI.SimplifyDemandedBits(Cond, DemandedMask, KnownZero, KnownOne,
24435 // If we changed the computation somewhere in the DAG, this change
24436 // will affect all users of Cond.
24437 // Make sure it is fine and update all the nodes so that we do not
24438 // use the generic VSELECT anymore. Otherwise, we may perform
24439 // wrong optimizations as we messed up with the actual expectation
24440 // for the vector boolean values.
24441 if (Cond != TLO.Old) {
24442 // Check all uses of that condition operand to check whether it will be
24443 // consumed by non-BLEND instructions, which may depend on all bits are
24445 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
24447 if (I->getOpcode() != ISD::VSELECT)
24448 // TODO: Add other opcodes eventually lowered into BLEND.
24451 // Update all the users of the condition, before committing the change,
24452 // so that the VSELECT optimizations that expect the correct vector
24453 // boolean value will not be triggered.
24454 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
24456 DAG.ReplaceAllUsesOfValueWith(
24458 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(*I), I->getValueType(0),
24459 Cond, I->getOperand(1), I->getOperand(2)));
24460 DCI.CommitTargetLoweringOpt(TLO);
24463 // At this point, only Cond is changed. Change the condition
24464 // just for N to keep the opportunity to optimize all other
24465 // users their own way.
24466 DAG.ReplaceAllUsesOfValueWith(
24468 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(N), N->getValueType(0),
24469 TLO.New, N->getOperand(1), N->getOperand(2)));
24477 // Check whether a boolean test is testing a boolean value generated by
24478 // X86ISD::SETCC. If so, return the operand of that SETCC and proper condition
24481 // Simplify the following patterns:
24482 // (Op (CMP (SETCC Cond EFLAGS) 1) EQ) or
24483 // (Op (CMP (SETCC Cond EFLAGS) 0) NEQ)
24484 // to (Op EFLAGS Cond)
24486 // (Op (CMP (SETCC Cond EFLAGS) 0) EQ) or
24487 // (Op (CMP (SETCC Cond EFLAGS) 1) NEQ)
24488 // to (Op EFLAGS !Cond)
24490 // where Op could be BRCOND or CMOV.
24492 static SDValue checkBoolTestSetCCCombine(SDValue Cmp, X86::CondCode &CC) {
24493 // Quit if not CMP and SUB with its value result used.
24494 if (Cmp.getOpcode() != X86ISD::CMP &&
24495 (Cmp.getOpcode() != X86ISD::SUB || Cmp.getNode()->hasAnyUseOfValue(0)))
24498 // Quit if not used as a boolean value.
24499 if (CC != X86::COND_E && CC != X86::COND_NE)
24502 // Check CMP operands. One of them should be 0 or 1 and the other should be
24503 // an SetCC or extended from it.
24504 SDValue Op1 = Cmp.getOperand(0);
24505 SDValue Op2 = Cmp.getOperand(1);
24508 const ConstantSDNode* C = nullptr;
24509 bool needOppositeCond = (CC == X86::COND_E);
24510 bool checkAgainstTrue = false; // Is it a comparison against 1?
24512 if ((C = dyn_cast<ConstantSDNode>(Op1)))
24514 else if ((C = dyn_cast<ConstantSDNode>(Op2)))
24516 else // Quit if all operands are not constants.
24519 if (C->getZExtValue() == 1) {
24520 needOppositeCond = !needOppositeCond;
24521 checkAgainstTrue = true;
24522 } else if (C->getZExtValue() != 0)
24523 // Quit if the constant is neither 0 or 1.
24526 bool truncatedToBoolWithAnd = false;
24527 // Skip (zext $x), (trunc $x), or (and $x, 1) node.
24528 while (SetCC.getOpcode() == ISD::ZERO_EXTEND ||
24529 SetCC.getOpcode() == ISD::TRUNCATE ||
24530 SetCC.getOpcode() == ISD::AND) {
24531 if (SetCC.getOpcode() == ISD::AND) {
24533 if (isOneConstant(SetCC.getOperand(0)))
24535 if (isOneConstant(SetCC.getOperand(1)))
24539 SetCC = SetCC.getOperand(OpIdx);
24540 truncatedToBoolWithAnd = true;
24542 SetCC = SetCC.getOperand(0);
24545 switch (SetCC.getOpcode()) {
24546 case X86ISD::SETCC_CARRY:
24547 // Since SETCC_CARRY gives output based on R = CF ? ~0 : 0, it's unsafe to
24548 // simplify it if the result of SETCC_CARRY is not canonicalized to 0 or 1,
24549 // i.e. it's a comparison against true but the result of SETCC_CARRY is not
24550 // truncated to i1 using 'and'.
24551 if (checkAgainstTrue && !truncatedToBoolWithAnd)
24553 assert(X86::CondCode(SetCC.getConstantOperandVal(0)) == X86::COND_B &&
24554 "Invalid use of SETCC_CARRY!");
24556 case X86ISD::SETCC:
24557 // Set the condition code or opposite one if necessary.
24558 CC = X86::CondCode(SetCC.getConstantOperandVal(0));
24559 if (needOppositeCond)
24560 CC = X86::GetOppositeBranchCondition(CC);
24561 return SetCC.getOperand(1);
24562 case X86ISD::CMOV: {
24563 // Check whether false/true value has canonical one, i.e. 0 or 1.
24564 ConstantSDNode *FVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(0));
24565 ConstantSDNode *TVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(1));
24566 // Quit if true value is not a constant.
24569 // Quit if false value is not a constant.
24571 SDValue Op = SetCC.getOperand(0);
24572 // Skip 'zext' or 'trunc' node.
24573 if (Op.getOpcode() == ISD::ZERO_EXTEND ||
24574 Op.getOpcode() == ISD::TRUNCATE)
24575 Op = Op.getOperand(0);
24576 // A special case for rdrand/rdseed, where 0 is set if false cond is
24578 if ((Op.getOpcode() != X86ISD::RDRAND &&
24579 Op.getOpcode() != X86ISD::RDSEED) || Op.getResNo() != 0)
24582 // Quit if false value is not the constant 0 or 1.
24583 bool FValIsFalse = true;
24584 if (FVal && FVal->getZExtValue() != 0) {
24585 if (FVal->getZExtValue() != 1)
24587 // If FVal is 1, opposite cond is needed.
24588 needOppositeCond = !needOppositeCond;
24589 FValIsFalse = false;
24591 // Quit if TVal is not the constant opposite of FVal.
24592 if (FValIsFalse && TVal->getZExtValue() != 1)
24594 if (!FValIsFalse && TVal->getZExtValue() != 0)
24596 CC = X86::CondCode(SetCC.getConstantOperandVal(2));
24597 if (needOppositeCond)
24598 CC = X86::GetOppositeBranchCondition(CC);
24599 return SetCC.getOperand(3);
24606 /// Check whether Cond is an AND/OR of SETCCs off of the same EFLAGS.
24608 /// (X86or (X86setcc) (X86setcc))
24609 /// (X86cmp (and (X86setcc) (X86setcc)), 0)
24610 static bool checkBoolTestAndOrSetCCCombine(SDValue Cond, X86::CondCode &CC0,
24611 X86::CondCode &CC1, SDValue &Flags,
24613 if (Cond->getOpcode() == X86ISD::CMP) {
24614 if (!isNullConstant(Cond->getOperand(1)))
24617 Cond = Cond->getOperand(0);
24622 SDValue SetCC0, SetCC1;
24623 switch (Cond->getOpcode()) {
24624 default: return false;
24631 SetCC0 = Cond->getOperand(0);
24632 SetCC1 = Cond->getOperand(1);
24636 // Make sure we have SETCC nodes, using the same flags value.
24637 if (SetCC0.getOpcode() != X86ISD::SETCC ||
24638 SetCC1.getOpcode() != X86ISD::SETCC ||
24639 SetCC0->getOperand(1) != SetCC1->getOperand(1))
24642 CC0 = (X86::CondCode)SetCC0->getConstantOperandVal(0);
24643 CC1 = (X86::CondCode)SetCC1->getConstantOperandVal(0);
24644 Flags = SetCC0->getOperand(1);
24648 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
24649 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
24650 TargetLowering::DAGCombinerInfo &DCI,
24651 const X86Subtarget *Subtarget) {
24654 // If the flag operand isn't dead, don't touch this CMOV.
24655 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
24658 SDValue FalseOp = N->getOperand(0);
24659 SDValue TrueOp = N->getOperand(1);
24660 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
24661 SDValue Cond = N->getOperand(3);
24663 if (CC == X86::COND_E || CC == X86::COND_NE) {
24664 switch (Cond.getOpcode()) {
24668 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
24669 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
24670 return (CC == X86::COND_E) ? FalseOp : TrueOp;
24676 Flags = checkBoolTestSetCCCombine(Cond, CC);
24677 if (Flags.getNode() &&
24678 // Extra check as FCMOV only supports a subset of X86 cond.
24679 (FalseOp.getValueType() != MVT::f80 || hasFPCMov(CC))) {
24680 SDValue Ops[] = { FalseOp, TrueOp,
24681 DAG.getConstant(CC, DL, MVT::i8), Flags };
24682 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
24685 // If this is a select between two integer constants, try to do some
24686 // optimizations. Note that the operands are ordered the opposite of SELECT
24688 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
24689 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
24690 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
24691 // larger than FalseC (the false value).
24692 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
24693 CC = X86::GetOppositeBranchCondition(CC);
24694 std::swap(TrueC, FalseC);
24695 std::swap(TrueOp, FalseOp);
24698 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
24699 // This is efficient for any integer data type (including i8/i16) and
24701 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
24702 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
24703 DAG.getConstant(CC, DL, MVT::i8), Cond);
24705 // Zero extend the condition if needed.
24706 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
24708 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
24709 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
24710 DAG.getConstant(ShAmt, DL, MVT::i8));
24711 if (N->getNumValues() == 2) // Dead flag value?
24712 return DCI.CombineTo(N, Cond, SDValue());
24716 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
24717 // for any integer data type, including i8/i16.
24718 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
24719 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
24720 DAG.getConstant(CC, DL, MVT::i8), Cond);
24722 // Zero extend the condition if needed.
24723 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
24724 FalseC->getValueType(0), Cond);
24725 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
24726 SDValue(FalseC, 0));
24728 if (N->getNumValues() == 2) // Dead flag value?
24729 return DCI.CombineTo(N, Cond, SDValue());
24733 // Optimize cases that will turn into an LEA instruction. This requires
24734 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
24735 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
24736 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
24737 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
24739 bool isFastMultiplier = false;
24741 switch ((unsigned char)Diff) {
24743 case 1: // result = add base, cond
24744 case 2: // result = lea base( , cond*2)
24745 case 3: // result = lea base(cond, cond*2)
24746 case 4: // result = lea base( , cond*4)
24747 case 5: // result = lea base(cond, cond*4)
24748 case 8: // result = lea base( , cond*8)
24749 case 9: // result = lea base(cond, cond*8)
24750 isFastMultiplier = true;
24755 if (isFastMultiplier) {
24756 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
24757 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
24758 DAG.getConstant(CC, DL, MVT::i8), Cond);
24759 // Zero extend the condition if needed.
24760 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
24762 // Scale the condition by the difference.
24764 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
24765 DAG.getConstant(Diff, DL, Cond.getValueType()));
24767 // Add the base if non-zero.
24768 if (FalseC->getAPIntValue() != 0)
24769 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
24770 SDValue(FalseC, 0));
24771 if (N->getNumValues() == 2) // Dead flag value?
24772 return DCI.CombineTo(N, Cond, SDValue());
24779 // Handle these cases:
24780 // (select (x != c), e, c) -> select (x != c), e, x),
24781 // (select (x == c), c, e) -> select (x == c), x, e)
24782 // where the c is an integer constant, and the "select" is the combination
24783 // of CMOV and CMP.
24785 // The rationale for this change is that the conditional-move from a constant
24786 // needs two instructions, however, conditional-move from a register needs
24787 // only one instruction.
24789 // CAVEAT: By replacing a constant with a symbolic value, it may obscure
24790 // some instruction-combining opportunities. This opt needs to be
24791 // postponed as late as possible.
24793 if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) {
24794 // the DCI.xxxx conditions are provided to postpone the optimization as
24795 // late as possible.
24797 ConstantSDNode *CmpAgainst = nullptr;
24798 if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) &&
24799 (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) &&
24800 !isa<ConstantSDNode>(Cond.getOperand(0))) {
24802 if (CC == X86::COND_NE &&
24803 CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) {
24804 CC = X86::GetOppositeBranchCondition(CC);
24805 std::swap(TrueOp, FalseOp);
24808 if (CC == X86::COND_E &&
24809 CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) {
24810 SDValue Ops[] = { FalseOp, Cond.getOperand(0),
24811 DAG.getConstant(CC, DL, MVT::i8), Cond };
24812 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops);
24817 // Fold and/or of setcc's to double CMOV:
24818 // (CMOV F, T, ((cc1 | cc2) != 0)) -> (CMOV (CMOV F, T, cc1), T, cc2)
24819 // (CMOV F, T, ((cc1 & cc2) != 0)) -> (CMOV (CMOV T, F, !cc1), F, !cc2)
24821 // This combine lets us generate:
24822 // cmovcc1 (jcc1 if we don't have CMOV)
24828 // cmovne (jne if we don't have CMOV)
24829 // When we can't use the CMOV instruction, it might increase branch
24831 // When we can use CMOV, or when there is no mispredict, this improves
24832 // throughput and reduces register pressure.
24834 if (CC == X86::COND_NE) {
24836 X86::CondCode CC0, CC1;
24838 if (checkBoolTestAndOrSetCCCombine(Cond, CC0, CC1, Flags, isAndSetCC)) {
24840 std::swap(FalseOp, TrueOp);
24841 CC0 = X86::GetOppositeBranchCondition(CC0);
24842 CC1 = X86::GetOppositeBranchCondition(CC1);
24845 SDValue LOps[] = {FalseOp, TrueOp, DAG.getConstant(CC0, DL, MVT::i8),
24847 SDValue LCMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), LOps);
24848 SDValue Ops[] = {LCMOV, TrueOp, DAG.getConstant(CC1, DL, MVT::i8), Flags};
24849 SDValue CMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
24850 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SDValue(CMOV.getNode(), 1));
24858 /// PerformMulCombine - Optimize a single multiply with constant into two
24859 /// in order to implement it with two cheaper instructions, e.g.
24860 /// LEA + SHL, LEA + LEA.
24861 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
24862 TargetLowering::DAGCombinerInfo &DCI) {
24863 // An imul is usually smaller than the alternative sequence.
24864 if (DAG.getMachineFunction().getFunction()->optForMinSize())
24867 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
24870 EVT VT = N->getValueType(0);
24871 if (VT != MVT::i64 && VT != MVT::i32)
24874 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
24877 uint64_t MulAmt = C->getZExtValue();
24878 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
24881 uint64_t MulAmt1 = 0;
24882 uint64_t MulAmt2 = 0;
24883 if ((MulAmt % 9) == 0) {
24885 MulAmt2 = MulAmt / 9;
24886 } else if ((MulAmt % 5) == 0) {
24888 MulAmt2 = MulAmt / 5;
24889 } else if ((MulAmt % 3) == 0) {
24891 MulAmt2 = MulAmt / 3;
24897 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
24899 if (isPowerOf2_64(MulAmt2) &&
24900 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
24901 // If second multiplifer is pow2, issue it first. We want the multiply by
24902 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
24904 std::swap(MulAmt1, MulAmt2);
24906 if (isPowerOf2_64(MulAmt1))
24907 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
24908 DAG.getConstant(Log2_64(MulAmt1), DL, MVT::i8));
24910 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
24911 DAG.getConstant(MulAmt1, DL, VT));
24913 if (isPowerOf2_64(MulAmt2))
24914 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
24915 DAG.getConstant(Log2_64(MulAmt2), DL, MVT::i8));
24917 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
24918 DAG.getConstant(MulAmt2, DL, VT));
24922 assert(MulAmt != 0 && MulAmt != (VT == MVT::i64 ? UINT64_MAX : UINT32_MAX)
24923 && "Both cases that could cause potential overflows should have "
24924 "already been handled.");
24925 if (isPowerOf2_64(MulAmt - 1))
24926 // (mul x, 2^N + 1) => (add (shl x, N), x)
24927 NewMul = DAG.getNode(ISD::ADD, DL, VT, N->getOperand(0),
24928 DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
24929 DAG.getConstant(Log2_64(MulAmt - 1), DL,
24932 else if (isPowerOf2_64(MulAmt + 1))
24933 // (mul x, 2^N - 1) => (sub (shl x, N), x)
24934 NewMul = DAG.getNode(ISD::SUB, DL, VT, DAG.getNode(ISD::SHL, DL, VT,
24936 DAG.getConstant(Log2_64(MulAmt + 1),
24937 DL, MVT::i8)), N->getOperand(0));
24941 // Do not add new nodes to DAG combiner worklist.
