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/LibCallSemantics.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::i8, Expand);
300 setOperationAction(ISD::BR_CC , MVT::i16, Expand);
301 setOperationAction(ISD::BR_CC , MVT::i32, Expand);
302 setOperationAction(ISD::BR_CC , MVT::i64, Expand);
303 setOperationAction(ISD::SELECT_CC , MVT::f32, Expand);
304 setOperationAction(ISD::SELECT_CC , MVT::f64, Expand);
305 setOperationAction(ISD::SELECT_CC , MVT::f80, Expand);
306 setOperationAction(ISD::SELECT_CC , MVT::i8, Expand);
307 setOperationAction(ISD::SELECT_CC , MVT::i16, Expand);
308 setOperationAction(ISD::SELECT_CC , MVT::i32, Expand);
309 setOperationAction(ISD::SELECT_CC , MVT::i64, Expand);
310 if (Subtarget->is64Bit())
311 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
312 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
313 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
314 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
315 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
317 if (Subtarget->is32Bit() && Subtarget->isTargetKnownWindowsMSVC()) {
318 // On 32 bit MSVC, `fmodf(f32)` is not defined - only `fmod(f64)`
319 // is. We should promote the value to 64-bits to solve this.
320 // This is what the CRT headers do - `fmodf` is an inline header
321 // function casting to f64 and calling `fmod`.
322 setOperationAction(ISD::FREM , MVT::f32 , Promote);
324 setOperationAction(ISD::FREM , MVT::f32 , Expand);
327 setOperationAction(ISD::FREM , MVT::f64 , Expand);
328 setOperationAction(ISD::FREM , MVT::f80 , Expand);
329 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
331 // Promote the i8 variants and force them on up to i32 which has a shorter
333 setOperationAction(ISD::CTTZ , MVT::i8 , Promote);
334 AddPromotedToType (ISD::CTTZ , MVT::i8 , MVT::i32);
335 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i8 , Promote);
336 AddPromotedToType (ISD::CTTZ_ZERO_UNDEF , MVT::i8 , MVT::i32);
337 if (Subtarget->hasBMI()) {
338 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i16 , Expand);
339 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32 , Expand);
340 if (Subtarget->is64Bit())
341 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
343 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
344 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
345 if (Subtarget->is64Bit())
346 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
349 if (Subtarget->hasLZCNT()) {
350 // When promoting the i8 variants, force them to i32 for a shorter
352 setOperationAction(ISD::CTLZ , MVT::i8 , Promote);
353 AddPromotedToType (ISD::CTLZ , MVT::i8 , MVT::i32);
354 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Promote);
355 AddPromotedToType (ISD::CTLZ_ZERO_UNDEF, MVT::i8 , MVT::i32);
356 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Expand);
357 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Expand);
358 if (Subtarget->is64Bit())
359 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
361 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
362 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
363 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
364 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Custom);
365 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Custom);
366 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Custom);
367 if (Subtarget->is64Bit()) {
368 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
369 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Custom);
373 // Special handling for half-precision floating point conversions.
374 // If we don't have F16C support, then lower half float conversions
375 // into library calls.
376 if (Subtarget->useSoftFloat() || !Subtarget->hasF16C()) {
377 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
378 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
381 // There's never any support for operations beyond MVT::f32.
382 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
383 setOperationAction(ISD::FP16_TO_FP, MVT::f80, Expand);
384 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
385 setOperationAction(ISD::FP_TO_FP16, MVT::f80, Expand);
387 setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
388 setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
389 setLoadExtAction(ISD::EXTLOAD, MVT::f80, MVT::f16, Expand);
390 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
391 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
392 setTruncStoreAction(MVT::f80, MVT::f16, Expand);
394 if (Subtarget->hasPOPCNT()) {
395 setOperationAction(ISD::CTPOP , MVT::i8 , Promote);
397 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
398 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
399 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
400 if (Subtarget->is64Bit())
401 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
404 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
406 if (!Subtarget->hasMOVBE())
407 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
409 // These should be promoted to a larger select which is supported.
410 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
411 // X86 wants to expand cmov itself.
412 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
413 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
414 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
415 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
416 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
417 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
418 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
419 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
420 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
421 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
422 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
423 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
424 setOperationAction(ISD::SETCCE , MVT::i8 , Custom);
425 setOperationAction(ISD::SETCCE , MVT::i16 , Custom);
426 setOperationAction(ISD::SETCCE , MVT::i32 , Custom);
427 if (Subtarget->is64Bit()) {
428 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
429 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
430 setOperationAction(ISD::SETCCE , MVT::i64 , Custom);
432 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
433 // NOTE: EH_SJLJ_SETJMP/_LONGJMP supported here is NOT intended to support
434 // SjLj exception handling but a light-weight setjmp/longjmp replacement to
435 // support continuation, user-level threading, and etc.. As a result, no
436 // other SjLj exception interfaces are implemented and please don't build
437 // your own exception handling based on them.
438 // LLVM/Clang supports zero-cost DWARF exception handling.
439 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
440 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
443 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
444 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
445 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
446 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
447 if (Subtarget->is64Bit())
448 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
449 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
450 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
451 if (Subtarget->is64Bit()) {
452 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
453 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
454 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
455 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
456 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
458 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
459 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
460 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
461 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
462 if (Subtarget->is64Bit()) {
463 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
464 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
465 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
468 if (Subtarget->hasSSE1())
469 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
471 setOperationAction(ISD::ATOMIC_FENCE , MVT::Other, Custom);
473 // Expand certain atomics
474 for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) {
475 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Custom);
476 setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
477 setOperationAction(ISD::ATOMIC_STORE, VT, Custom);
480 if (Subtarget->hasCmpxchg16b()) {
481 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i128, Custom);
484 // FIXME - use subtarget debug flags
485 if (!Subtarget->isTargetDarwin() && !Subtarget->isTargetELF() &&
486 !Subtarget->isTargetCygMing() && !Subtarget->isTargetWin64()) {
487 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
490 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
491 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
493 setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
494 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
496 setOperationAction(ISD::TRAP, MVT::Other, Legal);
497 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Legal);
499 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
500 setOperationAction(ISD::VASTART , MVT::Other, Custom);
501 setOperationAction(ISD::VAEND , MVT::Other, Expand);
502 if (Subtarget->is64Bit()) {
503 setOperationAction(ISD::VAARG , MVT::Other, Custom);
504 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
506 // TargetInfo::CharPtrBuiltinVaList
507 setOperationAction(ISD::VAARG , MVT::Other, Expand);
508 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
511 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
512 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
514 setOperationAction(ISD::DYNAMIC_STACKALLOC, PtrVT, Custom);
516 // GC_TRANSITION_START and GC_TRANSITION_END need custom lowering.
517 setOperationAction(ISD::GC_TRANSITION_START, MVT::Other, Custom);
518 setOperationAction(ISD::GC_TRANSITION_END, MVT::Other, Custom);
520 if (!Subtarget->useSoftFloat() && X86ScalarSSEf64) {
521 // f32 and f64 use SSE.
522 // Set up the FP register classes.
523 addRegisterClass(MVT::f32, &X86::FR32RegClass);
524 addRegisterClass(MVT::f64, &X86::FR64RegClass);
526 // Use ANDPD to simulate FABS.
527 setOperationAction(ISD::FABS , MVT::f64, Custom);
528 setOperationAction(ISD::FABS , MVT::f32, Custom);
530 // Use XORP to simulate FNEG.
531 setOperationAction(ISD::FNEG , MVT::f64, Custom);
532 setOperationAction(ISD::FNEG , MVT::f32, Custom);
534 // Use ANDPD and ORPD to simulate FCOPYSIGN.
535 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
536 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
538 // Lower this to FGETSIGNx86 plus an AND.
539 setOperationAction(ISD::FGETSIGN, MVT::i64, Custom);
540 setOperationAction(ISD::FGETSIGN, MVT::i32, Custom);
542 // We don't support sin/cos/fmod
543 setOperationAction(ISD::FSIN , MVT::f64, Expand);
544 setOperationAction(ISD::FCOS , MVT::f64, Expand);
545 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
546 setOperationAction(ISD::FSIN , MVT::f32, Expand);
547 setOperationAction(ISD::FCOS , MVT::f32, Expand);
548 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
550 // Expand FP immediates into loads from the stack, except for the special
552 addLegalFPImmediate(APFloat(+0.0)); // xorpd
553 addLegalFPImmediate(APFloat(+0.0f)); // xorps
554 } else if (!Subtarget->useSoftFloat() && X86ScalarSSEf32) {
555 // Use SSE for f32, x87 for f64.
556 // Set up the FP register classes.
557 addRegisterClass(MVT::f32, &X86::FR32RegClass);
558 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
560 // Use ANDPS to simulate FABS.
561 setOperationAction(ISD::FABS , MVT::f32, Custom);
563 // Use XORP to simulate FNEG.
564 setOperationAction(ISD::FNEG , MVT::f32, Custom);
566 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
568 // Use ANDPS and ORPS to simulate FCOPYSIGN.
569 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
570 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
572 // We don't support sin/cos/fmod
573 setOperationAction(ISD::FSIN , MVT::f32, Expand);
574 setOperationAction(ISD::FCOS , MVT::f32, Expand);
575 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
577 // Special cases we handle for FP constants.
578 addLegalFPImmediate(APFloat(+0.0f)); // xorps
579 addLegalFPImmediate(APFloat(+0.0)); // FLD0
580 addLegalFPImmediate(APFloat(+1.0)); // FLD1
581 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
582 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
584 if (!TM.Options.UnsafeFPMath) {
585 setOperationAction(ISD::FSIN , MVT::f64, Expand);
586 setOperationAction(ISD::FCOS , MVT::f64, Expand);
587 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
589 } else if (!Subtarget->useSoftFloat()) {
590 // f32 and f64 in x87.
591 // Set up the FP register classes.
592 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
593 addRegisterClass(MVT::f32, &X86::RFP32RegClass);
595 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
596 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
597 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
598 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
600 if (!TM.Options.UnsafeFPMath) {
601 setOperationAction(ISD::FSIN , MVT::f64, Expand);
602 setOperationAction(ISD::FSIN , MVT::f32, Expand);
603 setOperationAction(ISD::FCOS , MVT::f64, Expand);
604 setOperationAction(ISD::FCOS , MVT::f32, Expand);
605 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
606 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
608 addLegalFPImmediate(APFloat(+0.0)); // FLD0
609 addLegalFPImmediate(APFloat(+1.0)); // FLD1
610 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
611 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
612 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
613 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
614 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
615 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
618 // We don't support FMA.
619 setOperationAction(ISD::FMA, MVT::f64, Expand);
620 setOperationAction(ISD::FMA, MVT::f32, Expand);
622 // Long double always uses X87.
623 if (!Subtarget->useSoftFloat()) {
624 addRegisterClass(MVT::f80, &X86::RFP80RegClass);
625 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
626 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
628 APFloat TmpFlt = APFloat::getZero(APFloat::x87DoubleExtended);
629 addLegalFPImmediate(TmpFlt); // FLD0
631 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
634 APFloat TmpFlt2(+1.0);
635 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
637 addLegalFPImmediate(TmpFlt2); // FLD1
638 TmpFlt2.changeSign();
639 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
642 if (!TM.Options.UnsafeFPMath) {
643 setOperationAction(ISD::FSIN , MVT::f80, Expand);
644 setOperationAction(ISD::FCOS , MVT::f80, Expand);
645 setOperationAction(ISD::FSINCOS, MVT::f80, Expand);
648 setOperationAction(ISD::FFLOOR, MVT::f80, Expand);
649 setOperationAction(ISD::FCEIL, MVT::f80, Expand);
650 setOperationAction(ISD::FTRUNC, MVT::f80, Expand);
651 setOperationAction(ISD::FRINT, MVT::f80, Expand);
652 setOperationAction(ISD::FNEARBYINT, MVT::f80, Expand);
653 setOperationAction(ISD::FMA, MVT::f80, Expand);
656 // Always use a library call for pow.
657 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
658 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
659 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
661 setOperationAction(ISD::FLOG, MVT::f80, Expand);
662 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
663 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
664 setOperationAction(ISD::FEXP, MVT::f80, Expand);
665 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
666 setOperationAction(ISD::FMINNUM, MVT::f80, Expand);
667 setOperationAction(ISD::FMAXNUM, MVT::f80, Expand);
669 // First set operation action for all vector types to either promote
670 // (for widening) or expand (for scalarization). Then we will selectively
671 // turn on ones that can be effectively codegen'd.
672 for (MVT VT : MVT::vector_valuetypes()) {
673 setOperationAction(ISD::ADD , VT, Expand);
674 setOperationAction(ISD::SUB , VT, Expand);
675 setOperationAction(ISD::FADD, VT, Expand);
676 setOperationAction(ISD::FNEG, VT, Expand);
677 setOperationAction(ISD::FSUB, VT, Expand);
678 setOperationAction(ISD::MUL , VT, Expand);
679 setOperationAction(ISD::FMUL, VT, Expand);
680 setOperationAction(ISD::SDIV, VT, Expand);
681 setOperationAction(ISD::UDIV, VT, Expand);
682 setOperationAction(ISD::FDIV, VT, Expand);
683 setOperationAction(ISD::SREM, VT, Expand);
684 setOperationAction(ISD::UREM, VT, Expand);
685 setOperationAction(ISD::LOAD, VT, Expand);
686 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
687 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT,Expand);
688 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
689 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT,Expand);
690 setOperationAction(ISD::INSERT_SUBVECTOR, VT,Expand);
691 setOperationAction(ISD::FABS, VT, Expand);
692 setOperationAction(ISD::FSIN, VT, Expand);
693 setOperationAction(ISD::FSINCOS, VT, Expand);
694 setOperationAction(ISD::FCOS, VT, Expand);
695 setOperationAction(ISD::FSINCOS, VT, Expand);
696 setOperationAction(ISD::FREM, VT, Expand);
697 setOperationAction(ISD::FMA, VT, Expand);
698 setOperationAction(ISD::FPOWI, VT, Expand);
699 setOperationAction(ISD::FSQRT, VT, Expand);
700 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
701 setOperationAction(ISD::FFLOOR, VT, Expand);
702 setOperationAction(ISD::FCEIL, VT, Expand);
703 setOperationAction(ISD::FTRUNC, VT, Expand);
704 setOperationAction(ISD::FRINT, VT, Expand);
705 setOperationAction(ISD::FNEARBYINT, VT, Expand);
706 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
707 setOperationAction(ISD::MULHS, VT, Expand);
708 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
709 setOperationAction(ISD::MULHU, VT, Expand);
710 setOperationAction(ISD::SDIVREM, VT, Expand);
711 setOperationAction(ISD::UDIVREM, VT, Expand);
712 setOperationAction(ISD::FPOW, VT, Expand);
713 setOperationAction(ISD::CTPOP, VT, Expand);
714 setOperationAction(ISD::CTTZ, VT, Expand);
715 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
716 setOperationAction(ISD::CTLZ, VT, Expand);
717 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
718 setOperationAction(ISD::SHL, VT, Expand);
719 setOperationAction(ISD::SRA, VT, Expand);
720 setOperationAction(ISD::SRL, VT, Expand);
721 setOperationAction(ISD::ROTL, VT, Expand);
722 setOperationAction(ISD::ROTR, VT, Expand);
723 setOperationAction(ISD::BSWAP, VT, Expand);
724 setOperationAction(ISD::SETCC, VT, Expand);
725 setOperationAction(ISD::FLOG, VT, Expand);
726 setOperationAction(ISD::FLOG2, VT, Expand);
727 setOperationAction(ISD::FLOG10, VT, Expand);
728 setOperationAction(ISD::FEXP, VT, Expand);
729 setOperationAction(ISD::FEXP2, VT, Expand);
730 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
731 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
732 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
733 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
734 setOperationAction(ISD::SIGN_EXTEND_INREG, VT,Expand);
735 setOperationAction(ISD::TRUNCATE, VT, Expand);
736 setOperationAction(ISD::SIGN_EXTEND, VT, Expand);
737 setOperationAction(ISD::ZERO_EXTEND, VT, Expand);
738 setOperationAction(ISD::ANY_EXTEND, VT, Expand);
739 setOperationAction(ISD::VSELECT, VT, Expand);
740 setOperationAction(ISD::SELECT_CC, VT, Expand);
741 for (MVT InnerVT : MVT::vector_valuetypes()) {
742 setTruncStoreAction(InnerVT, VT, Expand);
744 setLoadExtAction(ISD::SEXTLOAD, InnerVT, VT, Expand);
745 setLoadExtAction(ISD::ZEXTLOAD, InnerVT, VT, Expand);
747 // N.b. ISD::EXTLOAD legality is basically ignored except for i1-like
748 // types, we have to deal with them whether we ask for Expansion or not.
749 // Setting Expand causes its own optimisation problems though, so leave
751 if (VT.getVectorElementType() == MVT::i1)
752 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
754 // EXTLOAD for MVT::f16 vectors is not legal because f16 vectors are
755 // split/scalarized right now.
756 if (VT.getVectorElementType() == MVT::f16)
757 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
761 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
762 // with -msoft-float, disable use of MMX as well.
763 if (!Subtarget->useSoftFloat() && Subtarget->hasMMX()) {
764 addRegisterClass(MVT::x86mmx, &X86::VR64RegClass);
765 // No operations on x86mmx supported, everything uses intrinsics.
768 // MMX-sized vectors (other than x86mmx) are expected to be expanded
769 // into smaller operations.
770 for (MVT MMXTy : {MVT::v8i8, MVT::v4i16, MVT::v2i32, MVT::v1i64}) {
771 setOperationAction(ISD::MULHS, MMXTy, Expand);
772 setOperationAction(ISD::AND, MMXTy, Expand);
773 setOperationAction(ISD::OR, MMXTy, Expand);
774 setOperationAction(ISD::XOR, MMXTy, Expand);
775 setOperationAction(ISD::SCALAR_TO_VECTOR, MMXTy, Expand);
776 setOperationAction(ISD::SELECT, MMXTy, Expand);
777 setOperationAction(ISD::BITCAST, MMXTy, Expand);
779 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v1i64, Expand);
781 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE1()) {
782 addRegisterClass(MVT::v4f32, &X86::VR128RegClass);
784 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
785 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
786 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
787 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
788 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
789 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
790 setOperationAction(ISD::FABS, MVT::v4f32, Custom);
791 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
792 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
793 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
794 setOperationAction(ISD::VSELECT, MVT::v4f32, Custom);
795 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
796 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
797 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Custom);
800 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE2()) {
801 addRegisterClass(MVT::v2f64, &X86::VR128RegClass);
803 // FIXME: Unfortunately, -soft-float and -no-implicit-float mean XMM
804 // registers cannot be used even for integer operations.
805 addRegisterClass(MVT::v16i8, &X86::VR128RegClass);
806 addRegisterClass(MVT::v8i16, &X86::VR128RegClass);
807 addRegisterClass(MVT::v4i32, &X86::VR128RegClass);
808 addRegisterClass(MVT::v2i64, &X86::VR128RegClass);
810 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
811 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
812 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
813 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
814 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
815 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
816 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
817 setOperationAction(ISD::UMUL_LOHI, MVT::v4i32, Custom);
818 setOperationAction(ISD::SMUL_LOHI, MVT::v4i32, Custom);
819 setOperationAction(ISD::MULHU, MVT::v8i16, Legal);
820 setOperationAction(ISD::MULHS, MVT::v8i16, Legal);
821 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
822 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
823 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
824 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
825 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
826 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
827 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
828 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
829 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
830 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
831 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
832 setOperationAction(ISD::FABS, MVT::v2f64, Custom);
834 setOperationAction(ISD::SMAX, MVT::v8i16, Legal);
835 setOperationAction(ISD::UMAX, MVT::v16i8, Legal);
836 setOperationAction(ISD::SMIN, MVT::v8i16, Legal);
837 setOperationAction(ISD::UMIN, MVT::v16i8, Legal);
839 setOperationAction(ISD::SETCC, MVT::v2i64, Custom);
840 setOperationAction(ISD::SETCC, MVT::v16i8, Custom);
841 setOperationAction(ISD::SETCC, MVT::v8i16, Custom);
842 setOperationAction(ISD::SETCC, MVT::v4i32, Custom);
844 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
845 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
846 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
847 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
848 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
850 setOperationAction(ISD::CTPOP, MVT::v16i8, Custom);
851 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom);
852 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
853 setOperationAction(ISD::CTPOP, MVT::v2i64, Custom);
855 setOperationAction(ISD::CTTZ, MVT::v16i8, Custom);
856 setOperationAction(ISD::CTTZ, MVT::v8i16, Custom);
857 setOperationAction(ISD::CTTZ, MVT::v4i32, Custom);
858 // ISD::CTTZ v2i64 - scalarization is faster.
859 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i8, Custom);
860 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i16, Custom);
861 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i32, Custom);
862 // ISD::CTTZ_ZERO_UNDEF v2i64 - scalarization is faster.
864 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
865 for (auto VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32 }) {
866 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
867 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
868 setOperationAction(ISD::VSELECT, VT, Custom);
869 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
872 // We support custom legalizing of sext and anyext loads for specific
873 // memory vector types which we can load as a scalar (or sequence of
874 // scalars) and extend in-register to a legal 128-bit vector type. For sext
875 // loads these must work with a single scalar load.
876 for (MVT VT : MVT::integer_vector_valuetypes()) {
877 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i8, Custom);
878 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i16, Custom);
879 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v8i8, Custom);
880 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i8, Custom);
881 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i16, Custom);
882 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i32, Custom);
883 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i8, Custom);
884 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i16, Custom);
885 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8i8, Custom);
888 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
889 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
890 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
891 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
892 setOperationAction(ISD::VSELECT, MVT::v2f64, Custom);
893 setOperationAction(ISD::VSELECT, MVT::v2i64, Custom);
894 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
895 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
897 if (Subtarget->is64Bit()) {
898 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
899 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
902 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
903 for (auto VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32 }) {
904 setOperationAction(ISD::AND, VT, Promote);
905 AddPromotedToType (ISD::AND, VT, MVT::v2i64);
906 setOperationAction(ISD::OR, VT, Promote);
907 AddPromotedToType (ISD::OR, VT, MVT::v2i64);
908 setOperationAction(ISD::XOR, VT, Promote);
909 AddPromotedToType (ISD::XOR, VT, MVT::v2i64);
910 setOperationAction(ISD::LOAD, VT, Promote);
911 AddPromotedToType (ISD::LOAD, VT, MVT::v2i64);
912 setOperationAction(ISD::SELECT, VT, Promote);
913 AddPromotedToType (ISD::SELECT, VT, MVT::v2i64);
916 // Custom lower v2i64 and v2f64 selects.
917 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
918 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
919 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
920 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
922 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
923 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
925 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
927 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Custom);
928 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
929 // As there is no 64-bit GPR available, we need build a special custom
930 // sequence to convert from v2i32 to v2f32.
931 if (!Subtarget->is64Bit())
932 setOperationAction(ISD::UINT_TO_FP, MVT::v2f32, Custom);
934 setOperationAction(ISD::FP_EXTEND, MVT::v2f32, Custom);
935 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Custom);
937 for (MVT VT : MVT::fp_vector_valuetypes())
938 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2f32, Legal);
940 setOperationAction(ISD::BITCAST, MVT::v2i32, Custom);
941 setOperationAction(ISD::BITCAST, MVT::v4i16, Custom);
942 setOperationAction(ISD::BITCAST, MVT::v8i8, Custom);
945 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE41()) {
946 for (MVT RoundedTy : {MVT::f32, MVT::f64, MVT::v4f32, MVT::v2f64}) {
947 setOperationAction(ISD::FFLOOR, RoundedTy, Legal);
948 setOperationAction(ISD::FCEIL, RoundedTy, Legal);
949 setOperationAction(ISD::FTRUNC, RoundedTy, Legal);
950 setOperationAction(ISD::FRINT, RoundedTy, Legal);
951 setOperationAction(ISD::FNEARBYINT, RoundedTy, Legal);
954 setOperationAction(ISD::SMAX, MVT::v16i8, Legal);
955 setOperationAction(ISD::SMAX, MVT::v4i32, Legal);
956 setOperationAction(ISD::UMAX, MVT::v8i16, Legal);
957 setOperationAction(ISD::UMAX, MVT::v4i32, Legal);
958 setOperationAction(ISD::SMIN, MVT::v16i8, Legal);
959 setOperationAction(ISD::SMIN, MVT::v4i32, Legal);
960 setOperationAction(ISD::UMIN, MVT::v8i16, Legal);
961 setOperationAction(ISD::UMIN, MVT::v4i32, Legal);
963 // FIXME: Do we need to handle scalar-to-vector here?
964 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
966 // We directly match byte blends in the backend as they match the VSELECT
968 setOperationAction(ISD::VSELECT, MVT::v16i8, Legal);
970 // SSE41 brings specific instructions for doing vector sign extend even in
971 // cases where we don't have SRA.
972 for (MVT VT : MVT::integer_vector_valuetypes()) {
973 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i8, Custom);
974 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i16, Custom);
975 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i32, Custom);
978 // SSE41 also has vector sign/zero extending loads, PMOV[SZ]X
979 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
980 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
981 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
982 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
983 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
984 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
986 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
987 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
988 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
989 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
990 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
991 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
993 // i8 and i16 vectors are custom because the source register and source
994 // source memory operand types are not the same width. f32 vectors are
995 // custom since the immediate controlling the insert encodes additional
997 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
998 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
999 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
1000 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
1002 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
1003 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
1004 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
1005 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
1007 // FIXME: these should be Legal, but that's only for the case where
1008 // the index is constant. For now custom expand to deal with that.