24942 DCI.CombineTo(N, NewMul, false);
24947 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
24948 SDValue N0 = N->getOperand(0);
24949 SDValue N1 = N->getOperand(1);
24950 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
24951 EVT VT = N0.getValueType();
24953 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
24954 // since the result of setcc_c is all zero's or all ones.
24955 if (VT.isInteger() && !VT.isVector() &&
24956 N1C && N0.getOpcode() == ISD::AND &&
24957 N0.getOperand(1).getOpcode() == ISD::Constant) {
24958 SDValue N00 = N0.getOperand(0);
24959 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
24960 APInt ShAmt = N1C->getAPIntValue();
24961 Mask = Mask.shl(ShAmt);
24962 bool MaskOK = false;
24963 // We can handle cases concerning bit-widening nodes containing setcc_c if
24964 // we carefully interrogate the mask to make sure we are semantics
24966 // The transform is not safe if the result of C1 << C2 exceeds the bitwidth
24967 // of the underlying setcc_c operation if the setcc_c was zero extended.
24968 // Consider the following example:
24969 // zext(setcc_c) -> i32 0x0000FFFF
24970 // c1 -> i32 0x0000FFFF
24971 // c2 -> i32 0x00000001
24972 // (shl (and (setcc_c), c1), c2) -> i32 0x0001FFFE
24973 // (and setcc_c, (c1 << c2)) -> i32 0x0000FFFE
24974 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
24976 } else if (N00.getOpcode() == ISD::SIGN_EXTEND &&
24977 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
24979 } else if ((N00.getOpcode() == ISD::ZERO_EXTEND ||
24980 N00.getOpcode() == ISD::ANY_EXTEND) &&
24981 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
24982 MaskOK = Mask.isIntN(N00.getOperand(0).getValueSizeInBits());
24984 if (MaskOK && Mask != 0) {
24986 return DAG.getNode(ISD::AND, DL, VT, N00, DAG.getConstant(Mask, DL, VT));
24990 // Hardware support for vector shifts is sparse which makes us scalarize the
24991 // vector operations in many cases. Also, on sandybridge ADD is faster than
24993 // (shl V, 1) -> add V,V
24994 if (auto *N1BV = dyn_cast<BuildVectorSDNode>(N1))
24995 if (auto *N1SplatC = N1BV->getConstantSplatNode()) {
24996 assert(N0.getValueType().isVector() && "Invalid vector shift type");
24997 // We shift all of the values by one. In many cases we do not have
24998 // hardware support for this operation. This is better expressed as an ADD
25000 if (N1SplatC->getAPIntValue() == 1)
25001 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N0);
25007 /// \brief Returns a vector of 0s if the node in input is a vector logical
25008 /// shift by a constant amount which is known to be bigger than or equal
25009 /// to the vector element size in bits.
25010 static SDValue performShiftToAllZeros(SDNode *N, SelectionDAG &DAG,
25011 const X86Subtarget *Subtarget) {
25012 EVT VT = N->getValueType(0);
25014 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16 &&
25015 (!Subtarget->hasInt256() ||
25016 (VT != MVT::v4i64 && VT != MVT::v8i32 && VT != MVT::v16i16)))
25019 SDValue Amt = N->getOperand(1);
25021 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Amt))
25022 if (auto *AmtSplat = AmtBV->getConstantSplatNode()) {
25023 APInt ShiftAmt = AmtSplat->getAPIntValue();
25024 unsigned MaxAmount =
25025 VT.getSimpleVT().getVectorElementType().getSizeInBits();
25027 // SSE2/AVX2 logical shifts always return a vector of 0s
25028 // if the shift amount is bigger than or equal to
25029 // the element size. The constant shift amount will be
25030 // encoded as a 8-bit immediate.
25031 if (ShiftAmt.trunc(8).uge(MaxAmount))
25032 return getZeroVector(VT.getSimpleVT(), Subtarget, DAG, DL);
25038 /// PerformShiftCombine - Combine shifts.
25039 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
25040 TargetLowering::DAGCombinerInfo &DCI,
25041 const X86Subtarget *Subtarget) {
25042 if (N->getOpcode() == ISD::SHL)
25043 if (SDValue V = PerformSHLCombine(N, DAG))
25046 // Try to fold this logical shift into a zero vector.
25047 if (N->getOpcode() != ISD::SRA)
25048 if (SDValue V = performShiftToAllZeros(N, DAG, Subtarget))
25054 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
25055 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
25056 // and friends. Likewise for OR -> CMPNEQSS.
25057 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
25058 TargetLowering::DAGCombinerInfo &DCI,
25059 const X86Subtarget *Subtarget) {
25062 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
25063 // we're requiring SSE2 for both.
25064 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
25065 SDValue N0 = N->getOperand(0);
25066 SDValue N1 = N->getOperand(1);
25067 SDValue CMP0 = N0->getOperand(1);
25068 SDValue CMP1 = N1->getOperand(1);
25071 // The SETCCs should both refer to the same CMP.
25072 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
25075 SDValue CMP00 = CMP0->getOperand(0);
25076 SDValue CMP01 = CMP0->getOperand(1);
25077 EVT VT = CMP00.getValueType();
25079 if (VT == MVT::f32 || VT == MVT::f64) {
25080 bool ExpectingFlags = false;
25081 // Check for any users that want flags:
25082 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
25083 !ExpectingFlags && UI != UE; ++UI)
25084 switch (UI->getOpcode()) {
25089 ExpectingFlags = true;
25091 case ISD::CopyToReg:
25092 case ISD::SIGN_EXTEND:
25093 case ISD::ZERO_EXTEND:
25094 case ISD::ANY_EXTEND:
25098 if (!ExpectingFlags) {
25099 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
25100 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
25102 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
25103 X86::CondCode tmp = cc0;
25108 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
25109 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
25110 // FIXME: need symbolic constants for these magic numbers.
25111 // See X86ATTInstPrinter.cpp:printSSECC().
25112 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
25113 if (Subtarget->hasAVX512()) {
25114 SDValue FSetCC = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CMP00,
25116 DAG.getConstant(x86cc, DL, MVT::i8));
25117 if (N->getValueType(0) != MVT::i1)
25118 return DAG.getNode(ISD::ZERO_EXTEND, DL, N->getValueType(0),
25122 SDValue OnesOrZeroesF = DAG.getNode(X86ISD::FSETCC, DL,
25123 CMP00.getValueType(), CMP00, CMP01,
25124 DAG.getConstant(x86cc, DL,
25127 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
25128 MVT IntVT = is64BitFP ? MVT::i64 : MVT::i32;
25130 if (is64BitFP && !Subtarget->is64Bit()) {
25131 // On a 32-bit target, we cannot bitcast the 64-bit float to a
25132 // 64-bit integer, since that's not a legal type. Since
25133 // OnesOrZeroesF is all ones of all zeroes, we don't need all the
25134 // bits, but can do this little dance to extract the lowest 32 bits
25135 // and work with those going forward.
25136 SDValue Vector64 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64,
25138 SDValue Vector32 = DAG.getBitcast(MVT::v4f32, Vector64);
25139 OnesOrZeroesF = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32,
25140 Vector32, DAG.getIntPtrConstant(0, DL));
25144 SDValue OnesOrZeroesI = DAG.getBitcast(IntVT, OnesOrZeroesF);
25145 SDValue ANDed = DAG.getNode(ISD::AND, DL, IntVT, OnesOrZeroesI,
25146 DAG.getConstant(1, DL, IntVT));
25147 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8,
25149 return OneBitOfTruth;
25157 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
25158 /// so it can be folded inside ANDNP.
25159 static bool CanFoldXORWithAllOnes(const SDNode *N) {
25160 EVT VT = N->getValueType(0);
25162 // Match direct AllOnes for 128 and 256-bit vectors
25163 if (ISD::isBuildVectorAllOnes(N))
25166 // Look through a bit convert.
25167 if (N->getOpcode() == ISD::BITCAST)
25168 N = N->getOperand(0).getNode();
25170 // Sometimes the operand may come from a insert_subvector building a 256-bit
25172 if (VT.is256BitVector() &&
25173 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
25174 SDValue V1 = N->getOperand(0);
25175 SDValue V2 = N->getOperand(1);
25177 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
25178 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
25179 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
25180 ISD::isBuildVectorAllOnes(V2.getNode()))
25187 // On AVX/AVX2 the type v8i1 is legalized to v8i16, which is an XMM sized
25188 // register. In most cases we actually compare or select YMM-sized registers
25189 // and mixing the two types creates horrible code. This method optimizes
25190 // some of the transition sequences.
25191 static SDValue WidenMaskArithmetic(SDNode *N, SelectionDAG &DAG,
25192 TargetLowering::DAGCombinerInfo &DCI,
25193 const X86Subtarget *Subtarget) {
25194 EVT VT = N->getValueType(0);
25195 if (!VT.is256BitVector())
25198 assert((N->getOpcode() == ISD::ANY_EXTEND ||
25199 N->getOpcode() == ISD::ZERO_EXTEND ||
25200 N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node");
25202 SDValue Narrow = N->getOperand(0);
25203 EVT NarrowVT = Narrow->getValueType(0);
25204 if (!NarrowVT.is128BitVector())
25207 if (Narrow->getOpcode() != ISD::XOR &&
25208 Narrow->getOpcode() != ISD::AND &&
25209 Narrow->getOpcode() != ISD::OR)
25212 SDValue N0 = Narrow->getOperand(0);
25213 SDValue N1 = Narrow->getOperand(1);
25216 // The Left side has to be a trunc.
25217 if (N0.getOpcode() != ISD::TRUNCATE)
25220 // The type of the truncated inputs.
25221 EVT WideVT = N0->getOperand(0)->getValueType(0);
25225 // The right side has to be a 'trunc' or a constant vector.
25226 bool RHSTrunc = N1.getOpcode() == ISD::TRUNCATE;
25227 ConstantSDNode *RHSConstSplat = nullptr;
25228 if (auto *RHSBV = dyn_cast<BuildVectorSDNode>(N1))
25229 RHSConstSplat = RHSBV->getConstantSplatNode();
25230 if (!RHSTrunc && !RHSConstSplat)
25233 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25235 if (!TLI.isOperationLegalOrPromote(Narrow->getOpcode(), WideVT))
25238 // Set N0 and N1 to hold the inputs to the new wide operation.
25239 N0 = N0->getOperand(0);
25240 if (RHSConstSplat) {
25241 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT.getVectorElementType(),
25242 SDValue(RHSConstSplat, 0));
25243 SmallVector<SDValue, 8> C(WideVT.getVectorNumElements(), N1);
25244 N1 = DAG.getNode(ISD::BUILD_VECTOR, DL, WideVT, C);
25245 } else if (RHSTrunc) {
25246 N1 = N1->getOperand(0);
25249 // Generate the wide operation.
25250 SDValue Op = DAG.getNode(Narrow->getOpcode(), DL, WideVT, N0, N1);
25251 unsigned Opcode = N->getOpcode();
25253 case ISD::ANY_EXTEND:
25255 case ISD::ZERO_EXTEND: {
25256 unsigned InBits = NarrowVT.getScalarSizeInBits();
25257 APInt Mask = APInt::getAllOnesValue(InBits);
25258 Mask = Mask.zext(VT.getScalarSizeInBits());
25259 return DAG.getNode(ISD::AND, DL, VT,
25260 Op, DAG.getConstant(Mask, DL, VT));
25262 case ISD::SIGN_EXTEND:
25263 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT,
25264 Op, DAG.getValueType(NarrowVT));
25266 llvm_unreachable("Unexpected opcode");
25270 static SDValue VectorZextCombine(SDNode *N, SelectionDAG &DAG,
25271 TargetLowering::DAGCombinerInfo &DCI,
25272 const X86Subtarget *Subtarget) {
25273 SDValue N0 = N->getOperand(0);
25274 SDValue N1 = N->getOperand(1);
25277 // A vector zext_in_reg may be represented as a shuffle,
25278 // feeding into a bitcast (this represents anyext) feeding into
25279 // an and with a mask.
25280 // We'd like to try to combine that into a shuffle with zero
25281 // plus a bitcast, removing the and.
25282 if (N0.getOpcode() != ISD::BITCAST ||
25283 N0.getOperand(0).getOpcode() != ISD::VECTOR_SHUFFLE)
25286 // The other side of the AND should be a splat of 2^C, where C
25287 // is the number of bits in the source type.
25288 if (N1.getOpcode() == ISD::BITCAST)
25289 N1 = N1.getOperand(0);
25290 if (N1.getOpcode() != ISD::BUILD_VECTOR)
25292 BuildVectorSDNode *Vector = cast<BuildVectorSDNode>(N1);
25294 ShuffleVectorSDNode *Shuffle = cast<ShuffleVectorSDNode>(N0.getOperand(0));
25295 EVT SrcType = Shuffle->getValueType(0);
25297 // We expect a single-source shuffle
25298 if (Shuffle->getOperand(1)->getOpcode() != ISD::UNDEF)
25301 unsigned SrcSize = SrcType.getScalarSizeInBits();
25303 APInt SplatValue, SplatUndef;
25304 unsigned SplatBitSize;
25306 if (!Vector->isConstantSplat(SplatValue, SplatUndef,
25307 SplatBitSize, HasAnyUndefs))
25310 unsigned ResSize = N1.getValueType().getScalarSizeInBits();
25311 // Make sure the splat matches the mask we expect
25312 if (SplatBitSize > ResSize ||
25313 (SplatValue + 1).exactLogBase2() != (int)SrcSize)
25316 // Make sure the input and output size make sense
25317 if (SrcSize >= ResSize || ResSize % SrcSize)
25320 // We expect a shuffle of the form <0, u, u, u, 1, u, u, u...>
25321 // The number of u's between each two values depends on the ratio between
25322 // the source and dest type.
25323 unsigned ZextRatio = ResSize / SrcSize;
25324 bool IsZext = true;
25325 for (unsigned i = 0; i < SrcType.getVectorNumElements(); ++i) {
25326 if (i % ZextRatio) {
25327 if (Shuffle->getMaskElt(i) > 0) {
25333 if (Shuffle->getMaskElt(i) != (int)(i / ZextRatio)) {
25334 // Expected element number
25344 // Ok, perform the transformation - replace the shuffle with
25345 // a shuffle of the form <0, k, k, k, 1, k, k, k> with zero
25346 // (instead of undef) where the k elements come from the zero vector.
25347 SmallVector<int, 8> Mask;
25348 unsigned NumElems = SrcType.getVectorNumElements();
25349 for (unsigned i = 0; i < NumElems; ++i)
25351 Mask.push_back(NumElems);
25353 Mask.push_back(i / ZextRatio);
25355 SDValue NewShuffle = DAG.getVectorShuffle(Shuffle->getValueType(0), DL,
25356 Shuffle->getOperand(0), DAG.getConstant(0, DL, SrcType), Mask);
25357 return DAG.getBitcast(N0.getValueType(), NewShuffle);
25360 /// If both input operands of a logic op are being cast from floating point
25361 /// types, try to convert this into a floating point logic node to avoid
25362 /// unnecessary moves from SSE to integer registers.
25363 static SDValue convertIntLogicToFPLogic(SDNode *N, SelectionDAG &DAG,
25364 const X86Subtarget *Subtarget) {
25365 unsigned FPOpcode = ISD::DELETED_NODE;
25366 if (N->getOpcode() == ISD::AND)
25367 FPOpcode = X86ISD::FAND;
25368 else if (N->getOpcode() == ISD::OR)
25369 FPOpcode = X86ISD::FOR;
25370 else if (N->getOpcode() == ISD::XOR)
25371 FPOpcode = X86ISD::FXOR;
25373 assert(FPOpcode != ISD::DELETED_NODE &&
25374 "Unexpected input node for FP logic conversion");
25376 EVT VT = N->getValueType(0);
25377 SDValue N0 = N->getOperand(0);
25378 SDValue N1 = N->getOperand(1);
25380 if (N0.getOpcode() == ISD::BITCAST && N1.getOpcode() == ISD::BITCAST &&
25381 ((Subtarget->hasSSE1() && VT == MVT::i32) ||
25382 (Subtarget->hasSSE2() && VT == MVT::i64))) {
25383 SDValue N00 = N0.getOperand(0);
25384 SDValue N10 = N1.getOperand(0);
25385 EVT N00Type = N00.getValueType();
25386 EVT N10Type = N10.getValueType();
25387 if (N00Type.isFloatingPoint() && N10Type.isFloatingPoint()) {
25388 SDValue FPLogic = DAG.getNode(FPOpcode, DL, N00Type, N00, N10);
25389 return DAG.getBitcast(VT, FPLogic);
25395 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
25396 TargetLowering::DAGCombinerInfo &DCI,
25397 const X86Subtarget *Subtarget) {
25398 if (DCI.isBeforeLegalizeOps())
25401 if (SDValue Zext = VectorZextCombine(N, DAG, DCI, Subtarget))
25404 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
25407 if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
25410 EVT VT = N->getValueType(0);
25411 SDValue N0 = N->getOperand(0);
25412 SDValue N1 = N->getOperand(1);
25415 // Create BEXTR instructions
25416 // BEXTR is ((X >> imm) & (2**size-1))
25417 if (VT == MVT::i32 || VT == MVT::i64) {
25418 // Check for BEXTR.
25419 if ((Subtarget->hasBMI() || Subtarget->hasTBM()) &&
25420 (N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::SRL)) {
25421 ConstantSDNode *MaskNode = dyn_cast<ConstantSDNode>(N1);
25422 ConstantSDNode *ShiftNode = dyn_cast<ConstantSDNode>(N0.getOperand(1));
25423 if (MaskNode && ShiftNode) {
25424 uint64_t Mask = MaskNode->getZExtValue();
25425 uint64_t Shift = ShiftNode->getZExtValue();
25426 if (isMask_64(Mask)) {
25427 uint64_t MaskSize = countPopulation(Mask);
25428 if (Shift + MaskSize <= VT.getSizeInBits())
25429 return DAG.getNode(X86ISD::BEXTR, DL, VT, N0.getOperand(0),
25430 DAG.getConstant(Shift | (MaskSize << 8), DL,
25439 // Want to form ANDNP nodes:
25440 // 1) In the hopes of then easily combining them with OR and AND nodes
25441 // to form PBLEND/PSIGN.
25442 // 2) To match ANDN packed intrinsics
25443 if (VT != MVT::v2i64 && VT != MVT::v4i64)
25446 // Check LHS for vnot
25447 if (N0.getOpcode() == ISD::XOR &&
25448 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
25449 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
25450 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
25452 // Check RHS for vnot
25453 if (N1.getOpcode() == ISD::XOR &&
25454 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
25455 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
25456 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
25461 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
25462 TargetLowering::DAGCombinerInfo &DCI,
25463 const X86Subtarget *Subtarget) {
25464 if (DCI.isBeforeLegalizeOps())
25467 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
25470 if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
25473 SDValue N0 = N->getOperand(0);
25474 SDValue N1 = N->getOperand(1);
25475 EVT VT = N->getValueType(0);
25477 // look for psign/blend
25478 if (VT == MVT::v2i64 || VT == MVT::v4i64) {
25479 if (!Subtarget->hasSSSE3() ||
25480 (VT == MVT::v4i64 && !Subtarget->hasInt256()))
25483 // Canonicalize pandn to RHS
25484 if (N0.getOpcode() == X86ISD::ANDNP)
25486 // or (and (m, y), (pandn m, x))
25487 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
25488 SDValue Mask = N1.getOperand(0);
25489 SDValue X = N1.getOperand(1);
25491 if (N0.getOperand(0) == Mask)
25492 Y = N0.getOperand(1);
25493 if (N0.getOperand(1) == Mask)
25494 Y = N0.getOperand(0);
25496 // Check to see if the mask appeared in both the AND and ANDNP and
25500 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
25501 // Look through mask bitcast.
25502 if (Mask.getOpcode() == ISD::BITCAST)
25503 Mask = Mask.getOperand(0);
25504 if (X.getOpcode() == ISD::BITCAST)
25505 X = X.getOperand(0);
25506 if (Y.getOpcode() == ISD::BITCAST)
25507 Y = Y.getOperand(0);
25509 EVT MaskVT = Mask.getValueType();
25511 // Validate that the Mask operand is a vector sra node.
25512 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
25513 // there is no psrai.b
25514 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
25515 unsigned SraAmt = ~0;
25516 if (Mask.getOpcode() == ISD::SRA) {
25517 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Mask.getOperand(1)))
25518 if (auto *AmtConst = AmtBV->getConstantSplatNode())
25519 SraAmt = AmtConst->getZExtValue();
25520 } else if (Mask.getOpcode() == X86ISD::VSRAI) {
25521 SDValue SraC = Mask.getOperand(1);
25522 SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
25524 if ((SraAmt + 1) != EltBits)
25529 // Now we know we at least have a plendvb with the mask val. See if
25530 // we can form a psignb/w/d.
25531 // psign = x.type == y.type == mask.type && y = sub(0, x);
25532 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
25533 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
25534 X.getValueType() == MaskVT && Y.getValueType() == MaskVT) {
25535 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
25536 "Unsupported VT for PSIGN");
25537 Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask.getOperand(0));
25538 return DAG.getBitcast(VT, Mask);
25540 // PBLENDVB only available on SSE 4.1
25541 if (!Subtarget->hasSSE41())
25544 MVT BlendVT = (VT == MVT::v4i64) ? MVT::v32i8 : MVT::v16i8;
25546 X = DAG.getBitcast(BlendVT, X);
25547 Y = DAG.getBitcast(BlendVT, Y);
25548 Mask = DAG.getBitcast(BlendVT, Mask);
25549 Mask = DAG.getNode(ISD::VSELECT, DL, BlendVT, Mask, Y, X);
25550 return DAG.getBitcast(VT, Mask);
25554 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
25557 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
25558 bool OptForSize = DAG.getMachineFunction().getFunction()->optForSize();
25560 // SHLD/SHRD instructions have lower register pressure, but on some
25561 // platforms they have higher latency than the equivalent
25562 // series of shifts/or that would otherwise be generated.
25563 // Don't fold (or (x << c) | (y >> (64 - c))) if SHLD/SHRD instructions
25564 // have higher latencies and we are not optimizing for size.
25565 if (!OptForSize && Subtarget->isSHLDSlow())
25568 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
25570 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
25572 if (!N0.hasOneUse() || !N1.hasOneUse())
25575 SDValue ShAmt0 = N0.getOperand(1);
25576 if (ShAmt0.getValueType() != MVT::i8)
25578 SDValue ShAmt1 = N1.getOperand(1);
25579 if (ShAmt1.getValueType() != MVT::i8)
25581 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
25582 ShAmt0 = ShAmt0.getOperand(0);
25583 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
25584 ShAmt1 = ShAmt1.getOperand(0);
25587 unsigned Opc = X86ISD::SHLD;
25588 SDValue Op0 = N0.getOperand(0);
25589 SDValue Op1 = N1.getOperand(0);
25590 if (ShAmt0.getOpcode() == ISD::SUB) {
25591 Opc = X86ISD::SHRD;
25592 std::swap(Op0, Op1);
25593 std::swap(ShAmt0, ShAmt1);
25596 unsigned Bits = VT.getSizeInBits();
25597 if (ShAmt1.getOpcode() == ISD::SUB) {
25598 SDValue Sum = ShAmt1.getOperand(0);
25599 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
25600 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
25601 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
25602 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
25603 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
25604 return DAG.getNode(Opc, DL, VT,
25606 DAG.getNode(ISD::TRUNCATE, DL,
25609 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
25610 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
25612 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
25613 return DAG.getNode(Opc, DL, VT,
25614 N0.getOperand(0), N1.getOperand(0),
25615 DAG.getNode(ISD::TRUNCATE, DL,
25622 // Generate NEG and CMOV for integer abs.
25623 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
25624 EVT VT = N->getValueType(0);
25626 // Since X86 does not have CMOV for 8-bit integer, we don't convert
25627 // 8-bit integer abs to NEG and CMOV.
25628 if (VT.isInteger() && VT.getSizeInBits() == 8)
25631 SDValue N0 = N->getOperand(0);
25632 SDValue N1 = N->getOperand(1);
25635 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
25636 // and change it to SUB and CMOV.
25637 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
25638 N0.getOpcode() == ISD::ADD &&
25639 N0.getOperand(1) == N1 &&
25640 N1.getOpcode() == ISD::SRA &&
25641 N1.getOperand(0) == N0.getOperand(0))
25642 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
25643 if (Y1C->getAPIntValue() == VT.getSizeInBits()-1) {
25644 // Generate SUB & CMOV.
25645 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32),
25646 DAG.getConstant(0, DL, VT), N0.getOperand(0));
25648 SDValue Ops[] = { N0.getOperand(0), Neg,
25649 DAG.getConstant(X86::COND_GE, DL, MVT::i8),
25650 SDValue(Neg.getNode(), 1) };
25651 return DAG.getNode(X86ISD::CMOV, DL, DAG.getVTList(VT, MVT::Glue), Ops);
25656 // Try to turn tests against the signbit in the form of:
25657 // XOR(TRUNCATE(SRL(X, size(X)-1)), 1)
25660 static SDValue foldXorTruncShiftIntoCmp(SDNode *N, SelectionDAG &DAG) {
25661 // This is only worth doing if the output type is i8.
25662 if (N->getValueType(0) != MVT::i8)
25665 SDValue N0 = N->getOperand(0);
25666 SDValue N1 = N->getOperand(1);
25668 // We should be performing an xor against a truncated shift.
25669 if (N0.getOpcode() != ISD::TRUNCATE || !N0.hasOneUse())
25672 // Make sure we are performing an xor against one.
25673 if (!isOneConstant(N1))
25676 // SetCC on x86 zero extends so only act on this if it's a logical shift.
25677 SDValue Shift = N0.getOperand(0);
25678 if (Shift.getOpcode() != ISD::SRL || !Shift.hasOneUse())
25681 // Make sure we are truncating from one of i16, i32 or i64.
25682 EVT ShiftTy = Shift.getValueType();
25683 if (ShiftTy != MVT::i16 && ShiftTy != MVT::i32 && ShiftTy != MVT::i64)
25686 // Make sure the shift amount extracts the sign bit.
25687 if (!isa<ConstantSDNode>(Shift.getOperand(1)) ||
25688 Shift.getConstantOperandVal(1) != ShiftTy.getSizeInBits() - 1)
25691 // Create a greater-than comparison against -1.
25692 // N.B. Using SETGE against 0 works but we want a canonical looking
25693 // comparison, using SETGT matches up with what TranslateX86CC.
25695 SDValue ShiftOp = Shift.getOperand(0);
25696 EVT ShiftOpTy = ShiftOp.getValueType();
25697 SDValue Cond = DAG.getSetCC(DL, MVT::i8, ShiftOp,
25698 DAG.getConstant(-1, DL, ShiftOpTy), ISD::SETGT);
25702 static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
25703 TargetLowering::DAGCombinerInfo &DCI,
25704 const X86Subtarget *Subtarget) {
25705 if (DCI.isBeforeLegalizeOps())
25708 if (SDValue RV = foldXorTruncShiftIntoCmp(N, DAG))
25711 if (Subtarget->hasCMov())
25712 if (SDValue RV = performIntegerAbsCombine(N, DAG))
25715 if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
25721 /// This function detects the AVG pattern between vectors of unsigned i8/i16,
25722 /// which is c = (a + b + 1) / 2, and replace this operation with the efficient
25723 /// X86ISD::AVG instruction.
25724 static SDValue detectAVGPattern(SDValue In, EVT VT, SelectionDAG &DAG,
25725 const X86Subtarget *Subtarget, SDLoc DL) {
25726 if (!VT.isVector() || !VT.isSimple())
25728 EVT InVT = In.getValueType();
25729 unsigned NumElems = VT.getVectorNumElements();
25731 EVT ScalarVT = VT.getVectorElementType();
25732 if (!((ScalarVT == MVT::i8 || ScalarVT == MVT::i16) &&
25733 isPowerOf2_32(NumElems)))
25736 // InScalarVT is the intermediate type in AVG pattern and it should be greater
25737 // than the original input type (i8/i16).
25738 EVT InScalarVT = InVT.getVectorElementType();
25739 if (InScalarVT.getSizeInBits() <= ScalarVT.getSizeInBits())
25742 if (Subtarget->hasAVX512()) {
25743 if (VT.getSizeInBits() > 512)
25745 } else if (Subtarget->hasAVX2()) {
25746 if (VT.getSizeInBits() > 256)
25749 if (VT.getSizeInBits() > 128)
25753 // Detect the following pattern:
25755 // %1 = zext <N x i8> %a to <N x i32>
25756 // %2 = zext <N x i8> %b to <N x i32>
25757 // %3 = add nuw nsw <N x i32> %1, <i32 1 x N>
25758 // %4 = add nuw nsw <N x i32> %3, %2
25759 // %5 = lshr <N x i32> %N, <i32 1 x N>
25760 // %6 = trunc <N x i32> %5 to <N x i8>
25762 // In AVX512, the last instruction can also be a trunc store.
25764 if (In.getOpcode() != ISD::SRL)
25767 // A lambda checking the given SDValue is a constant vector and each element
25768 // is in the range [Min, Max].
25769 auto IsConstVectorInRange = [](SDValue V, unsigned Min, unsigned Max) {
25770 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(V);
25771 if (!BV || !BV->isConstant())
25773 for (unsigned i = 0, e = V.getNumOperands(); i < e; i++) {
25774 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V.getOperand(i));
25777 uint64_t Val = C->getZExtValue();
25778 if (Val < Min || Val > Max)
25784 // Check if each element of the vector is left-shifted by one.
25785 auto LHS = In.getOperand(0);
25786 auto RHS = In.getOperand(1);
25787 if (!IsConstVectorInRange(RHS, 1, 1))
25789 if (LHS.getOpcode() != ISD::ADD)
25792 // Detect a pattern of a + b + 1 where the order doesn't matter.
25793 SDValue Operands[3];
25794 Operands[0] = LHS.getOperand(0);
25795 Operands[1] = LHS.getOperand(1);
25797 // Take care of the case when one of the operands is a constant vector whose
25798 // element is in the range [1, 256].
25799 if (IsConstVectorInRange(Operands[1], 1, ScalarVT == MVT::i8 ? 256 : 65536) &&
25800 Operands[0].getOpcode() == ISD::ZERO_EXTEND &&
25801 Operands[0].getOperand(0).getValueType() == VT) {
25802 // The pattern is detected. Subtract one from the constant vector, then
25803 // demote it and emit X86ISD::AVG instruction.
25804 SDValue One = DAG.getConstant(1, DL, InScalarVT);
25805 SDValue Ones = DAG.getNode(ISD::BUILD_VECTOR, DL, InVT,
25806 SmallVector<SDValue, 8>(NumElems, One));
25807 Operands[1] = DAG.getNode(ISD::SUB, DL, InVT, Operands[1], Ones);
25808 Operands[1] = DAG.getNode(ISD::TRUNCATE, DL, VT, Operands[1]);
25809 return DAG.getNode(X86ISD::AVG, DL, VT, Operands[0].getOperand(0),
25813 if (Operands[0].getOpcode() == ISD::ADD)
25814 std::swap(Operands[0], Operands[1]);
25815 else if (Operands[1].getOpcode() != ISD::ADD)
25817 Operands[2] = Operands[1].getOperand(0);
25818 Operands[1] = Operands[1].getOperand(1);
25820 // Now we have three operands of two additions. Check that one of them is a
25821 // constant vector with ones, and the other two are promoted from i8/i16.
25822 for (int i = 0; i < 3; ++i) {
25823 if (!IsConstVectorInRange(Operands[i], 1, 1))
25825 std::swap(Operands[i], Operands[2]);
25827 // Check if Operands[0] and Operands[1] are results of type promotion.
25828 for (int j = 0; j < 2; ++j)
25829 if (Operands[j].getOpcode() != ISD::ZERO_EXTEND ||
25830 Operands[j].getOperand(0).getValueType() != VT)
25833 // The pattern is detected, emit X86ISD::AVG instruction.
25834 return DAG.getNode(X86ISD::AVG, DL, VT, Operands[0].getOperand(0),
25835 Operands[1].getOperand(0));
25841 static SDValue PerformTRUNCATECombine(SDNode *N, SelectionDAG &DAG,
25842 const X86Subtarget *Subtarget) {
25843 return detectAVGPattern(N->getOperand(0), N->getValueType(0), DAG, Subtarget,
25847 /// PerformLOADCombine - Do target-specific dag combines on LOAD nodes.
25848 static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
25849 TargetLowering::DAGCombinerInfo &DCI,
25850 const X86Subtarget *Subtarget) {
25851 LoadSDNode *Ld = cast<LoadSDNode>(N);
25852 EVT RegVT = Ld->getValueType(0);
25853 EVT MemVT = Ld->getMemoryVT();
25855 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25857 // For chips with slow 32-byte unaligned loads, break the 32-byte operation
25858 // into two 16-byte operations.