1009 if (Subtarget->is64Bit()) {
1010 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
1011 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
1015 if (Subtarget->hasSSE2()) {
1016 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v2i64, Custom);
1017 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v4i32, Custom);
1018 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v8i16, Custom);
1020 setOperationAction(ISD::SRL, MVT::v8i16, Custom);
1021 setOperationAction(ISD::SRL, MVT::v16i8, Custom);
1023 setOperationAction(ISD::SHL, MVT::v8i16, Custom);
1024 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
1026 setOperationAction(ISD::SRA, MVT::v8i16, Custom);
1027 setOperationAction(ISD::SRA, MVT::v16i8, Custom);
1029 // In the customized shift lowering, the legal cases in AVX2 will be
1031 setOperationAction(ISD::SRL, MVT::v2i64, Custom);
1032 setOperationAction(ISD::SRL, MVT::v4i32, Custom);
1034 setOperationAction(ISD::SHL, MVT::v2i64, Custom);
1035 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
1037 setOperationAction(ISD::SRA, MVT::v2i64, Custom);
1038 setOperationAction(ISD::SRA, MVT::v4i32, Custom);
1041 if (Subtarget->hasXOP()) {
1042 setOperationAction(ISD::ROTL, MVT::v16i8, Custom);
1043 setOperationAction(ISD::ROTL, MVT::v8i16, Custom);
1044 setOperationAction(ISD::ROTL, MVT::v4i32, Custom);
1045 setOperationAction(ISD::ROTL, MVT::v2i64, Custom);
1046 setOperationAction(ISD::ROTL, MVT::v32i8, Custom);
1047 setOperationAction(ISD::ROTL, MVT::v16i16, Custom);
1048 setOperationAction(ISD::ROTL, MVT::v8i32, Custom);
1049 setOperationAction(ISD::ROTL, MVT::v4i64, Custom);
1052 if (!Subtarget->useSoftFloat() && Subtarget->hasFp256()) {
1053 addRegisterClass(MVT::v32i8, &X86::VR256RegClass);
1054 addRegisterClass(MVT::v16i16, &X86::VR256RegClass);
1055 addRegisterClass(MVT::v8i32, &X86::VR256RegClass);
1056 addRegisterClass(MVT::v8f32, &X86::VR256RegClass);
1057 addRegisterClass(MVT::v4i64, &X86::VR256RegClass);
1058 addRegisterClass(MVT::v4f64, &X86::VR256RegClass);
1060 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
1061 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
1062 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
1064 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
1065 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
1066 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
1067 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
1068 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
1069 setOperationAction(ISD::FFLOOR, MVT::v8f32, Legal);
1070 setOperationAction(ISD::FCEIL, MVT::v8f32, Legal);
1071 setOperationAction(ISD::FTRUNC, MVT::v8f32, Legal);
1072 setOperationAction(ISD::FRINT, MVT::v8f32, Legal);
1073 setOperationAction(ISD::FNEARBYINT, MVT::v8f32, Legal);
1074 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
1075 setOperationAction(ISD::FABS, MVT::v8f32, Custom);
1077 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
1078 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
1079 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
1080 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
1081 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
1082 setOperationAction(ISD::FFLOOR, MVT::v4f64, Legal);
1083 setOperationAction(ISD::FCEIL, MVT::v4f64, Legal);
1084 setOperationAction(ISD::FTRUNC, MVT::v4f64, Legal);
1085 setOperationAction(ISD::FRINT, MVT::v4f64, Legal);
1086 setOperationAction(ISD::FNEARBYINT, MVT::v4f64, Legal);
1087 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
1088 setOperationAction(ISD::FABS, MVT::v4f64, Custom);
1090 // (fp_to_int:v8i16 (v8f32 ..)) requires the result type to be promoted
1091 // even though v8i16 is a legal type.
1092 setOperationAction(ISD::FP_TO_SINT, MVT::v8i16, Promote);
1093 setOperationAction(ISD::FP_TO_UINT, MVT::v8i16, Promote);
1094 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1096 setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Promote);
1097 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1098 setOperationAction(ISD::FP_ROUND, MVT::v4f32, Legal);
1100 setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Custom);
1101 setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Custom);
1103 for (MVT VT : MVT::fp_vector_valuetypes())
1104 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4f32, Legal);
1106 setOperationAction(ISD::SRL, MVT::v16i16, Custom);
1107 setOperationAction(ISD::SRL, MVT::v32i8, Custom);
1109 setOperationAction(ISD::SHL, MVT::v16i16, Custom);
1110 setOperationAction(ISD::SHL, MVT::v32i8, Custom);
1112 setOperationAction(ISD::SRA, MVT::v16i16, Custom);
1113 setOperationAction(ISD::SRA, MVT::v32i8, Custom);
1115 setOperationAction(ISD::SETCC, MVT::v32i8, Custom);
1116 setOperationAction(ISD::SETCC, MVT::v16i16, Custom);
1117 setOperationAction(ISD::SETCC, MVT::v8i32, Custom);
1118 setOperationAction(ISD::SETCC, MVT::v4i64, Custom);
1120 setOperationAction(ISD::SELECT, MVT::v4f64, Custom);
1121 setOperationAction(ISD::SELECT, MVT::v4i64, Custom);
1122 setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
1124 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i64, Custom);
1125 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i32, Custom);
1126 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1127 setOperationAction(ISD::ZERO_EXTEND, MVT::v4i64, Custom);
1128 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i32, Custom);
1129 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i16, Custom);
1130 setOperationAction(ISD::ANY_EXTEND, MVT::v4i64, Custom);
1131 setOperationAction(ISD::ANY_EXTEND, MVT::v8i32, Custom);
1132 setOperationAction(ISD::ANY_EXTEND, MVT::v16i16, Custom);
1133 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1134 setOperationAction(ISD::TRUNCATE, MVT::v8i16, Custom);
1135 setOperationAction(ISD::TRUNCATE, MVT::v4i32, Custom);
1137 setOperationAction(ISD::CTPOP, MVT::v32i8, Custom);
1138 setOperationAction(ISD::CTPOP, MVT::v16i16, Custom);
1139 setOperationAction(ISD::CTPOP, MVT::v8i32, Custom);
1140 setOperationAction(ISD::CTPOP, MVT::v4i64, Custom);
1142 setOperationAction(ISD::CTTZ, MVT::v32i8, Custom);
1143 setOperationAction(ISD::CTTZ, MVT::v16i16, Custom);
1144 setOperationAction(ISD::CTTZ, MVT::v8i32, Custom);
1145 setOperationAction(ISD::CTTZ, MVT::v4i64, Custom);
1146 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v32i8, Custom);
1147 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i16, Custom);
1148 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i32, Custom);
1149 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i64, Custom);
1151 if (Subtarget->hasFMA() || Subtarget->hasFMA4() || Subtarget->hasAVX512()) {
1152 setOperationAction(ISD::FMA, MVT::v8f32, Legal);
1153 setOperationAction(ISD::FMA, MVT::v4f64, Legal);
1154 setOperationAction(ISD::FMA, MVT::v4f32, Legal);
1155 setOperationAction(ISD::FMA, MVT::v2f64, Legal);
1156 setOperationAction(ISD::FMA, MVT::f32, Legal);
1157 setOperationAction(ISD::FMA, MVT::f64, Legal);
1160 if (Subtarget->hasInt256()) {
1161 setOperationAction(ISD::ADD, MVT::v4i64, Legal);
1162 setOperationAction(ISD::ADD, MVT::v8i32, Legal);
1163 setOperationAction(ISD::ADD, MVT::v16i16, Legal);
1164 setOperationAction(ISD::ADD, MVT::v32i8, Legal);
1166 setOperationAction(ISD::SUB, MVT::v4i64, Legal);
1167 setOperationAction(ISD::SUB, MVT::v8i32, Legal);
1168 setOperationAction(ISD::SUB, MVT::v16i16, Legal);
1169 setOperationAction(ISD::SUB, MVT::v32i8, Legal);
1171 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1172 setOperationAction(ISD::MUL, MVT::v8i32, Legal);
1173 setOperationAction(ISD::MUL, MVT::v16i16, Legal);
1174 setOperationAction(ISD::MUL, MVT::v32i8, Custom);
1176 setOperationAction(ISD::UMUL_LOHI, MVT::v8i32, Custom);
1177 setOperationAction(ISD::SMUL_LOHI, MVT::v8i32, Custom);
1178 setOperationAction(ISD::MULHU, MVT::v16i16, Legal);
1179 setOperationAction(ISD::MULHS, MVT::v16i16, Legal);
1181 setOperationAction(ISD::SMAX, MVT::v32i8, Legal);
1182 setOperationAction(ISD::SMAX, MVT::v16i16, Legal);
1183 setOperationAction(ISD::SMAX, MVT::v8i32, Legal);
1184 setOperationAction(ISD::UMAX, MVT::v32i8, Legal);
1185 setOperationAction(ISD::UMAX, MVT::v16i16, Legal);
1186 setOperationAction(ISD::UMAX, MVT::v8i32, Legal);
1187 setOperationAction(ISD::SMIN, MVT::v32i8, Legal);
1188 setOperationAction(ISD::SMIN, MVT::v16i16, Legal);
1189 setOperationAction(ISD::SMIN, MVT::v8i32, Legal);
1190 setOperationAction(ISD::UMIN, MVT::v32i8, Legal);
1191 setOperationAction(ISD::UMIN, MVT::v16i16, Legal);
1192 setOperationAction(ISD::UMIN, MVT::v8i32, Legal);
1194 // The custom lowering for UINT_TO_FP for v8i32 becomes interesting
1195 // when we have a 256bit-wide blend with immediate.
1196 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Custom);
1198 // AVX2 also has wider vector sign/zero extending loads, VPMOV[SZ]X
1199 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1200 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1201 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1202 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1203 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1204 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1206 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1207 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1208 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1209 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1210 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1211 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1213 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
1214 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
1215 setOperationAction(ISD::ADD, MVT::v16i16, Custom);
1216 setOperationAction(ISD::ADD, MVT::v32i8, Custom);
1218 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
1219 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
1220 setOperationAction(ISD::SUB, MVT::v16i16, Custom);
1221 setOperationAction(ISD::SUB, MVT::v32i8, Custom);
1223 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1224 setOperationAction(ISD::MUL, MVT::v8i32, Custom);
1225 setOperationAction(ISD::MUL, MVT::v16i16, Custom);
1226 setOperationAction(ISD::MUL, MVT::v32i8, Custom);
1228 setOperationAction(ISD::SMAX, MVT::v32i8, Custom);
1229 setOperationAction(ISD::SMAX, MVT::v16i16, Custom);
1230 setOperationAction(ISD::SMAX, MVT::v8i32, Custom);
1231 setOperationAction(ISD::UMAX, MVT::v32i8, Custom);
1232 setOperationAction(ISD::UMAX, MVT::v16i16, Custom);
1233 setOperationAction(ISD::UMAX, MVT::v8i32, Custom);
1234 setOperationAction(ISD::SMIN, MVT::v32i8, Custom);
1235 setOperationAction(ISD::SMIN, MVT::v16i16, Custom);
1236 setOperationAction(ISD::SMIN, MVT::v8i32, Custom);
1237 setOperationAction(ISD::UMIN, MVT::v32i8, Custom);
1238 setOperationAction(ISD::UMIN, MVT::v16i16, Custom);
1239 setOperationAction(ISD::UMIN, MVT::v8i32, Custom);
1242 // In the customized shift lowering, the legal cases in AVX2 will be
1244 setOperationAction(ISD::SRL, MVT::v4i64, Custom);
1245 setOperationAction(ISD::SRL, MVT::v8i32, Custom);
1247 setOperationAction(ISD::SHL, MVT::v4i64, Custom);
1248 setOperationAction(ISD::SHL, MVT::v8i32, Custom);
1250 setOperationAction(ISD::SRA, MVT::v4i64, Custom);
1251 setOperationAction(ISD::SRA, MVT::v8i32, Custom);
1253 // Custom lower several nodes for 256-bit types.
1254 for (MVT VT : MVT::vector_valuetypes()) {
1255 if (VT.getScalarSizeInBits() >= 32) {
1256 setOperationAction(ISD::MLOAD, VT, Legal);
1257 setOperationAction(ISD::MSTORE, VT, Legal);
1259 // Extract subvector is special because the value type
1260 // (result) is 128-bit but the source is 256-bit wide.
1261 if (VT.is128BitVector()) {
1262 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1264 // Do not attempt to custom lower other non-256-bit vectors
1265 if (!VT.is256BitVector())
1268 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1269 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1270 setOperationAction(ISD::VSELECT, VT, Custom);
1271 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1272 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1273 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1274 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1275 setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
1278 if (Subtarget->hasInt256())
1279 setOperationAction(ISD::VSELECT, MVT::v32i8, Legal);
1281 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
1282 for (auto VT : { MVT::v32i8, MVT::v16i16, MVT::v8i32 }) {
1283 setOperationAction(ISD::AND, VT, Promote);
1284 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
1285 setOperationAction(ISD::OR, VT, Promote);
1286 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
1287 setOperationAction(ISD::XOR, VT, Promote);
1288 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
1289 setOperationAction(ISD::LOAD, VT, Promote);
1290 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
1291 setOperationAction(ISD::SELECT, VT, Promote);
1292 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
1296 if (!Subtarget->useSoftFloat() && Subtarget->hasAVX512()) {
1297 addRegisterClass(MVT::v16i32, &X86::VR512RegClass);
1298 addRegisterClass(MVT::v16f32, &X86::VR512RegClass);
1299 addRegisterClass(MVT::v8i64, &X86::VR512RegClass);
1300 addRegisterClass(MVT::v8f64, &X86::VR512RegClass);
1302 addRegisterClass(MVT::i1, &X86::VK1RegClass);
1303 addRegisterClass(MVT::v8i1, &X86::VK8RegClass);
1304 addRegisterClass(MVT::v16i1, &X86::VK16RegClass);
1306 for (MVT VT : MVT::fp_vector_valuetypes())
1307 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8f32, Legal);
1309 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i32, MVT::v16i8, Legal);
1310 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i32, MVT::v16i8, Legal);
1311 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i32, MVT::v16i16, Legal);
1312 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i32, MVT::v16i16, Legal);
1313 setLoadExtAction(ISD::ZEXTLOAD, MVT::v32i16, MVT::v32i8, Legal);
1314 setLoadExtAction(ISD::SEXTLOAD, MVT::v32i16, MVT::v32i8, Legal);
1315 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i8, Legal);
1316 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i8, Legal);
1317 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i16, Legal);
1318 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i16, Legal);
1319 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i32, Legal);
1320 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i32, Legal);
1322 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
1323 setOperationAction(ISD::SETCC, MVT::i1, Custom);
1324 setOperationAction(ISD::SELECT_CC, MVT::i1, Expand);
1325 setOperationAction(ISD::XOR, MVT::i1, Legal);
1326 setOperationAction(ISD::OR, MVT::i1, Legal);
1327 setOperationAction(ISD::AND, MVT::i1, Legal);
1328 setOperationAction(ISD::SUB, MVT::i1, Custom);
1329 setOperationAction(ISD::ADD, MVT::i1, Custom);
1330 setOperationAction(ISD::MUL, MVT::i1, Custom);
1331 setOperationAction(ISD::LOAD, MVT::v16f32, Legal);
1332 setOperationAction(ISD::LOAD, MVT::v8f64, Legal);
1333 setOperationAction(ISD::LOAD, MVT::v8i64, Legal);
1334 setOperationAction(ISD::LOAD, MVT::v16i32, Legal);
1335 setOperationAction(ISD::LOAD, MVT::v16i1, Legal);
1337 setOperationAction(ISD::FADD, MVT::v16f32, Legal);
1338 setOperationAction(ISD::FSUB, MVT::v16f32, Legal);
1339 setOperationAction(ISD::FMUL, MVT::v16f32, Legal);
1340 setOperationAction(ISD::FDIV, MVT::v16f32, Legal);
1341 setOperationAction(ISD::FSQRT, MVT::v16f32, Legal);
1342 setOperationAction(ISD::FNEG, MVT::v16f32, Custom);
1344 setOperationAction(ISD::FADD, MVT::v8f64, Legal);
1345 setOperationAction(ISD::FSUB, MVT::v8f64, Legal);
1346 setOperationAction(ISD::FMUL, MVT::v8f64, Legal);
1347 setOperationAction(ISD::FDIV, MVT::v8f64, Legal);
1348 setOperationAction(ISD::FSQRT, MVT::v8f64, Legal);
1349 setOperationAction(ISD::FNEG, MVT::v8f64, Custom);
1350 setOperationAction(ISD::FMA, MVT::v8f64, Legal);
1351 setOperationAction(ISD::FMA, MVT::v16f32, Legal);
1353 setOperationAction(ISD::FP_TO_SINT, MVT::v16i32, Legal);
1354 setOperationAction(ISD::FP_TO_UINT, MVT::v16i32, Legal);
1355 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1356 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1357 setOperationAction(ISD::SINT_TO_FP, MVT::v16i32, Legal);
1358 setOperationAction(ISD::SINT_TO_FP, MVT::v8i1, Custom);
1359 setOperationAction(ISD::SINT_TO_FP, MVT::v16i1, Custom);
1360 setOperationAction(ISD::SINT_TO_FP, MVT::v16i8, Promote);
1361 setOperationAction(ISD::SINT_TO_FP, MVT::v16i16, Promote);
1362 setOperationAction(ISD::UINT_TO_FP, MVT::v16i32, Legal);
1363 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1364 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1365 setOperationAction(ISD::UINT_TO_FP, MVT::v16i8, Custom);
1366 setOperationAction(ISD::UINT_TO_FP, MVT::v16i16, Custom);
1367 setOperationAction(ISD::FP_ROUND, MVT::v8f32, Legal);
1368 setOperationAction(ISD::FP_EXTEND, MVT::v8f32, Legal);
1370 setTruncStoreAction(MVT::v8i64, MVT::v8i8, Legal);
1371 setTruncStoreAction(MVT::v8i64, MVT::v8i16, Legal);
1372 setTruncStoreAction(MVT::v8i64, MVT::v8i32, Legal);
1373 setTruncStoreAction(MVT::v16i32, MVT::v16i8, Legal);
1374 setTruncStoreAction(MVT::v16i32, MVT::v16i16, Legal);
1375 if (Subtarget->hasVLX()){
1376 setTruncStoreAction(MVT::v4i64, MVT::v4i8, Legal);
1377 setTruncStoreAction(MVT::v4i64, MVT::v4i16, Legal);
1378 setTruncStoreAction(MVT::v4i64, MVT::v4i32, Legal);
1379 setTruncStoreAction(MVT::v8i32, MVT::v8i8, Legal);
1380 setTruncStoreAction(MVT::v8i32, MVT::v8i16, Legal);
1382 setTruncStoreAction(MVT::v2i64, MVT::v2i8, Legal);
1383 setTruncStoreAction(MVT::v2i64, MVT::v2i16, Legal);
1384 setTruncStoreAction(MVT::v2i64, MVT::v2i32, Legal);
1385 setTruncStoreAction(MVT::v4i32, MVT::v4i8, Legal);
1386 setTruncStoreAction(MVT::v4i32, MVT::v4i16, Legal);
1388 setOperationAction(ISD::TRUNCATE, MVT::i1, Custom);
1389 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1390 setOperationAction(ISD::TRUNCATE, MVT::v8i32, Custom);
1391 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i1, Custom);
1392 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i1, Custom);
1393 if (Subtarget->hasDQI()) {
1394 setOperationAction(ISD::TRUNCATE, MVT::v2i1, Custom);
1395 setOperationAction(ISD::TRUNCATE, MVT::v4i1, Custom);
1397 setOperationAction(ISD::SINT_TO_FP, MVT::v8i64, Legal);
1398 setOperationAction(ISD::UINT_TO_FP, MVT::v8i64, Legal);
1399 setOperationAction(ISD::FP_TO_SINT, MVT::v8i64, Legal);
1400 setOperationAction(ISD::FP_TO_UINT, MVT::v8i64, Legal);
1401 if (Subtarget->hasVLX()) {
1402 setOperationAction(ISD::SINT_TO_FP, MVT::v4i64, Legal);
1403 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Legal);
1404 setOperationAction(ISD::UINT_TO_FP, MVT::v4i64, Legal);
1405 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Legal);
1406 setOperationAction(ISD::FP_TO_SINT, MVT::v4i64, Legal);
1407 setOperationAction(ISD::FP_TO_SINT, MVT::v2i64, Legal);
1408 setOperationAction(ISD::FP_TO_UINT, MVT::v4i64, Legal);
1409 setOperationAction(ISD::FP_TO_UINT, MVT::v2i64, Legal);
1412 if (Subtarget->hasVLX()) {
1413 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1414 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1415 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1416 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1417 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
1418 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1419 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
1420 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1422 setOperationAction(ISD::TRUNCATE, MVT::v8i1, Custom);
1423 setOperationAction(ISD::TRUNCATE, MVT::v16i1, Custom);
1424 setOperationAction(ISD::TRUNCATE, MVT::v16i16, Custom);
1425 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i32, Custom);
1426 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i64, Custom);
1427 setOperationAction(ISD::ANY_EXTEND, MVT::v16i32, Custom);
1428 setOperationAction(ISD::ANY_EXTEND, MVT::v8i64, Custom);
1429 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i32, Custom);
1430 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i64, Custom);
1431 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i8, Custom);
1432 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i16, Custom);
1433 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1434 if (Subtarget->hasDQI()) {
1435 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i32, Custom);
1436 setOperationAction(ISD::SIGN_EXTEND, MVT::v2i64, Custom);
1438 setOperationAction(ISD::FFLOOR, MVT::v16f32, Legal);
1439 setOperationAction(ISD::FFLOOR, MVT::v8f64, Legal);
1440 setOperationAction(ISD::FCEIL, MVT::v16f32, Legal);
1441 setOperationAction(ISD::FCEIL, MVT::v8f64, Legal);
1442 setOperationAction(ISD::FTRUNC, MVT::v16f32, Legal);
1443 setOperationAction(ISD::FTRUNC, MVT::v8f64, Legal);
1444 setOperationAction(ISD::FRINT, MVT::v16f32, Legal);
1445 setOperationAction(ISD::FRINT, MVT::v8f64, Legal);
1446 setOperationAction(ISD::FNEARBYINT, MVT::v16f32, Legal);
1447 setOperationAction(ISD::FNEARBYINT, MVT::v8f64, Legal);
1449 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f64, Custom);
1450 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i64, Custom);
1451 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16f32, Custom);
1452 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i32, Custom);
1453 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i1, Legal);
1455 setOperationAction(ISD::SETCC, MVT::v16i1, Custom);
1456 setOperationAction(ISD::SETCC, MVT::v8i1, Custom);
1458 setOperationAction(ISD::MUL, MVT::v8i64, Custom);
1460 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i1, Custom);
1461 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i1, Custom);
1462 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i1, Custom);
1463 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i1, Custom);
1464 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i1, Custom);
1465 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i1, Custom);
1466 setOperationAction(ISD::SELECT, MVT::v8f64, Custom);
1467 setOperationAction(ISD::SELECT, MVT::v8i64, Custom);
1468 setOperationAction(ISD::SELECT, MVT::v16f32, Custom);
1469 setOperationAction(ISD::SELECT, MVT::v16i1, Custom);
1470 setOperationAction(ISD::SELECT, MVT::v8i1, Custom);
1472 setOperationAction(ISD::SMAX, MVT::v16i32, Legal);
1473 setOperationAction(ISD::SMAX, MVT::v8i64, Legal);
1474 setOperationAction(ISD::UMAX, MVT::v16i32, Legal);
1475 setOperationAction(ISD::UMAX, MVT::v8i64, Legal);
1476 setOperationAction(ISD::SMIN, MVT::v16i32, Legal);
1477 setOperationAction(ISD::SMIN, MVT::v8i64, Legal);
1478 setOperationAction(ISD::UMIN, MVT::v16i32, Legal);
1479 setOperationAction(ISD::UMIN, MVT::v8i64, Legal);
1481 setOperationAction(ISD::ADD, MVT::v8i64, Legal);
1482 setOperationAction(ISD::ADD, MVT::v16i32, Legal);
1484 setOperationAction(ISD::SUB, MVT::v8i64, Legal);
1485 setOperationAction(ISD::SUB, MVT::v16i32, Legal);
1487 setOperationAction(ISD::MUL, MVT::v16i32, Legal);
1489 setOperationAction(ISD::SRL, MVT::v8i64, Custom);
1490 setOperationAction(ISD::SRL, MVT::v16i32, Custom);
1492 setOperationAction(ISD::SHL, MVT::v8i64, Custom);
1493 setOperationAction(ISD::SHL, MVT::v16i32, Custom);
1495 setOperationAction(ISD::SRA, MVT::v8i64, Custom);
1496 setOperationAction(ISD::SRA, MVT::v16i32, Custom);
1498 setOperationAction(ISD::AND, MVT::v8i64, Legal);
1499 setOperationAction(ISD::OR, MVT::v8i64, Legal);
1500 setOperationAction(ISD::XOR, MVT::v8i64, Legal);
1501 setOperationAction(ISD::AND, MVT::v16i32, Legal);
1502 setOperationAction(ISD::OR, MVT::v16i32, Legal);
1503 setOperationAction(ISD::XOR, MVT::v16i32, Legal);
1505 if (Subtarget->hasCDI()) {
1506 setOperationAction(ISD::CTLZ, MVT::v8i64, Legal);
1507 setOperationAction(ISD::CTLZ, MVT::v16i32, Legal);
1508 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i64, Legal);
1509 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v16i32, Legal);
1511 setOperationAction(ISD::CTLZ, MVT::v8i16, Custom);
1512 setOperationAction(ISD::CTLZ, MVT::v16i8, Custom);
1513 setOperationAction(ISD::CTLZ, MVT::v16i16, Custom);
1514 setOperationAction(ISD::CTLZ, MVT::v32i8, Custom);
1515 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i16, Custom);
1516 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v16i8, Custom);
1517 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v16i16, Custom);
1518 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v32i8, Custom);
1520 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i64, Custom);
1521 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i32, Custom);
1523 if (Subtarget->hasVLX()) {
1524 setOperationAction(ISD::CTLZ, MVT::v4i64, Legal);
1525 setOperationAction(ISD::CTLZ, MVT::v8i32, Legal);
1526 setOperationAction(ISD::CTLZ, MVT::v2i64, Legal);
1527 setOperationAction(ISD::CTLZ, MVT::v4i32, Legal);
1528 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i64, Legal);
1529 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i32, Legal);
1530 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v2i64, Legal);
1531 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i32, Legal);
1533 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i64, Custom);
1534 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i32, Custom);
1535 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v2i64, Custom);
1536 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i32, Custom);
1538 setOperationAction(ISD::CTLZ, MVT::v4i64, Custom);
1539 setOperationAction(ISD::CTLZ, MVT::v8i32, Custom);
1540 setOperationAction(ISD::CTLZ, MVT::v2i64, Custom);
1541 setOperationAction(ISD::CTLZ, MVT::v4i32, Custom);
1542 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i64, Custom);
1543 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i32, Custom);
1544 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v2i64, Custom);
1545 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i32, Custom);
1547 } // Subtarget->hasCDI()
1549 if (Subtarget->hasDQI()) {
1550 setOperationAction(ISD::MUL, MVT::v2i64, Legal);
1551 setOperationAction(ISD::MUL, MVT::v4i64, Legal);
1552 setOperationAction(ISD::MUL, MVT::v8i64, Legal);
1554 // Custom lower several nodes.
1555 for (MVT VT : MVT::vector_valuetypes()) {
1556 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1558 setOperationAction(ISD::AND, VT, Legal);
1559 setOperationAction(ISD::OR, VT, Legal);
1560 setOperationAction(ISD::XOR, VT, Legal);
1562 if (EltSize >= 32 && VT.getSizeInBits() <= 512) {
1563 setOperationAction(ISD::MGATHER, VT, Custom);
1564 setOperationAction(ISD::MSCATTER, VT, Custom);
1566 // Extract subvector is special because the value type
1567 // (result) is 256/128-bit but the source is 512-bit wide.