25859 ISD::LoadExtType Ext = Ld->getExtensionType();
25861 unsigned AddressSpace = Ld->getAddressSpace();
25862 unsigned Alignment = Ld->getAlignment();
25863 if (RegVT.is256BitVector() && !DCI.isBeforeLegalizeOps() &&
25864 Ext == ISD::NON_EXTLOAD &&
25865 TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), RegVT,
25866 AddressSpace, Alignment, &Fast) && !Fast) {
25867 unsigned NumElems = RegVT.getVectorNumElements();
25871 SDValue Ptr = Ld->getBasePtr();
25872 SDValue Increment =
25873 DAG.getConstant(16, dl, TLI.getPointerTy(DAG.getDataLayout()));
25875 EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
25877 SDValue Load1 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
25878 Ld->getPointerInfo(), Ld->isVolatile(),
25879 Ld->isNonTemporal(), Ld->isInvariant(),
25881 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
25882 SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
25883 Ld->getPointerInfo(), Ld->isVolatile(),
25884 Ld->isNonTemporal(), Ld->isInvariant(),
25885 std::min(16U, Alignment));
25886 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
25888 Load2.getValue(1));
25890 SDValue NewVec = DAG.getUNDEF(RegVT);
25891 NewVec = Insert128BitVector(NewVec, Load1, 0, DAG, dl);
25892 NewVec = Insert128BitVector(NewVec, Load2, NumElems/2, DAG, dl);
25893 return DCI.CombineTo(N, NewVec, TF, true);
25899 /// PerformMLOADCombine - Resolve extending loads
25900 static SDValue PerformMLOADCombine(SDNode *N, SelectionDAG &DAG,
25901 TargetLowering::DAGCombinerInfo &DCI,
25902 const X86Subtarget *Subtarget) {
25903 MaskedLoadSDNode *Mld = cast<MaskedLoadSDNode>(N);
25904 if (Mld->getExtensionType() != ISD::SEXTLOAD)
25907 EVT VT = Mld->getValueType(0);
25908 unsigned NumElems = VT.getVectorNumElements();
25909 EVT LdVT = Mld->getMemoryVT();
25912 assert(LdVT != VT && "Cannot extend to the same type");
25913 unsigned ToSz = VT.getVectorElementType().getSizeInBits();
25914 unsigned FromSz = LdVT.getVectorElementType().getSizeInBits();
25915 // From, To sizes and ElemCount must be pow of two
25916 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
25917 "Unexpected size for extending masked load");
25919 unsigned SizeRatio = ToSz / FromSz;
25920 assert(SizeRatio * NumElems * FromSz == VT.getSizeInBits());
25922 // Create a type on which we perform the shuffle
25923 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
25924 LdVT.getScalarType(), NumElems*SizeRatio);
25925 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
25927 // Convert Src0 value
25928 SDValue WideSrc0 = DAG.getBitcast(WideVecVT, Mld->getSrc0());
25929 if (Mld->getSrc0().getOpcode() != ISD::UNDEF) {
25930 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
25931 for (unsigned i = 0; i != NumElems; ++i)
25932 ShuffleVec[i] = i * SizeRatio;
25934 // Can't shuffle using an illegal type.
25935 assert(DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) &&
25936 "WideVecVT should be legal");
25937 WideSrc0 = DAG.getVectorShuffle(WideVecVT, dl, WideSrc0,
25938 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
25940 // Prepare the new mask
25942 SDValue Mask = Mld->getMask();
25943 if (Mask.getValueType() == VT) {
25944 // Mask and original value have the same type
25945 NewMask = DAG.getBitcast(WideVecVT, Mask);
25946 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
25947 for (unsigned i = 0; i != NumElems; ++i)
25948 ShuffleVec[i] = i * SizeRatio;
25949 for (unsigned i = NumElems; i != NumElems * SizeRatio; ++i)
25950 ShuffleVec[i] = NumElems * SizeRatio;
25951 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
25952 DAG.getConstant(0, dl, WideVecVT),
25956 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
25957 unsigned WidenNumElts = NumElems*SizeRatio;
25958 unsigned MaskNumElts = VT.getVectorNumElements();
25959 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
25962 unsigned NumConcat = WidenNumElts / MaskNumElts;
25963 SmallVector<SDValue, 16> Ops(NumConcat);
25964 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
25966 for (unsigned i = 1; i != NumConcat; ++i)
25969 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
25972 SDValue WideLd = DAG.getMaskedLoad(WideVecVT, dl, Mld->getChain(),
25973 Mld->getBasePtr(), NewMask, WideSrc0,
25974 Mld->getMemoryVT(), Mld->getMemOperand(),
25976 SDValue NewVec = DAG.getNode(X86ISD::VSEXT, dl, VT, WideLd);
25977 return DCI.CombineTo(N, NewVec, WideLd.getValue(1), true);
25979 /// PerformMSTORECombine - Resolve truncating stores
25980 static SDValue PerformMSTORECombine(SDNode *N, SelectionDAG &DAG,
25981 const X86Subtarget *Subtarget) {
25982 MaskedStoreSDNode *Mst = cast<MaskedStoreSDNode>(N);
25983 if (!Mst->isTruncatingStore())
25986 EVT VT = Mst->getValue().getValueType();
25987 unsigned NumElems = VT.getVectorNumElements();
25988 EVT StVT = Mst->getMemoryVT();
25991 assert(StVT != VT && "Cannot truncate to the same type");
25992 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
25993 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
25995 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25997 // The truncating store is legal in some cases. For example
25998 // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
25999 // are designated for truncate store.
26000 // In this case we don't need any further transformations.
26001 if (TLI.isTruncStoreLegal(VT, StVT))
26004 // From, To sizes and ElemCount must be pow of two
26005 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
26006 "Unexpected size for truncating masked store");
26007 // We are going to use the original vector elt for storing.
26008 // Accumulated smaller vector elements must be a multiple of the store size.
26009 assert (((NumElems * FromSz) % ToSz) == 0 &&
26010 "Unexpected ratio for truncating masked store");
26012 unsigned SizeRatio = FromSz / ToSz;
26013 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
26015 // Create a type on which we perform the shuffle
26016 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
26017 StVT.getScalarType(), NumElems*SizeRatio);
26019 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
26021 SDValue WideVec = DAG.getBitcast(WideVecVT, Mst->getValue());
26022 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
26023 for (unsigned i = 0; i != NumElems; ++i)
26024 ShuffleVec[i] = i * SizeRatio;
26026 // Can't shuffle using an illegal type.
26027 assert(DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) &&
26028 "WideVecVT should be legal");
26030 SDValue TruncatedVal = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
26031 DAG.getUNDEF(WideVecVT),
26035 SDValue Mask = Mst->getMask();
26036 if (Mask.getValueType() == VT) {
26037 // Mask and original value have the same type
26038 NewMask = DAG.getBitcast(WideVecVT, Mask);
26039 for (unsigned i = 0; i != NumElems; ++i)
26040 ShuffleVec[i] = i * SizeRatio;
26041 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
26042 ShuffleVec[i] = NumElems*SizeRatio;
26043 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
26044 DAG.getConstant(0, dl, WideVecVT),
26048 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
26049 unsigned WidenNumElts = NumElems*SizeRatio;
26050 unsigned MaskNumElts = VT.getVectorNumElements();
26051 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
26054 unsigned NumConcat = WidenNumElts / MaskNumElts;
26055 SmallVector<SDValue, 16> Ops(NumConcat);
26056 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
26058 for (unsigned i = 1; i != NumConcat; ++i)
26061 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
26064 return DAG.getMaskedStore(Mst->getChain(), dl, TruncatedVal,
26065 Mst->getBasePtr(), NewMask, StVT,
26066 Mst->getMemOperand(), false);
26068 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
26069 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
26070 const X86Subtarget *Subtarget) {
26071 StoreSDNode *St = cast<StoreSDNode>(N);
26072 EVT VT = St->getValue().getValueType();
26073 EVT StVT = St->getMemoryVT();
26075 SDValue StoredVal = St->getOperand(1);
26076 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
26078 // If we are saving a concatenation of two XMM registers and 32-byte stores
26079 // are slow, such as on Sandy Bridge, perform two 16-byte stores.
26081 unsigned AddressSpace = St->getAddressSpace();
26082 unsigned Alignment = St->getAlignment();
26083 if (VT.is256BitVector() && StVT == VT &&
26084 TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), VT,
26085 AddressSpace, Alignment, &Fast) && !Fast) {
26086 unsigned NumElems = VT.getVectorNumElements();
26090 SDValue Value0 = Extract128BitVector(StoredVal, 0, DAG, dl);
26091 SDValue Value1 = Extract128BitVector(StoredVal, NumElems/2, DAG, dl);
26094 DAG.getConstant(16, dl, TLI.getPointerTy(DAG.getDataLayout()));
26095 SDValue Ptr0 = St->getBasePtr();
26096 SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
26098 SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
26099 St->getPointerInfo(), St->isVolatile(),
26100 St->isNonTemporal(), Alignment);
26101 SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
26102 St->getPointerInfo(), St->isVolatile(),
26103 St->isNonTemporal(),
26104 std::min(16U, Alignment));
26105 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
26108 // Optimize trunc store (of multiple scalars) to shuffle and store.
26109 // First, pack all of the elements in one place. Next, store to memory
26110 // in fewer chunks.
26111 if (St->isTruncatingStore() && VT.isVector()) {
26112 // Check if we can detect an AVG pattern from the truncation. If yes,
26113 // replace the trunc store by a normal store with the result of X86ISD::AVG
26116 detectAVGPattern(St->getValue(), St->getMemoryVT(), DAG, Subtarget, dl);
26118 return DAG.getStore(St->getChain(), dl, Avg, St->getBasePtr(),
26119 St->getPointerInfo(), St->isVolatile(),
26120 St->isNonTemporal(), St->getAlignment());
26122 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
26123 unsigned NumElems = VT.getVectorNumElements();
26124 assert(StVT != VT && "Cannot truncate to the same type");
26125 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
26126 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
26128 // The truncating store is legal in some cases. For example
26129 // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
26130 // are designated for truncate store.
26131 // In this case we don't need any further transformations.
26132 if (TLI.isTruncStoreLegal(VT, StVT))
26135 // From, To sizes and ElemCount must be pow of two
26136 if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
26137 // We are going to use the original vector elt for storing.
26138 // Accumulated smaller vector elements must be a multiple of the store size.
26139 if (0 != (NumElems * FromSz) % ToSz) return SDValue();
26141 unsigned SizeRatio = FromSz / ToSz;
26143 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
26145 // Create a type on which we perform the shuffle
26146 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
26147 StVT.getScalarType(), NumElems*SizeRatio);
26149 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
26151 SDValue WideVec = DAG.getBitcast(WideVecVT, St->getValue());
26152 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
26153 for (unsigned i = 0; i != NumElems; ++i)
26154 ShuffleVec[i] = i * SizeRatio;
26156 // Can't shuffle using an illegal type.
26157 if (!TLI.isTypeLegal(WideVecVT))
26160 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
26161 DAG.getUNDEF(WideVecVT),
26163 // At this point all of the data is stored at the bottom of the
26164 // register. We now need to save it to mem.
26166 // Find the largest store unit
26167 MVT StoreType = MVT::i8;
26168 for (MVT Tp : MVT::integer_valuetypes()) {
26169 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToSz)
26173 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
26174 if (TLI.isTypeLegal(MVT::f64) && StoreType.getSizeInBits() < 64 &&
26175 (64 <= NumElems * ToSz))
26176 StoreType = MVT::f64;
26178 // Bitcast the original vector into a vector of store-size units
26179 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
26180 StoreType, VT.getSizeInBits()/StoreType.getSizeInBits());
26181 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
26182 SDValue ShuffWide = DAG.getBitcast(StoreVecVT, Shuff);
26183 SmallVector<SDValue, 8> Chains;
26184 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits() / 8, dl,
26185 TLI.getPointerTy(DAG.getDataLayout()));
26186 SDValue Ptr = St->getBasePtr();
26188 // Perform one or more big stores into memory.
26189 for (unsigned i=0, e=(ToSz*NumElems)/StoreType.getSizeInBits(); i!=e; ++i) {
26190 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
26191 StoreType, ShuffWide,
26192 DAG.getIntPtrConstant(i, dl));
26193 SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr,
26194 St->getPointerInfo(), St->isVolatile(),
26195 St->isNonTemporal(), St->getAlignment());
26196 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
26197 Chains.push_back(Ch);
26200 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
26203 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
26204 // the FP state in cases where an emms may be missing.
26205 // A preferable solution to the general problem is to figure out the right
26206 // places to insert EMMS. This qualifies as a quick hack.
26208 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
26209 if (VT.getSizeInBits() != 64)
26212 const Function *F = DAG.getMachineFunction().getFunction();
26213 bool NoImplicitFloatOps = F->hasFnAttribute(Attribute::NoImplicitFloat);
26215 !Subtarget->useSoftFloat() && !NoImplicitFloatOps && Subtarget->hasSSE2();
26216 if ((VT.isVector() ||
26217 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
26218 isa<LoadSDNode>(St->getValue()) &&
26219 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
26220 St->getChain().hasOneUse() && !St->isVolatile()) {
26221 SDNode* LdVal = St->getValue().getNode();
26222 LoadSDNode *Ld = nullptr;
26223 int TokenFactorIndex = -1;
26224 SmallVector<SDValue, 8> Ops;
26225 SDNode* ChainVal = St->getChain().getNode();
26226 // Must be a store of a load. We currently handle two cases: the load
26227 // is a direct child, and it's under an intervening TokenFactor. It is
26228 // possible to dig deeper under nested TokenFactors.
26229 if (ChainVal == LdVal)
26230 Ld = cast<LoadSDNode>(St->getChain());
26231 else if (St->getValue().hasOneUse() &&
26232 ChainVal->getOpcode() == ISD::TokenFactor) {
26233 for (unsigned i = 0, e = ChainVal->getNumOperands(); i != e; ++i) {
26234 if (ChainVal->getOperand(i).getNode() == LdVal) {
26235 TokenFactorIndex = i;
26236 Ld = cast<LoadSDNode>(St->getValue());
26238 Ops.push_back(ChainVal->getOperand(i));
26242 if (!Ld || !ISD::isNormalLoad(Ld))
26245 // If this is not the MMX case, i.e. we are just turning i64 load/store
26246 // into f64 load/store, avoid the transformation if there are multiple
26247 // uses of the loaded value.
26248 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
26253 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
26254 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
26256 if (Subtarget->is64Bit() || F64IsLegal) {
26257 MVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
26258 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
26259 Ld->getPointerInfo(), Ld->isVolatile(),
26260 Ld->isNonTemporal(), Ld->isInvariant(),
26261 Ld->getAlignment());
26262 SDValue NewChain = NewLd.getValue(1);
26263 if (TokenFactorIndex != -1) {
26264 Ops.push_back(NewChain);
26265 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
26267 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
26268 St->getPointerInfo(),
26269 St->isVolatile(), St->isNonTemporal(),
26270 St->getAlignment());
26273 // Otherwise, lower to two pairs of 32-bit loads / stores.
26274 SDValue LoAddr = Ld->getBasePtr();
26275 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
26276 DAG.getConstant(4, LdDL, MVT::i32));
26278 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
26279 Ld->getPointerInfo(),
26280 Ld->isVolatile(), Ld->isNonTemporal(),
26281 Ld->isInvariant(), Ld->getAlignment());
26282 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
26283 Ld->getPointerInfo().getWithOffset(4),
26284 Ld->isVolatile(), Ld->isNonTemporal(),
26286 MinAlign(Ld->getAlignment(), 4));
26288 SDValue NewChain = LoLd.getValue(1);
26289 if (TokenFactorIndex != -1) {
26290 Ops.push_back(LoLd);
26291 Ops.push_back(HiLd);
26292 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
26295 LoAddr = St->getBasePtr();
26296 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
26297 DAG.getConstant(4, StDL, MVT::i32));
26299 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
26300 St->getPointerInfo(),
26301 St->isVolatile(), St->isNonTemporal(),
26302 St->getAlignment());
26303 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
26304 St->getPointerInfo().getWithOffset(4),
26306 St->isNonTemporal(),
26307 MinAlign(St->getAlignment(), 4));
26308 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
26311 // This is similar to the above case, but here we handle a scalar 64-bit
26312 // integer store that is extracted from a vector on a 32-bit target.
26313 // If we have SSE2, then we can treat it like a floating-point double
26314 // to get past legalization. The execution dependencies fixup pass will
26315 // choose the optimal machine instruction for the store if this really is
26316 // an integer or v2f32 rather than an f64.
26317 if (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit() &&
26318 St->getOperand(1).getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
26319 SDValue OldExtract = St->getOperand(1);
26320 SDValue ExtOp0 = OldExtract.getOperand(0);
26321 unsigned VecSize = ExtOp0.getValueSizeInBits();
26322 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, VecSize / 64);
26323 SDValue BitCast = DAG.getBitcast(VecVT, ExtOp0);
26324 SDValue NewExtract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
26325 BitCast, OldExtract.getOperand(1));
26326 return DAG.getStore(St->getChain(), dl, NewExtract, St->getBasePtr(),
26327 St->getPointerInfo(), St->isVolatile(),
26328 St->isNonTemporal(), St->getAlignment());
26334 /// Return 'true' if this vector operation is "horizontal"
26335 /// and return the operands for the horizontal operation in LHS and RHS. A
26336 /// horizontal operation performs the binary operation on successive elements
26337 /// of its first operand, then on successive elements of its second operand,
26338 /// returning the resulting values in a vector. For example, if
26339 /// A = < float a0, float a1, float a2, float a3 >
26341 /// B = < float b0, float b1, float b2, float b3 >
26342 /// then the result of doing a horizontal operation on A and B is
26343 /// A horizontal-op B = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >.
26344 /// In short, LHS and RHS are inspected to see if LHS op RHS is of the form
26345 /// A horizontal-op B, for some already available A and B, and if so then LHS is
26346 /// set to A, RHS to B, and the routine returns 'true'.
26347 /// Note that the binary operation should have the property that if one of the
26348 /// operands is UNDEF then the result is UNDEF.