1568 if (VT.is128BitVector() || VT.is256BitVector()) {
1569 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1571 if (VT.getVectorElementType() == MVT::i1)
1572 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
1574 // Do not attempt to custom lower other non-512-bit vectors
1575 if (!VT.is512BitVector())
1578 if (EltSize >= 32) {
1579 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1580 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1581 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1582 setOperationAction(ISD::VSELECT, VT, Legal);
1583 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1584 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1585 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1586 setOperationAction(ISD::MLOAD, VT, Legal);
1587 setOperationAction(ISD::MSTORE, VT, Legal);
1590 for (auto VT : { MVT::v64i8, MVT::v32i16, MVT::v16i32 }) {
1591 setOperationAction(ISD::SELECT, VT, Promote);
1592 AddPromotedToType (ISD::SELECT, VT, MVT::v8i64);
1596 if (!Subtarget->useSoftFloat() && Subtarget->hasBWI()) {
1597 addRegisterClass(MVT::v32i16, &X86::VR512RegClass);
1598 addRegisterClass(MVT::v64i8, &X86::VR512RegClass);
1600 addRegisterClass(MVT::v32i1, &X86::VK32RegClass);
1601 addRegisterClass(MVT::v64i1, &X86::VK64RegClass);
1603 setOperationAction(ISD::LOAD, MVT::v32i16, Legal);
1604 setOperationAction(ISD::LOAD, MVT::v64i8, Legal);
1605 setOperationAction(ISD::SETCC, MVT::v32i1, Custom);
1606 setOperationAction(ISD::SETCC, MVT::v64i1, Custom);
1607 setOperationAction(ISD::ADD, MVT::v32i16, Legal);
1608 setOperationAction(ISD::ADD, MVT::v64i8, Legal);
1609 setOperationAction(ISD::SUB, MVT::v32i16, Legal);
1610 setOperationAction(ISD::SUB, MVT::v64i8, Legal);
1611 setOperationAction(ISD::MUL, MVT::v32i16, Legal);
1612 setOperationAction(ISD::MULHS, MVT::v32i16, Legal);
1613 setOperationAction(ISD::MULHU, MVT::v32i16, Legal);
1614 setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i1, Legal);
1615 setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i1, Legal);
1616 setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i16, Custom);
1617 setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i8, Custom);
1618 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v32i1, Custom);
1619 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v64i1, Custom);
1620 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v32i16, Custom);
1621 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v64i8, Custom);
1622 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v32i16, Custom);
1623 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v64i8, Custom);
1624 setOperationAction(ISD::SELECT, MVT::v32i1, Custom);
1625 setOperationAction(ISD::SELECT, MVT::v64i1, Custom);
1626 setOperationAction(ISD::SIGN_EXTEND, MVT::v32i8, Custom);
1627 setOperationAction(ISD::ZERO_EXTEND, MVT::v32i8, Custom);
1628 setOperationAction(ISD::SIGN_EXTEND, MVT::v32i16, Custom);
1629 setOperationAction(ISD::ZERO_EXTEND, MVT::v32i16, Custom);
1630 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v32i16, Custom);
1631 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v64i8, Custom);
1632 setOperationAction(ISD::SIGN_EXTEND, MVT::v64i8, Custom);
1633 setOperationAction(ISD::ZERO_EXTEND, MVT::v64i8, Custom);
1634 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v32i1, Custom);
1635 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v64i1, Custom);
1636 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v32i16, Custom);
1637 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v64i8, Custom);
1638 setOperationAction(ISD::VSELECT, MVT::v32i16, Legal);
1639 setOperationAction(ISD::VSELECT, MVT::v64i8, Legal);
1640 setOperationAction(ISD::TRUNCATE, MVT::v32i1, Custom);
1641 setOperationAction(ISD::TRUNCATE, MVT::v64i1, Custom);
1642 setOperationAction(ISD::TRUNCATE, MVT::v32i8, Custom);
1643 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v32i1, Custom);
1644 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v64i1, Custom);
1646 setOperationAction(ISD::SMAX, MVT::v64i8, Legal);
1647 setOperationAction(ISD::SMAX, MVT::v32i16, Legal);
1648 setOperationAction(ISD::UMAX, MVT::v64i8, Legal);
1649 setOperationAction(ISD::UMAX, MVT::v32i16, Legal);
1650 setOperationAction(ISD::SMIN, MVT::v64i8, Legal);
1651 setOperationAction(ISD::SMIN, MVT::v32i16, Legal);
1652 setOperationAction(ISD::UMIN, MVT::v64i8, Legal);
1653 setOperationAction(ISD::UMIN, MVT::v32i16, Legal);
1655 setTruncStoreAction(MVT::v32i16, MVT::v32i8, Legal);
1656 setTruncStoreAction(MVT::v16i16, MVT::v16i8, Legal);
1657 if (Subtarget->hasVLX())
1658 setTruncStoreAction(MVT::v8i16, MVT::v8i8, Legal);
1660 if (Subtarget->hasCDI()) {
1661 setOperationAction(ISD::CTLZ, MVT::v32i16, Custom);
1662 setOperationAction(ISD::CTLZ, MVT::v64i8, Custom);
1663 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v32i16, Custom);
1664 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v64i8, Custom);
1667 for (auto VT : { MVT::v64i8, MVT::v32i16 }) {
1668 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1669 setOperationAction(ISD::VSELECT, VT, Legal);
1673 if (!Subtarget->useSoftFloat() && Subtarget->hasVLX()) {
1674 addRegisterClass(MVT::v4i1, &X86::VK4RegClass);
1675 addRegisterClass(MVT::v2i1, &X86::VK2RegClass);
1677 setOperationAction(ISD::SETCC, MVT::v4i1, Custom);
1678 setOperationAction(ISD::SETCC, MVT::v2i1, Custom);
1679 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i1, Custom);
1680 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i1, Custom);
1681 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v8i1, Custom);
1682 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v4i1, Custom);
1683 setOperationAction(ISD::SELECT, MVT::v4i1, Custom);
1684 setOperationAction(ISD::SELECT, MVT::v2i1, Custom);
1685 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i1, Custom);
1686 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i1, Custom);
1687 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i1, Custom);
1688 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i1, Custom);
1690 setOperationAction(ISD::AND, MVT::v8i32, Legal);
1691 setOperationAction(ISD::OR, MVT::v8i32, Legal);
1692 setOperationAction(ISD::XOR, MVT::v8i32, Legal);
1693 setOperationAction(ISD::AND, MVT::v4i32, Legal);
1694 setOperationAction(ISD::OR, MVT::v4i32, Legal);
1695 setOperationAction(ISD::XOR, MVT::v4i32, Legal);
1696 setOperationAction(ISD::SRA, MVT::v2i64, Custom);
1697 setOperationAction(ISD::SRA, MVT::v4i64, Custom);
1699 setOperationAction(ISD::SMAX, MVT::v2i64, Legal);
1700 setOperationAction(ISD::SMAX, MVT::v4i64, Legal);
1701 setOperationAction(ISD::UMAX, MVT::v2i64, Legal);
1702 setOperationAction(ISD::UMAX, MVT::v4i64, Legal);
1703 setOperationAction(ISD::SMIN, MVT::v2i64, Legal);
1704 setOperationAction(ISD::SMIN, MVT::v4i64, Legal);
1705 setOperationAction(ISD::UMIN, MVT::v2i64, Legal);
1706 setOperationAction(ISD::UMIN, MVT::v4i64, Legal);
1709 // We want to custom lower some of our intrinsics.
1710 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1711 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
1712 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
1713 if (!Subtarget->is64Bit())
1714 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
1716 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1717 // handle type legalization for these operations here.
1719 // FIXME: We really should do custom legalization for addition and
1720 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1721 // than generic legalization for 64-bit multiplication-with-overflow, though.
1722 for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) {
1723 if (VT == MVT::i64 && !Subtarget->is64Bit())
1725 // Add/Sub/Mul with overflow operations are custom lowered.
1726 setOperationAction(ISD::SADDO, VT, Custom);
1727 setOperationAction(ISD::UADDO, VT, Custom);
1728 setOperationAction(ISD::SSUBO, VT, Custom);
1729 setOperationAction(ISD::USUBO, VT, Custom);
1730 setOperationAction(ISD::SMULO, VT, Custom);
1731 setOperationAction(ISD::UMULO, VT, Custom);
1734 if (!Subtarget->is64Bit()) {
1735 // These libcalls are not available in 32-bit.
1736 setLibcallName(RTLIB::SHL_I128, nullptr);
1737 setLibcallName(RTLIB::SRL_I128, nullptr);
1738 setLibcallName(RTLIB::SRA_I128, nullptr);
1741 // Combine sin / cos into one node or libcall if possible.
1742 if (Subtarget->hasSinCos()) {
1743 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
1744 setLibcallName(RTLIB::SINCOS_F64, "sincos");
1745 if (Subtarget->isTargetDarwin()) {
1746 // For MacOSX, we don't want the normal expansion of a libcall to sincos.
1747 // We want to issue a libcall to __sincos_stret to avoid memory traffic.
1748 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
1749 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
1753 if (Subtarget->isTargetWin64()) {
1754 setOperationAction(ISD::SDIV, MVT::i128, Custom);
1755 setOperationAction(ISD::UDIV, MVT::i128, Custom);
1756 setOperationAction(ISD::SREM, MVT::i128, Custom);
1757 setOperationAction(ISD::UREM, MVT::i128, Custom);
1758 setOperationAction(ISD::SDIVREM, MVT::i128, Custom);
1759 setOperationAction(ISD::UDIVREM, MVT::i128, Custom);
1762 // We have target-specific dag combine patterns for the following nodes:
1763 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1764 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1765 setTargetDAGCombine(ISD::BITCAST);
1766 setTargetDAGCombine(ISD::VSELECT);
1767 setTargetDAGCombine(ISD::SELECT);
1768 setTargetDAGCombine(ISD::SHL);
1769 setTargetDAGCombine(ISD::SRA);
1770 setTargetDAGCombine(ISD::SRL);
1771 setTargetDAGCombine(ISD::OR);
1772 setTargetDAGCombine(ISD::AND);
1773 setTargetDAGCombine(ISD::ADD);
1774 setTargetDAGCombine(ISD::FADD);
1775 setTargetDAGCombine(ISD::FSUB);
1776 setTargetDAGCombine(ISD::FMA);
1777 setTargetDAGCombine(ISD::SUB);
1778 setTargetDAGCombine(ISD::LOAD);
1779 setTargetDAGCombine(ISD::MLOAD);
1780 setTargetDAGCombine(ISD::STORE);
1781 setTargetDAGCombine(ISD::MSTORE);
1782 setTargetDAGCombine(ISD::ZERO_EXTEND);
1783 setTargetDAGCombine(ISD::ANY_EXTEND);
1784 setTargetDAGCombine(ISD::SIGN_EXTEND);
1785 setTargetDAGCombine(ISD::SIGN_EXTEND_INREG);
1786 setTargetDAGCombine(ISD::SINT_TO_FP);
1787 setTargetDAGCombine(ISD::UINT_TO_FP);
1788 setTargetDAGCombine(ISD::SETCC);
1789 setTargetDAGCombine(ISD::BUILD_VECTOR);
1790 setTargetDAGCombine(ISD::MUL);
1791 setTargetDAGCombine(ISD::XOR);
1793 computeRegisterProperties(Subtarget->getRegisterInfo());
1795 MaxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1796 MaxStoresPerMemsetOptSize = 8;
1797 MaxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1798 MaxStoresPerMemcpyOptSize = 4;
1799 MaxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1800 MaxStoresPerMemmoveOptSize = 4;
1801 setPrefLoopAlignment(4); // 2^4 bytes.
1803 // A predictable cmov does not hurt on an in-order CPU.
1804 // FIXME: Use a CPU attribute to trigger this, not a CPU model.
1805 PredictableSelectIsExpensive = !Subtarget->isAtom();
1806 EnableExtLdPromotion = true;
1807 setPrefFunctionAlignment(4); // 2^4 bytes.
1809 verifyIntrinsicTables();
1812 // This has so far only been implemented for 64-bit MachO.
1813 bool X86TargetLowering::useLoadStackGuardNode() const {
1814 return Subtarget->isTargetMachO() && Subtarget->is64Bit();
1817 TargetLoweringBase::LegalizeTypeAction
1818 X86TargetLowering::getPreferredVectorAction(EVT VT) const {
1819 if (ExperimentalVectorWideningLegalization &&
1820 VT.getVectorNumElements() != 1 &&
1821 VT.getVectorElementType().getSimpleVT() != MVT::i1)
1822 return TypeWidenVector;
1824 return TargetLoweringBase::getPreferredVectorAction(VT);
1827 EVT X86TargetLowering::getSetCCResultType(const DataLayout &DL, LLVMContext &,
1830 return Subtarget->hasAVX512() ? MVT::i1: MVT::i8;
1832 if (VT.isSimple()) {
1833 MVT VVT = VT.getSimpleVT();
1834 const unsigned NumElts = VVT.getVectorNumElements();
1835 const MVT EltVT = VVT.getVectorElementType();
1836 if (VVT.is512BitVector()) {
1837 if (Subtarget->hasAVX512())
1838 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1839 EltVT == MVT::f32 || EltVT == MVT::f64)
1841 case 8: return MVT::v8i1;
1842 case 16: return MVT::v16i1;
1844 if (Subtarget->hasBWI())
1845 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1847 case 32: return MVT::v32i1;
1848 case 64: return MVT::v64i1;
1852 if (VVT.is256BitVector() || VVT.is128BitVector()) {
1853 if (Subtarget->hasVLX())
1854 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1855 EltVT == MVT::f32 || EltVT == MVT::f64)
1857 case 2: return MVT::v2i1;
1858 case 4: return MVT::v4i1;
1859 case 8: return MVT::v8i1;
1861 if (Subtarget->hasBWI() && Subtarget->hasVLX())
1862 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1864 case 8: return MVT::v8i1;
1865 case 16: return MVT::v16i1;
1866 case 32: return MVT::v32i1;
1871 return VT.changeVectorElementTypeToInteger();
1874 /// Helper for getByValTypeAlignment to determine
1875 /// the desired ByVal argument alignment.
1876 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1879 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1880 if (VTy->getBitWidth() == 128)
1882 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1883 unsigned EltAlign = 0;
1884 getMaxByValAlign(ATy->getElementType(), EltAlign);
1885 if (EltAlign > MaxAlign)
1886 MaxAlign = EltAlign;
1887 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1888 for (auto *EltTy : STy->elements()) {
1889 unsigned EltAlign = 0;
1890 getMaxByValAlign(EltTy, EltAlign);
1891 if (EltAlign > MaxAlign)
1892 MaxAlign = EltAlign;
1899 /// Return the desired alignment for ByVal aggregate
1900 /// function arguments in the caller parameter area. For X86, aggregates
1901 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1902 /// are at 4-byte boundaries.
1903 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty,
1904 const DataLayout &DL) const {
1905 if (Subtarget->is64Bit()) {
1906 // Max of 8 and alignment of type.
1907 unsigned TyAlign = DL.getABITypeAlignment(Ty);
1914 if (Subtarget->hasSSE1())
1915 getMaxByValAlign(Ty, Align);
1919 /// Returns the target specific optimal type for load
1920 /// and store operations as a result of memset, memcpy, and memmove
1921 /// lowering. If DstAlign is zero that means it's safe to destination
1922 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1923 /// means there isn't a need to check it against alignment requirement,
1924 /// probably because the source does not need to be loaded. If 'IsMemset' is
1925 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
1926 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
1927 /// source is constant so it does not need to be loaded.
1928 /// It returns EVT::Other if the type should be determined using generic
1929 /// target-independent logic.
1931 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1932 unsigned DstAlign, unsigned SrcAlign,
1933 bool IsMemset, bool ZeroMemset,
1935 MachineFunction &MF) const {
1936 const Function *F = MF.getFunction();
1937 if ((!IsMemset || ZeroMemset) &&
1938 !F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1940 (!Subtarget->isUnalignedMem16Slow() ||
1941 ((DstAlign == 0 || DstAlign >= 16) &&
1942 (SrcAlign == 0 || SrcAlign >= 16)))) {
1944 // FIXME: Check if unaligned 32-byte accesses are slow.
1945 if (Subtarget->hasInt256())
1947 if (Subtarget->hasFp256())
1950 if (Subtarget->hasSSE2())
1952 if (Subtarget->hasSSE1())
1954 } else if (!MemcpyStrSrc && Size >= 8 &&
1955 !Subtarget->is64Bit() &&
1956 Subtarget->hasSSE2()) {
1957 // Do not use f64 to lower memcpy if source is string constant. It's
1958 // better to use i32 to avoid the loads.
1962 // This is a compromise. If we reach here, unaligned accesses may be slow on
1963 // this target. However, creating smaller, aligned accesses could be even
1964 // slower and would certainly be a lot more code.
1965 if (Subtarget->is64Bit() && Size >= 8)
1970 bool X86TargetLowering::isSafeMemOpType(MVT VT) const {
1972 return X86ScalarSSEf32;
1973 else if (VT == MVT::f64)
1974 return X86ScalarSSEf64;
1979 X86TargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
1984 switch (VT.getSizeInBits()) {
1986 // 8-byte and under are always assumed to be fast.
1990 *Fast = !Subtarget->isUnalignedMem16Slow();
1993 *Fast = !Subtarget->isUnalignedMem32Slow();
1995 // TODO: What about AVX-512 (512-bit) accesses?
1998 // Misaligned accesses of any size are always allowed.
2002 /// Return the entry encoding for a jump table in the
2003 /// current function. The returned value is a member of the
2004 /// MachineJumpTableInfo::JTEntryKind enum.
2005 unsigned X86TargetLowering::getJumpTableEncoding() const {
2006 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
2008 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2009 Subtarget->isPICStyleGOT())
2010 return MachineJumpTableInfo::EK_Custom32;
2012 // Otherwise, use the normal jump table encoding heuristics.
2013 return TargetLowering::getJumpTableEncoding();
2016 bool X86TargetLowering::useSoftFloat() const {
2017 return Subtarget->useSoftFloat();
2021 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
2022 const MachineBasicBlock *MBB,
2023 unsigned uid,MCContext &Ctx) const{
2024 assert(MBB->getParent()->getTarget().getRelocationModel() == Reloc::PIC_ &&
2025 Subtarget->isPICStyleGOT());
2026 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
2028 return MCSymbolRefExpr::create(MBB->getSymbol(),
2029 MCSymbolRefExpr::VK_GOTOFF, Ctx);
2032 /// Returns relocation base for the given PIC jumptable.
2033 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
2034 SelectionDAG &DAG) const {
2035 if (!Subtarget->is64Bit())
2036 // This doesn't have SDLoc associated with it, but is not really the
2037 // same as a Register.
2038 return DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(),
2039 getPointerTy(DAG.getDataLayout()));
2043 /// This returns the relocation base for the given PIC jumptable,
2044 /// the same as getPICJumpTableRelocBase, but as an MCExpr.
2045 const MCExpr *X86TargetLowering::
2046 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
2047 MCContext &Ctx) const {
2048 // X86-64 uses RIP relative addressing based on the jump table label.
2049 if (Subtarget->isPICStyleRIPRel())
2050 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
2052 // Otherwise, the reference is relative to the PIC base.
2053 return MCSymbolRefExpr::create(MF->getPICBaseSymbol(), Ctx);
2056 std::pair<const TargetRegisterClass *, uint8_t>
2057 X86TargetLowering::findRepresentativeClass(const TargetRegisterInfo *TRI,
2059 const TargetRegisterClass *RRC = nullptr;
2061 switch (VT.SimpleTy) {
2063 return TargetLowering::findRepresentativeClass(TRI, VT);
2064 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
2065 RRC = Subtarget->is64Bit() ? &X86::GR64RegClass : &X86::GR32RegClass;
2068 RRC = &X86::VR64RegClass;
2070 case MVT::f32: case MVT::f64:
2071 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
2072 case MVT::v4f32: case MVT::v2f64:
2073 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
2075 RRC = &X86::VR128RegClass;
2078 return std::make_pair(RRC, Cost);
2081 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
2082 unsigned &Offset) const {
2083 if (!Subtarget->isTargetLinux())
2086 if (Subtarget->is64Bit()) {
2087 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
2089 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
2101 Value *X86TargetLowering::getSafeStackPointerLocation(IRBuilder<> &IRB) const {
2102 if (!Subtarget->isTargetAndroid())
2103 return TargetLowering::getSafeStackPointerLocation(IRB);
2105 // Android provides a fixed TLS slot for the SafeStack pointer. See the
2106 // definition of TLS_SLOT_SAFESTACK in
2107 // https://android.googlesource.com/platform/bionic/+/master/libc/private/bionic_tls.h
2108 unsigned AddressSpace, Offset;
2109 if (Subtarget->is64Bit()) {
2110 // %fs:0x48, unless we're using a Kernel code model, in which case it's %gs:
2112 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
2122 return ConstantExpr::getIntToPtr(
2123 ConstantInt::get(Type::getInt32Ty(IRB.getContext()), Offset),
2124 Type::getInt8PtrTy(IRB.getContext())->getPointerTo(AddressSpace));
2127 bool X86TargetLowering::isNoopAddrSpaceCast(unsigned SrcAS,
2128 unsigned DestAS) const {
2129 assert(SrcAS != DestAS && "Expected different address spaces!");
2131 return SrcAS < 256 && DestAS < 256;
2134 //===----------------------------------------------------------------------===//
2135 // Return Value Calling Convention Implementation
2136 //===----------------------------------------------------------------------===//
2138 #include "X86GenCallingConv.inc"
2140 bool X86TargetLowering::CanLowerReturn(
2141 CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg,
2142 const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const {
2143 SmallVector<CCValAssign, 16> RVLocs;
2144 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2145 return CCInfo.CheckReturn(Outs, RetCC_X86);
2148 const MCPhysReg *X86TargetLowering::getScratchRegisters(CallingConv::ID) const {
2149 static const MCPhysReg ScratchRegs[] = { X86::R11, 0 };
2154 X86TargetLowering::LowerReturn(SDValue Chain,
2155 CallingConv::ID CallConv, bool isVarArg,
2156 const SmallVectorImpl<ISD::OutputArg> &Outs,
2157 const SmallVectorImpl<SDValue> &OutVals,
2158 SDLoc dl, SelectionDAG &DAG) const {
2159 MachineFunction &MF = DAG.getMachineFunction();
2160 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2162 SmallVector<CCValAssign, 16> RVLocs;
2163 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, *DAG.getContext());
2164 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
2167 SmallVector<SDValue, 6> RetOps;
2168 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2169 // Operand #1 = Bytes To Pop
2170 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(), dl,
2173 // Copy the result values into the output registers.
2174 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2175 CCValAssign &VA = RVLocs[i];
2176 assert(VA.isRegLoc() && "Can only return in registers!");
2177 SDValue ValToCopy = OutVals[i];
2178 EVT ValVT = ValToCopy.getValueType();
2180 // Promote values to the appropriate types.
2181 if (VA.getLocInfo() == CCValAssign::SExt)
2182 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
2183 else if (VA.getLocInfo() == CCValAssign::ZExt)
2184 ValToCopy = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), ValToCopy);
2185 else if (VA.getLocInfo() == CCValAssign::AExt) {
2186 if (ValVT.isVector() && ValVT.getVectorElementType() == MVT::i1)
2187 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
2189 ValToCopy = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), ValToCopy);
2191 else if (VA.getLocInfo() == CCValAssign::BCvt)
2192 ValToCopy = DAG.getBitcast(VA.getLocVT(), ValToCopy);
2194 assert(VA.getLocInfo() != CCValAssign::FPExt &&
2195 "Unexpected FP-extend for return value.");
2197 // If this is x86-64, and we disabled SSE, we can't return FP values,
2198 // or SSE or MMX vectors.
2199 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
2200 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
2201 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
2202 report_fatal_error("SSE register return with SSE disabled");
2204 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
2205 // llvm-gcc has never done it right and no one has noticed, so this
2206 // should be OK for now.
2207 if (ValVT == MVT::f64 &&
2208 (Subtarget->is64Bit() && !Subtarget->hasSSE2()))
2209 report_fatal_error("SSE2 register return with SSE2 disabled");
2211 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
2212 // the RET instruction and handled by the FP Stackifier.
2213 if (VA.getLocReg() == X86::FP0 ||
2214 VA.getLocReg() == X86::FP1) {
2215 // If this is a copy from an xmm register to ST(0), use an FPExtend to
2216 // change the value to the FP stack register class.
2217 if (isScalarFPTypeInSSEReg(VA.getValVT()))
2218 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
2219 RetOps.push_back(ValToCopy);
2220 // Don't emit a copytoreg.
2224 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
2225 // which is returned in RAX / RDX.
2226 if (Subtarget->is64Bit()) {
2227 if (ValVT == MVT::x86mmx) {
2228 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
2229 ValToCopy = DAG.getBitcast(MVT::i64, ValToCopy);
2230 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
2232 // If we don't have SSE2 available, convert to v4f32 so the generated
2233 // register is legal.
2234 if (!Subtarget->hasSSE2())
2235 ValToCopy = DAG.getBitcast(MVT::v4f32, ValToCopy);
2240 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
2241 Flag = Chain.getValue(1);
2242 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2245 // All x86 ABIs require that for returning structs by value we copy
2246 // the sret argument into %rax/%eax (depending on ABI) for the return.
2247 // We saved the argument into a virtual register in the entry block,
2248 // so now we copy the value out and into %rax/%eax.
2250 // Checking Function.hasStructRetAttr() here is insufficient because the IR
2251 // may not have an explicit sret argument. If FuncInfo.CanLowerReturn is
2252 // false, then an sret argument may be implicitly inserted in the SelDAG. In
2253 // either case FuncInfo->setSRetReturnReg() will have been called.
2254 if (unsigned SRetReg = FuncInfo->getSRetReturnReg()) {
2255 SDValue Val = DAG.getCopyFromReg(Chain, dl, SRetReg,
2256 getPointerTy(MF.getDataLayout()));
2259 = (Subtarget->is64Bit() && !Subtarget->isTarget64BitILP32()) ?
2260 X86::RAX : X86::EAX;
2261 Chain = DAG.getCopyToReg(Chain, dl, RetValReg, Val, Flag);
2262 Flag = Chain.getValue(1);
2264 // RAX/EAX now acts like a return value.
2266 DAG.getRegister(RetValReg, getPointerTy(DAG.getDataLayout())));
2269 RetOps[0] = Chain; // Update chain.
2271 // Add the flag if we have it.
2273 RetOps.push_back(Flag);
2275 return DAG.getNode(X86ISD::RET_FLAG, dl, MVT::Other, RetOps);
2278 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2279 if (N->getNumValues() != 1)
2281 if (!N->hasNUsesOfValue(1, 0))
2284 SDValue TCChain = Chain;
2285 SDNode *Copy = *N->use_begin();
2286 if (Copy->getOpcode() == ISD::CopyToReg) {
2287 // If the copy has a glue operand, we conservatively assume it isn't safe to
2288 // perform a tail call.