26349 static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool IsCommutative) {
26350 // Look for the following pattern: if
26351 // A = < float a0, float a1, float a2, float a3 >
26352 // B = < float b0, float b1, float b2, float b3 >
26354 // LHS = VECTOR_SHUFFLE A, B, <0, 2, 4, 6>
26355 // RHS = VECTOR_SHUFFLE A, B, <1, 3, 5, 7>
26356 // then LHS op RHS = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >
26357 // which is A horizontal-op B.
26359 // At least one of the operands should be a vector shuffle.
26360 if (LHS.getOpcode() != ISD::VECTOR_SHUFFLE &&
26361 RHS.getOpcode() != ISD::VECTOR_SHUFFLE)
26364 MVT VT = LHS.getSimpleValueType();
26366 assert((VT.is128BitVector() || VT.is256BitVector()) &&
26367 "Unsupported vector type for horizontal add/sub");
26369 // Handle 128 and 256-bit vector lengths. AVX defines horizontal add/sub to
26370 // operate independently on 128-bit lanes.
26371 unsigned NumElts = VT.getVectorNumElements();
26372 unsigned NumLanes = VT.getSizeInBits()/128;
26373 unsigned NumLaneElts = NumElts / NumLanes;
26374 assert((NumLaneElts % 2 == 0) &&
26375 "Vector type should have an even number of elements in each lane");
26376 unsigned HalfLaneElts = NumLaneElts/2;
26378 // View LHS in the form
26379 // LHS = VECTOR_SHUFFLE A, B, LMask
26380 // If LHS is not a shuffle then pretend it is the shuffle
26381 // LHS = VECTOR_SHUFFLE LHS, undef, <0, 1, ..., N-1>
26382 // NOTE: in what follows a default initialized SDValue represents an UNDEF of
26385 SmallVector<int, 16> LMask(NumElts);
26386 if (LHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
26387 if (LHS.getOperand(0).getOpcode() != ISD::UNDEF)
26388 A = LHS.getOperand(0);
26389 if (LHS.getOperand(1).getOpcode() != ISD::UNDEF)
26390 B = LHS.getOperand(1);
26391 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(LHS.getNode())->getMask();
26392 std::copy(Mask.begin(), Mask.end(), LMask.begin());
26394 if (LHS.getOpcode() != ISD::UNDEF)
26396 for (unsigned i = 0; i != NumElts; ++i)
26400 // Likewise, view RHS in the form
26401 // RHS = VECTOR_SHUFFLE C, D, RMask
26403 SmallVector<int, 16> RMask(NumElts);
26404 if (RHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
26405 if (RHS.getOperand(0).getOpcode() != ISD::UNDEF)
26406 C = RHS.getOperand(0);
26407 if (RHS.getOperand(1).getOpcode() != ISD::UNDEF)
26408 D = RHS.getOperand(1);
26409 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(RHS.getNode())->getMask();
26410 std::copy(Mask.begin(), Mask.end(), RMask.begin());
26412 if (RHS.getOpcode() != ISD::UNDEF)
26414 for (unsigned i = 0; i != NumElts; ++i)
26418 // Check that the shuffles are both shuffling the same vectors.
26419 if (!(A == C && B == D) && !(A == D && B == C))
26422 // If everything is UNDEF then bail out: it would be better to fold to UNDEF.
26423 if (!A.getNode() && !B.getNode())
26426 // If A and B occur in reverse order in RHS, then "swap" them (which means
26427 // rewriting the mask).
26429 ShuffleVectorSDNode::commuteMask(RMask);
26431 // At this point LHS and RHS are equivalent to
26432 // LHS = VECTOR_SHUFFLE A, B, LMask
26433 // RHS = VECTOR_SHUFFLE A, B, RMask
26434 // Check that the masks correspond to performing a horizontal operation.
26435 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
26436 for (unsigned i = 0; i != NumLaneElts; ++i) {
26437 int LIdx = LMask[i+l], RIdx = RMask[i+l];
26439 // Ignore any UNDEF components.
26440 if (LIdx < 0 || RIdx < 0 ||
26441 (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) ||
26442 (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts)))
26445 // Check that successive elements are being operated on. If not, this is
26446 // not a horizontal operation.
26447 unsigned Src = (i/HalfLaneElts); // each lane is split between srcs
26448 int Index = 2*(i%HalfLaneElts) + NumElts*Src + l;
26449 if (!(LIdx == Index && RIdx == Index + 1) &&
26450 !(IsCommutative && LIdx == Index + 1 && RIdx == Index))
26455 LHS = A.getNode() ? A : B; // If A is 'UNDEF', use B for it.
26456 RHS = B.getNode() ? B : A; // If B is 'UNDEF', use A for it.
26460 /// Do target-specific dag combines on floating point adds.
26461 static SDValue PerformFADDCombine(SDNode *N, SelectionDAG &DAG,
26462 const X86Subtarget *Subtarget) {
26463 EVT VT = N->getValueType(0);
26464 SDValue LHS = N->getOperand(0);
26465 SDValue RHS = N->getOperand(1);
26467 // Try to synthesize horizontal adds from adds of shuffles.
26468 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
26469 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
26470 isHorizontalBinOp(LHS, RHS, true))
26471 return DAG.getNode(X86ISD::FHADD, SDLoc(N), VT, LHS, RHS);
26475 /// Do target-specific dag combines on floating point subs.
26476 static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG,
26477 const X86Subtarget *Subtarget) {
26478 EVT VT = N->getValueType(0);
26479 SDValue LHS = N->getOperand(0);
26480 SDValue RHS = N->getOperand(1);
26482 // Try to synthesize horizontal subs from subs of shuffles.
26483 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
26484 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
26485 isHorizontalBinOp(LHS, RHS, false))
26486 return DAG.getNode(X86ISD::FHSUB, SDLoc(N), VT, LHS, RHS);
26490 /// Do target-specific dag combines on floating point negations.
26491 static SDValue PerformFNEGCombine(SDNode *N, SelectionDAG &DAG,
26492 const X86Subtarget *Subtarget) {
26493 EVT VT = N->getValueType(0);
26494 EVT SVT = VT.getScalarType();
26495 SDValue Arg = N->getOperand(0);
26498 // Let legalize expand this if it isn't a legal type yet.
26499 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
26502 // If we're negating a FMUL node on a target with FMA, then we can avoid the
26503 // use of a constant by performing (-0 - A*B) instead.
26504 // FIXME: Check rounding control flags as well once it becomes available.
26505 if (Arg.getOpcode() == ISD::FMUL && (SVT == MVT::f32 || SVT == MVT::f64) &&
26506 Arg->getFlags()->hasNoSignedZeros() && Subtarget->hasAnyFMA()) {
26507 SDValue Zero = DAG.getConstantFP(0.0, DL, VT);
26508 return DAG.getNode(X86ISD::FNMSUB, DL, VT, Arg.getOperand(0),
26509 Arg.getOperand(1), Zero);
26512 // If we're negating a FMA node, then we can adjust the
26513 // instruction to include the extra negation.
26514 if (Arg.hasOneUse()) {
26515 switch (Arg.getOpcode()) {
26516 case X86ISD::FMADD:
26517 return DAG.getNode(X86ISD::FNMSUB, DL, VT, Arg.getOperand(0),
26518 Arg.getOperand(1), Arg.getOperand(2));
26519 case X86ISD::FMSUB:
26520 return DAG.getNode(X86ISD::FNMADD, DL, VT, Arg.getOperand(0),
26521 Arg.getOperand(1), Arg.getOperand(2));
26522 case X86ISD::FNMADD:
26523 return DAG.getNode(X86ISD::FMSUB, DL, VT, Arg.getOperand(0),
26524 Arg.getOperand(1), Arg.getOperand(2));
26525 case X86ISD::FNMSUB:
26526 return DAG.getNode(X86ISD::FMADD, DL, VT, Arg.getOperand(0),
26527 Arg.getOperand(1), Arg.getOperand(2));
26533 static SDValue lowerX86FPLogicOp(SDNode *N, SelectionDAG &DAG,
26534 const X86Subtarget *Subtarget) {
26535 EVT VT = N->getValueType(0);
26536 if (VT.is512BitVector() && !Subtarget->hasDQI()) {
26537 // VXORPS, VORPS, VANDPS, VANDNPS are supported only under DQ extention.
26538 // These logic operations may be executed in the integer domain.
26540 MVT IntScalar = MVT::getIntegerVT(VT.getScalarSizeInBits());
26541 MVT IntVT = MVT::getVectorVT(IntScalar, VT.getVectorNumElements());
26543 SDValue Op0 = DAG.getNode(ISD::BITCAST, dl, IntVT, N->getOperand(0));
26544 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, IntVT, N->getOperand(1));
26545 unsigned IntOpcode = 0;
26546 switch (N->getOpcode()) {
26547 default: llvm_unreachable("Unexpected FP logic op");
26548 case X86ISD::FOR: IntOpcode = ISD::OR; break;
26549 case X86ISD::FXOR: IntOpcode = ISD::XOR; break;
26550 case X86ISD::FAND: IntOpcode = ISD::AND; break;
26551 case X86ISD::FANDN: IntOpcode = X86ISD::ANDNP; break;
26553 SDValue IntOp = DAG.getNode(IntOpcode, dl, IntVT, Op0, Op1);
26554 return DAG.getNode(ISD::BITCAST, dl, VT, IntOp);
26558 /// Do target-specific dag combines on X86ISD::FOR and X86ISD::FXOR nodes.
26559 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG,
26560 const X86Subtarget *Subtarget) {
26561 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
26563 // F[X]OR(0.0, x) -> x
26564 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
26565 if (C->getValueAPF().isPosZero())
26566 return N->getOperand(1);
26568 // F[X]OR(x, 0.0) -> x
26569 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
26570 if (C->getValueAPF().isPosZero())
26571 return N->getOperand(0);
26573 return lowerX86FPLogicOp(N, DAG, Subtarget);
26576 /// Do target-specific dag combines on X86ISD::FMIN and X86ISD::FMAX nodes.
26577 static SDValue PerformFMinFMaxCombine(SDNode *N, SelectionDAG &DAG) {
26578 assert(N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD::FMAX);
26580 // Only perform optimizations if UnsafeMath is used.
26581 if (!DAG.getTarget().Options.UnsafeFPMath)
26584 // If we run in unsafe-math mode, then convert the FMAX and FMIN nodes
26585 // into FMINC and FMAXC, which are Commutative operations.
26586 unsigned NewOp = 0;
26587 switch (N->getOpcode()) {
26588 default: llvm_unreachable("unknown opcode");
26589 case X86ISD::FMIN: NewOp = X86ISD::FMINC; break;
26590 case X86ISD::FMAX: NewOp = X86ISD::FMAXC; break;
26593 return DAG.getNode(NewOp, SDLoc(N), N->getValueType(0),
26594 N->getOperand(0), N->getOperand(1));
26597 /// Do target-specific dag combines on X86ISD::FAND nodes.
26598 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG,
26599 const X86Subtarget *Subtarget) {
26600 // FAND(0.0, x) -> 0.0
26601 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
26602 if (C->getValueAPF().isPosZero())
26603 return N->getOperand(0);
26605 // FAND(x, 0.0) -> 0.0
26606 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
26607 if (C->getValueAPF().isPosZero())
26608 return N->getOperand(1);
26610 return lowerX86FPLogicOp(N, DAG, Subtarget);
26613 /// Do target-specific dag combines on X86ISD::FANDN nodes
26614 static SDValue PerformFANDNCombine(SDNode *N, SelectionDAG &DAG,
26615 const X86Subtarget *Subtarget) {
26616 // FANDN(0.0, x) -> x
26617 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
26618 if (C->getValueAPF().isPosZero())
26619 return N->getOperand(1);
26621 // FANDN(x, 0.0) -> 0.0
26622 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
26623 if (C->getValueAPF().isPosZero())
26624 return N->getOperand(1);
26626 return lowerX86FPLogicOp(N, DAG, Subtarget);
26629 static SDValue PerformBTCombine(SDNode *N,
26631 TargetLowering::DAGCombinerInfo &DCI) {
26632 // BT ignores high bits in the bit index operand.
26633 SDValue Op1 = N->getOperand(1);
26634 if (Op1.hasOneUse()) {
26635 unsigned BitWidth = Op1.getValueSizeInBits();
26636 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
26637 APInt KnownZero, KnownOne;
26638 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
26639 !DCI.isBeforeLegalizeOps());
26640 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
26641 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
26642 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
26643 DCI.CommitTargetLoweringOpt(TLO);
26648 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
26649 SDValue Op = N->getOperand(0);
26650 if (Op.getOpcode() == ISD::BITCAST)
26651 Op = Op.getOperand(0);
26652 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
26653 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
26654 VT.getVectorElementType().getSizeInBits() ==
26655 OpVT.getVectorElementType().getSizeInBits()) {
26656 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
26661 static SDValue PerformSIGN_EXTEND_INREGCombine(SDNode *N, SelectionDAG &DAG,
26662 const X86Subtarget *Subtarget) {
26663 EVT VT = N->getValueType(0);
26664 if (!VT.isVector())
26667 SDValue N0 = N->getOperand(0);
26668 SDValue N1 = N->getOperand(1);
26669 EVT ExtraVT = cast<VTSDNode>(N1)->getVT();
26672 // The SIGN_EXTEND_INREG to v4i64 is expensive operation on the
26673 // both SSE and AVX2 since there is no sign-extended shift right
26674 // operation on a vector with 64-bit elements.
26675 //(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) ->
26676 // (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT)))
26677 if (VT == MVT::v4i64 && (N0.getOpcode() == ISD::ANY_EXTEND ||
26678 N0.getOpcode() == ISD::SIGN_EXTEND)) {
26679 SDValue N00 = N0.getOperand(0);
26681 // EXTLOAD has a better solution on AVX2,
26682 // it may be replaced with X86ISD::VSEXT node.
26683 if (N00.getOpcode() == ISD::LOAD && Subtarget->hasInt256())
26684 if (!ISD::isNormalLoad(N00.getNode()))
26687 if (N00.getValueType() == MVT::v4i32 && ExtraVT.getSizeInBits() < 128) {
26688 SDValue Tmp = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32,
26690 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i64, Tmp);
26696 /// sext(add_nsw(x, C)) --> add(sext(x), C_sext)
26697 /// Promoting a sign extension ahead of an 'add nsw' exposes opportunities
26698 /// to combine math ops, use an LEA, or use a complex addressing mode. This can
26699 /// eliminate extend, add, and shift instructions.
26700 static SDValue promoteSextBeforeAddNSW(SDNode *Sext, SelectionDAG &DAG,
26701 const X86Subtarget *Subtarget) {
26702 // TODO: This should be valid for other integer types.
26703 EVT VT = Sext->getValueType(0);
26704 if (VT != MVT::i64)
26707 // We need an 'add nsw' feeding into the 'sext'.
26708 SDValue Add = Sext->getOperand(0);
26709 if (Add.getOpcode() != ISD::ADD || !Add->getFlags()->hasNoSignedWrap())
26712 // Having a constant operand to the 'add' ensures that we are not increasing
26713 // the instruction count because the constant is extended for free below.
26714 // A constant operand can also become the displacement field of an LEA.
26715 auto *AddOp1 = dyn_cast<ConstantSDNode>(Add.getOperand(1));
26719 // Don't make the 'add' bigger if there's no hope of combining it with some
26720 // other 'add' or 'shl' instruction.
26721 // TODO: It may be profitable to generate simpler LEA instructions in place
26722 // of single 'add' instructions, but the cost model for selecting an LEA
26723 // currently has a high threshold.
26724 bool HasLEAPotential = false;
26725 for (auto *User : Sext->uses()) {
26726 if (User->getOpcode() == ISD::ADD || User->getOpcode() == ISD::SHL) {
26727 HasLEAPotential = true;
26731 if (!HasLEAPotential)
26734 // Everything looks good, so pull the 'sext' ahead of the 'add'.
26735 int64_t AddConstant = AddOp1->getSExtValue();
26736 SDValue AddOp0 = Add.getOperand(0);
26737 SDValue NewSext = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(Sext), VT, AddOp0);
26738 SDValue NewConstant = DAG.getConstant(AddConstant, SDLoc(Add), VT);
26740 // The wider add is guaranteed to not wrap because both operands are
26743 Flags.setNoSignedWrap(true);
26744 return DAG.getNode(ISD::ADD, SDLoc(Add), VT, NewSext, NewConstant, &Flags);
26747 static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
26748 TargetLowering::DAGCombinerInfo &DCI,
26749 const X86Subtarget *Subtarget) {
26750 SDValue N0 = N->getOperand(0);
26751 EVT VT = N->getValueType(0);
26752 EVT SVT = VT.getScalarType();
26753 EVT InVT = N0.getValueType();
26754 EVT InSVT = InVT.getScalarType();
26757 // (i8,i32 sext (sdivrem (i8 x, i8 y)) ->
26758 // (i8,i32 (sdivrem_sext_hreg (i8 x, i8 y)
26759 // This exposes the sext to the sdivrem lowering, so that it directly extends
26760 // from AH (which we otherwise need to do contortions to access).