2289 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2291 TCChain = Copy->getOperand(0);
2292 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
2295 bool HasRet = false;
2296 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2298 if (UI->getOpcode() != X86ISD::RET_FLAG)
2300 // If we are returning more than one value, we can definitely
2301 // not make a tail call see PR19530
2302 if (UI->getNumOperands() > 4)
2304 if (UI->getNumOperands() == 4 &&
2305 UI->getOperand(UI->getNumOperands()-1).getValueType() != MVT::Glue)
2318 X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
2319 ISD::NodeType ExtendKind) const {
2321 // TODO: Is this also valid on 32-bit?
2322 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
2323 ReturnMVT = MVT::i8;
2325 ReturnMVT = MVT::i32;
2327 EVT MinVT = getRegisterType(Context, ReturnMVT);
2328 return VT.bitsLT(MinVT) ? MinVT : VT;
2331 /// Lower the result values of a call into the
2332 /// appropriate copies out of appropriate physical registers.
2335 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2336 CallingConv::ID CallConv, bool isVarArg,
2337 const SmallVectorImpl<ISD::InputArg> &Ins,
2338 SDLoc dl, SelectionDAG &DAG,
2339 SmallVectorImpl<SDValue> &InVals) const {
2341 // Assign locations to each value returned by this call.
2342 SmallVector<CCValAssign, 16> RVLocs;
2343 bool Is64Bit = Subtarget->is64Bit();
2344 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2346 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2348 // Copy all of the result registers out of their specified physreg.
2349 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2350 CCValAssign &VA = RVLocs[i];
2351 EVT CopyVT = VA.getLocVT();
2353 // If this is x86-64, and we disabled SSE, we can't return FP values
2354 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
2355 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
2356 report_fatal_error("SSE register return with SSE disabled");
2359 // If we prefer to use the value in xmm registers, copy it out as f80 and
2360 // use a truncate to move it from fp stack reg to xmm reg.
2361 bool RoundAfterCopy = false;
2362 if ((VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) &&
2363 isScalarFPTypeInSSEReg(VA.getValVT())) {
2365 RoundAfterCopy = (CopyVT != VA.getLocVT());
2368 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
2369 CopyVT, InFlag).getValue(1);
2370 SDValue Val = Chain.getValue(0);
2373 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
2374 // This truncation won't change the value.
2375 DAG.getIntPtrConstant(1, dl));
2377 if (VA.isExtInLoc() && VA.getValVT().getScalarType() == MVT::i1)
2378 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
2380 InFlag = Chain.getValue(2);
2381 InVals.push_back(Val);
2387 //===----------------------------------------------------------------------===//
2388 // C & StdCall & Fast Calling Convention implementation
2389 //===----------------------------------------------------------------------===//
2390 // StdCall calling convention seems to be standard for many Windows' API
2391 // routines and around. It differs from C calling convention just a little:
2392 // callee should clean up the stack, not caller. Symbols should be also
2393 // decorated in some fancy way :) It doesn't support any vector arguments.
2394 // For info on fast calling convention see Fast Calling Convention (tail call)
2395 // implementation LowerX86_32FastCCCallTo.
2397 /// CallIsStructReturn - Determines whether a call uses struct return
2399 enum StructReturnType {
2404 static StructReturnType
2405 callIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
2407 return NotStructReturn;
2409 const ISD::ArgFlagsTy &Flags = Outs[0].Flags;
2410 if (!Flags.isSRet())
2411 return NotStructReturn;
2412 if (Flags.isInReg())
2413 return RegStructReturn;
2414 return StackStructReturn;
2417 /// Determines whether a function uses struct return semantics.
2418 static StructReturnType
2419 argsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
2421 return NotStructReturn;
2423 const ISD::ArgFlagsTy &Flags = Ins[0].Flags;
2424 if (!Flags.isSRet())
2425 return NotStructReturn;
2426 if (Flags.isInReg())
2427 return RegStructReturn;
2428 return StackStructReturn;
2431 /// Make a copy of an aggregate at address specified by "Src" to address
2432 /// "Dst" with size and alignment information specified by the specific
2433 /// parameter attribute. The copy will be passed as a byval function parameter.
2435 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2436 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2438 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), dl, MVT::i32);
2440 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2441 /*isVolatile*/false, /*AlwaysInline=*/true,
2442 /*isTailCall*/false,
2443 MachinePointerInfo(), MachinePointerInfo());
2446 /// Return true if the calling convention is one that we can guarantee TCO for.
2447 static bool canGuaranteeTCO(CallingConv::ID CC) {
2448 return (CC == CallingConv::Fast || CC == CallingConv::GHC ||
2449 CC == CallingConv::HiPE || CC == CallingConv::HHVM);
2452 /// Return true if we might ever do TCO for calls with this calling convention.
2453 static bool mayTailCallThisCC(CallingConv::ID CC) {
2455 // C calling conventions:
2456 case CallingConv::C:
2457 case CallingConv::X86_64_Win64:
2458 case CallingConv::X86_64_SysV:
2459 // Callee pop conventions:
2460 case CallingConv::X86_ThisCall:
2461 case CallingConv::X86_StdCall:
2462 case CallingConv::X86_VectorCall:
2463 case CallingConv::X86_FastCall:
2466 return canGuaranteeTCO(CC);
2470 /// Return true if the function is being made into a tailcall target by
2471 /// changing its ABI.
2472 static bool shouldGuaranteeTCO(CallingConv::ID CC, bool GuaranteedTailCallOpt) {
2473 return GuaranteedTailCallOpt && canGuaranteeTCO(CC);
2476 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2478 CI->getParent()->getParent()->getFnAttribute("disable-tail-calls");
2479 if (!CI->isTailCall() || Attr.getValueAsString() == "true")
2483 CallingConv::ID CalleeCC = CS.getCallingConv();
2484 if (!mayTailCallThisCC(CalleeCC))
2491 X86TargetLowering::LowerMemArgument(SDValue Chain,
2492 CallingConv::ID CallConv,
2493 const SmallVectorImpl<ISD::InputArg> &Ins,
2494 SDLoc dl, SelectionDAG &DAG,
2495 const CCValAssign &VA,
2496 MachineFrameInfo *MFI,
2498 // Create the nodes corresponding to a load from this parameter slot.
2499 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2500 bool AlwaysUseMutable = shouldGuaranteeTCO(
2501 CallConv, DAG.getTarget().Options.GuaranteedTailCallOpt);
2502 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
2505 // If value is passed by pointer we have address passed instead of the value
2507 bool ExtendedInMem = VA.isExtInLoc() &&
2508 VA.getValVT().getScalarType() == MVT::i1;
2510 if (VA.getLocInfo() == CCValAssign::Indirect || ExtendedInMem)
2511 ValVT = VA.getLocVT();
2513 ValVT = VA.getValVT();
2515 // FIXME: For now, all byval parameter objects are marked mutable. This can be
2516 // changed with more analysis.
2517 // In case of tail call optimization mark all arguments mutable. Since they
2518 // could be overwritten by lowering of arguments in case of a tail call.
2519 if (Flags.isByVal()) {
2520 unsigned Bytes = Flags.getByValSize();
2521 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2522 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
2523 return DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2525 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2526 VA.getLocMemOffset(), isImmutable);
2527 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2528 SDValue Val = DAG.getLoad(
2529 ValVT, dl, Chain, FIN,
2530 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI), false,
2532 return ExtendedInMem ?
2533 DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val) : Val;
2537 // FIXME: Get this from tablegen.
2538 static ArrayRef<MCPhysReg> get64BitArgumentGPRs(CallingConv::ID CallConv,
2539 const X86Subtarget *Subtarget) {
2540 assert(Subtarget->is64Bit());
2542 if (Subtarget->isCallingConvWin64(CallConv)) {
2543 static const MCPhysReg GPR64ArgRegsWin64[] = {
2544 X86::RCX, X86::RDX, X86::R8, X86::R9
2546 return makeArrayRef(std::begin(GPR64ArgRegsWin64), std::end(GPR64ArgRegsWin64));
2549 static const MCPhysReg GPR64ArgRegs64Bit[] = {
2550 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
2552 return makeArrayRef(std::begin(GPR64ArgRegs64Bit), std::end(GPR64ArgRegs64Bit));
2555 // FIXME: Get this from tablegen.
2556 static ArrayRef<MCPhysReg> get64BitArgumentXMMs(MachineFunction &MF,
2557 CallingConv::ID CallConv,
2558 const X86Subtarget *Subtarget) {
2559 assert(Subtarget->is64Bit());
2560 if (Subtarget->isCallingConvWin64(CallConv)) {
2561 // The XMM registers which might contain var arg parameters are shadowed
2562 // in their paired GPR. So we only need to save the GPR to their home
2564 // TODO: __vectorcall will change this.
2568 const Function *Fn = MF.getFunction();
2569 bool NoImplicitFloatOps = Fn->hasFnAttribute(Attribute::NoImplicitFloat);
2570 bool isSoftFloat = Subtarget->useSoftFloat();
2571 assert(!(isSoftFloat && NoImplicitFloatOps) &&
2572 "SSE register cannot be used when SSE is disabled!");
2573 if (isSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
2574 // Kernel mode asks for SSE to be disabled, so there are no XMM argument
2578 static const MCPhysReg XMMArgRegs64Bit[] = {
2579 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2580 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2582 return makeArrayRef(std::begin(XMMArgRegs64Bit), std::end(XMMArgRegs64Bit));
2585 SDValue X86TargetLowering::LowerFormalArguments(
2586 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
2587 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc dl, SelectionDAG &DAG,
2588 SmallVectorImpl<SDValue> &InVals) const {
2589 MachineFunction &MF = DAG.getMachineFunction();
2590 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2591 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
2593 const Function* Fn = MF.getFunction();
2594 if (Fn->hasExternalLinkage() &&
2595 Subtarget->isTargetCygMing() &&
2596 Fn->getName() == "main")
2597 FuncInfo->setForceFramePointer(true);
2599 MachineFrameInfo *MFI = MF.getFrameInfo();
2600 bool Is64Bit = Subtarget->is64Bit();
2601 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2603 assert(!(isVarArg && canGuaranteeTCO(CallConv)) &&
2604 "Var args not supported with calling convention fastcc, ghc or hipe");
2606 // Assign locations to all of the incoming arguments.
2607 SmallVector<CCValAssign, 16> ArgLocs;
2608 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2610 // Allocate shadow area for Win64
2612 CCInfo.AllocateStack(32, 8);
2614 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
2616 unsigned LastVal = ~0U;
2618 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2619 CCValAssign &VA = ArgLocs[i];
2620 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
2622 assert(VA.getValNo() != LastVal &&
2623 "Don't support value assigned to multiple locs yet");
2625 LastVal = VA.getValNo();
2627 if (VA.isRegLoc()) {
2628 EVT RegVT = VA.getLocVT();
2629 const TargetRegisterClass *RC;
2630 if (RegVT == MVT::i32)
2631 RC = &X86::GR32RegClass;
2632 else if (Is64Bit && RegVT == MVT::i64)
2633 RC = &X86::GR64RegClass;
2634 else if (RegVT == MVT::f32)
2635 RC = &X86::FR32RegClass;
2636 else if (RegVT == MVT::f64)
2637 RC = &X86::FR64RegClass;
2638 else if (RegVT.is512BitVector())
2639 RC = &X86::VR512RegClass;
2640 else if (RegVT.is256BitVector())
2641 RC = &X86::VR256RegClass;
2642 else if (RegVT.is128BitVector())
2643 RC = &X86::VR128RegClass;
2644 else if (RegVT == MVT::x86mmx)
2645 RC = &X86::VR64RegClass;
2646 else if (RegVT == MVT::i1)
2647 RC = &X86::VK1RegClass;
2648 else if (RegVT == MVT::v8i1)
2649 RC = &X86::VK8RegClass;
2650 else if (RegVT == MVT::v16i1)
2651 RC = &X86::VK16RegClass;
2652 else if (RegVT == MVT::v32i1)
2653 RC = &X86::VK32RegClass;
2654 else if (RegVT == MVT::v64i1)
2655 RC = &X86::VK64RegClass;
2657 llvm_unreachable("Unknown argument type!");
2659 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2660 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2662 // If this is an 8 or 16-bit value, it is really passed promoted to 32
2663 // bits. Insert an assert[sz]ext to capture this, then truncate to the
2665 if (VA.getLocInfo() == CCValAssign::SExt)
2666 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2667 DAG.getValueType(VA.getValVT()));
2668 else if (VA.getLocInfo() == CCValAssign::ZExt)
2669 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2670 DAG.getValueType(VA.getValVT()));
2671 else if (VA.getLocInfo() == CCValAssign::BCvt)
2672 ArgValue = DAG.getBitcast(VA.getValVT(), ArgValue);
2674 if (VA.isExtInLoc()) {
2675 // Handle MMX values passed in XMM regs.
2676 if (RegVT.isVector() && VA.getValVT().getScalarType() != MVT::i1)
2677 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(), ArgValue);
2679 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2682 assert(VA.isMemLoc());
2683 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
2686 // If value is passed via pointer - do a load.
2687 if (VA.getLocInfo() == CCValAssign::Indirect)
2688 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
2689 MachinePointerInfo(), false, false, false, 0);
2691 InVals.push_back(ArgValue);
2694 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2695 // All x86 ABIs require that for returning structs by value we copy the
2696 // sret argument into %rax/%eax (depending on ABI) for the return. Save
2697 // the argument into a virtual register so that we can access it from the
2699 if (Ins[i].Flags.isSRet()) {
2700 unsigned Reg = FuncInfo->getSRetReturnReg();
2702 MVT PtrTy = getPointerTy(DAG.getDataLayout());
2703 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrTy));
2704 FuncInfo->setSRetReturnReg(Reg);
2706 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[i]);
2707 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
2712 unsigned StackSize = CCInfo.getNextStackOffset();
2713 // Align stack specially for tail calls.
2714 if (shouldGuaranteeTCO(CallConv,
2715 MF.getTarget().Options.GuaranteedTailCallOpt))
2716 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
2718 // If the function takes variable number of arguments, make a frame index for
2719 // the start of the first vararg value... for expansion of llvm.va_start. We
2720 // can skip this if there are no va_start calls.
2721 if (MFI->hasVAStart() &&
2722 (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
2723 CallConv != CallingConv::X86_ThisCall))) {
2724 FuncInfo->setVarArgsFrameIndex(
2725 MFI->CreateFixedObject(1, StackSize, true));
2728 // Figure out if XMM registers are in use.
2729 assert(!(Subtarget->useSoftFloat() &&
2730 Fn->hasFnAttribute(Attribute::NoImplicitFloat)) &&
2731 "SSE register cannot be used when SSE is disabled!");
2733 // 64-bit calling conventions support varargs and register parameters, so we
2734 // have to do extra work to spill them in the prologue.
2735 if (Is64Bit && isVarArg && MFI->hasVAStart()) {
2736 // Find the first unallocated argument registers.
2737 ArrayRef<MCPhysReg> ArgGPRs = get64BitArgumentGPRs(CallConv, Subtarget);
2738 ArrayRef<MCPhysReg> ArgXMMs = get64BitArgumentXMMs(MF, CallConv, Subtarget);
2739 unsigned NumIntRegs = CCInfo.getFirstUnallocated(ArgGPRs);
2740 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(ArgXMMs);
2741 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
2742 "SSE register cannot be used when SSE is disabled!");
2744 // Gather all the live in physical registers.
2745 SmallVector<SDValue, 6> LiveGPRs;
2746 SmallVector<SDValue, 8> LiveXMMRegs;
2748 for (MCPhysReg Reg : ArgGPRs.slice(NumIntRegs)) {
2749 unsigned GPR = MF.addLiveIn(Reg, &X86::GR64RegClass);
2751 DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64));
2753 if (!ArgXMMs.empty()) {
2754 unsigned AL = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2755 ALVal = DAG.getCopyFromReg(Chain, dl, AL, MVT::i8);
2756 for (MCPhysReg Reg : ArgXMMs.slice(NumXMMRegs)) {
2757 unsigned XMMReg = MF.addLiveIn(Reg, &X86::VR128RegClass);
2758 LiveXMMRegs.push_back(
2759 DAG.getCopyFromReg(Chain, dl, XMMReg, MVT::v4f32));
2764 // Get to the caller-allocated home save location. Add 8 to account
2765 // for the return address.
2766 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2767 FuncInfo->setRegSaveFrameIndex(
2768 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
2769 // Fixup to set vararg frame on shadow area (4 x i64).
2771 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
2773 // For X86-64, if there are vararg parameters that are passed via
2774 // registers, then we must store them to their spots on the stack so
2775 // they may be loaded by deferencing the result of va_next.
2776 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
2777 FuncInfo->setVarArgsFPOffset(ArgGPRs.size() * 8 + NumXMMRegs * 16);
2778 FuncInfo->setRegSaveFrameIndex(MFI->CreateStackObject(
2779 ArgGPRs.size() * 8 + ArgXMMs.size() * 16, 16, false));
2782 // Store the integer parameter registers.
2783 SmallVector<SDValue, 8> MemOps;
2784 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
2785 getPointerTy(DAG.getDataLayout()));
2786 unsigned Offset = FuncInfo->getVarArgsGPOffset();
2787 for (SDValue Val : LiveGPRs) {
2788 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
2789 RSFIN, DAG.getIntPtrConstant(Offset, dl));
2791 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2792 MachinePointerInfo::getFixedStack(
2793 DAG.getMachineFunction(),
2794 FuncInfo->getRegSaveFrameIndex(), Offset),
2796 MemOps.push_back(Store);
2800 if (!ArgXMMs.empty() && NumXMMRegs != ArgXMMs.size()) {
2801 // Now store the XMM (fp + vector) parameter registers.
2802 SmallVector<SDValue, 12> SaveXMMOps;
2803 SaveXMMOps.push_back(Chain);
2804 SaveXMMOps.push_back(ALVal);
2805 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2806 FuncInfo->getRegSaveFrameIndex(), dl));
2807 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2808 FuncInfo->getVarArgsFPOffset(), dl));
2809 SaveXMMOps.insert(SaveXMMOps.end(), LiveXMMRegs.begin(),
2811 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
2812 MVT::Other, SaveXMMOps));
2815 if (!MemOps.empty())
2816 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2819 if (isVarArg && MFI->hasMustTailInVarArgFunc()) {
2820 // Find the largest legal vector type.
2821 MVT VecVT = MVT::Other;
2822 // FIXME: Only some x86_32 calling conventions support AVX512.
2823 if (Subtarget->hasAVX512() &&
2824 (Is64Bit || (CallConv == CallingConv::X86_VectorCall ||
2825 CallConv == CallingConv::Intel_OCL_BI)))
2826 VecVT = MVT::v16f32;
2827 else if (Subtarget->hasAVX())
2829 else if (Subtarget->hasSSE2())
2832 // We forward some GPRs and some vector types.
2833 SmallVector<MVT, 2> RegParmTypes;
2834 MVT IntVT = Is64Bit ? MVT::i64 : MVT::i32;
2835 RegParmTypes.push_back(IntVT);
2836 if (VecVT != MVT::Other)
2837 RegParmTypes.push_back(VecVT);
2839 // Compute the set of forwarded registers. The rest are scratch.
2840 SmallVectorImpl<ForwardedRegister> &Forwards =
2841 FuncInfo->getForwardedMustTailRegParms();
2842 CCInfo.analyzeMustTailForwardedRegisters(Forwards, RegParmTypes, CC_X86);
2844 // Conservatively forward AL on x86_64, since it might be used for varargs.
2845 if (Is64Bit && !CCInfo.isAllocated(X86::AL)) {
2846 unsigned ALVReg = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2847 Forwards.push_back(ForwardedRegister(ALVReg, X86::AL, MVT::i8));
2850 // Copy all forwards from physical to virtual registers.
2851 for (ForwardedRegister &F : Forwards) {
2852 // FIXME: Can we use a less constrained schedule?
2853 SDValue RegVal = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
2854 F.VReg = MF.getRegInfo().createVirtualRegister(getRegClassFor(F.VT));
2855 Chain = DAG.getCopyToReg(Chain, dl, F.VReg, RegVal);
2859 // Some CCs need callee pop.
2860 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
2861 MF.getTarget().Options.GuaranteedTailCallOpt)) {
2862 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
2864 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
2865 // If this is an sret function, the return should pop the hidden pointer.
2866 if (!Is64Bit && !canGuaranteeTCO(CallConv) &&
2867 !Subtarget->getTargetTriple().isOSMSVCRT() &&
2868 argsAreStructReturn(Ins) == StackStructReturn)
2869 FuncInfo->setBytesToPopOnReturn(4);
2873 // RegSaveFrameIndex is X86-64 only.
2874 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
2875 if (CallConv == CallingConv::X86_FastCall ||
2876 CallConv == CallingConv::X86_ThisCall)
2877 // fastcc functions can't have varargs.
2878 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
2881 FuncInfo->setArgumentStackSize(StackSize);
2883 if (WinEHFuncInfo *EHInfo = MF.getWinEHFuncInfo()) {
2884 EHPersonality Personality = classifyEHPersonality(Fn->getPersonalityFn());
2885 if (Personality == EHPersonality::CoreCLR) {
2887 // TODO: Add a mechanism to frame lowering that will allow us to indicate
2888 // that we'd prefer this slot be allocated towards the bottom of the frame
2889 // (i.e. near the stack pointer after allocating the frame). Every
2890 // funclet needs a copy of this slot in its (mostly empty) frame, and the
2891 // offset from the bottom of this and each funclet's frame must be the
2892 // same, so the size of funclets' (mostly empty) frames is dictated by
2893 // how far this slot is from the bottom (since they allocate just enough
2894 // space to accomodate holding this slot at the correct offset).
2895 int PSPSymFI = MFI->CreateStackObject(8, 8, /*isSS=*/false);
2896 EHInfo->PSPSymFrameIdx = PSPSymFI;
2904 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
2905 SDValue StackPtr, SDValue Arg,
2906 SDLoc dl, SelectionDAG &DAG,
2907 const CCValAssign &VA,
2908 ISD::ArgFlagsTy Flags) const {
2909 unsigned LocMemOffset = VA.getLocMemOffset();
2910 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
2911 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
2913 if (Flags.isByVal())
2914 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
2916 return DAG.getStore(
2917 Chain, dl, Arg, PtrOff,
2918 MachinePointerInfo::getStack(DAG.getMachineFunction(), LocMemOffset),
2922 /// Emit a load of return address if tail call
2923 /// optimization is performed and it is required.
2925 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
2926 SDValue &OutRetAddr, SDValue Chain,
2927 bool IsTailCall, bool Is64Bit,
2928 int FPDiff, SDLoc dl) const {
2929 // Adjust the Return address stack slot.
2930 EVT VT = getPointerTy(DAG.getDataLayout());
2931 OutRetAddr = getReturnAddressFrameIndex(DAG);
2933 // Load the "old" Return address.
2934 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
2935 false, false, false, 0);
2936 return SDValue(OutRetAddr.getNode(), 1);
2939 /// Emit a store of the return address if tail call
2940 /// optimization is performed and it is required (FPDiff!=0).
2941 static SDValue EmitTailCallStoreRetAddr(SelectionDAG &DAG, MachineFunction &MF,
2942 SDValue Chain, SDValue RetAddrFrIdx,
2943 EVT PtrVT, unsigned SlotSize,
2944 int FPDiff, SDLoc dl) {
2945 // Store the return address to the appropriate stack slot.
2946 if (!FPDiff) return Chain;
2947 // Calculate the new stack slot for the return address.
2948 int NewReturnAddrFI =
2949 MF.getFrameInfo()->CreateFixedObject(SlotSize, (int64_t)FPDiff - SlotSize,
2951 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, PtrVT);
2952 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
2953 MachinePointerInfo::getFixedStack(
2954 DAG.getMachineFunction(), NewReturnAddrFI),
2959 /// Returns a vector_shuffle mask for an movs{s|d}, movd
2960 /// operation of specified width.
2961 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
2963 unsigned NumElems = VT.getVectorNumElements();
2964 SmallVector<int, 8> Mask;
2965 Mask.push_back(NumElems);
2966 for (unsigned i = 1; i != NumElems; ++i)
2968 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
2972 X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2973 SmallVectorImpl<SDValue> &InVals) const {
2974 SelectionDAG &DAG = CLI.DAG;
2976 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
2977 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
2978 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
2979 SDValue Chain = CLI.Chain;
2980 SDValue Callee = CLI.Callee;
2981 CallingConv::ID CallConv = CLI.CallConv;
2982 bool &isTailCall = CLI.IsTailCall;
2983 bool isVarArg = CLI.IsVarArg;
2985 MachineFunction &MF = DAG.getMachineFunction();
2986 bool Is64Bit = Subtarget->is64Bit();
2987 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2988 StructReturnType SR = callIsStructReturn(Outs);
2989 bool IsSibcall = false;
2990 X86MachineFunctionInfo *X86Info = MF.getInfo<X86MachineFunctionInfo>();
2991 auto Attr = MF.getFunction()->getFnAttribute("disable-tail-calls");
2993 if (Attr.getValueAsString() == "true")
2996 if (Subtarget->isPICStyleGOT() &&
2997 !MF.getTarget().Options.GuaranteedTailCallOpt) {
2998 // If we are using a GOT, disable tail calls to external symbols with
2999 // default visibility. Tail calling such a symbol requires using a GOT
3000 // relocation, which forces early binding of the symbol. This breaks code
3001 // that require lazy function symbol resolution. Using musttail or
3002 // GuaranteedTailCallOpt will override this.
3003 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
3004 if (!G || (!G->getGlobal()->hasLocalLinkage() &&
3005 G->getGlobal()->hasDefaultVisibility()))
3009 bool IsMustTail = CLI.CS && CLI.CS->isMustTailCall();
3011 // Force this to be a tail call. The verifier rules are enough to ensure
3012 // that we can lower this successfully without moving the return address
3015 } else if (isTailCall) {
3016 // Check if it's really possible to do a tail call.
3017 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
3018 isVarArg, SR != NotStructReturn,
3019 MF.getFunction()->hasStructRetAttr(), CLI.RetTy,
3020 Outs, OutVals, Ins, DAG);
3022 // Sibcalls are automatically detected tailcalls which do not require
3024 if (!MF.getTarget().Options.GuaranteedTailCallOpt && isTailCall)
3031 assert(!(isVarArg && canGuaranteeTCO(CallConv)) &&
3032 "Var args not supported with calling convention fastcc, ghc or hipe");
3034 // Analyze operands of the call, assigning locations to each operand.
3035 SmallVector<CCValAssign, 16> ArgLocs;
3036 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
3038 // Allocate shadow area for Win64
3040 CCInfo.AllocateStack(32, 8);
3042 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3044 // Get a count of how many bytes are to be pushed on the stack.
3045 unsigned NumBytes = CCInfo.getAlignedCallFrameSize();
3047 // This is a sibcall. The memory operands are available in caller's
3048 // own caller's stack.
3050 else if (MF.getTarget().Options.GuaranteedTailCallOpt &&
3051 canGuaranteeTCO(CallConv))
3052 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
3055 if (isTailCall && !IsSibcall && !IsMustTail) {
3056 // Lower arguments at fp - stackoffset + fpdiff.