26761 if (N0.getOpcode() == ISD::SDIVREM && N0.getResNo() == 1 &&
26762 InVT == MVT::i8 && VT == MVT::i32) {
26763 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
26764 SDValue R = DAG.getNode(X86ISD::SDIVREM8_SEXT_HREG, DL, NodeTys,
26765 N0.getOperand(0), N0.getOperand(1));
26766 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
26767 return R.getValue(1);
26770 if (!DCI.isBeforeLegalizeOps()) {
26771 if (InVT == MVT::i1) {
26772 SDValue Zero = DAG.getConstant(0, DL, VT);
26774 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), DL, VT);
26775 return DAG.getNode(ISD::SELECT, DL, VT, N0, AllOnes, Zero);
26780 if (VT.isVector() && Subtarget->hasSSE2()) {
26781 auto ExtendVecSize = [&DAG](SDLoc DL, SDValue N, unsigned Size) {
26782 EVT InVT = N.getValueType();
26783 EVT OutVT = EVT::getVectorVT(*DAG.getContext(), InVT.getScalarType(),
26784 Size / InVT.getScalarSizeInBits());
26785 SmallVector<SDValue, 8> Opnds(Size / InVT.getSizeInBits(),
26786 DAG.getUNDEF(InVT));
26788 return DAG.getNode(ISD::CONCAT_VECTORS, DL, OutVT, Opnds);
26791 // If target-size is less than 128-bits, extend to a type that would extend
26792 // to 128 bits, extend that and extract the original target vector.
26793 if (VT.getSizeInBits() < 128 && !(128 % VT.getSizeInBits()) &&
26794 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
26795 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
26796 unsigned Scale = 128 / VT.getSizeInBits();
26798 EVT::getVectorVT(*DAG.getContext(), SVT, 128 / SVT.getSizeInBits());
26799 SDValue Ex = ExtendVecSize(DL, N0, Scale * InVT.getSizeInBits());
26800 SDValue SExt = DAG.getNode(ISD::SIGN_EXTEND, DL, ExVT, Ex);
26801 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, SExt,
26802 DAG.getIntPtrConstant(0, DL));
26805 // If target-size is 128-bits, then convert to ISD::SIGN_EXTEND_VECTOR_INREG
26806 // which ensures lowering to X86ISD::VSEXT (pmovsx*).
26807 if (VT.getSizeInBits() == 128 &&
26808 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
26809 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
26810 SDValue ExOp = ExtendVecSize(DL, N0, 128);
26811 return DAG.getSignExtendVectorInReg(ExOp, DL, VT);
26814 // On pre-AVX2 targets, split into 128-bit nodes of
26815 // ISD::SIGN_EXTEND_VECTOR_INREG.
26816 if (!Subtarget->hasInt256() && !(VT.getSizeInBits() % 128) &&
26817 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
26818 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
26819 unsigned NumVecs = VT.getSizeInBits() / 128;
26820 unsigned NumSubElts = 128 / SVT.getSizeInBits();
26821 EVT SubVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumSubElts);
26822 EVT InSubVT = EVT::getVectorVT(*DAG.getContext(), InSVT, NumSubElts);
26824 SmallVector<SDValue, 8> Opnds;
26825 for (unsigned i = 0, Offset = 0; i != NumVecs;
26826 ++i, Offset += NumSubElts) {
26827 SDValue SrcVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InSubVT, N0,
26828 DAG.getIntPtrConstant(Offset, DL));
26829 SrcVec = ExtendVecSize(DL, SrcVec, 128);
26830 SrcVec = DAG.getSignExtendVectorInReg(SrcVec, DL, SubVT);
26831 Opnds.push_back(SrcVec);
26833 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Opnds);
26837 if (Subtarget->hasAVX() && VT.is256BitVector())
26838 if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
26841 if (SDValue NewAdd = promoteSextBeforeAddNSW(N, DAG, Subtarget))
26847 static SDValue PerformFMACombine(SDNode *N, SelectionDAG &DAG,
26848 const X86Subtarget* Subtarget) {
26850 EVT VT = N->getValueType(0);
26852 // Let legalize expand this if it isn't a legal type yet.
26853 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
26856 EVT ScalarVT = VT.getScalarType();
26857 if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) || !Subtarget->hasAnyFMA())
26860 SDValue A = N->getOperand(0);
26861 SDValue B = N->getOperand(1);
26862 SDValue C = N->getOperand(2);
26864 bool NegA = (A.getOpcode() == ISD::FNEG);
26865 bool NegB = (B.getOpcode() == ISD::FNEG);
26866 bool NegC = (C.getOpcode() == ISD::FNEG);
26868 // Negative multiplication when NegA xor NegB
26869 bool NegMul = (NegA != NegB);
26871 A = A.getOperand(0);
26873 B = B.getOperand(0);
26875 C = C.getOperand(0);
26879 Opcode = (!NegC) ? X86ISD::FMADD : X86ISD::FMSUB;
26881 Opcode = (!NegC) ? X86ISD::FNMADD : X86ISD::FNMSUB;
26883 return DAG.getNode(Opcode, dl, VT, A, B, C);
26886 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG,
26887 TargetLowering::DAGCombinerInfo &DCI,
26888 const X86Subtarget *Subtarget) {
26889 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
26890 // (and (i32 x86isd::setcc_carry), 1)
26891 // This eliminates the zext. This transformation is necessary because
26892 // ISD::SETCC is always legalized to i8.
26894 SDValue N0 = N->getOperand(0);
26895 EVT VT = N->getValueType(0);
26897 if (N0.getOpcode() == ISD::AND &&
26899 N0.getOperand(0).hasOneUse()) {
26900 SDValue N00 = N0.getOperand(0);
26901 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
26902 if (!isOneConstant(N0.getOperand(1)))
26904 return DAG.getNode(ISD::AND, dl, VT,
26905 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
26906 N00.getOperand(0), N00.getOperand(1)),
26907 DAG.getConstant(1, dl, VT));
26911 if (N0.getOpcode() == ISD::TRUNCATE &&
26913 N0.getOperand(0).hasOneUse()) {
26914 SDValue N00 = N0.getOperand(0);
26915 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
26916 return DAG.getNode(ISD::AND, dl, VT,
26917 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
26918 N00.getOperand(0), N00.getOperand(1)),
26919 DAG.getConstant(1, dl, VT));
26923 if (VT.is256BitVector())
26924 if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
26927 // (i8,i32 zext (udivrem (i8 x, i8 y)) ->
26928 // (i8,i32 (udivrem_zext_hreg (i8 x, i8 y)
26929 // This exposes the zext to the udivrem lowering, so that it directly extends
26930 // from AH (which we otherwise need to do contortions to access).
26931 if (N0.getOpcode() == ISD::UDIVREM &&
26932 N0.getResNo() == 1 && N0.getValueType() == MVT::i8 &&
26933 (VT == MVT::i32 || VT == MVT::i64)) {
26934 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
26935 SDValue R = DAG.getNode(X86ISD::UDIVREM8_ZEXT_HREG, dl, NodeTys,
26936 N0.getOperand(0), N0.getOperand(1));
26937 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
26938 return R.getValue(1);
26944 // Optimize x == -y --> x+y == 0
26945 // x != -y --> x+y != 0
26946 static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG,
26947 const X86Subtarget* Subtarget) {
26948 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
26949 SDValue LHS = N->getOperand(0);
26950 SDValue RHS = N->getOperand(1);
26951 EVT VT = N->getValueType(0);
26954 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB)
26955 if (isNullConstant(LHS.getOperand(0)) && LHS.hasOneUse()) {
26956 SDValue addV = DAG.getNode(ISD::ADD, DL, LHS.getValueType(), RHS,
26957 LHS.getOperand(1));
26958 return DAG.getSetCC(DL, N->getValueType(0), addV,
26959 DAG.getConstant(0, DL, addV.getValueType()), CC);
26961 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && RHS.getOpcode() == ISD::SUB)
26962 if (isNullConstant(RHS.getOperand(0)) && RHS.hasOneUse()) {
26963 SDValue addV = DAG.getNode(ISD::ADD, DL, RHS.getValueType(), LHS,
26964 RHS.getOperand(1));
26965 return DAG.getSetCC(DL, N->getValueType(0), addV,
26966 DAG.getConstant(0, DL, addV.getValueType()), CC);
26969 if (VT.getScalarType() == MVT::i1 &&
26970 (CC == ISD::SETNE || CC == ISD::SETEQ || ISD::isSignedIntSetCC(CC))) {
26972 (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
26973 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
26974 bool IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
26976 if (!IsSEXT0 || !IsVZero1) {
26977 // Swap the operands and update the condition code.
26978 std::swap(LHS, RHS);
26979 CC = ISD::getSetCCSwappedOperands(CC);
26981 IsSEXT0 = (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
26982 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
26983 IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
26986 if (IsSEXT0 && IsVZero1) {
26987 assert(VT == LHS.getOperand(0).getValueType() &&
26988 "Uexpected operand type");
26989 if (CC == ISD::SETGT)
26990 return DAG.getConstant(0, DL, VT);
26991 if (CC == ISD::SETLE)
26992 return DAG.getConstant(1, DL, VT);
26993 if (CC == ISD::SETEQ || CC == ISD::SETGE)
26994 return DAG.getNOT(DL, LHS.getOperand(0), VT);
26996 assert((CC == ISD::SETNE || CC == ISD::SETLT) &&
26997 "Unexpected condition code!");
26998 return LHS.getOperand(0);
27005 static SDValue PerformBLENDICombine(SDNode *N, SelectionDAG &DAG) {
27006 SDValue V0 = N->getOperand(0);
27007 SDValue V1 = N->getOperand(1);
27009 EVT VT = N->getValueType(0);
27011 // Canonicalize a v2f64 blend with a mask of 2 by swapping the vector
27012 // operands and changing the mask to 1. This saves us a bunch of
27013 // pattern-matching possibilities related to scalar math ops in SSE/AVX.
27014 // x86InstrInfo knows how to commute this back after instruction selection
27015 // if it would help register allocation.
27017 // TODO: If optimizing for size or a processor that doesn't suffer from
27018 // partial register update stalls, this should be transformed into a MOVSD
27019 // instruction because a MOVSD is 1-2 bytes smaller than a BLENDPD.
27021 if (VT == MVT::v2f64)
27022 if (auto *Mask = dyn_cast<ConstantSDNode>(N->getOperand(2)))
27023 if (Mask->getZExtValue() == 2 && !isShuffleFoldableLoad(V0)) {
27024 SDValue NewMask = DAG.getConstant(1, DL, MVT::i8);
27025 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V0, NewMask);
27031 static SDValue PerformGatherScatterCombine(SDNode *N, SelectionDAG &DAG) {
27033 // Gather and Scatter instructions use k-registers for masks. The type of
27034 // the masks is v*i1. So the mask will be truncated anyway.
27035 // The SIGN_EXTEND_INREG my be dropped.
27036 SDValue Mask = N->getOperand(2);
27037 if (Mask.getOpcode() == ISD::SIGN_EXTEND_INREG) {
27038 SmallVector<SDValue, 5> NewOps(N->op_begin(), N->op_end());
27039 NewOps[2] = Mask.getOperand(0);
27040 DAG.UpdateNodeOperands(N, NewOps);
27045 // Helper function of PerformSETCCCombine. It is to materialize "setb reg"
27046 // as "sbb reg,reg", since it can be extended without zext and produces
27047 // an all-ones bit which is more useful than 0/1 in some cases.
27048 static SDValue MaterializeSETB(SDLoc DL, SDValue EFLAGS, SelectionDAG &DAG,
27051 return DAG.getNode(ISD::AND, DL, VT,
27052 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
27053 DAG.getConstant(X86::COND_B, DL, MVT::i8),
27055 DAG.getConstant(1, DL, VT));
27056 assert (VT == MVT::i1 && "Unexpected type for SECCC node");
27057 return DAG.getNode(ISD::TRUNCATE, DL, MVT::i1,
27058 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
27059 DAG.getConstant(X86::COND_B, DL, MVT::i8),
27063 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
27064 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG,
27065 TargetLowering::DAGCombinerInfo &DCI,
27066 const X86Subtarget *Subtarget) {
27068 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0));
27069 SDValue EFLAGS = N->getOperand(1);
27071 if (CC == X86::COND_A) {
27072 // Try to convert COND_A into COND_B in an attempt to facilitate
27073 // materializing "setb reg".
27075 // Do not flip "e > c", where "c" is a constant, because Cmp instruction
27076 // cannot take an immediate as its first operand.
27078 if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() &&
27079 EFLAGS.getValueType().isInteger() &&
27080 !isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
27081 SDValue NewSub = DAG.getNode(X86ISD::SUB, SDLoc(EFLAGS),
27082 EFLAGS.getNode()->getVTList(),
27083 EFLAGS.getOperand(1), EFLAGS.getOperand(0));
27084 SDValue NewEFLAGS = SDValue(NewSub.getNode(), EFLAGS.getResNo());
27085 return MaterializeSETB(DL, NewEFLAGS, DAG, N->getSimpleValueType(0));
27089 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
27090 // a zext and produces an all-ones bit which is more useful than 0/1 in some
27092 if (CC == X86::COND_B)
27093 return MaterializeSETB(DL, EFLAGS, DAG, N->getSimpleValueType(0));
27095 if (SDValue Flags = checkBoolTestSetCCCombine(EFLAGS, CC)) {
27096 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
27097 return DAG.getNode(X86ISD::SETCC, DL, N->getVTList(), Cond, Flags);
27103 // Optimize branch condition evaluation.
27105 static SDValue PerformBrCondCombine(SDNode *N, SelectionDAG &DAG,
27106 TargetLowering::DAGCombinerInfo &DCI,
27107 const X86Subtarget *Subtarget) {
27109 SDValue Chain = N->getOperand(0);
27110 SDValue Dest = N->getOperand(1);
27111 SDValue EFLAGS = N->getOperand(3);
27112 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(2));
27114 if (SDValue Flags = checkBoolTestSetCCCombine(EFLAGS, CC)) {
27115 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
27116 return DAG.getNode(X86ISD::BRCOND, DL, N->getVTList(), Chain, Dest, Cond,
27123 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
27124 SelectionDAG &DAG) {
27125 // Take advantage of vector comparisons producing 0 or -1 in each lane to
27126 // optimize away operation when it's from a constant.
27128 // The general transformation is:
27129 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
27130 // AND(VECTOR_CMP(x,y), constant2)
27131 // constant2 = UNARYOP(constant)
27133 // Early exit if this isn't a vector operation, the operand of the
27134 // unary operation isn't a bitwise AND, or if the sizes of the operations
27135 // aren't the same.
27136 EVT VT = N->getValueType(0);
27137 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
27138 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
27139 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
27142 // Now check that the other operand of the AND is a constant. We could
27143 // make the transformation for non-constant splats as well, but it's unclear
27144 // that would be a benefit as it would not eliminate any operations, just
27145 // perform one more step in scalar code before moving to the vector unit.
27146 if (BuildVectorSDNode *BV =
27147 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
27148 // Bail out if the vector isn't a constant.
27149 if (!BV->isConstant())
27152 // Everything checks out. Build up the new and improved node.
27154 EVT IntVT = BV->getValueType(0);
27155 // Create a new constant of the appropriate type for the transformed
27157 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
27158 // The AND node needs bitcasts to/from an integer vector type around it.
27159 SDValue MaskConst = DAG.getBitcast(IntVT, SourceConst);
27160 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
27161 N->getOperand(0)->getOperand(0), MaskConst);
27162 SDValue Res = DAG.getBitcast(VT, NewAnd);
27169 static SDValue PerformUINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
27170 const X86Subtarget *Subtarget) {
27171 SDValue Op0 = N->getOperand(0);
27172 EVT VT = N->getValueType(0);
27173 EVT InVT = Op0.getValueType();
27174 EVT InSVT = InVT.getScalarType();
27175 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
27177 // UINT_TO_FP(vXi8) -> SINT_TO_FP(ZEXT(vXi8 to vXi32))
27178 // UINT_TO_FP(vXi16) -> SINT_TO_FP(ZEXT(vXi16 to vXi32))
27179 if (InVT.isVector() && (InSVT == MVT::i8 || InSVT == MVT::i16)) {
27181 EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
27182 InVT.getVectorNumElements());
27183 SDValue P = DAG.getNode(ISD::ZERO_EXTEND, dl, DstVT, Op0);
27185 if (TLI.isOperationLegal(ISD::UINT_TO_FP, DstVT))
27186 return DAG.getNode(ISD::UINT_TO_FP, dl, VT, P);
27188 return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P);
27194 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
27195 const X86Subtarget *Subtarget) {
27196 // First try to optimize away the conversion entirely when it's
27197 // conditionally from a constant. Vectors only.
27198 if (SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG))
27201 // Now move on to more general possibilities.