3057 unsigned NumBytesCallerPushed = X86Info->getBytesToPopOnReturn();
3059 FPDiff = NumBytesCallerPushed - NumBytes;
3061 // Set the delta of movement of the returnaddr stackslot.
3062 // But only set if delta is greater than previous delta.
3063 if (FPDiff < X86Info->getTCReturnAddrDelta())
3064 X86Info->setTCReturnAddrDelta(FPDiff);
3067 unsigned NumBytesToPush = NumBytes;
3068 unsigned NumBytesToPop = NumBytes;
3070 // If we have an inalloca argument, all stack space has already been allocated
3071 // for us and be right at the top of the stack. We don't support multiple
3072 // arguments passed in memory when using inalloca.
3073 if (!Outs.empty() && Outs.back().Flags.isInAlloca()) {
3075 if (!ArgLocs.back().isMemLoc())
3076 report_fatal_error("cannot use inalloca attribute on a register "
3078 if (ArgLocs.back().getLocMemOffset() != 0)
3079 report_fatal_error("any parameter with the inalloca attribute must be "
3080 "the only memory argument");
3084 Chain = DAG.getCALLSEQ_START(
3085 Chain, DAG.getIntPtrConstant(NumBytesToPush, dl, true), dl);
3087 SDValue RetAddrFrIdx;
3088 // Load return address for tail calls.
3089 if (isTailCall && FPDiff)
3090 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
3091 Is64Bit, FPDiff, dl);
3093 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3094 SmallVector<SDValue, 8> MemOpChains;
3097 // Walk the register/memloc assignments, inserting copies/loads. In the case
3098 // of tail call optimization arguments are handle later.
3099 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3100 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3101 // Skip inalloca arguments, they have already been written.
3102 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3103 if (Flags.isInAlloca())
3106 CCValAssign &VA = ArgLocs[i];
3107 EVT RegVT = VA.getLocVT();
3108 SDValue Arg = OutVals[i];
3109 bool isByVal = Flags.isByVal();
3111 // Promote the value if needed.
3112 switch (VA.getLocInfo()) {
3113 default: llvm_unreachable("Unknown loc info!");
3114 case CCValAssign::Full: break;
3115 case CCValAssign::SExt:
3116 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
3118 case CCValAssign::ZExt:
3119 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
3121 case CCValAssign::AExt:
3122 if (Arg.getValueType().isVector() &&
3123 Arg.getValueType().getVectorElementType() == MVT::i1)
3124 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
3125 else if (RegVT.is128BitVector()) {
3126 // Special case: passing MMX values in XMM registers.
3127 Arg = DAG.getBitcast(MVT::i64, Arg);
3128 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
3129 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
3131 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
3133 case CCValAssign::BCvt:
3134 Arg = DAG.getBitcast(RegVT, Arg);
3136 case CCValAssign::Indirect: {
3137 // Store the argument.
3138 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
3139 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
3140 Chain = DAG.getStore(
3141 Chain, dl, Arg, SpillSlot,
3142 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3149 if (VA.isRegLoc()) {
3150 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3151 if (isVarArg && IsWin64) {
3152 // Win64 ABI requires argument XMM reg to be copied to the corresponding
3153 // shadow reg if callee is a varargs function.
3154 unsigned ShadowReg = 0;
3155 switch (VA.getLocReg()) {
3156 case X86::XMM0: ShadowReg = X86::RCX; break;
3157 case X86::XMM1: ShadowReg = X86::RDX; break;
3158 case X86::XMM2: ShadowReg = X86::R8; break;
3159 case X86::XMM3: ShadowReg = X86::R9; break;
3162 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
3164 } else if (!IsSibcall && (!isTailCall || isByVal)) {
3165 assert(VA.isMemLoc());
3166 if (!StackPtr.getNode())
3167 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
3168 getPointerTy(DAG.getDataLayout()));
3169 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
3170 dl, DAG, VA, Flags));
3174 if (!MemOpChains.empty())
3175 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
3177 if (Subtarget->isPICStyleGOT()) {
3178 // ELF / PIC requires GOT in the EBX register before function calls via PLT
3181 RegsToPass.push_back(std::make_pair(
3182 unsigned(X86::EBX), DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(),
3183 getPointerTy(DAG.getDataLayout()))));
3185 // If we are tail calling and generating PIC/GOT style code load the
3186 // address of the callee into ECX. The value in ecx is used as target of
3187 // the tail jump. This is done to circumvent the ebx/callee-saved problem
3188 // for tail calls on PIC/GOT architectures. Normally we would just put the
3189 // address of GOT into ebx and then call target@PLT. But for tail calls
3190 // ebx would be restored (since ebx is callee saved) before jumping to the
3193 // Note: The actual moving to ECX is done further down.
3194 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
3195 if (G && !G->getGlobal()->hasLocalLinkage() &&
3196 G->getGlobal()->hasDefaultVisibility())
3197 Callee = LowerGlobalAddress(Callee, DAG);
3198 else if (isa<ExternalSymbolSDNode>(Callee))
3199 Callee = LowerExternalSymbol(Callee, DAG);
3203 if (Is64Bit && isVarArg && !IsWin64 && !IsMustTail) {
3204 // From AMD64 ABI document:
3205 // For calls that may call functions that use varargs or stdargs
3206 // (prototype-less calls or calls to functions containing ellipsis (...) in
3207 // the declaration) %al is used as hidden argument to specify the number
3208 // of SSE registers used. The contents of %al do not need to match exactly
3209 // the number of registers, but must be an ubound on the number of SSE
3210 // registers used and is in the range 0 - 8 inclusive.
3212 // Count the number of XMM registers allocated.
3213 static const MCPhysReg XMMArgRegs[] = {
3214 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
3215 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
3217 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs);
3218 assert((Subtarget->hasSSE1() || !NumXMMRegs)
3219 && "SSE registers cannot be used when SSE is disabled");
3221 RegsToPass.push_back(std::make_pair(unsigned(X86::AL),
3222 DAG.getConstant(NumXMMRegs, dl,
3226 if (isVarArg && IsMustTail) {
3227 const auto &Forwards = X86Info->getForwardedMustTailRegParms();
3228 for (const auto &F : Forwards) {
3229 SDValue Val = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
3230 RegsToPass.push_back(std::make_pair(unsigned(F.PReg), Val));
3234 // For tail calls lower the arguments to the 'real' stack slots. Sibcalls
3235 // don't need this because the eligibility check rejects calls that require
3236 // shuffling arguments passed in memory.
3237 if (!IsSibcall && isTailCall) {
3238 // Force all the incoming stack arguments to be loaded from the stack
3239 // before any new outgoing arguments are stored to the stack, because the
3240 // outgoing stack slots may alias the incoming argument stack slots, and
3241 // the alias isn't otherwise explicit. This is slightly more conservative
3242 // than necessary, because it means that each store effectively depends
3243 // on every argument instead of just those arguments it would clobber.
3244 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
3246 SmallVector<SDValue, 8> MemOpChains2;
3249 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3250 CCValAssign &VA = ArgLocs[i];
3253 assert(VA.isMemLoc());
3254 SDValue Arg = OutVals[i];
3255 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3256 // Skip inalloca arguments. They don't require any work.
3257 if (Flags.isInAlloca())
3259 // Create frame index.
3260 int32_t Offset = VA.getLocMemOffset()+FPDiff;
3261 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
3262 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
3263 FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
3265 if (Flags.isByVal()) {
3266 // Copy relative to framepointer.
3267 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset(), dl);
3268 if (!StackPtr.getNode())
3269 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
3270 getPointerTy(DAG.getDataLayout()));
3271 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
3274 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
3278 // Store relative to framepointer.
3279 MemOpChains2.push_back(DAG.getStore(
3280 ArgChain, dl, Arg, FIN,
3281 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3286 if (!MemOpChains2.empty())
3287 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains2);
3289 // Store the return address to the appropriate stack slot.
3290 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx,
3291 getPointerTy(DAG.getDataLayout()),
3292 RegInfo->getSlotSize(), FPDiff, dl);
3295 // Build a sequence of copy-to-reg nodes chained together with token chain
3296 // and flag operands which copy the outgoing args into registers.
3298 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3299 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3300 RegsToPass[i].second, InFlag);
3301 InFlag = Chain.getValue(1);
3304 if (DAG.getTarget().getCodeModel() == CodeModel::Large) {
3305 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
3306 // In the 64-bit large code model, we have to make all calls
3307 // through a register, since the call instruction's 32-bit
3308 // pc-relative offset may not be large enough to hold the whole
3310 } else if (Callee->getOpcode() == ISD::GlobalAddress) {
3311 // If the callee is a GlobalAddress node (quite common, every direct call
3312 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
3314 GlobalAddressSDNode* G = cast<GlobalAddressSDNode>(Callee);
3316 // We should use extra load for direct calls to dllimported functions in
3318 const GlobalValue *GV = G->getGlobal();
3319 if (!GV->hasDLLImportStorageClass()) {
3320 unsigned char OpFlags = 0;
3321 bool ExtraLoad = false;
3322 unsigned WrapperKind = ISD::DELETED_NODE;
3324 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
3325 // external symbols most go through the PLT in PIC mode. If the symbol
3326 // has hidden or protected visibility, or if it is static or local, then
3327 // we don't need to use the PLT - we can directly call it.
3328 if (Subtarget->isTargetELF() &&
3329 DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
3330 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
3331 OpFlags = X86II::MO_PLT;
3332 } else if (Subtarget->isPICStyleStubAny() &&
3333 !GV->isStrongDefinitionForLinker() &&
3334 (!Subtarget->getTargetTriple().isMacOSX() ||
3335 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3336 // PC-relative references to external symbols should go through $stub,
3337 // unless we're building with the leopard linker or later, which
3338 // automatically synthesizes these stubs.
3339 OpFlags = X86II::MO_DARWIN_STUB;
3340 } else if (Subtarget->isPICStyleRIPRel() && isa<Function>(GV) &&
3341 cast<Function>(GV)->hasFnAttribute(Attribute::NonLazyBind)) {
3342 // If the function is marked as non-lazy, generate an indirect call
3343 // which loads from the GOT directly. This avoids runtime overhead
3344 // at the cost of eager binding (and one extra byte of encoding).
3345 OpFlags = X86II::MO_GOTPCREL;
3346 WrapperKind = X86ISD::WrapperRIP;
3350 Callee = DAG.getTargetGlobalAddress(
3351 GV, dl, getPointerTy(DAG.getDataLayout()), G->getOffset(), OpFlags);
3353 // Add a wrapper if needed.
3354 if (WrapperKind != ISD::DELETED_NODE)
3355 Callee = DAG.getNode(X86ISD::WrapperRIP, dl,
3356 getPointerTy(DAG.getDataLayout()), Callee);
3357 // Add extra indirection if needed.
3359 Callee = DAG.getLoad(
3360 getPointerTy(DAG.getDataLayout()), dl, DAG.getEntryNode(), Callee,
3361 MachinePointerInfo::getGOT(DAG.getMachineFunction()), false, false,
3364 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3365 unsigned char OpFlags = 0;
3367 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
3368 // external symbols should go through the PLT.
3369 if (Subtarget->isTargetELF() &&
3370 DAG.getTarget().getRelocationModel() == Reloc::PIC_) {
3371 OpFlags = X86II::MO_PLT;
3372 } else if (Subtarget->isPICStyleStubAny() &&
3373 (!Subtarget->getTargetTriple().isMacOSX() ||
3374 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3375 // PC-relative references to external symbols should go through $stub,
3376 // unless we're building with the leopard linker or later, which
3377 // automatically synthesizes these stubs.
3378 OpFlags = X86II::MO_DARWIN_STUB;
3381 Callee = DAG.getTargetExternalSymbol(
3382 S->getSymbol(), getPointerTy(DAG.getDataLayout()), OpFlags);
3383 } else if (Subtarget->isTarget64BitILP32() &&
3384 Callee->getValueType(0) == MVT::i32) {
3385 // Zero-extend the 32-bit Callee address into a 64-bit according to x32 ABI
3386 Callee = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, Callee);
3389 // Returns a chain & a flag for retval copy to use.
3390 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3391 SmallVector<SDValue, 8> Ops;
3393 if (!IsSibcall && isTailCall) {
3394 Chain = DAG.getCALLSEQ_END(Chain,
3395 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3396 DAG.getIntPtrConstant(0, dl, true), InFlag, dl);
3397 InFlag = Chain.getValue(1);
3400 Ops.push_back(Chain);
3401 Ops.push_back(Callee);
3404 Ops.push_back(DAG.getConstant(FPDiff, dl, MVT::i32));
3406 // Add argument registers to the end of the list so that they are known live
3408 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3409 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3410 RegsToPass[i].second.getValueType()));
3412 // Add a register mask operand representing the call-preserved registers.
3413 const uint32_t *Mask = RegInfo->getCallPreservedMask(MF, CallConv);
3414 assert(Mask && "Missing call preserved mask for calling convention");
3416 // If this is an invoke in a 32-bit function using a funclet-based
3417 // personality, assume the function clobbers all registers. If an exception
3418 // is thrown, the runtime will not restore CSRs.
3419 // FIXME: Model this more precisely so that we can register allocate across
3420 // the normal edge and spill and fill across the exceptional edge.
3421 if (!Is64Bit && CLI.CS && CLI.CS->isInvoke()) {
3422 const Function *CallerFn = MF.getFunction();
3423 EHPersonality Pers =
3424 CallerFn->hasPersonalityFn()
3425 ? classifyEHPersonality(CallerFn->getPersonalityFn())
3426 : EHPersonality::Unknown;
3427 if (isFuncletEHPersonality(Pers))
3428 Mask = RegInfo->getNoPreservedMask();
3431 Ops.push_back(DAG.getRegisterMask(Mask));
3433 if (InFlag.getNode())
3434 Ops.push_back(InFlag);
3438 //// If this is the first return lowered for this function, add the regs
3439 //// to the liveout set for the function.
3440 // This isn't right, although it's probably harmless on x86; liveouts
3441 // should be computed from returns not tail calls. Consider a void
3442 // function making a tail call to a function returning int.
3443 MF.getFrameInfo()->setHasTailCall();
3444 return DAG.getNode(X86ISD::TC_RETURN, dl, NodeTys, Ops);
3447 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops);
3448 InFlag = Chain.getValue(1);
3450 // Create the CALLSEQ_END node.
3451 unsigned NumBytesForCalleeToPop;
3452 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
3453 DAG.getTarget().Options.GuaranteedTailCallOpt))
3454 NumBytesForCalleeToPop = NumBytes; // Callee pops everything
3455 else if (!Is64Bit && !canGuaranteeTCO(CallConv) &&
3456 !Subtarget->getTargetTriple().isOSMSVCRT() &&
3457 SR == StackStructReturn)
3458 // If this is a call to a struct-return function, the callee
3459 // pops the hidden struct pointer, so we have to push it back.
3460 // This is common for Darwin/X86, Linux & Mingw32 targets.
3461 // For MSVC Win32 targets, the caller pops the hidden struct pointer.
3462 NumBytesForCalleeToPop = 4;
3464 NumBytesForCalleeToPop = 0; // Callee pops nothing.
3466 // Returns a flag for retval copy to use.
3468 Chain = DAG.getCALLSEQ_END(Chain,
3469 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3470 DAG.getIntPtrConstant(NumBytesForCalleeToPop, dl,
3473 InFlag = Chain.getValue(1);
3476 // Handle result values, copying them out of physregs into vregs that we
3478 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3479 Ins, dl, DAG, InVals);
3482 //===----------------------------------------------------------------------===//
3483 // Fast Calling Convention (tail call) implementation
3484 //===----------------------------------------------------------------------===//
3486 // Like std call, callee cleans arguments, convention except that ECX is
3487 // reserved for storing the tail called function address. Only 2 registers are
3488 // free for argument passing (inreg). Tail call optimization is performed
3490 // * tailcallopt is enabled
3491 // * caller/callee are fastcc
3492 // On X86_64 architecture with GOT-style position independent code only local
3493 // (within module) calls are supported at the moment.
3494 // To keep the stack aligned according to platform abi the function
3495 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
3496 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
3497 // If a tail called function callee has more arguments than the caller the
3498 // caller needs to make sure that there is room to move the RETADDR to. This is
3499 // achieved by reserving an area the size of the argument delta right after the
3500 // original RETADDR, but before the saved framepointer or the spilled registers
3501 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
3513 /// Make the stack size align e.g 16n + 12 aligned for a 16-byte align
3516 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
3517 SelectionDAG& DAG) const {
3518 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3519 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
3520 unsigned StackAlignment = TFI.getStackAlignment();
3521 uint64_t AlignMask = StackAlignment - 1;
3522 int64_t Offset = StackSize;
3523 unsigned SlotSize = RegInfo->getSlotSize();
3524 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
3525 // Number smaller than 12 so just add the difference.
3526 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
3528 // Mask out lower bits, add stackalignment once plus the 12 bytes.
3529 Offset = ((~AlignMask) & Offset) + StackAlignment +
3530 (StackAlignment-SlotSize);
3535 /// Return true if the given stack call argument is already available in the
3536 /// same position (relatively) of the caller's incoming argument stack.
3538 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
3539 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
3540 const X86InstrInfo *TII) {
3541 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
3543 if (Arg.getOpcode() == ISD::CopyFromReg) {
3544 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
3545 if (!TargetRegisterInfo::isVirtualRegister(VR))
3547 MachineInstr *Def = MRI->getVRegDef(VR);
3550 if (!Flags.isByVal()) {
3551 if (!TII->isLoadFromStackSlot(Def, FI))
3554 unsigned Opcode = Def->getOpcode();
3555 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r ||
3556 Opcode == X86::LEA64_32r) &&
3557 Def->getOperand(1).isFI()) {
3558 FI = Def->getOperand(1).getIndex();
3559 Bytes = Flags.getByValSize();
3563 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
3564 if (Flags.isByVal())
3565 // ByVal argument is passed in as a pointer but it's now being
3566 // dereferenced. e.g.
3567 // define @foo(%struct.X* %A) {
3568 // tail call @bar(%struct.X* byval %A)
3571 SDValue Ptr = Ld->getBasePtr();
3572 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
3575 FI = FINode->getIndex();
3576 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
3577 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
3578 FI = FINode->getIndex();
3579 Bytes = Flags.getByValSize();
3583 assert(FI != INT_MAX);
3584 if (!MFI->isFixedObjectIndex(FI))
3586 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
3589 /// Check whether the call is eligible for tail call optimization. Targets
3590 /// that want to do tail call optimization should implement this function.
3591 bool X86TargetLowering::IsEligibleForTailCallOptimization(
3592 SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg,
3593 bool isCalleeStructRet, bool isCallerStructRet, Type *RetTy,
3594 const SmallVectorImpl<ISD::OutputArg> &Outs,
3595 const SmallVectorImpl<SDValue> &OutVals,
3596 const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const {
3597 if (!mayTailCallThisCC(CalleeCC))
3600 // If -tailcallopt is specified, make fastcc functions tail-callable.
3601 MachineFunction &MF = DAG.getMachineFunction();
3602 const Function *CallerF = MF.getFunction();
3604 // If the function return type is x86_fp80 and the callee return type is not,
3605 // then the FP_EXTEND of the call result is not a nop. It's not safe to
3606 // perform a tailcall optimization here.
3607 if (CallerF->getReturnType()->isX86_FP80Ty() && !RetTy->isX86_FP80Ty())
3610 CallingConv::ID CallerCC = CallerF->getCallingConv();
3611 bool CCMatch = CallerCC == CalleeCC;
3612 bool IsCalleeWin64 = Subtarget->isCallingConvWin64(CalleeCC);
3613 bool IsCallerWin64 = Subtarget->isCallingConvWin64(CallerCC);
3615 // Win64 functions have extra shadow space for argument homing. Don't do the
3616 // sibcall if the caller and callee have mismatched expectations for this
3618 if (IsCalleeWin64 != IsCallerWin64)
3621 if (DAG.getTarget().Options.GuaranteedTailCallOpt) {
3622 if (canGuaranteeTCO(CalleeCC) && CCMatch)
3627 // Look for obvious safe cases to perform tail call optimization that do not
3628 // require ABI changes. This is what gcc calls sibcall.
3630 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
3631 // emit a special epilogue.
3632 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3633 if (RegInfo->needsStackRealignment(MF))
3636 // Also avoid sibcall optimization if either caller or callee uses struct
3637 // return semantics.
3638 if (isCalleeStructRet || isCallerStructRet)
3641 // Do not sibcall optimize vararg calls unless all arguments are passed via
3643 if (isVarArg && !Outs.empty()) {
3644 // Optimizing for varargs on Win64 is unlikely to be safe without
3645 // additional testing.
3646 if (IsCalleeWin64 || IsCallerWin64)
3649 SmallVector<CCValAssign, 16> ArgLocs;
3650 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3653 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3654 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
3655 if (!ArgLocs[i].isRegLoc())
3659 // If the call result is in ST0 / ST1, it needs to be popped off the x87
3660 // stack. Therefore, if it's not used by the call it is not safe to optimize
3661 // this into a sibcall.
3662 bool Unused = false;
3663 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
3670 SmallVector<CCValAssign, 16> RVLocs;
3671 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(), RVLocs,
3673 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
3674 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3675 CCValAssign &VA = RVLocs[i];
3676 if (VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1)
3681 // If the calling conventions do not match, then we'd better make sure the
3682 // results are returned in the same way as what the caller expects.
3684 SmallVector<CCValAssign, 16> RVLocs1;
3685 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
3687 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
3689 SmallVector<CCValAssign, 16> RVLocs2;
3690 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
3692 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
3694 if (RVLocs1.size() != RVLocs2.size())
3696 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
3697 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
3699 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
3701 if (RVLocs1[i].isRegLoc()) {
3702 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
3705 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
3711 unsigned StackArgsSize = 0;
3713 // If the callee takes no arguments then go on to check the results of the
3715 if (!Outs.empty()) {
3716 // Check if stack adjustment is needed. For now, do not do this if any
3717 // argument is passed on the stack.
3718 SmallVector<CCValAssign, 16> ArgLocs;
3719 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3722 // Allocate shadow area for Win64
3724 CCInfo.AllocateStack(32, 8);
3726 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3727 StackArgsSize = CCInfo.getNextStackOffset();
3729 if (CCInfo.getNextStackOffset()) {
3730 // Check if the arguments are already laid out in the right way as
3731 // the caller's fixed stack objects.
3732 MachineFrameInfo *MFI = MF.getFrameInfo();
3733 const MachineRegisterInfo *MRI = &MF.getRegInfo();
3734 const X86InstrInfo *TII = Subtarget->getInstrInfo();
3735 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3736 CCValAssign &VA = ArgLocs[i];
3737 SDValue Arg = OutVals[i];
3738 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3739 if (VA.getLocInfo() == CCValAssign::Indirect)
3741 if (!VA.isRegLoc()) {
3742 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
3749 // If the tailcall address may be in a register, then make sure it's
3750 // possible to register allocate for it. In 32-bit, the call address can
3751 // only target EAX, EDX, or ECX since the tail call must be scheduled after
3752 // callee-saved registers are restored. These happen to be the same
3753 // registers used to pass 'inreg' arguments so watch out for those.
3754 if (!Subtarget->is64Bit() &&
3755 ((!isa<GlobalAddressSDNode>(Callee) &&
3756 !isa<ExternalSymbolSDNode>(Callee)) ||
3757 DAG.getTarget().getRelocationModel() == Reloc::PIC_)) {
3758 unsigned NumInRegs = 0;
3759 // In PIC we need an extra register to formulate the address computation
3761 unsigned MaxInRegs =
3762 (DAG.getTarget().getRelocationModel() == Reloc::PIC_) ? 2 : 3;
3764 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3765 CCValAssign &VA = ArgLocs[i];
3768 unsigned Reg = VA.getLocReg();
3771 case X86::EAX: case X86::EDX: case X86::ECX:
3772 if (++NumInRegs == MaxInRegs)
3780 bool CalleeWillPop =
3781 X86::isCalleePop(CalleeCC, Subtarget->is64Bit(), isVarArg,
3782 MF.getTarget().Options.GuaranteedTailCallOpt);
3784 if (unsigned BytesToPop =
3785 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn()) {
3786 // If we have bytes to pop, the callee must pop them.
3787 bool CalleePopMatches = CalleeWillPop && BytesToPop == StackArgsSize;
3788 if (!CalleePopMatches)
3790 } else if (CalleeWillPop && StackArgsSize > 0) {
3791 // If we don't have bytes to pop, make sure the callee doesn't pop any.
3799 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
3800 const TargetLibraryInfo *libInfo) const {
3801 return X86::createFastISel(funcInfo, libInfo);
3804 //===----------------------------------------------------------------------===//
3805 // Other Lowering Hooks
3806 //===----------------------------------------------------------------------===//
3808 static bool MayFoldLoad(SDValue Op) {
3809 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
3812 static bool MayFoldIntoStore(SDValue Op) {
3813 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
3816 static bool isTargetShuffle(unsigned Opcode) {
3818 default: return false;
3819 case X86ISD::BLENDI:
3820 case X86ISD::PSHUFB:
3821 case X86ISD::PSHUFD:
3822 case X86ISD::PSHUFHW:
3823 case X86ISD::PSHUFLW:
3825 case X86ISD::PALIGNR:
3826 case X86ISD::MOVLHPS:
3827 case X86ISD::MOVLHPD:
3828 case X86ISD::MOVHLPS:
3829 case X86ISD::MOVLPS:
3830 case X86ISD::MOVLPD:
3831 case X86ISD::MOVSHDUP:
3832 case X86ISD::MOVSLDUP:
3833 case X86ISD::MOVDDUP:
3836 case X86ISD::UNPCKL:
3837 case X86ISD::UNPCKH:
3838 case X86ISD::VPERMILPI:
3839 case X86ISD::VPERM2X128:
3840 case X86ISD::VPERMI:
3841 case X86ISD::VPERMV:
3842 case X86ISD::VPERMV3:
3847 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, MVT VT,
3848 SDValue V1, unsigned TargetMask,
3849 SelectionDAG &DAG) {
3851 default: llvm_unreachable("Unknown x86 shuffle node");
3852 case X86ISD::PSHUFD:
3853 case X86ISD::PSHUFHW:
3854 case X86ISD::PSHUFLW:
3855 case X86ISD::VPERMILPI:
3856 case X86ISD::VPERMI:
3857 return DAG.getNode(Opc, dl, VT, V1,
3858 DAG.getConstant(TargetMask, dl, MVT::i8));
3862 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, MVT VT,
3863 SDValue V1, SDValue V2, SelectionDAG &DAG) {
3865 default: llvm_unreachable("Unknown x86 shuffle node");
3866 case X86ISD::MOVLHPS:
3867 case X86ISD::MOVLHPD:
3868 case X86ISD::MOVHLPS:
3869 case X86ISD::MOVLPS:
3870 case X86ISD::MOVLPD:
3873 case X86ISD::UNPCKL:
3874 case X86ISD::UNPCKH:
3875 return DAG.getNode(Opc, dl, VT, V1, V2);
3879 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
3880 MachineFunction &MF = DAG.getMachineFunction();
3881 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3882 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
3883 int ReturnAddrIndex = FuncInfo->getRAIndex();
3885 if (ReturnAddrIndex == 0) {
3886 // Set up a frame object for the return address.