27202 SDValue Op0 = N->getOperand(0);
27203 EVT VT = N->getValueType(0);
27204 EVT InVT = Op0.getValueType();
27205 EVT InSVT = InVT.getScalarType();
27207 // SINT_TO_FP(vXi8) -> SINT_TO_FP(SEXT(vXi8 to vXi32))
27208 // SINT_TO_FP(vXi16) -> SINT_TO_FP(SEXT(vXi16 to vXi32))
27209 if (InVT.isVector() && (InSVT == MVT::i8 || InSVT == MVT::i16)) {
27211 EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
27212 InVT.getVectorNumElements());
27213 SDValue P = DAG.getNode(ISD::SIGN_EXTEND, dl, DstVT, Op0);
27214 return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P);
27217 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
27218 // a 32-bit target where SSE doesn't support i64->FP operations.
27219 if (!Subtarget->useSoftFloat() && Op0.getOpcode() == ISD::LOAD) {
27220 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
27221 EVT LdVT = Ld->getValueType(0);
27223 // This transformation is not supported if the result type is f16
27224 if (VT == MVT::f16)
27227 if (!Ld->isVolatile() && !VT.isVector() &&
27228 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
27229 !Subtarget->is64Bit() && LdVT == MVT::i64) {
27230 SDValue FILDChain = Subtarget->getTargetLowering()->BuildFILD(
27231 SDValue(N, 0), LdVT, Ld->getChain(), Op0, DAG);
27232 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
27239 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
27240 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
27241 X86TargetLowering::DAGCombinerInfo &DCI) {
27242 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
27243 // the result is either zero or one (depending on the input carry bit).
27244 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
27245 if (X86::isZeroNode(N->getOperand(0)) &&
27246 X86::isZeroNode(N->getOperand(1)) &&
27247 // We don't have a good way to replace an EFLAGS use, so only do this when
27249 SDValue(N, 1).use_empty()) {
27251 EVT VT = N->getValueType(0);
27252 SDValue CarryOut = DAG.getConstant(0, DL, N->getValueType(1));
27253 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
27254 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
27255 DAG.getConstant(X86::COND_B, DL,
27258 DAG.getConstant(1, DL, VT));
27259 return DCI.CombineTo(N, Res1, CarryOut);
27265 // fold (add Y, (sete X, 0)) -> adc 0, Y
27266 // (add Y, (setne X, 0)) -> sbb -1, Y
27267 // (sub (sete X, 0), Y) -> sbb 0, Y
27268 // (sub (setne X, 0), Y) -> adc -1, Y
27269 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
27272 // Look through ZExts.
27273 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
27274 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
27277 SDValue SetCC = Ext.getOperand(0);
27278 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
27281 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
27282 if (CC != X86::COND_E && CC != X86::COND_NE)
27285 SDValue Cmp = SetCC.getOperand(1);
27286 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
27287 !X86::isZeroNode(Cmp.getOperand(1)) ||
27288 !Cmp.getOperand(0).getValueType().isInteger())
27291 SDValue CmpOp0 = Cmp.getOperand(0);
27292 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
27293 DAG.getConstant(1, DL, CmpOp0.getValueType()));
27295 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
27296 if (CC == X86::COND_NE)
27297 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
27298 DL, OtherVal.getValueType(), OtherVal,
27299 DAG.getConstant(-1ULL, DL, OtherVal.getValueType()),
27301 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
27302 DL, OtherVal.getValueType(), OtherVal,
27303 DAG.getConstant(0, DL, OtherVal.getValueType()), NewCmp);
27306 /// PerformADDCombine - Do target-specific dag combines on integer adds.
27307 static SDValue PerformAddCombine(SDNode *N, SelectionDAG &DAG,
27308 const X86Subtarget *Subtarget) {
27309 EVT VT = N->getValueType(0);
27310 SDValue Op0 = N->getOperand(0);
27311 SDValue Op1 = N->getOperand(1);
27313 // Try to synthesize horizontal adds from adds of shuffles.
27314 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
27315 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
27316 isHorizontalBinOp(Op0, Op1, true))
27317 return DAG.getNode(X86ISD::HADD, SDLoc(N), VT, Op0, Op1);
27319 return OptimizeConditionalInDecrement(N, DAG);
27322 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG,
27323 const X86Subtarget *Subtarget) {
27324 SDValue Op0 = N->getOperand(0);
27325 SDValue Op1 = N->getOperand(1);
27327 // X86 can't encode an immediate LHS of a sub. See if we can push the
27328 // negation into a preceding instruction.
27329 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
27330 // If the RHS of the sub is a XOR with one use and a constant, invert the
27331 // immediate. Then add one to the LHS of the sub so we can turn
27332 // X-Y -> X+~Y+1, saving one register.
27333 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
27334 isa<ConstantSDNode>(Op1.getOperand(1))) {
27335 APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue();
27336 EVT VT = Op0.getValueType();
27337 SDValue NewXor = DAG.getNode(ISD::XOR, SDLoc(Op1), VT,
27339 DAG.getConstant(~XorC, SDLoc(Op1), VT));
27340 return DAG.getNode(ISD::ADD, SDLoc(N), VT, NewXor,
27341 DAG.getConstant(C->getAPIntValue() + 1, SDLoc(N), VT));
27345 // Try to synthesize horizontal adds from adds of shuffles.
27346 EVT VT = N->getValueType(0);
27347 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
27348 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
27349 isHorizontalBinOp(Op0, Op1, true))
27350 return DAG.getNode(X86ISD::HSUB, SDLoc(N), VT, Op0, Op1);
27352 return OptimizeConditionalInDecrement(N, DAG);
27355 /// performVZEXTCombine - Performs build vector combines
27356 static SDValue performVZEXTCombine(SDNode *N, SelectionDAG &DAG,
27357 TargetLowering::DAGCombinerInfo &DCI,
27358 const X86Subtarget *Subtarget) {
27360 MVT VT = N->getSimpleValueType(0);
27361 SDValue Op = N->getOperand(0);
27362 MVT OpVT = Op.getSimpleValueType();
27363 MVT OpEltVT = OpVT.getVectorElementType();
27364 unsigned InputBits = OpEltVT.getSizeInBits() * VT.getVectorNumElements();
27366 // (vzext (bitcast (vzext (x)) -> (vzext x)
27368 while (V.getOpcode() == ISD::BITCAST)
27369 V = V.getOperand(0);
27371 if (V != Op && V.getOpcode() == X86ISD::VZEXT) {
27372 MVT InnerVT = V.getSimpleValueType();
27373 MVT InnerEltVT = InnerVT.getVectorElementType();
27375 // If the element sizes match exactly, we can just do one larger vzext. This
27376 // is always an exact type match as vzext operates on integer types.
27377 if (OpEltVT == InnerEltVT) {
27378 assert(OpVT == InnerVT && "Types must match for vzext!");
27379 return DAG.getNode(X86ISD::VZEXT, DL, VT, V.getOperand(0));
27382 // The only other way we can combine them is if only a single element of the
27383 // inner vzext is used in the input to the outer vzext.
27384 if (InnerEltVT.getSizeInBits() < InputBits)
27387 // In this case, the inner vzext is completely dead because we're going to
27388 // only look at bits inside of the low element. Just do the outer vzext on
27389 // a bitcast of the input to the inner.
27390 return DAG.getNode(X86ISD::VZEXT, DL, VT, DAG.getBitcast(OpVT, V));
27393 // Check if we can bypass extracting and re-inserting an element of an input
27394 // vector. Essentially:
27395 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast x)
27396 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR &&
27397 V.getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
27398 V.getOperand(0).getSimpleValueType().getSizeInBits() == InputBits) {
27399 SDValue ExtractedV = V.getOperand(0);
27400 SDValue OrigV = ExtractedV.getOperand(0);
27401 if (isNullConstant(ExtractedV.getOperand(1))) {
27402 MVT OrigVT = OrigV.getSimpleValueType();
27403 // Extract a subvector if necessary...
27404 if (OrigVT.getSizeInBits() > OpVT.getSizeInBits()) {
27405 int Ratio = OrigVT.getSizeInBits() / OpVT.getSizeInBits();
27406 OrigVT = MVT::getVectorVT(OrigVT.getVectorElementType(),
27407 OrigVT.getVectorNumElements() / Ratio);
27408 OrigV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigVT, OrigV,
27409 DAG.getIntPtrConstant(0, DL));
27411 Op = DAG.getBitcast(OpVT, OrigV);
27412 return DAG.getNode(X86ISD::VZEXT, DL, VT, Op);
27419 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
27420 DAGCombinerInfo &DCI) const {
27421 SelectionDAG &DAG = DCI.DAG;
27422 switch (N->getOpcode()) {
27424 case ISD::EXTRACT_VECTOR_ELT:
27425 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, DCI);
27428 case X86ISD::SHRUNKBLEND:
27429 return PerformSELECTCombine(N, DAG, DCI, Subtarget);
27430 case ISD::BITCAST: return PerformBITCASTCombine(N, DAG, Subtarget);
27431 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI, Subtarget);
27432 case ISD::ADD: return PerformAddCombine(N, DAG, Subtarget);
27433 case ISD::SUB: return PerformSubCombine(N, DAG, Subtarget);
27434 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
27435 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
27438 case ISD::SRL: return PerformShiftCombine(N, DAG, DCI, Subtarget);
27439 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
27440 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
27441 case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget);
27442 case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget);
27443 case ISD::MLOAD: return PerformMLOADCombine(N, DAG, DCI, Subtarget);
27444 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
27445 case ISD::MSTORE: return PerformMSTORECombine(N, DAG, Subtarget);
27446 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, Subtarget);
27447 case ISD::UINT_TO_FP: return PerformUINT_TO_FPCombine(N, DAG, Subtarget);
27448 case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
27449 case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
27450 case ISD::FNEG: return PerformFNEGCombine(N, DAG, Subtarget);
27451 case ISD::TRUNCATE: return PerformTRUNCATECombine(N, DAG, Subtarget);
27453 case X86ISD::FOR: return PerformFORCombine(N, DAG, Subtarget);
27455 case X86ISD::FMAX: return PerformFMinFMaxCombine(N, DAG);
27456 case X86ISD::FAND: return PerformFANDCombine(N, DAG, Subtarget);
27457 case X86ISD::FANDN: return PerformFANDNCombine(N, DAG, Subtarget);
27458 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
27459 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
27460 case ISD::ANY_EXTEND:
27461 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG, DCI, Subtarget);
27462 case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
27463 case ISD::SIGN_EXTEND_INREG:
27464 return PerformSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
27465 case ISD::SETCC: return PerformISDSETCCCombine(N, DAG, Subtarget);
27466 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
27467 case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
27468 case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);
27469 case X86ISD::SHUFP: // Handle all target specific shuffles
27470 case X86ISD::PALIGNR:
27471 case X86ISD::UNPCKH:
27472 case X86ISD::UNPCKL:
27473 case X86ISD::MOVHLPS:
27474 case X86ISD::MOVLHPS:
27475 case X86ISD::PSHUFB:
27476 case X86ISD::PSHUFD:
27477 case X86ISD::PSHUFHW:
27478 case X86ISD::PSHUFLW:
27479 case X86ISD::MOVSS:
27480 case X86ISD::MOVSD:
27481 case X86ISD::VPERMILPI:
27482 case X86ISD::VPERM2X128:
27483 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
27484 case ISD::FMA: return PerformFMACombine(N, DAG, Subtarget);
27485 case X86ISD::BLENDI: return PerformBLENDICombine(N, DAG);
27487 case ISD::MSCATTER: return PerformGatherScatterCombine(N, DAG);
27493 /// isTypeDesirableForOp - Return true if the target has native support for
27494 /// the specified value type and it is 'desirable' to use the type for the
27495 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
27496 /// instruction encodings are longer and some i16 instructions are slow.
27497 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
27498 if (!isTypeLegal(VT))
27500 if (VT != MVT::i16)
27507 case ISD::SIGN_EXTEND:
27508 case ISD::ZERO_EXTEND:
27509 case ISD::ANY_EXTEND:
27522 /// IsDesirableToPromoteOp - This method query the target whether it is
27523 /// beneficial for dag combiner to promote the specified node. If true, it
27524 /// should return the desired promotion type by reference.
27525 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
27526 EVT VT = Op.getValueType();
27527 if (VT != MVT::i16)
27530 bool Promote = false;
27531 bool Commute = false;
27532 switch (Op.getOpcode()) {
27535 LoadSDNode *LD = cast<LoadSDNode>(Op);
27536 // If the non-extending load has a single use and it's not live out, then it
27537 // might be folded.
27538 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
27539 Op.hasOneUse()*/) {
27540 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
27541 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
27542 // The only case where we'd want to promote LOAD (rather then it being
27543 // promoted as an operand is when it's only use is liveout.
27544 if (UI->getOpcode() != ISD::CopyToReg)
27551 case ISD::SIGN_EXTEND:
27552 case ISD::ZERO_EXTEND:
27553 case ISD::ANY_EXTEND:
27558 SDValue N0 = Op.getOperand(0);
27559 // Look out for (store (shl (load), x)).
27560 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
27573 SDValue N0 = Op.getOperand(0);
27574 SDValue N1 = Op.getOperand(1);
27575 if (!Commute && MayFoldLoad(N1))
27577 // Avoid disabling potential load folding opportunities.
27578 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
27580 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
27590 //===----------------------------------------------------------------------===//
27591 // X86 Inline Assembly Support
27592 //===----------------------------------------------------------------------===//
27594 // Helper to match a string separated by whitespace.
27595 static bool matchAsm(StringRef S, ArrayRef<const char *> Pieces) {
27596 S = S.substr(S.find_first_not_of(" \t")); // Skip leading whitespace.
27598 for (StringRef Piece : Pieces) {
27599 if (!S.startswith(Piece)) // Check if the piece matches.
27602 S = S.substr(Piece.size());
27603 StringRef::size_type Pos = S.find_first_not_of(" \t");
27604 if (Pos == 0) // We matched a prefix.
27613 static bool clobbersFlagRegisters(const SmallVector<StringRef, 4> &AsmPieces) {
27615 if (AsmPieces.size() == 3 || AsmPieces.size() == 4) {
27616 if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{cc}") &&
27617 std::count(AsmPieces.begin(), AsmPieces.end(), "~{flags}") &&
27618 std::count(AsmPieces.begin(), AsmPieces.end(), "~{fpsr}")) {
27620 if (AsmPieces.size() == 3)
27622 else if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{dirflag}"))
27629 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
27630 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
27632 std::string AsmStr = IA->getAsmString();
27634 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
27635 if (!Ty || Ty->getBitWidth() % 16 != 0)
27638 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
27639 SmallVector<StringRef, 4> AsmPieces;
27640 SplitString(AsmStr, AsmPieces, ";\n");
27642 switch (AsmPieces.size()) {
27643 default: return false;
27645 // FIXME: this should verify that we are targeting a 486 or better. If not,
27646 // we will turn this bswap into something that will be lowered to logical
27647 // ops instead of emitting the bswap asm. For now, we don't support 486 or
27648 // lower so don't worry about this.
27650 if (matchAsm(AsmPieces[0], {"bswap", "$0"}) ||
27651 matchAsm(AsmPieces[0], {"bswapl", "$0"}) ||
27652 matchAsm(AsmPieces[0], {"bswapq", "$0"}) ||
27653 matchAsm(AsmPieces[0], {"bswap", "${0:q}"}) ||
27654 matchAsm(AsmPieces[0], {"bswapl", "${0:q}"}) ||
27655 matchAsm(AsmPieces[0], {"bswapq", "${0:q}"})) {
27656 // No need to check constraints, nothing other than the equivalent of
27657 // "=r,0" would be valid here.
27658 return IntrinsicLowering::LowerToByteSwap(CI);
27661 // rorw $$8, ${0:w} --> llvm.bswap.i16
27662 if (CI->getType()->isIntegerTy(16) &&
27663 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
27664 (matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) ||
27665 matchAsm(AsmPieces[0], {"rolw", "$$8,", "${0:w}"}))) {
27667 StringRef ConstraintsStr = IA->getConstraintString();
27668 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
27669 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
27670 if (clobbersFlagRegisters(AsmPieces))
27671 return IntrinsicLowering::LowerToByteSwap(CI);
27675 if (CI->getType()->isIntegerTy(32) &&
27676 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
27677 matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) &&
27678 matchAsm(AsmPieces[1], {"rorl", "$$16,", "$0"}) &&
27679 matchAsm(AsmPieces[2], {"rorw", "$$8,", "${0:w}"})) {
27681 StringRef ConstraintsStr = IA->getConstraintString();
27682 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
27683 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
27684 if (clobbersFlagRegisters(AsmPieces))
27685 return IntrinsicLowering::LowerToByteSwap(CI);
27688 if (CI->getType()->isIntegerTy(64)) {
27689 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
27690 if (Constraints.size() >= 2 &&
27691 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
27692 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
27693 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
27694 if (matchAsm(AsmPieces[0], {"bswap", "%eax"}) &&
27695 matchAsm(AsmPieces[1], {"bswap", "%edx"}) &&
27696 matchAsm(AsmPieces[2], {"xchgl", "%eax,", "%edx"}))
27697 return IntrinsicLowering::LowerToByteSwap(CI);
27705 /// getConstraintType - Given a constraint letter, return the type of
27706 /// constraint it is for this target.