3887 unsigned SlotSize = RegInfo->getSlotSize();
3888 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize,
3891 FuncInfo->setRAIndex(ReturnAddrIndex);
3894 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy(DAG.getDataLayout()));
3897 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
3898 bool hasSymbolicDisplacement) {
3899 // Offset should fit into 32 bit immediate field.
3900 if (!isInt<32>(Offset))
3903 // If we don't have a symbolic displacement - we don't have any extra
3905 if (!hasSymbolicDisplacement)
3908 // FIXME: Some tweaks might be needed for medium code model.
3909 if (M != CodeModel::Small && M != CodeModel::Kernel)
3912 // For small code model we assume that latest object is 16MB before end of 31
3913 // bits boundary. We may also accept pretty large negative constants knowing
3914 // that all objects are in the positive half of address space.
3915 if (M == CodeModel::Small && Offset < 16*1024*1024)
3918 // For kernel code model we know that all object resist in the negative half
3919 // of 32bits address space. We may not accept negative offsets, since they may
3920 // be just off and we may accept pretty large positive ones.
3921 if (M == CodeModel::Kernel && Offset >= 0)
3927 /// Determines whether the callee is required to pop its own arguments.
3928 /// Callee pop is necessary to support tail calls.
3929 bool X86::isCalleePop(CallingConv::ID CallingConv,
3930 bool is64Bit, bool IsVarArg, bool GuaranteeTCO) {
3931 // If GuaranteeTCO is true, we force some calls to be callee pop so that we
3932 // can guarantee TCO.
3933 if (!IsVarArg && shouldGuaranteeTCO(CallingConv, GuaranteeTCO))
3936 switch (CallingConv) {
3939 case CallingConv::X86_StdCall:
3940 case CallingConv::X86_FastCall:
3941 case CallingConv::X86_ThisCall:
3942 case CallingConv::X86_VectorCall:
3947 /// \brief Return true if the condition is an unsigned comparison operation.
3948 static bool isX86CCUnsigned(unsigned X86CC) {
3950 default: llvm_unreachable("Invalid integer condition!");
3951 case X86::COND_E: return true;
3952 case X86::COND_G: return false;
3953 case X86::COND_GE: return false;
3954 case X86::COND_L: return false;
3955 case X86::COND_LE: return false;
3956 case X86::COND_NE: return true;
3957 case X86::COND_B: return true;
3958 case X86::COND_A: return true;
3959 case X86::COND_BE: return true;
3960 case X86::COND_AE: return true;
3964 static X86::CondCode TranslateIntegerX86CC(ISD::CondCode SetCCOpcode) {
3965 switch (SetCCOpcode) {
3966 default: llvm_unreachable("Invalid integer condition!");
3967 case ISD::SETEQ: return X86::COND_E;
3968 case ISD::SETGT: return X86::COND_G;
3969 case ISD::SETGE: return X86::COND_GE;
3970 case ISD::SETLT: return X86::COND_L;
3971 case ISD::SETLE: return X86::COND_LE;
3972 case ISD::SETNE: return X86::COND_NE;
3973 case ISD::SETULT: return X86::COND_B;
3974 case ISD::SETUGT: return X86::COND_A;
3975 case ISD::SETULE: return X86::COND_BE;
3976 case ISD::SETUGE: return X86::COND_AE;
3980 /// Do a one-to-one translation of a ISD::CondCode to the X86-specific
3981 /// condition code, returning the condition code and the LHS/RHS of the
3982 /// comparison to make.
3983 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, SDLoc DL, bool isFP,
3984 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
3986 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
3987 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
3988 // X > -1 -> X == 0, jump !sign.
3989 RHS = DAG.getConstant(0, DL, RHS.getValueType());
3990 return X86::COND_NS;
3992 if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
3993 // X < 0 -> X == 0, jump on sign.
3996 if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
3998 RHS = DAG.getConstant(0, DL, RHS.getValueType());
3999 return X86::COND_LE;
4003 return TranslateIntegerX86CC(SetCCOpcode);
4006 // First determine if it is required or is profitable to flip the operands.
4008 // If LHS is a foldable load, but RHS is not, flip the condition.
4009 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
4010 !ISD::isNON_EXTLoad(RHS.getNode())) {
4011 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
4012 std::swap(LHS, RHS);
4015 switch (SetCCOpcode) {
4021 std::swap(LHS, RHS);
4025 // On a floating point condition, the flags are set as follows:
4027 // 0 | 0 | 0 | X > Y
4028 // 0 | 0 | 1 | X < Y
4029 // 1 | 0 | 0 | X == Y
4030 // 1 | 1 | 1 | unordered
4031 switch (SetCCOpcode) {
4032 default: llvm_unreachable("Condcode should be pre-legalized away");
4034 case ISD::SETEQ: return X86::COND_E;
4035 case ISD::SETOLT: // flipped
4037 case ISD::SETGT: return X86::COND_A;
4038 case ISD::SETOLE: // flipped
4040 case ISD::SETGE: return X86::COND_AE;
4041 case ISD::SETUGT: // flipped
4043 case ISD::SETLT: return X86::COND_B;
4044 case ISD::SETUGE: // flipped
4046 case ISD::SETLE: return X86::COND_BE;
4048 case ISD::SETNE: return X86::COND_NE;
4049 case ISD::SETUO: return X86::COND_P;
4050 case ISD::SETO: return X86::COND_NP;
4052 case ISD::SETUNE: return X86::COND_INVALID;
4056 /// Is there a floating point cmov for the specific X86 condition code?
4057 /// Current x86 isa includes the following FP cmov instructions:
4058 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
4059 static bool hasFPCMov(unsigned X86CC) {
4075 /// Returns true if the target can instruction select the
4076 /// specified FP immediate natively. If false, the legalizer will
4077 /// materialize the FP immediate as a load from a constant pool.
4078 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4079 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
4080 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
4086 bool X86TargetLowering::shouldReduceLoadWidth(SDNode *Load,
4087 ISD::LoadExtType ExtTy,
4089 // "ELF Handling for Thread-Local Storage" specifies that R_X86_64_GOTTPOFF
4090 // relocation target a movq or addq instruction: don't let the load shrink.
4091 SDValue BasePtr = cast<LoadSDNode>(Load)->getBasePtr();
4092 if (BasePtr.getOpcode() == X86ISD::WrapperRIP)
4093 if (const auto *GA = dyn_cast<GlobalAddressSDNode>(BasePtr.getOperand(0)))
4094 return GA->getTargetFlags() != X86II::MO_GOTTPOFF;
4098 /// \brief Returns true if it is beneficial to convert a load of a constant
4099 /// to just the constant itself.
4100 bool X86TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
4102 assert(Ty->isIntegerTy());
4104 unsigned BitSize = Ty->getPrimitiveSizeInBits();
4105 if (BitSize == 0 || BitSize > 64)
4110 bool X86TargetLowering::isExtractSubvectorCheap(EVT ResVT,
4111 unsigned Index) const {
4112 if (!isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, ResVT))
4115 return (Index == 0 || Index == ResVT.getVectorNumElements());
4118 bool X86TargetLowering::isCheapToSpeculateCttz() const {
4119 // Speculate cttz only if we can directly use TZCNT.
4120 return Subtarget->hasBMI();
4123 bool X86TargetLowering::isCheapToSpeculateCtlz() const {
4124 // Speculate ctlz only if we can directly use LZCNT.
4125 return Subtarget->hasLZCNT();
4128 /// Return true if every element in Mask, beginning
4129 /// from position Pos and ending in Pos+Size is undef.
4130 static bool isUndefInRange(ArrayRef<int> Mask, unsigned Pos, unsigned Size) {
4131 for (unsigned i = Pos, e = Pos + Size; i != e; ++i)
4137 /// Return true if Val is undef or if its value falls within the
4138 /// specified range (L, H].
4139 static bool isUndefOrInRange(int Val, int Low, int Hi) {
4140 return (Val < 0) || (Val >= Low && Val < Hi);
4143 /// Val is either less than zero (undef) or equal to the specified value.
4144 static bool isUndefOrEqual(int Val, int CmpVal) {
4145 return (Val < 0 || Val == CmpVal);
4148 /// Return true if every element in Mask, beginning
4149 /// from position Pos and ending in Pos+Size, falls within the specified
4150 /// sequential range (Low, Low+Size]. or is undef.
4151 static bool isSequentialOrUndefInRange(ArrayRef<int> Mask,
4152 unsigned Pos, unsigned Size, int Low) {
4153 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
4154 if (!isUndefOrEqual(Mask[i], Low))
4159 /// Return true if the specified EXTRACT_SUBVECTOR operand specifies a vector
4160 /// extract that is suitable for instruction that extract 128 or 256 bit vectors
4161 static bool isVEXTRACTIndex(SDNode *N, unsigned vecWidth) {
4162 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4163 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4166 // The index should be aligned on a vecWidth-bit boundary.
4168 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4170 MVT VT = N->getSimpleValueType(0);
4171 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4172 bool Result = (Index * ElSize) % vecWidth == 0;
4177 /// Return true if the specified INSERT_SUBVECTOR
4178 /// operand specifies a subvector insert that is suitable for input to
4179 /// insertion of 128 or 256-bit subvectors
4180 static bool isVINSERTIndex(SDNode *N, unsigned vecWidth) {
4181 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4182 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4184 // The index should be aligned on a vecWidth-bit boundary.
4186 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4188 MVT VT = N->getSimpleValueType(0);
4189 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4190 bool Result = (Index * ElSize) % vecWidth == 0;
4195 bool X86::isVINSERT128Index(SDNode *N) {
4196 return isVINSERTIndex(N, 128);
4199 bool X86::isVINSERT256Index(SDNode *N) {
4200 return isVINSERTIndex(N, 256);
4203 bool X86::isVEXTRACT128Index(SDNode *N) {
4204 return isVEXTRACTIndex(N, 128);
4207 bool X86::isVEXTRACT256Index(SDNode *N) {
4208 return isVEXTRACTIndex(N, 256);
4211 static unsigned getExtractVEXTRACTImmediate(SDNode *N, unsigned vecWidth) {
4212 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4213 assert(isa<ConstantSDNode>(N->getOperand(1).getNode()) &&
4214 "Illegal extract subvector for VEXTRACT");
4217 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4219 MVT VecVT = N->getOperand(0).getSimpleValueType();
4220 MVT ElVT = VecVT.getVectorElementType();
4222 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4223 return Index / NumElemsPerChunk;
4226 static unsigned getInsertVINSERTImmediate(SDNode *N, unsigned vecWidth) {
4227 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4228 assert(isa<ConstantSDNode>(N->getOperand(2).getNode()) &&
4229 "Illegal insert subvector for VINSERT");
4232 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4234 MVT VecVT = N->getSimpleValueType(0);
4235 MVT ElVT = VecVT.getVectorElementType();
4237 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4238 return Index / NumElemsPerChunk;
4241 /// Return the appropriate immediate to extract the specified
4242 /// EXTRACT_SUBVECTOR index with VEXTRACTF128 and VINSERTI128 instructions.
4243 unsigned X86::getExtractVEXTRACT128Immediate(SDNode *N) {
4244 return getExtractVEXTRACTImmediate(N, 128);
4247 /// Return the appropriate immediate to extract the specified
4248 /// EXTRACT_SUBVECTOR index with VEXTRACTF64x4 and VINSERTI64x4 instructions.
4249 unsigned X86::getExtractVEXTRACT256Immediate(SDNode *N) {
4250 return getExtractVEXTRACTImmediate(N, 256);
4253 /// Return the appropriate immediate to insert at the specified
4254 /// INSERT_SUBVECTOR index with VINSERTF128 and VINSERTI128 instructions.
4255 unsigned X86::getInsertVINSERT128Immediate(SDNode *N) {
4256 return getInsertVINSERTImmediate(N, 128);
4259 /// Return the appropriate immediate to insert at the specified
4260 /// INSERT_SUBVECTOR index with VINSERTF46x4 and VINSERTI64x4 instructions.
4261 unsigned X86::getInsertVINSERT256Immediate(SDNode *N) {
4262 return getInsertVINSERTImmediate(N, 256);
4265 /// Returns true if V is a constant integer zero.
4266 static bool isZero(SDValue V) {
4267 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
4268 return C && C->isNullValue();
4271 /// Returns true if Elt is a constant zero or a floating point constant +0.0.
4272 bool X86::isZeroNode(SDValue Elt) {
4275 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Elt))
4276 return CFP->getValueAPF().isPosZero();
4280 // Build a vector of constants
4281 // Use an UNDEF node if MaskElt == -1.
4282 // Spilt 64-bit constants in the 32-bit mode.
4283 static SDValue getConstVector(ArrayRef<int> Values, MVT VT,
4285 SDLoc dl, bool IsMask = false) {
4287 SmallVector<SDValue, 32> Ops;
4290 MVT ConstVecVT = VT;
4291 unsigned NumElts = VT.getVectorNumElements();
4292 bool In64BitMode = DAG.getTargetLoweringInfo().isTypeLegal(MVT::i64);
4293 if (!In64BitMode && VT.getVectorElementType() == MVT::i64) {
4294 ConstVecVT = MVT::getVectorVT(MVT::i32, NumElts * 2);
4298 MVT EltVT = ConstVecVT.getVectorElementType();
4299 for (unsigned i = 0; i < NumElts; ++i) {
4300 bool IsUndef = Values[i] < 0 && IsMask;
4301 SDValue OpNode = IsUndef ? DAG.getUNDEF(EltVT) :
4302 DAG.getConstant(Values[i], dl, EltVT);
4303 Ops.push_back(OpNode);
4305 Ops.push_back(IsUndef ? DAG.getUNDEF(EltVT) :
4306 DAG.getConstant(0, dl, EltVT));
4308 SDValue ConstsNode = DAG.getNode(ISD::BUILD_VECTOR, dl, ConstVecVT, Ops);
4310 ConstsNode = DAG.getBitcast(VT, ConstsNode);
4314 /// Returns a vector of specified type with all zero elements.
4315 static SDValue getZeroVector(EVT VT, const X86Subtarget *Subtarget,
4316 SelectionDAG &DAG, SDLoc dl) {
4317 assert(VT.isVector() && "Expected a vector type");
4319 // Always build SSE zero vectors as <4 x i32> bitcasted
4320 // to their dest type. This ensures they get CSE'd.
4322 if (VT.is128BitVector()) { // SSE
4323 if (Subtarget->hasSSE2()) { // SSE2
4324 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4325 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4327 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
4328 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
4330 } else if (VT.is256BitVector()) { // AVX
4331 if (Subtarget->hasInt256()) { // AVX2
4332 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4333 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4334 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4336 // 256-bit logic and arithmetic instructions in AVX are all
4337 // floating-point, no support for integer ops. Emit fp zeroed vectors.
4338 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
4339 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4340 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops);
4342 } else if (VT.is512BitVector()) { // AVX-512
4343 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4344 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
4345 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4346 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
4347 } else if (VT.getVectorElementType() == MVT::i1) {
4349 assert((Subtarget->hasBWI() || VT.getVectorNumElements() <= 16)
4350 && "Unexpected vector type");
4351 assert((Subtarget->hasVLX() || VT.getVectorNumElements() >= 8)
4352 && "Unexpected vector type");
4353 SDValue Cst = DAG.getConstant(0, dl, MVT::i1);
4354 SmallVector<SDValue, 64> Ops(VT.getVectorNumElements(), Cst);
4355 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
4357 llvm_unreachable("Unexpected vector type");
4359 return DAG.getBitcast(VT, Vec);
4362 static SDValue ExtractSubVector(SDValue Vec, unsigned IdxVal,
4363 SelectionDAG &DAG, SDLoc dl,
4364 unsigned vectorWidth) {
4365 assert((vectorWidth == 128 || vectorWidth == 256) &&
4366 "Unsupported vector width");
4367 EVT VT = Vec.getValueType();
4368 EVT ElVT = VT.getVectorElementType();
4369 unsigned Factor = VT.getSizeInBits()/vectorWidth;
4370 EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT,
4371 VT.getVectorNumElements()/Factor);
4373 // Extract from UNDEF is UNDEF.
4374 if (Vec.getOpcode() == ISD::UNDEF)
4375 return DAG.getUNDEF(ResultVT);
4377 // Extract the relevant vectorWidth bits. Generate an EXTRACT_SUBVECTOR
4378 unsigned ElemsPerChunk = vectorWidth / ElVT.getSizeInBits();
4379 assert(isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2");
4381 // This is the index of the first element of the vectorWidth-bit chunk
4382 // we want. Since ElemsPerChunk is a power of 2 just need to clear bits.
4383 IdxVal &= ~(ElemsPerChunk - 1);
4385 // If the input is a buildvector just emit a smaller one.
4386 if (Vec.getOpcode() == ISD::BUILD_VECTOR)
4387 return DAG.getNode(ISD::BUILD_VECTOR, dl, ResultVT,
4388 makeArrayRef(Vec->op_begin() + IdxVal, ElemsPerChunk));
4390 SDValue VecIdx = DAG.getIntPtrConstant(IdxVal, dl);
4391 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec, VecIdx);
4394 /// Generate a DAG to grab 128-bits from a vector > 128 bits. This
4395 /// sets things up to match to an AVX VEXTRACTF128 / VEXTRACTI128
4396 /// or AVX-512 VEXTRACTF32x4 / VEXTRACTI32x4
4397 /// instructions or a simple subregister reference. Idx is an index in the
4398 /// 128 bits we want. It need not be aligned to a 128-bit boundary. That makes
4399 /// lowering EXTRACT_VECTOR_ELT operations easier.
4400 static SDValue Extract128BitVector(SDValue Vec, unsigned IdxVal,
4401 SelectionDAG &DAG, SDLoc dl) {
4402 assert((Vec.getValueType().is256BitVector() ||
4403 Vec.getValueType().is512BitVector()) && "Unexpected vector size!");
4404 return ExtractSubVector(Vec, IdxVal, DAG, dl, 128);
4407 /// Generate a DAG to grab 256-bits from a 512-bit vector.
4408 static SDValue Extract256BitVector(SDValue Vec, unsigned IdxVal,
4409 SelectionDAG &DAG, SDLoc dl) {
4410 assert(Vec.getValueType().is512BitVector() && "Unexpected vector size!");
4411 return ExtractSubVector(Vec, IdxVal, DAG, dl, 256);
4414 static SDValue InsertSubVector(SDValue Result, SDValue Vec,
4415 unsigned IdxVal, SelectionDAG &DAG,
4416 SDLoc dl, unsigned vectorWidth) {
4417 assert((vectorWidth == 128 || vectorWidth == 256) &&
4418 "Unsupported vector width");
4419 // Inserting UNDEF is Result
4420 if (Vec.getOpcode() == ISD::UNDEF)
4422 EVT VT = Vec.getValueType();
4423 EVT ElVT = VT.getVectorElementType();
4424 EVT ResultVT = Result.getValueType();
4426 // Insert the relevant vectorWidth bits.
4427 unsigned ElemsPerChunk = vectorWidth/ElVT.getSizeInBits();
4428 assert(isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2");
4430 // This is the index of the first element of the vectorWidth-bit chunk
4431 // we want. Since ElemsPerChunk is a power of 2 just need to clear bits.
4432 IdxVal &= ~(ElemsPerChunk - 1);
4434 SDValue VecIdx = DAG.getIntPtrConstant(IdxVal, dl);
4435 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec, VecIdx);
4438 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
4439 /// sets things up to match to an AVX VINSERTF128/VINSERTI128 or
4440 /// AVX-512 VINSERTF32x4/VINSERTI32x4 instructions or a
4441 /// simple superregister reference. Idx is an index in the 128 bits
4442 /// we want. It need not be aligned to a 128-bit boundary. That makes
4443 /// lowering INSERT_VECTOR_ELT operations easier.
4444 static SDValue Insert128BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4445 SelectionDAG &DAG, SDLoc dl) {
4446 assert(Vec.getValueType().is128BitVector() && "Unexpected vector size!");
4448 // For insertion into the zero index (low half) of a 256-bit vector, it is
4449 // more efficient to generate a blend with immediate instead of an insert*128.
4450 // We are still creating an INSERT_SUBVECTOR below with an undef node to
4451 // extend the subvector to the size of the result vector. Make sure that
4452 // we are not recursing on that node by checking for undef here.
4453 if (IdxVal == 0 && Result.getValueType().is256BitVector() &&
4454 Result.getOpcode() != ISD::UNDEF) {
4455 EVT ResultVT = Result.getValueType();
4456 SDValue ZeroIndex = DAG.getIntPtrConstant(0, dl);
4457 SDValue Undef = DAG.getUNDEF(ResultVT);
4458 SDValue Vec256 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Undef,
4461 // The blend instruction, and therefore its mask, depend on the data type.
4462 MVT ScalarType = ResultVT.getVectorElementType().getSimpleVT();
4463 if (ScalarType.isFloatingPoint()) {
4464 // Choose either vblendps (float) or vblendpd (double).
4465 unsigned ScalarSize = ScalarType.getSizeInBits();
4466 assert((ScalarSize == 64 || ScalarSize == 32) && "Unknown float type");
4467 unsigned MaskVal = (ScalarSize == 64) ? 0x03 : 0x0f;
4468 SDValue Mask = DAG.getConstant(MaskVal, dl, MVT::i8);
4469 return DAG.getNode(X86ISD::BLENDI, dl, ResultVT, Result, Vec256, Mask);
4472 const X86Subtarget &Subtarget =
4473 static_cast<const X86Subtarget &>(DAG.getSubtarget());
4475 // AVX2 is needed for 256-bit integer blend support.
4476 // Integers must be cast to 32-bit because there is only vpblendd;
4477 // vpblendw can't be used for this because it has a handicapped mask.
4479 // If we don't have AVX2, then cast to float. Using a wrong domain blend
4480 // is still more efficient than using the wrong domain vinsertf128 that
4481 // will be created by InsertSubVector().
4482 MVT CastVT = Subtarget.hasAVX2() ? MVT::v8i32 : MVT::v8f32;
4484 SDValue Mask = DAG.getConstant(0x0f, dl, MVT::i8);
4485 Vec256 = DAG.getBitcast(CastVT, Vec256);
4486 Vec256 = DAG.getNode(X86ISD::BLENDI, dl, CastVT, Result, Vec256, Mask);
4487 return DAG.getBitcast(ResultVT, Vec256);
4490 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 128);
4493 static SDValue Insert256BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4494 SelectionDAG &DAG, SDLoc dl) {
4495 assert(Vec.getValueType().is256BitVector() && "Unexpected vector size!");
4496 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 256);
4499 /// Concat two 128-bit vectors into a 256 bit vector using VINSERTF128
4500 /// instructions. This is used because creating CONCAT_VECTOR nodes of
4501 /// BUILD_VECTORS returns a larger BUILD_VECTOR while we're trying to lower
4502 /// large BUILD_VECTORS.
4503 static SDValue Concat128BitVectors(SDValue V1, SDValue V2, EVT VT,
4504 unsigned NumElems, SelectionDAG &DAG,
4506 SDValue V = Insert128BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4507 return Insert128BitVector(V, V2, NumElems/2, DAG, dl);
4510 static SDValue Concat256BitVectors(SDValue V1, SDValue V2, EVT VT,
4511 unsigned NumElems, SelectionDAG &DAG,
4513 SDValue V = Insert256BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4514 return Insert256BitVector(V, V2, NumElems/2, DAG, dl);
4517 /// Returns a vector of specified type with all bits set.
4518 /// Always build ones vectors as <4 x i32> or <8 x i32>. For 256-bit types with
4519 /// no AVX2 supprt, use two <4 x i32> inserted in a <8 x i32> appropriately.
4520 /// Then bitcast to their original type, ensuring they get CSE'd.
4521 static SDValue getOnesVector(EVT VT, const X86Subtarget *Subtarget,
4522 SelectionDAG &DAG, SDLoc dl) {
4523 assert(VT.isVector() && "Expected a vector type");
4525 SDValue Cst = DAG.getConstant(~0U, dl, MVT::i32);
4527 if (VT.is512BitVector()) {
4528 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
4529 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4530 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
4531 } else if (VT.is256BitVector()) {
4532 if (Subtarget->hasInt256()) { // AVX2
4533 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4534 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4536 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4537 Vec = Concat128BitVectors(Vec, Vec, MVT::v8i32, 8, DAG, dl);
4539 } else if (VT.is128BitVector()) {
4540 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4542 llvm_unreachable("Unexpected vector type");
4544 return DAG.getBitcast(VT, Vec);
4547 /// Returns a vector_shuffle node for an unpackl operation.
4548 static SDValue getUnpackl(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4550 unsigned NumElems = VT.getVectorNumElements();
4551 SmallVector<int, 8> Mask;
4552 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
4554 Mask.push_back(i + NumElems);
4556 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4559 /// Returns a vector_shuffle node for an unpackh operation.
4560 static SDValue getUnpackh(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4562 unsigned NumElems = VT.getVectorNumElements();
4563 SmallVector<int, 8> Mask;
4564 for (unsigned i = 0, Half = NumElems/2; i != Half; ++i) {
4565 Mask.push_back(i + Half);
4566 Mask.push_back(i + NumElems + Half);
4568 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4571 /// Return a vector_shuffle of the specified vector of zero or undef vector.
4572 /// This produces a shuffle where the low element of V2 is swizzled into the
4573 /// zero/undef vector, landing at element Idx.
4574 /// This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
4575 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
4577 const X86Subtarget *Subtarget,
4578 SelectionDAG &DAG) {
4579 MVT VT = V2.getSimpleValueType();
4581 ? getZeroVector(VT, Subtarget, DAG, SDLoc(V2)) : DAG.getUNDEF(VT);
4582 unsigned NumElems = VT.getVectorNumElements();
4583 SmallVector<int, 16> MaskVec;
4584 for (unsigned i = 0; i != NumElems; ++i)
4585 // If this is the insertion idx, put the low elt of V2 here.
4586 MaskVec.push_back(i == Idx ? NumElems : i);
4587 return DAG.getVectorShuffle(VT, SDLoc(V2), V1, V2, &MaskVec[0]);
4590 /// Calculates the shuffle mask corresponding to the target-specific opcode.
4591 /// Returns true if the Mask could be calculated. Sets IsUnary to true if only
4592 /// uses one source. Note that this will set IsUnary for shuffles which use a
4593 /// single input multiple times, and in those cases it will
4594 /// adjust the mask to only have indices within that single input.
4595 /// FIXME: Add support for Decode*Mask functions that return SM_SentinelZero.
4596 static bool getTargetShuffleMask(SDNode *N, MVT VT,
4597 SmallVectorImpl<int> &Mask, bool &IsUnary) {
4598 unsigned NumElems = VT.getVectorNumElements();
4602 bool IsFakeUnary = false;
4603 switch(N->getOpcode()) {
4604 case X86ISD::BLENDI:
4605 ImmN = N->getOperand(N->getNumOperands()-1);
4606 DecodeBLENDMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4609 ImmN = N->getOperand(N->getNumOperands()-1);
4610 DecodeSHUFPMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4611 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4613 case X86ISD::UNPCKH:
4614 DecodeUNPCKHMask(VT, Mask);
4615 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4617 case X86ISD::UNPCKL:
4618 DecodeUNPCKLMask(VT, Mask);
4619 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4621 case X86ISD::MOVHLPS:
4622 DecodeMOVHLPSMask(NumElems, Mask);
4623 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4625 case X86ISD::MOVLHPS:
4626 DecodeMOVLHPSMask(NumElems, Mask);
4627 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4629 case X86ISD::PALIGNR:
4630 ImmN = N->getOperand(N->getNumOperands()-1);
4631 DecodePALIGNRMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4633 case X86ISD::PSHUFD:
4634 case X86ISD::VPERMILPI:
4635 ImmN = N->getOperand(N->getNumOperands()-1);
4636 DecodePSHUFMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4639 case X86ISD::PSHUFHW:
4640 ImmN = N->getOperand(N->getNumOperands()-1);
4641 DecodePSHUFHWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4644 case X86ISD::PSHUFLW:
4645 ImmN = N->getOperand(N->getNumOperands()-1);
4646 DecodePSHUFLWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4649 case X86ISD::PSHUFB: {
4651 SDValue MaskNode = N->getOperand(1);
4652 while (MaskNode->getOpcode() == ISD::BITCAST)
4653 MaskNode = MaskNode->getOperand(0);
4655 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4656 // If we have a build-vector, then things are easy.
4657 MVT VT = MaskNode.getSimpleValueType();
4658 assert(VT.isVector() &&
4659 "Can't produce a non-vector with a build_vector!");
4660 if (!VT.isInteger())
4663 int NumBytesPerElement = VT.getVectorElementType().getSizeInBits() / 8;
4665 SmallVector<uint64_t, 32> RawMask;
4666 for (int i = 0, e = MaskNode->getNumOperands(); i < e; ++i) {
4667 SDValue Op = MaskNode->getOperand(i);
4668 if (Op->getOpcode() == ISD::UNDEF) {
4669 RawMask.push_back((uint64_t)SM_SentinelUndef);
4672 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
4675 APInt MaskElement = CN->getAPIntValue();
4677 // We now have to decode the element which could be any integer size and
4678 // extract each byte of it.
4679 for (int j = 0; j < NumBytesPerElement; ++j) {
4680 // Note that this is x86 and so always little endian: the low byte is
4681 // the first byte of the mask.
4682 RawMask.push_back(MaskElement.getLoBits(8).getZExtValue());
4683 MaskElement = MaskElement.lshr(8);
4686 DecodePSHUFBMask(RawMask, Mask);
4690 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4694 SDValue Ptr = MaskLoad->getBasePtr();
4695 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4696 Ptr->getOpcode() == X86ISD::WrapperRIP)
4697 Ptr = Ptr->getOperand(0);
4699 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4700 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4703 if (auto *C = dyn_cast<Constant>(MaskCP->getConstVal())) {
4704 DecodePSHUFBMask(C, Mask);
4712 case X86ISD::VPERMI:
4713 ImmN = N->getOperand(N->getNumOperands()-1);
4714 DecodeVPERMMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4719 DecodeScalarMoveMask(VT, /* IsLoad */ false, Mask);
4721 case X86ISD::VPERM2X128:
4722 ImmN = N->getOperand(N->getNumOperands()-1);
4723 DecodeVPERM2X128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4724 if (Mask.empty()) return false;
4725 // Mask only contains negative index if an element is zero.
4726 if (std::any_of(Mask.begin(), Mask.end(),
4727 [](int M){ return M == SM_SentinelZero; }))
4730 case X86ISD::MOVSLDUP:
4731 DecodeMOVSLDUPMask(VT, Mask);
4734 case X86ISD::MOVSHDUP:
4735 DecodeMOVSHDUPMask(VT, Mask);
4738 case X86ISD::MOVDDUP:
4739 DecodeMOVDDUPMask(VT, Mask);
4742 case X86ISD::MOVLHPD:
4743 case X86ISD::MOVLPD:
4744 case X86ISD::MOVLPS:
4745 // Not yet implemented
4747 case X86ISD::VPERMV: {
4749 SDValue MaskNode = N->getOperand(0);
4750 while (MaskNode->getOpcode() == ISD::BITCAST)
4751 MaskNode = MaskNode->getOperand(0);
4753 unsigned MaskLoBits = Log2_64(VT.getVectorNumElements());
4754 SmallVector<uint64_t, 32> RawMask;
4755 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4756 // If we have a build-vector, then things are easy.
4757 assert(MaskNode.getSimpleValueType().isInteger() &&
4758 MaskNode.getSimpleValueType().getVectorNumElements() ==
4759 VT.getVectorNumElements());
4761 for (unsigned i = 0; i < MaskNode->getNumOperands(); ++i) {
4762 SDValue Op = MaskNode->getOperand(i);
4763 if (Op->getOpcode() == ISD::UNDEF)
4764 RawMask.push_back((uint64_t)SM_SentinelUndef);
4765 else if (isa<ConstantSDNode>(Op)) {
4766 APInt MaskElement = cast<ConstantSDNode>(Op)->getAPIntValue();
4767 RawMask.push_back(MaskElement.getLoBits(MaskLoBits).getZExtValue());
4771 DecodeVPERMVMask(RawMask, Mask);
4774 if (MaskNode->getOpcode() == X86ISD::VBROADCAST) {
4775 unsigned NumEltsInMask = MaskNode->getNumOperands();
4776 MaskNode = MaskNode->getOperand(0);
4777 auto *CN = dyn_cast<ConstantSDNode>(MaskNode);
4779 APInt MaskEltValue = CN->getAPIntValue();
4780 for (unsigned i = 0; i < NumEltsInMask; ++i)
4781 RawMask.push_back(MaskEltValue.getLoBits(MaskLoBits).getZExtValue());
4782 DecodeVPERMVMask(RawMask, Mask);
4785 // It may be a scalar load
4788 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4792 SDValue Ptr = MaskLoad->getBasePtr();
4793 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4794 Ptr->getOpcode() == X86ISD::WrapperRIP)
4795 Ptr = Ptr->getOperand(0);
4797 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4798 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4801 auto *C = dyn_cast<Constant>(MaskCP->getConstVal());
4803 DecodeVPERMVMask(C, VT, Mask);
4810 case X86ISD::VPERMV3: {
4812 SDValue MaskNode = N->getOperand(1);
4813 while (MaskNode->getOpcode() == ISD::BITCAST)
4814 MaskNode = MaskNode->getOperand(1);
4816 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4817 // If we have a build-vector, then things are easy.
4818 assert(MaskNode.getSimpleValueType().isInteger() &&
4819 MaskNode.getSimpleValueType().getVectorNumElements() ==
4820 VT.getVectorNumElements());
4822 SmallVector<uint64_t, 32> RawMask;
4823 unsigned MaskLoBits = Log2_64(VT.getVectorNumElements()*2);
4825 for (unsigned i = 0; i < MaskNode->getNumOperands(); ++i) {
4826 SDValue Op = MaskNode->getOperand(i);
4827 if (Op->getOpcode() == ISD::UNDEF)
4828 RawMask.push_back((uint64_t)SM_SentinelUndef);
4830 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
4833 APInt MaskElement = CN->getAPIntValue();
4834 RawMask.push_back(MaskElement.getLoBits(MaskLoBits).getZExtValue());
4837 DecodeVPERMV3Mask(RawMask, Mask);
4841 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4845 SDValue Ptr = MaskLoad->getBasePtr();
4846 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4847 Ptr->getOpcode() == X86ISD::WrapperRIP)
4848 Ptr = Ptr->getOperand(0);
4850 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4851 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4854 auto *C = dyn_cast<Constant>(MaskCP->getConstVal());
4856 DecodeVPERMV3Mask(C, VT, Mask);
4863 default: llvm_unreachable("unknown target shuffle node");
4866 // If we have a fake unary shuffle, the shuffle mask is spread across two
4867 // inputs that are actually the same node. Re-map the mask to always point
4868 // into the first input.
4871 if (M >= (int)Mask.size())
4877 /// Returns the scalar element that will make up the ith
4878 /// element of the result of the vector shuffle.
4879 static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG,
4882 return SDValue(); // Limit search depth.
4884 SDValue V = SDValue(N, 0);
4885 EVT VT = V.getValueType();
4886 unsigned Opcode = V.getOpcode();
4888 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
4889 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
4890 int Elt = SV->getMaskElt(Index);
4893 return DAG.getUNDEF(VT.getVectorElementType());
4895 unsigned NumElems = VT.getVectorNumElements();
4896 SDValue NewV = (Elt < (int)NumElems) ? SV->getOperand(0)
4897 : SV->getOperand(1);
4898 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG, Depth+1);
4901 // Recurse into target specific vector shuffles to find scalars.
4902 if (isTargetShuffle(Opcode)) {
4903 MVT ShufVT = V.getSimpleValueType();
4904 unsigned NumElems = ShufVT.getVectorNumElements();
4905 SmallVector<int, 16> ShuffleMask;
4908 if (!getTargetShuffleMask(N, ShufVT, ShuffleMask, IsUnary))
4911 int Elt = ShuffleMask[Index];
4913 return DAG.getUNDEF(ShufVT.getVectorElementType());
4915 SDValue NewV = (Elt < (int)NumElems) ? N->getOperand(0)
4917 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG,
4921 // Actual nodes that may contain scalar elements
4922 if (Opcode == ISD::BITCAST) {
4923 V = V.getOperand(0);
4924 EVT SrcVT = V.getValueType();
4925 unsigned NumElems = VT.getVectorNumElements();
4927 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
4931 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
4932 return (Index == 0) ? V.getOperand(0)
4933 : DAG.getUNDEF(VT.getVectorElementType());
4935 if (V.getOpcode() == ISD::BUILD_VECTOR)
4936 return V.getOperand(Index);
4941 /// Custom lower build_vector of v16i8.
4942 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
4943 unsigned NumNonZero, unsigned NumZero,
4945 const X86Subtarget* Subtarget,
4946 const TargetLowering &TLI) {
4954 // SSE4.1 - use PINSRB to insert each byte directly.
4955 if (Subtarget->hasSSE41()) {
4956 for (unsigned i = 0; i < 16; ++i) {
4957 bool isNonZero = (NonZeros & (1 << i)) != 0;
4961 V = getZeroVector(MVT::v16i8, Subtarget, DAG, dl);
4963 V = DAG.getUNDEF(MVT::v16i8);
4966 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4967 MVT::v16i8, V, Op.getOperand(i),
4968 DAG.getIntPtrConstant(i, dl));
4975 // Pre-SSE4.1 - merge byte pairs and insert with PINSRW.
4976 for (unsigned i = 0; i < 16; ++i) {
4977 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
4978 if (ThisIsNonZero && First) {
4980 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
4982 V = DAG.getUNDEF(MVT::v8i16);
4987 SDValue ThisElt, LastElt;
4988 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
4989 if (LastIsNonZero) {
4990 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
4991 MVT::i16, Op.getOperand(i-1));
4993 if (ThisIsNonZero) {
4994 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
4995 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
4996 ThisElt, DAG.getConstant(8, dl, MVT::i8));
4998 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
5002 if (ThisElt.getNode())
5003 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
5004 DAG.getIntPtrConstant(i/2, dl));
5008 return DAG.getBitcast(MVT::v16i8, V);
5011 /// Custom lower build_vector of v8i16.
5012 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
5013 unsigned NumNonZero, unsigned NumZero,
5015 const X86Subtarget* Subtarget,
5016 const TargetLowering &TLI) {
5023 for (unsigned i = 0; i < 8; ++i) {
5024 bool isNonZero = (NonZeros & (1 << i)) != 0;
5028 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
5030 V = DAG.getUNDEF(MVT::v8i16);
5033 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
5034 MVT::v8i16, V, Op.getOperand(i),
5035 DAG.getIntPtrConstant(i, dl));
5042 /// Custom lower build_vector of v4i32 or v4f32.
5043 static SDValue LowerBuildVectorv4x32(SDValue Op, SelectionDAG &DAG,
5044 const X86Subtarget *Subtarget,
5045 const TargetLowering &TLI) {
5046 // Find all zeroable elements.
5047 std::bitset<4> Zeroable;
5048 for (int i=0; i < 4; ++i) {
5049 SDValue Elt = Op->getOperand(i);
5050 Zeroable[i] = (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt));
5052 assert(Zeroable.size() - Zeroable.count() > 1 &&
5053 "We expect at least two non-zero elements!");
5055 // We only know how to deal with build_vector nodes where elements are either
5056 // zeroable or extract_vector_elt with constant index.
5057 SDValue FirstNonZero;
5058 unsigned FirstNonZeroIdx;
5059 for (unsigned i=0; i < 4; ++i) {
5062 SDValue Elt = Op->getOperand(i);
5063 if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5064 !isa<ConstantSDNode>(Elt.getOperand(1)))
5066 // Make sure that this node is extracting from a 128-bit vector.
5067 MVT VT = Elt.getOperand(0).getSimpleValueType();
5068 if (!VT.is128BitVector())
5070 if (!FirstNonZero.getNode()) {
5072 FirstNonZeroIdx = i;
5076 assert(FirstNonZero.getNode() && "Unexpected build vector of all zeros!");
5077 SDValue V1 = FirstNonZero.getOperand(0);
5078 MVT VT = V1.getSimpleValueType();
5080 // See if this build_vector can be lowered as a blend with zero.
5082 unsigned EltMaskIdx, EltIdx;
5084 for (EltIdx = 0; EltIdx < 4; ++EltIdx) {
5085 if (Zeroable[EltIdx]) {
5086 // The zero vector will be on the right hand side.
5087 Mask[EltIdx] = EltIdx+4;
5091 Elt = Op->getOperand(EltIdx);
5092 // By construction, Elt is a EXTRACT_VECTOR_ELT with constant index.
5093 EltMaskIdx = cast<ConstantSDNode>(Elt.getOperand(1))->getZExtValue();
5094 if (Elt.getOperand(0) != V1 || EltMaskIdx != EltIdx)
5096 Mask[EltIdx] = EltIdx;
5100 // Let the shuffle legalizer deal with blend operations.
5101 SDValue VZero = getZeroVector(VT, Subtarget, DAG, SDLoc(Op));
5102 if (V1.getSimpleValueType() != VT)
5103 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), VT, V1);
5104 return DAG.getVectorShuffle(VT, SDLoc(V1), V1, VZero, &Mask[0]);
5107 // See if we can lower this build_vector to a INSERTPS.
5108 if (!Subtarget->hasSSE41())
5111 SDValue V2 = Elt.getOperand(0);
5112 if (Elt == FirstNonZero && EltIdx == FirstNonZeroIdx)
5115 bool CanFold = true;
5116 for (unsigned i = EltIdx + 1; i < 4 && CanFold; ++i) {
5120 SDValue Current = Op->getOperand(i);
5121 SDValue SrcVector = Current->getOperand(0);
5124 CanFold = SrcVector == V1 &&
5125 cast<ConstantSDNode>(Current.getOperand(1))->getZExtValue() == i;
5131 assert(V1.getNode() && "Expected at least two non-zero elements!");
5132 if (V1.getSimpleValueType() != MVT::v4f32)
5133 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), MVT::v4f32, V1);
5134 if (V2.getSimpleValueType() != MVT::v4f32)
5135 V2 = DAG.getNode(ISD::BITCAST, SDLoc(V2), MVT::v4f32, V2);
5137 // Ok, we can emit an INSERTPS instruction.
5138 unsigned ZMask = Zeroable.to_ulong();
5140 unsigned InsertPSMask = EltMaskIdx << 6 | EltIdx << 4 | ZMask;
5141 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
5143 SDValue Result = DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
5144 DAG.getIntPtrConstant(InsertPSMask, DL));
5145 return DAG.getBitcast(VT, Result);
5148 /// Return a vector logical shift node.
5149 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
5150 unsigned NumBits, SelectionDAG &DAG,
5151 const TargetLowering &TLI, SDLoc dl) {
5152 assert(VT.is128BitVector() && "Unknown type for VShift");
5153 MVT ShVT = MVT::v2i64;
5154 unsigned Opc = isLeft ? X86ISD::VSHLDQ : X86ISD::VSRLDQ;
5155 SrcOp = DAG.getBitcast(ShVT, SrcOp);
5156 MVT ScalarShiftTy = TLI.getScalarShiftAmountTy(DAG.getDataLayout(), VT);
5157 assert(NumBits % 8 == 0 && "Only support byte sized shifts");
5158 SDValue ShiftVal = DAG.getConstant(NumBits/8, dl, ScalarShiftTy);
5159 return DAG.getBitcast(VT, DAG.getNode(Opc, dl, ShVT, SrcOp, ShiftVal));
5163 LowerAsSplatVectorLoad(SDValue SrcOp, MVT VT, SDLoc dl, SelectionDAG &DAG) {
5165 // Check if the scalar load can be widened into a vector load. And if
5166 // the address is "base + cst" see if the cst can be "absorbed" into
5167 // the shuffle mask.
5168 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
5169 SDValue Ptr = LD->getBasePtr();
5170 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
5172 EVT PVT = LD->getValueType(0);
5173 if (PVT != MVT::i32 && PVT != MVT::f32)
5178 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
5179 FI = FINode->getIndex();
5181 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
5182 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
5183 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
5184 Offset = Ptr.getConstantOperandVal(1);
5185 Ptr = Ptr.getOperand(0);
5190 // FIXME: 256-bit vector instructions don't require a strict alignment,
5191 // improve this code to support it better.
5192 unsigned RequiredAlign = VT.getSizeInBits()/8;
5193 SDValue Chain = LD->getChain();
5194 // Make sure the stack object alignment is at least 16 or 32.
5195 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5196 if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
5197 if (MFI->isFixedObjectIndex(FI)) {
5198 // Can't change the alignment. FIXME: It's possible to compute
5199 // the exact stack offset and reference FI + adjust offset instead.
5200 // If someone *really* cares about this. That's the way to implement it.
5203 MFI->setObjectAlignment(FI, RequiredAlign);
5207 // (Offset % 16 or 32) must be multiple of 4. Then address is then
5208 // Ptr + (Offset & ~15).
5211 if ((Offset % RequiredAlign) & 3)
5213 int64_t StartOffset = Offset & ~int64_t(RequiredAlign - 1);
5216 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
5217 DAG.getConstant(StartOffset, DL, Ptr.getValueType()));
5220 int EltNo = (Offset - StartOffset) >> 2;
5221 unsigned NumElems = VT.getVectorNumElements();
5223 EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
5224 SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
5225 LD->getPointerInfo().getWithOffset(StartOffset),
5226 false, false, false, 0);
5228 SmallVector<int, 8> Mask(NumElems, EltNo);
5230 return DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), &Mask[0]);
5236 /// Given the initializing elements 'Elts' of a vector of type 'VT', see if the
5237 /// elements can be replaced by a single large load which has the same value as
5238 /// a build_vector or insert_subvector whose loaded operands are 'Elts'.
5240 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
5242 /// FIXME: we'd also like to handle the case where the last elements are zero
5243 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
5244 /// There's even a handy isZeroNode for that purpose.
5245 static SDValue EltsFromConsecutiveLoads(EVT VT, ArrayRef<SDValue> Elts,
5246 SDLoc &DL, SelectionDAG &DAG,
5247 bool isAfterLegalize) {
5248 unsigned NumElems = Elts.size();
5250 LoadSDNode *LDBase = nullptr;
5251 unsigned LastLoadedElt = -1U;
5253 // For each element in the initializer, see if we've found a load or an undef.
5254 // If we don't find an initial load element, or later load elements are
5255 // non-consecutive, bail out.
5256 for (unsigned i = 0; i < NumElems; ++i) {
5257 SDValue Elt = Elts[i];
5258 // Look through a bitcast.
5259 if (Elt.getNode() && Elt.getOpcode() == ISD::BITCAST)
5260 Elt = Elt.getOperand(0);
5261 if (!Elt.getNode() ||
5262 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
5265 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
5267 LDBase = cast<LoadSDNode>(Elt.getNode());
5271 if (Elt.getOpcode() == ISD::UNDEF)
5274 LoadSDNode *LD = cast<LoadSDNode>(Elt);
5275 EVT LdVT = Elt.getValueType();
5276 // Each loaded element must be the correct fractional portion of the
5277 // requested vector load.
5278 if (LdVT.getSizeInBits() != VT.getSizeInBits() / NumElems)
5280 if (!DAG.isConsecutiveLoad(LD, LDBase, LdVT.getSizeInBits() / 8, i))
5285 // If we have found an entire vector of loads and undefs, then return a large
5286 // load of the entire vector width starting at the base pointer. If we found
5287 // consecutive loads for the low half, generate a vzext_load node.
5288 if (LastLoadedElt == NumElems - 1) {
5289 assert(LDBase && "Did not find base load for merging consecutive loads");
5290 EVT EltVT = LDBase->getValueType(0);
5291 // Ensure that the input vector size for the merged loads matches the
5292 // cumulative size of the input elements.
5293 if (VT.getSizeInBits() != EltVT.getSizeInBits() * NumElems)
5296 if (isAfterLegalize &&
5297 !DAG.getTargetLoweringInfo().isOperationLegal(ISD::LOAD, VT))
5300 SDValue NewLd = SDValue();
5302 NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
5303 LDBase->getPointerInfo(), LDBase->isVolatile(),
5304 LDBase->isNonTemporal(), LDBase->isInvariant(),
5305 LDBase->getAlignment());
5307 if (LDBase->hasAnyUseOfValue(1)) {
5308 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5310 SDValue(NewLd.getNode(), 1));
5311 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5312 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5313 SDValue(NewLd.getNode(), 1));
5319 //TODO: The code below fires only for for loading the low v2i32 / v2f32
5320 //of a v4i32 / v4f32. It's probably worth generalizing.
5321 EVT EltVT = VT.getVectorElementType();
5322 if (NumElems == 4 && LastLoadedElt == 1 && (EltVT.getSizeInBits() == 32) &&
5323 DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
5324 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
5325 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
5327 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys, Ops, MVT::i64,
5328 LDBase->getPointerInfo(),
5329 LDBase->getAlignment(),
5330 false/*isVolatile*/, true/*ReadMem*/,
5333 // Make sure the newly-created LOAD is in the same position as LDBase in
5334 // terms of dependency. We create a TokenFactor for LDBase and ResNode, and
5335 // update uses of LDBase's output chain to use the TokenFactor.
5336 if (LDBase->hasAnyUseOfValue(1)) {
5337 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5338 SDValue(LDBase, 1), SDValue(ResNode.getNode(), 1));
5339 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5340 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5341 SDValue(ResNode.getNode(), 1));
5344 return DAG.getBitcast(VT, ResNode);
5349 /// LowerVectorBroadcast - Attempt to use the vbroadcast instruction
5350 /// to generate a splat value for the following cases:
5351 /// 1. A splat BUILD_VECTOR which uses a single scalar load, or a constant.
5352 /// 2. A splat shuffle which uses a scalar_to_vector node which comes from
5353 /// a scalar load, or a constant.
5354 /// The VBROADCAST node is returned when a pattern is found,
5355 /// or SDValue() otherwise.
5356 static SDValue LowerVectorBroadcast(SDValue Op, const X86Subtarget* Subtarget,
5357 SelectionDAG &DAG) {
5358 // VBROADCAST requires AVX.
5359 // TODO: Splats could be generated for non-AVX CPUs using SSE
5360 // instructions, but there's less potential gain for only 128-bit vectors.
5361 if (!Subtarget->hasAVX())
5364 MVT VT = Op.getSimpleValueType();
5367 assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&
5368 "Unsupported vector type for broadcast.");
5373 switch (Op.getOpcode()) {
5375 // Unknown pattern found.
5378 case ISD::BUILD_VECTOR: {
5379 auto *BVOp = cast<BuildVectorSDNode>(Op.getNode());
5380 BitVector UndefElements;
5381 SDValue Splat = BVOp->getSplatValue(&UndefElements);
5383 // We need a splat of a single value to use broadcast, and it doesn't
5384 // make any sense if the value is only in one element of the vector.
5385 if (!Splat || (VT.getVectorNumElements() - UndefElements.count()) <= 1)
5389 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5390 Ld.getOpcode() == ISD::ConstantFP);
5392 // Make sure that all of the users of a non-constant load are from the
5393 // BUILD_VECTOR node.
5394 if (!ConstSplatVal && !BVOp->isOnlyUserOf(Ld.getNode()))
5399 case ISD::VECTOR_SHUFFLE: {
5400 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5402 // Shuffles must have a splat mask where the first element is
5404 if ((!SVOp->isSplat()) || SVOp->getMaskElt(0) != 0)
5407 SDValue Sc = Op.getOperand(0);
5408 if (Sc.getOpcode() != ISD::SCALAR_TO_VECTOR &&
5409 Sc.getOpcode() != ISD::BUILD_VECTOR) {
5411 if (!Subtarget->hasInt256())
5414 // Use the register form of the broadcast instruction available on AVX2.
5415 if (VT.getSizeInBits() >= 256)
5416 Sc = Extract128BitVector(Sc, 0, DAG, dl);
5417 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Sc);
5420 Ld = Sc.getOperand(0);
5421 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5422 Ld.getOpcode() == ISD::ConstantFP);
5424 // The scalar_to_vector node and the suspected
5425 // load node must have exactly one user.
5426 // Constants may have multiple users.
5428 // AVX-512 has register version of the broadcast
5429 bool hasRegVer = Subtarget->hasAVX512() && VT.is512BitVector() &&
5430 Ld.getValueType().getSizeInBits() >= 32;
5431 if (!ConstSplatVal && ((!Sc.hasOneUse() || !Ld.hasOneUse()) &&
5438 unsigned ScalarSize = Ld.getValueType().getSizeInBits();
5439 bool IsGE256 = (VT.getSizeInBits() >= 256);
5441 // When optimizing for size, generate up to 5 extra bytes for a broadcast
5442 // instruction to save 8 or more bytes of constant pool data.
5443 // TODO: If multiple splats are generated to load the same constant,
5444 // it may be detrimental to overall size. There needs to be a way to detect
5445 // that condition to know if this is truly a size win.
5446 bool OptForSize = DAG.getMachineFunction().getFunction()->optForSize();
5448 // Handle broadcasting a single constant scalar from the constant pool
5450 // On Sandybridge (no AVX2), it is still better to load a constant vector
5451 // from the constant pool and not to broadcast it from a scalar.
5452 // But override that restriction when optimizing for size.
5453 // TODO: Check if splatting is recommended for other AVX-capable CPUs.
5454 if (ConstSplatVal && (Subtarget->hasAVX2() || OptForSize)) {
5455 EVT CVT = Ld.getValueType();
5456 assert(!CVT.isVector() && "Must not broadcast a vector type");
5458 // Splat f32, i32, v4f64, v4i64 in all cases with AVX2.