27707 X86TargetLowering::ConstraintType
27708 X86TargetLowering::getConstraintType(StringRef Constraint) const {
27709 if (Constraint.size() == 1) {
27710 switch (Constraint[0]) {
27721 return C_RegisterClass;
27745 return TargetLowering::getConstraintType(Constraint);
27748 /// Examine constraint type and operand type and determine a weight value.
27749 /// This object must already have been set up with the operand type
27750 /// and the current alternative constraint selected.
27751 TargetLowering::ConstraintWeight
27752 X86TargetLowering::getSingleConstraintMatchWeight(
27753 AsmOperandInfo &info, const char *constraint) const {
27754 ConstraintWeight weight = CW_Invalid;
27755 Value *CallOperandVal = info.CallOperandVal;
27756 // If we don't have a value, we can't do a match,
27757 // but allow it at the lowest weight.
27758 if (!CallOperandVal)
27760 Type *type = CallOperandVal->getType();
27761 // Look at the constraint type.
27762 switch (*constraint) {
27764 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
27775 if (CallOperandVal->getType()->isIntegerTy())
27776 weight = CW_SpecificReg;
27781 if (type->isFloatingPointTy())
27782 weight = CW_SpecificReg;
27785 if (type->isX86_MMXTy() && Subtarget->hasMMX())
27786 weight = CW_SpecificReg;
27790 if (((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasSSE1()) ||
27791 ((type->getPrimitiveSizeInBits() == 256) && Subtarget->hasFp256()))
27792 weight = CW_Register;
27795 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
27796 if (C->getZExtValue() <= 31)
27797 weight = CW_Constant;
27801 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27802 if (C->getZExtValue() <= 63)
27803 weight = CW_Constant;
27807 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27808 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
27809 weight = CW_Constant;
27813 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27814 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
27815 weight = CW_Constant;
27819 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27820 if (C->getZExtValue() <= 3)
27821 weight = CW_Constant;
27825 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27826 if (C->getZExtValue() <= 0xff)
27827 weight = CW_Constant;
27832 if (isa<ConstantFP>(CallOperandVal)) {
27833 weight = CW_Constant;
27837 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27838 if ((C->getSExtValue() >= -0x80000000LL) &&
27839 (C->getSExtValue() <= 0x7fffffffLL))
27840 weight = CW_Constant;
27844 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27845 if (C->getZExtValue() <= 0xffffffff)
27846 weight = CW_Constant;
27853 /// LowerXConstraint - try to replace an X constraint, which matches anything,
27854 /// with another that has more specific requirements based on the type of the
27855 /// corresponding operand.
27856 const char *X86TargetLowering::
27857 LowerXConstraint(EVT ConstraintVT) const {
27858 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
27859 // 'f' like normal targets.
27860 if (ConstraintVT.isFloatingPoint()) {
27861 if (Subtarget->hasSSE2())
27863 if (Subtarget->hasSSE1())
27867 return TargetLowering::LowerXConstraint(ConstraintVT);
27870 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
27871 /// vector. If it is invalid, don't add anything to Ops.
27872 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
27873 std::string &Constraint,
27874 std::vector<SDValue>&Ops,
27875 SelectionDAG &DAG) const {
27878 // Only support length 1 constraints for now.
27879 if (Constraint.length() > 1) return;
27881 char ConstraintLetter = Constraint[0];
27882 switch (ConstraintLetter) {
27885 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27886 if (C->getZExtValue() <= 31) {
27887 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27888 Op.getValueType());
27894 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27895 if (C->getZExtValue() <= 63) {
27896 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27897 Op.getValueType());
27903 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27904 if (isInt<8>(C->getSExtValue())) {
27905 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27906 Op.getValueType());
27912 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27913 if (C->getZExtValue() == 0xff || C->getZExtValue() == 0xffff ||
27914 (Subtarget->is64Bit() && C->getZExtValue() == 0xffffffff)) {
27915 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op),
27916 Op.getValueType());
27922 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27923 if (C->getZExtValue() <= 3) {
27924 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27925 Op.getValueType());
27931 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27932 if (C->getZExtValue() <= 255) {
27933 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27934 Op.getValueType());
27940 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27941 if (C->getZExtValue() <= 127) {
27942 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27943 Op.getValueType());
27949 // 32-bit signed value
27950 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27951 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
27952 C->getSExtValue())) {
27953 // Widen to 64 bits here to get it sign extended.
27954 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op), MVT::i64);
27957 // FIXME gcc accepts some relocatable values here too, but only in certain
27958 // memory models; it's complicated.
27963 // 32-bit unsigned value
27964 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27965 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
27966 C->getZExtValue())) {
27967 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27968 Op.getValueType());
27972 // FIXME gcc accepts some relocatable values here too, but only in certain
27973 // memory models; it's complicated.
27977 // Literal immediates are always ok.
27978 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
27979 // Widen to 64 bits here to get it sign extended.
27980 Result = DAG.getTargetConstant(CST->getSExtValue(), SDLoc(Op), MVT::i64);
27984 // In any sort of PIC mode addresses need to be computed at runtime by
27985 // adding in a register or some sort of table lookup. These can't
27986 // be used as immediates.
27987 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
27990 // If we are in non-pic codegen mode, we allow the address of a global (with
27991 // an optional displacement) to be used with 'i'.
27992 GlobalAddressSDNode *GA = nullptr;
27993 int64_t Offset = 0;
27995 // Match either (GA), (GA+C), (GA+C1+C2), etc.
27997 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
27998 Offset += GA->getOffset();
28000 } else if (Op.getOpcode() == ISD::ADD) {
28001 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
28002 Offset += C->getZExtValue();
28003 Op = Op.getOperand(0);
28006 } else if (Op.getOpcode() == ISD::SUB) {
28007 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
28008 Offset += -C->getZExtValue();
28009 Op = Op.getOperand(0);
28014 // Otherwise, this isn't something we can handle, reject it.
28018 const GlobalValue *GV = GA->getGlobal();
28019 // If we require an extra load to get this address, as in PIC mode, we
28020 // can't accept it.
28021 if (isGlobalStubReference(
28022 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget())))
28025 Result = DAG.getTargetGlobalAddress(GV, SDLoc(Op),
28026 GA->getValueType(0), Offset);
28031 if (Result.getNode()) {
28032 Ops.push_back(Result);
28035 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
28038 std::pair<unsigned, const TargetRegisterClass *>
28039 X86TargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
28040 StringRef Constraint,
28042 // First, see if this is a constraint that directly corresponds to an LLVM
28044 if (Constraint.size() == 1) {
28045 // GCC Constraint Letters
28046 switch (Constraint[0]) {
28048 // TODO: Slight differences here in allocation order and leaving
28049 // RIP in the class. Do they matter any more here than they do
28050 // in the normal allocation?
28051 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
28052 if (Subtarget->is64Bit()) {
28053 if (VT == MVT::i32 || VT == MVT::f32)
28054 return std::make_pair(0U, &X86::GR32RegClass);
28055 if (VT == MVT::i16)
28056 return std::make_pair(0U, &X86::GR16RegClass);
28057 if (VT == MVT::i8 || VT == MVT::i1)
28058 return std::make_pair(0U, &X86::GR8RegClass);
28059 if (VT == MVT::i64 || VT == MVT::f64)
28060 return std::make_pair(0U, &X86::GR64RegClass);
28063 // 32-bit fallthrough
28064 case 'Q': // Q_REGS
28065 if (VT == MVT::i32 || VT == MVT::f32)
28066 return std::make_pair(0U, &X86::GR32_ABCDRegClass);
28067 if (VT == MVT::i16)
28068 return std::make_pair(0U, &X86::GR16_ABCDRegClass);
28069 if (VT == MVT::i8 || VT == MVT::i1)
28070 return std::make_pair(0U, &X86::GR8_ABCD_LRegClass);
28071 if (VT == MVT::i64)
28072 return std::make_pair(0U, &X86::GR64_ABCDRegClass);
28074 case 'r': // GENERAL_REGS
28075 case 'l': // INDEX_REGS
28076 if (VT == MVT::i8 || VT == MVT::i1)
28077 return std::make_pair(0U, &X86::GR8RegClass);
28078 if (VT == MVT::i16)
28079 return std::make_pair(0U, &X86::GR16RegClass);
28080 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
28081 return std::make_pair(0U, &X86::GR32RegClass);
28082 return std::make_pair(0U, &X86::GR64RegClass);
28083 case 'R': // LEGACY_REGS
28084 if (VT == MVT::i8 || VT == MVT::i1)
28085 return std::make_pair(0U, &X86::GR8_NOREXRegClass);
28086 if (VT == MVT::i16)
28087 return std::make_pair(0U, &X86::GR16_NOREXRegClass);
28088 if (VT == MVT::i32 || !Subtarget->is64Bit())
28089 return std::make_pair(0U, &X86::GR32_NOREXRegClass);
28090 return std::make_pair(0U, &X86::GR64_NOREXRegClass);
28091 case 'f': // FP Stack registers.
28092 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
28093 // value to the correct fpstack register class.
28094 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
28095 return std::make_pair(0U, &X86::RFP32RegClass);
28096 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
28097 return std::make_pair(0U, &X86::RFP64RegClass);
28098 return std::make_pair(0U, &X86::RFP80RegClass);
28099 case 'y': // MMX_REGS if MMX allowed.
28100 if (!Subtarget->hasMMX()) break;
28101 return std::make_pair(0U, &X86::VR64RegClass);
28102 case 'Y': // SSE_REGS if SSE2 allowed
28103 if (!Subtarget->hasSSE2()) break;
28105 case 'x': // SSE_REGS if SSE1 allowed or AVX_REGS if AVX allowed
28106 if (!Subtarget->hasSSE1()) break;
28108 switch (VT.SimpleTy) {
28110 // Scalar SSE types.
28113 return std::make_pair(0U, &X86::FR32RegClass);
28116 return std::make_pair(0U, &X86::FR64RegClass);
28117 // TODO: Handle f128 and i128 in FR128RegClass after it is tested well.
28125 return std::make_pair(0U, &X86::VR128RegClass);
28133 return std::make_pair(0U, &X86::VR256RegClass);
28138 return std::make_pair(0U, &X86::VR512RegClass);
28144 // Use the default implementation in TargetLowering to convert the register
28145 // constraint into a member of a register class.
28146 std::pair<unsigned, const TargetRegisterClass*> Res;
28147 Res = TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
28149 // Not found as a standard register?
28151 // Map st(0) -> st(7) -> ST0
28152 if (Constraint.size() == 7 && Constraint[0] == '{' &&
28153 tolower(Constraint[1]) == 's' &&
28154 tolower(Constraint[2]) == 't' &&
28155 Constraint[3] == '(' &&
28156 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
28157 Constraint[5] == ')' &&
28158 Constraint[6] == '}') {
28160 Res.first = X86::FP0+Constraint[4]-'0';
28161 Res.second = &X86::RFP80RegClass;
28165 // GCC allows "st(0)" to be called just plain "st".
28166 if (StringRef("{st}").equals_lower(Constraint)) {
28167 Res.first = X86::FP0;
28168 Res.second = &X86::RFP80RegClass;
28173 if (StringRef("{flags}").equals_lower(Constraint)) {
28174 Res.first = X86::EFLAGS;
28175 Res.second = &X86::CCRRegClass;
28179 // 'A' means EAX + EDX.
28180 if (Constraint == "A") {
28181 Res.first = X86::EAX;
28182 Res.second = &X86::GR32_ADRegClass;
28188 // Otherwise, check to see if this is a register class of the wrong value
28189 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
28190 // turn into {ax},{dx}.
28191 // MVT::Other is used to specify clobber names.
28192 if (Res.second->hasType(VT) || VT == MVT::Other)
28193 return Res; // Correct type already, nothing to do.
28195 // Get a matching integer of the correct size. i.e. "ax" with MVT::32 should
28196 // return "eax". This should even work for things like getting 64bit integer
28197 // registers when given an f64 type.
28198 const TargetRegisterClass *Class = Res.second;
28199 if (Class == &X86::GR8RegClass || Class == &X86::GR16RegClass ||
28200 Class == &X86::GR32RegClass || Class == &X86::GR64RegClass) {
28201 unsigned Size = VT.getSizeInBits();
28202 MVT::SimpleValueType SimpleTy = Size == 1 || Size == 8 ? MVT::i8
28203 : Size == 16 ? MVT::i16
28204 : Size == 32 ? MVT::i32
28205 : Size == 64 ? MVT::i64
28207 unsigned DestReg = getX86SubSuperRegisterOrZero(Res.first, SimpleTy);
28209 Res.first = DestReg;
28210 Res.second = SimpleTy == MVT::i8 ? &X86::GR8RegClass
28211 : SimpleTy == MVT::i16 ? &X86::GR16RegClass
28212 : SimpleTy == MVT::i32 ? &X86::GR32RegClass
28213 : &X86::GR64RegClass;
28214 assert(Res.second->contains(Res.first) && "Register in register class");
28216 // No register found/type mismatch.
28218 Res.second = nullptr;
28220 } else if (Class == &X86::FR32RegClass || Class == &X86::FR64RegClass ||
28221 Class == &X86::VR128RegClass || Class == &X86::VR256RegClass ||
28222 Class == &X86::FR32XRegClass || Class == &X86::FR64XRegClass ||
28223 Class == &X86::VR128XRegClass || Class == &X86::VR256XRegClass ||
28224 Class == &X86::VR512RegClass) {
28225 // Handle references to XMM physical registers that got mapped into the
28226 // wrong class. This can happen with constraints like {xmm0} where the
28227 // target independent register mapper will just pick the first match it can
28228 // find, ignoring the required type.
28230 // TODO: Handle f128 and i128 in FR128RegClass after it is tested well.
28231 if (VT == MVT::f32 || VT == MVT::i32)
28232 Res.second = &X86::FR32RegClass;
28233 else if (VT == MVT::f64 || VT == MVT::i64)
28234 Res.second = &X86::FR64RegClass;
28235 else if (X86::VR128RegClass.hasType(VT))
28236 Res.second = &X86::VR128RegClass;
28237 else if (X86::VR256RegClass.hasType(VT))
28238 Res.second = &X86::VR256RegClass;
28239 else if (X86::VR512RegClass.hasType(VT))
28240 Res.second = &X86::VR512RegClass;
28242 // Type mismatch and not a clobber: Return an error;
28244 Res.second = nullptr;
28251 int X86TargetLowering::getScalingFactorCost(const DataLayout &DL,
28252 const AddrMode &AM, Type *Ty,
28253 unsigned AS) const {
28254 // Scaling factors are not free at all.
28255 // An indexed folded instruction, i.e., inst (reg1, reg2, scale),
28256 // will take 2 allocations in the out of order engine instead of 1
28257 // for plain addressing mode, i.e. inst (reg1).
28259 // vaddps (%rsi,%drx), %ymm0, %ymm1
28260 // Requires two allocations (one for the load, one for the computation)
28262 // vaddps (%rsi), %ymm0, %ymm1
28263 // Requires just 1 allocation, i.e., freeing allocations for other operations
28264 // and having less micro operations to execute.
28266 // For some X86 architectures, this is even worse because for instance for
28267 // stores, the complex addressing mode forces the instruction to use the
28268 // "load" ports instead of the dedicated "store" port.
28269 // E.g., on Haswell:
28270 // vmovaps %ymm1, (%r8, %rdi) can use port 2 or 3.
28271 // vmovaps %ymm1, (%r8) can use port 2, 3, or 7.
28272 if (isLegalAddressingMode(DL, AM, Ty, AS))
28273 // Scale represents reg2 * scale, thus account for 1
28274 // as soon as we use a second register.
28275 return AM.Scale != 0;
28279 bool X86TargetLowering::isIntDivCheap(EVT VT, AttributeSet Attr) const {
28280 // Integer division on x86 is expensive. However, when aggressively optimizing
28281 // for code size, we prefer to use a div instruction, as it is usually smaller
28282 // than the alternative sequence.
28283 // The exception to this is vector division. Since x86 doesn't have vector
28284 // integer division, leaving the division as-is is a loss even in terms of
28285 // size, because it will have to be scalarized, while the alternative code
28286 // sequence can be performed in vector form.
28287 bool OptSize = Attr.hasAttribute(AttributeSet::FunctionIndex,
28288 Attribute::MinSize);
28289 return OptSize && !VT.isVector();
28292 void X86TargetLowering::markInRegArguments(SelectionDAG &DAG,
28293 TargetLowering::ArgListTy& Args) const {
28294 // The MCU psABI requires some arguments to be passed in-register.
28295 // For regular calls, the inreg arguments are marked by the front-end.
28296 // However, for compiler generated library calls, we have to patch this
28298 if (!Subtarget->isTargetMCU() || !Args.size())
28301 unsigned FreeRegs = 3;
28302 for (auto &Arg : Args) {
28303 // For library functions, we do not expect any fancy types.
28304 unsigned Size = DAG.getDataLayout().getTypeSizeInBits(Arg.Ty);
28305 unsigned SizeInRegs = (Size + 31) / 32;
28306 if (SizeInRegs > 2 || SizeInRegs > FreeRegs)
28309 Arg.isInReg = true;
28310 FreeRegs -= SizeInRegs;