5459 // For size optimization, also splat v2f64 and v2i64, and for size opt
5460 // with AVX2, also splat i8 and i16.
5461 // With pattern matching, the VBROADCAST node may become a VMOVDDUP.
5462 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5463 (OptForSize && (ScalarSize == 64 || Subtarget->hasAVX2()))) {
5464 const Constant *C = nullptr;
5465 if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld))
5466 C = CI->getConstantIntValue();
5467 else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld))
5468 C = CF->getConstantFPValue();
5470 assert(C && "Invalid constant type");
5472 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5474 DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout()));
5475 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
5477 CVT, dl, DAG.getEntryNode(), CP,
5478 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
5479 false, false, Alignment);
5481 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5485 bool IsLoad = ISD::isNormalLoad(Ld.getNode());
5487 // Handle AVX2 in-register broadcasts.
5488 if (!IsLoad && Subtarget->hasInt256() &&
5489 (ScalarSize == 32 || (IsGE256 && ScalarSize == 64)))
5490 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5492 // The scalar source must be a normal load.
5496 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5497 (Subtarget->hasVLX() && ScalarSize == 64))
5498 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5500 // The integer check is needed for the 64-bit into 128-bit so it doesn't match
5501 // double since there is no vbroadcastsd xmm
5502 if (Subtarget->hasInt256() && Ld.getValueType().isInteger()) {
5503 if (ScalarSize == 8 || ScalarSize == 16 || ScalarSize == 64)
5504 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5507 // Unsupported broadcast.
5511 /// \brief For an EXTRACT_VECTOR_ELT with a constant index return the real
5512 /// underlying vector and index.
5514 /// Modifies \p ExtractedFromVec to the real vector and returns the real
5516 static int getUnderlyingExtractedFromVec(SDValue &ExtractedFromVec,
5518 int Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue();
5519 if (!isa<ShuffleVectorSDNode>(ExtractedFromVec))
5522 // For 256-bit vectors, LowerEXTRACT_VECTOR_ELT_SSE4 may have already
5524 // (extract_vector_elt (v8f32 %vreg1), Constant<6>)
5526 // (extract_vector_elt (vector_shuffle<2,u,u,u>
5527 // (extract_subvector (v8f32 %vreg0), Constant<4>),
5530 // In this case the vector is the extract_subvector expression and the index
5531 // is 2, as specified by the shuffle.
5532 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(ExtractedFromVec);
5533 SDValue ShuffleVec = SVOp->getOperand(0);
5534 MVT ShuffleVecVT = ShuffleVec.getSimpleValueType();
5535 assert(ShuffleVecVT.getVectorElementType() ==
5536 ExtractedFromVec.getSimpleValueType().getVectorElementType());
5538 int ShuffleIdx = SVOp->getMaskElt(Idx);
5539 if (isUndefOrInRange(ShuffleIdx, 0, ShuffleVecVT.getVectorNumElements())) {
5540 ExtractedFromVec = ShuffleVec;
5546 static SDValue buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) {
5547 MVT VT = Op.getSimpleValueType();
5549 // Skip if insert_vec_elt is not supported.
5550 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5551 if (!TLI.isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT))
5555 unsigned NumElems = Op.getNumOperands();
5559 SmallVector<unsigned, 4> InsertIndices;
5560 SmallVector<int, 8> Mask(NumElems, -1);
5562 for (unsigned i = 0; i != NumElems; ++i) {
5563 unsigned Opc = Op.getOperand(i).getOpcode();
5565 if (Opc == ISD::UNDEF)
5568 if (Opc != ISD::EXTRACT_VECTOR_ELT) {
5569 // Quit if more than 1 elements need inserting.
5570 if (InsertIndices.size() > 1)
5573 InsertIndices.push_back(i);
5577 SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0);
5578 SDValue ExtIdx = Op.getOperand(i).getOperand(1);
5579 // Quit if non-constant index.
5580 if (!isa<ConstantSDNode>(ExtIdx))
5582 int Idx = getUnderlyingExtractedFromVec(ExtractedFromVec, ExtIdx);
5584 // Quit if extracted from vector of different type.
5585 if (ExtractedFromVec.getValueType() != VT)
5588 if (!VecIn1.getNode())
5589 VecIn1 = ExtractedFromVec;
5590 else if (VecIn1 != ExtractedFromVec) {
5591 if (!VecIn2.getNode())
5592 VecIn2 = ExtractedFromVec;
5593 else if (VecIn2 != ExtractedFromVec)
5594 // Quit if more than 2 vectors to shuffle
5598 if (ExtractedFromVec == VecIn1)
5600 else if (ExtractedFromVec == VecIn2)
5601 Mask[i] = Idx + NumElems;
5604 if (!VecIn1.getNode())
5607 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
5608 SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, &Mask[0]);
5609 for (unsigned i = 0, e = InsertIndices.size(); i != e; ++i) {
5610 unsigned Idx = InsertIndices[i];
5611 NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx),
5612 DAG.getIntPtrConstant(Idx, DL));
5618 static SDValue ConvertI1VectorToInteger(SDValue Op, SelectionDAG &DAG) {
5619 assert(ISD::isBuildVectorOfConstantSDNodes(Op.getNode()) &&
5620 Op.getScalarValueSizeInBits() == 1 &&
5621 "Can not convert non-constant vector");
5622 uint64_t Immediate = 0;
5623 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5624 SDValue In = Op.getOperand(idx);
5625 if (In.getOpcode() != ISD::UNDEF)
5626 Immediate |= cast<ConstantSDNode>(In)->getZExtValue() << idx;
5630 MVT::getIntegerVT(std::max((int)Op.getValueType().getSizeInBits(), 8));
5631 return DAG.getConstant(Immediate, dl, VT);
5633 // Lower BUILD_VECTOR operation for v8i1 and v16i1 types.
5635 X86TargetLowering::LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const {
5637 MVT VT = Op.getSimpleValueType();
5638 assert((VT.getVectorElementType() == MVT::i1) &&
5639 "Unexpected type in LowerBUILD_VECTORvXi1!");
5642 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5643 SDValue Cst = DAG.getTargetConstant(0, dl, MVT::i1);
5644 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5645 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5648 if (ISD::isBuildVectorAllOnes(Op.getNode())) {
5649 SDValue Cst = DAG.getTargetConstant(1, dl, MVT::i1);
5650 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5651 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5654 if (ISD::isBuildVectorOfConstantSDNodes(Op.getNode())) {
5655 SDValue Imm = ConvertI1VectorToInteger(Op, DAG);
5656 if (Imm.getValueSizeInBits() == VT.getSizeInBits())
5657 return DAG.getBitcast(VT, Imm);
5658 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, Imm);
5659 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec,
5660 DAG.getIntPtrConstant(0, dl));
5663 // Vector has one or more non-const elements
5664 uint64_t Immediate = 0;
5665 SmallVector<unsigned, 16> NonConstIdx;
5666 bool IsSplat = true;
5667 bool HasConstElts = false;
5669 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5670 SDValue In = Op.getOperand(idx);
5671 if (In.getOpcode() == ISD::UNDEF)
5673 if (!isa<ConstantSDNode>(In))
5674 NonConstIdx.push_back(idx);
5676 Immediate |= cast<ConstantSDNode>(In)->getZExtValue() << idx;
5677 HasConstElts = true;
5681 else if (In != Op.getOperand(SplatIdx))
5685 // for splat use " (select i1 splat_elt, all-ones, all-zeroes)"
5687 return DAG.getNode(ISD::SELECT, dl, VT, Op.getOperand(SplatIdx),
5688 DAG.getConstant(1, dl, VT),
5689 DAG.getConstant(0, dl, VT));
5691 // insert elements one by one
5695 MVT ImmVT = MVT::getIntegerVT(std::max((int)VT.getSizeInBits(), 8));
5696 Imm = DAG.getConstant(Immediate, dl, ImmVT);
5698 else if (HasConstElts)
5699 Imm = DAG.getConstant(0, dl, VT);
5701 Imm = DAG.getUNDEF(VT);
5702 if (Imm.getValueSizeInBits() == VT.getSizeInBits())
5703 DstVec = DAG.getBitcast(VT, Imm);
5705 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, Imm);
5706 DstVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec,
5707 DAG.getIntPtrConstant(0, dl));
5710 for (unsigned i = 0; i < NonConstIdx.size(); ++i) {
5711 unsigned InsertIdx = NonConstIdx[i];
5712 DstVec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
5713 Op.getOperand(InsertIdx),
5714 DAG.getIntPtrConstant(InsertIdx, dl));
5719 /// \brief Return true if \p N implements a horizontal binop and return the
5720 /// operands for the horizontal binop into V0 and V1.
5722 /// This is a helper function of LowerToHorizontalOp().
5723 /// This function checks that the build_vector \p N in input implements a
5724 /// horizontal operation. Parameter \p Opcode defines the kind of horizontal
5725 /// operation to match.
5726 /// For example, if \p Opcode is equal to ISD::ADD, then this function
5727 /// checks if \p N implements a horizontal arithmetic add; if instead \p Opcode
5728 /// is equal to ISD::SUB, then this function checks if this is a horizontal
5731 /// This function only analyzes elements of \p N whose indices are
5732 /// in range [BaseIdx, LastIdx).
5733 static bool isHorizontalBinOp(const BuildVectorSDNode *N, unsigned Opcode,
5735 unsigned BaseIdx, unsigned LastIdx,
5736 SDValue &V0, SDValue &V1) {
5737 EVT VT = N->getValueType(0);
5739 assert(BaseIdx * 2 <= LastIdx && "Invalid Indices in input!");
5740 assert(VT.isVector() && VT.getVectorNumElements() >= LastIdx &&
5741 "Invalid Vector in input!");
5743 bool IsCommutable = (Opcode == ISD::ADD || Opcode == ISD::FADD);
5744 bool CanFold = true;
5745 unsigned ExpectedVExtractIdx = BaseIdx;
5746 unsigned NumElts = LastIdx - BaseIdx;
5747 V0 = DAG.getUNDEF(VT);
5748 V1 = DAG.getUNDEF(VT);
5750 // Check if N implements a horizontal binop.
5751 for (unsigned i = 0, e = NumElts; i != e && CanFold; ++i) {
5752 SDValue Op = N->getOperand(i + BaseIdx);
5755 if (Op->getOpcode() == ISD::UNDEF) {
5756 // Update the expected vector extract index.
5757 if (i * 2 == NumElts)
5758 ExpectedVExtractIdx = BaseIdx;
5759 ExpectedVExtractIdx += 2;
5763 CanFold = Op->getOpcode() == Opcode && Op->hasOneUse();
5768 SDValue Op0 = Op.getOperand(0);
5769 SDValue Op1 = Op.getOperand(1);
5771 // Try to match the following pattern:
5772 // (BINOP (extract_vector_elt A, I), (extract_vector_elt A, I+1))
5773 CanFold = (Op0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5774 Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5775 Op0.getOperand(0) == Op1.getOperand(0) &&
5776 isa<ConstantSDNode>(Op0.getOperand(1)) &&
5777 isa<ConstantSDNode>(Op1.getOperand(1)));
5781 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5782 unsigned I1 = cast<ConstantSDNode>(Op1.getOperand(1))->getZExtValue();
5784 if (i * 2 < NumElts) {
5785 if (V0.getOpcode() == ISD::UNDEF) {
5786 V0 = Op0.getOperand(0);
5787 if (V0.getValueType() != VT)
5791 if (V1.getOpcode() == ISD::UNDEF) {
5792 V1 = Op0.getOperand(0);
5793 if (V1.getValueType() != VT)
5796 if (i * 2 == NumElts)
5797 ExpectedVExtractIdx = BaseIdx;
5800 SDValue Expected = (i * 2 < NumElts) ? V0 : V1;
5801 if (I0 == ExpectedVExtractIdx)
5802 CanFold = I1 == I0 + 1 && Op0.getOperand(0) == Expected;
5803 else if (IsCommutable && I1 == ExpectedVExtractIdx) {
5804 // Try to match the following dag sequence:
5805 // (BINOP (extract_vector_elt A, I+1), (extract_vector_elt A, I))
5806 CanFold = I0 == I1 + 1 && Op1.getOperand(0) == Expected;
5810 ExpectedVExtractIdx += 2;
5816 /// \brief Emit a sequence of two 128-bit horizontal add/sub followed by
5817 /// a concat_vector.
5819 /// This is a helper function of LowerToHorizontalOp().
5820 /// This function expects two 256-bit vectors called V0 and V1.
5821 /// At first, each vector is split into two separate 128-bit vectors.
5822 /// Then, the resulting 128-bit vectors are used to implement two
5823 /// horizontal binary operations.
5825 /// The kind of horizontal binary operation is defined by \p X86Opcode.
5827 /// \p Mode specifies how the 128-bit parts of V0 and V1 are passed in input to
5828 /// the two new horizontal binop.
5829 /// When Mode is set, the first horizontal binop dag node would take as input
5830 /// the lower 128-bit of V0 and the upper 128-bit of V0. The second
5831 /// horizontal binop dag node would take as input the lower 128-bit of V1
5832 /// and the upper 128-bit of V1.
5834 /// HADD V0_LO, V0_HI
5835 /// HADD V1_LO, V1_HI
5837 /// Otherwise, the first horizontal binop dag node takes as input the lower
5838 /// 128-bit of V0 and the lower 128-bit of V1, and the second horizontal binop
5839 /// dag node takes the upper 128-bit of V0 and the upper 128-bit of V1.
5841 /// HADD V0_LO, V1_LO
5842 /// HADD V0_HI, V1_HI
5844 /// If \p isUndefLO is set, then the algorithm propagates UNDEF to the lower
5845 /// 128-bits of the result. If \p isUndefHI is set, then UNDEF is propagated to
5846 /// the upper 128-bits of the result.
5847 static SDValue ExpandHorizontalBinOp(const SDValue &V0, const SDValue &V1,
5848 SDLoc DL, SelectionDAG &DAG,
5849 unsigned X86Opcode, bool Mode,
5850 bool isUndefLO, bool isUndefHI) {
5851 EVT VT = V0.getValueType();
5852 assert(VT.is256BitVector() && VT == V1.getValueType() &&
5853 "Invalid nodes in input!");
5855 unsigned NumElts = VT.getVectorNumElements();
5856 SDValue V0_LO = Extract128BitVector(V0, 0, DAG, DL);
5857 SDValue V0_HI = Extract128BitVector(V0, NumElts/2, DAG, DL);
5858 SDValue V1_LO = Extract128BitVector(V1, 0, DAG, DL);
5859 SDValue V1_HI = Extract128BitVector(V1, NumElts/2, DAG, DL);
5860 EVT NewVT = V0_LO.getValueType();
5862 SDValue LO = DAG.getUNDEF(NewVT);
5863 SDValue HI = DAG.getUNDEF(NewVT);
5866 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5867 if (!isUndefLO && V0->getOpcode() != ISD::UNDEF)
5868 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V0_HI);
5869 if (!isUndefHI && V1->getOpcode() != ISD::UNDEF)
5870 HI = DAG.getNode(X86Opcode, DL, NewVT, V1_LO, V1_HI);
5872 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5873 if (!isUndefLO && (V0_LO->getOpcode() != ISD::UNDEF ||
5874 V1_LO->getOpcode() != ISD::UNDEF))
5875 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V1_LO);
5877 if (!isUndefHI && (V0_HI->getOpcode() != ISD::UNDEF ||
5878 V1_HI->getOpcode() != ISD::UNDEF))
5879 HI = DAG.getNode(X86Opcode, DL, NewVT, V0_HI, V1_HI);
5882 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LO, HI);
5885 /// Try to fold a build_vector that performs an 'addsub' to an X86ISD::ADDSUB
5887 static SDValue LowerToAddSub(const BuildVectorSDNode *BV,
5888 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
5889 MVT VT = BV->getSimpleValueType(0);
5890 if ((!Subtarget->hasSSE3() || (VT != MVT::v4f32 && VT != MVT::v2f64)) &&
5891 (!Subtarget->hasAVX() || (VT != MVT::v8f32 && VT != MVT::v4f64)))
5895 unsigned NumElts = VT.getVectorNumElements();
5896 SDValue InVec0 = DAG.getUNDEF(VT);
5897 SDValue InVec1 = DAG.getUNDEF(VT);
5899 assert((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v4f32 ||
5900 VT == MVT::v2f64) && "build_vector with an invalid type found!");
5902 // Odd-numbered elements in the input build vector are obtained from
5903 // adding two integer/float elements.
5904 // Even-numbered elements in the input build vector are obtained from
5905 // subtracting two integer/float elements.
5906 unsigned ExpectedOpcode = ISD::FSUB;
5907 unsigned NextExpectedOpcode = ISD::FADD;
5908 bool AddFound = false;
5909 bool SubFound = false;
5911 for (unsigned i = 0, e = NumElts; i != e; ++i) {
5912 SDValue Op = BV->getOperand(i);
5914 // Skip 'undef' values.
5915 unsigned Opcode = Op.getOpcode();
5916 if (Opcode == ISD::UNDEF) {
5917 std::swap(ExpectedOpcode, NextExpectedOpcode);
5921 // Early exit if we found an unexpected opcode.
5922 if (Opcode != ExpectedOpcode)
5925 SDValue Op0 = Op.getOperand(0);
5926 SDValue Op1 = Op.getOperand(1);
5928 // Try to match the following pattern:
5929 // (BINOP (extract_vector_elt A, i), (extract_vector_elt B, i))
5930 // Early exit if we cannot match that sequence.
5931 if (Op0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5932 Op1.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5933 !isa<ConstantSDNode>(Op0.getOperand(1)) ||
5934 !isa<ConstantSDNode>(Op1.getOperand(1)) ||
5935 Op0.getOperand(1) != Op1.getOperand(1))
5938 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5942 // We found a valid add/sub node. Update the information accordingly.
5948 // Update InVec0 and InVec1.
5949 if (InVec0.getOpcode() == ISD::UNDEF) {
5950 InVec0 = Op0.getOperand(0);
5951 if (InVec0.getSimpleValueType() != VT)
5954 if (InVec1.getOpcode() == ISD::UNDEF) {
5955 InVec1 = Op1.getOperand(0);
5956 if (InVec1.getSimpleValueType() != VT)
5960 // Make sure that operands in input to each add/sub node always
5961 // come from a same pair of vectors.
5962 if (InVec0 != Op0.getOperand(0)) {
5963 if (ExpectedOpcode == ISD::FSUB)
5966 // FADD is commutable. Try to commute the operands
5967 // and then test again.
5968 std::swap(Op0, Op1);
5969 if (InVec0 != Op0.getOperand(0))
5973 if (InVec1 != Op1.getOperand(0))
5976 // Update the pair of expected opcodes.
5977 std::swap(ExpectedOpcode, NextExpectedOpcode);
5980 // Don't try to fold this build_vector into an ADDSUB if the inputs are undef.
5981 if (AddFound && SubFound && InVec0.getOpcode() != ISD::UNDEF &&
5982 InVec1.getOpcode() != ISD::UNDEF)
5983 return DAG.getNode(X86ISD::ADDSUB, DL, VT, InVec0, InVec1);
5988 /// Lower BUILD_VECTOR to a horizontal add/sub operation if possible.
5989 static SDValue LowerToHorizontalOp(const BuildVectorSDNode *BV,
5990 const X86Subtarget *Subtarget,
5991 SelectionDAG &DAG) {
5992 MVT VT = BV->getSimpleValueType(0);
5993 unsigned NumElts = VT.getVectorNumElements();
5994 unsigned NumUndefsLO = 0;
5995 unsigned NumUndefsHI = 0;
5996 unsigned Half = NumElts/2;
5998 // Count the number of UNDEF operands in the build_vector in input.
5999 for (unsigned i = 0, e = Half; i != e; ++i)
6000 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
6003 for (unsigned i = Half, e = NumElts; i != e; ++i)
6004 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
6007 // Early exit if this is either a build_vector of all UNDEFs or all the
6008 // operands but one are UNDEF.
6009 if (NumUndefsLO + NumUndefsHI + 1 >= NumElts)
6013 SDValue InVec0, InVec1;
6014 if ((VT == MVT::v4f32 || VT == MVT::v2f64) && Subtarget->hasSSE3()) {
6015 // Try to match an SSE3 float HADD/HSUB.
6016 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
6017 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
6019 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
6020 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
6021 } else if ((VT == MVT::v4i32 || VT == MVT::v8i16) && Subtarget->hasSSSE3()) {
6022 // Try to match an SSSE3 integer HADD/HSUB.
6023 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
6024 return DAG.getNode(X86ISD::HADD, DL, VT, InVec0, InVec1);
6026 if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
6027 return DAG.getNode(X86ISD::HSUB, DL, VT, InVec0, InVec1);
6030 if (!Subtarget->hasAVX())
6033 if ((VT == MVT::v8f32 || VT == MVT::v4f64)) {
6034 // Try to match an AVX horizontal add/sub of packed single/double
6035 // precision floating point values from 256-bit vectors.
6036 SDValue InVec2, InVec3;
6037 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, Half, InVec0, InVec1) &&
6038 isHorizontalBinOp(BV, ISD::FADD, DAG, Half, NumElts, InVec2, InVec3) &&
6039 ((InVec0.getOpcode() == ISD::UNDEF ||
6040 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6041 ((InVec1.getOpcode() == ISD::UNDEF ||
6042 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6043 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
6045 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, Half, InVec0, InVec1) &&
6046 isHorizontalBinOp(BV, ISD::FSUB, DAG, Half, NumElts, InVec2, InVec3) &&
6047 ((InVec0.getOpcode() == ISD::UNDEF ||
6048 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6049 ((InVec1.getOpcode() == ISD::UNDEF ||
6050 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6051 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
6052 } else if (VT == MVT::v8i32 || VT == MVT::v16i16) {
6053 // Try to match an AVX2 horizontal add/sub of signed integers.
6054 SDValue InVec2, InVec3;
6056 bool CanFold = true;
6058 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, Half, InVec0, InVec1) &&
6059 isHorizontalBinOp(BV, ISD::ADD, DAG, Half, NumElts, InVec2, InVec3) &&
6060 ((InVec0.getOpcode() == ISD::UNDEF ||
6061 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6062 ((InVec1.getOpcode() == ISD::UNDEF ||
6063 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6064 X86Opcode = X86ISD::HADD;
6065 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, Half, InVec0, InVec1) &&
6066 isHorizontalBinOp(BV, ISD::SUB, DAG, Half, NumElts, InVec2, InVec3) &&
6067 ((InVec0.getOpcode() == ISD::UNDEF ||
6068 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6069 ((InVec1.getOpcode() == ISD::UNDEF ||
6070 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6071 X86Opcode = X86ISD::HSUB;
6076 // Fold this build_vector into a single horizontal add/sub.
6077 // Do this only if the target has AVX2.
6078 if (Subtarget->hasAVX2())
6079 return DAG.getNode(X86Opcode, DL, VT, InVec0, InVec1);
6081 // Do not try to expand this build_vector into a pair of horizontal
6082 // add/sub if we can emit a pair of scalar add/sub.
6083 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
6086 // Convert this build_vector into a pair of horizontal binop followed by
6088 bool isUndefLO = NumUndefsLO == Half;
6089 bool isUndefHI = NumUndefsHI == Half;
6090 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, false,
6091 isUndefLO, isUndefHI);
6095 if ((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v8i32 ||
6096 VT == MVT::v16i16) && Subtarget->hasAVX()) {
6098 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
6099 X86Opcode = X86ISD::HADD;
6100 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
6101 X86Opcode = X86ISD::HSUB;
6102 else if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
6103 X86Opcode = X86ISD::FHADD;
6104 else if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
6105 X86Opcode = X86ISD::FHSUB;
6109 // Don't try to expand this build_vector into a pair of horizontal add/sub
6110 // if we can simply emit a pair of scalar add/sub.
6111 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
6114 // Convert this build_vector into two horizontal add/sub followed by
6116 bool isUndefLO = NumUndefsLO == Half;
6117 bool isUndefHI = NumUndefsHI == Half;
6118 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, true,
6119 isUndefLO, isUndefHI);
6126 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
6129 MVT VT = Op.getSimpleValueType();
6130 MVT ExtVT = VT.getVectorElementType();
6131 unsigned NumElems = Op.getNumOperands();
6133 // Generate vectors for predicate vectors.
6134 if (VT.getVectorElementType() == MVT::i1 && Subtarget->hasAVX512())
6135 return LowerBUILD_VECTORvXi1(Op, DAG);
6137 // Vectors containing all zeros can be matched by pxor and xorps later
6138 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
6139 // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
6140 // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
6141 if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32)
6144 return getZeroVector(VT, Subtarget, DAG, dl);
6147 // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
6148 // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use
6149 // vpcmpeqd on 256-bit vectors.
6150 if (Subtarget->hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) {
6151 if (VT == MVT::v4i32 || (VT == MVT::v8i32 && Subtarget->hasInt256()))
6154 if (!VT.is512BitVector())
6155 return getOnesVector(VT, Subtarget, DAG, dl);
6158 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(Op.getNode());
6159 if (SDValue AddSub = LowerToAddSub(BV, Subtarget, DAG))
6161 if (SDValue HorizontalOp = LowerToHorizontalOp(BV, Subtarget, DAG))
6162 return HorizontalOp;
6163 if (SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG))
6166 unsigned EVTBits = ExtVT.getSizeInBits();
6168 unsigned NumZero = 0;
6169 unsigned NumNonZero = 0;
6170 unsigned NonZeros = 0;
6171 bool IsAllConstants = true;
6172 SmallSet<SDValue, 8> Values;
6173 for (unsigned i = 0; i < NumElems; ++i) {
6174 SDValue Elt = Op.getOperand(i);
6175 if (Elt.getOpcode() == ISD::UNDEF)
6178 if (Elt.getOpcode() != ISD::Constant &&
6179 Elt.getOpcode() != ISD::ConstantFP)
6180 IsAllConstants = false;
6181 if (X86::isZeroNode(Elt))
6184 NonZeros |= (1 << i);
6189 // All undef vector. Return an UNDEF. All zero vectors were handled above.
6190 if (NumNonZero == 0)
6191 return DAG.getUNDEF(VT);
6193 // Special case for single non-zero, non-undef, element.
6194 if (NumNonZero == 1) {
6195 unsigned Idx = countTrailingZeros(NonZeros);
6196 SDValue Item = Op.getOperand(Idx);
6198 // If this is an insertion of an i64 value on x86-32, and if the top bits of
6199 // the value are obviously zero, truncate the value to i32 and do the
6200 // insertion that way. Only do this if the value is non-constant or if the
6201 // value is a constant being inserted into element 0. It is cheaper to do
6202 // a constant pool load than it is to do a movd + shuffle.
6203 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
6204 (!IsAllConstants || Idx == 0)) {
6205 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
6207 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
6208 MVT VecVT = MVT::v4i32;
6210 // Truncate the value (which may itself be a constant) to i32, and
6211 // convert it to a vector with movd (S2V+shuffle to zero extend).
6212 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
6213 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
6214 return DAG.getBitcast(VT, getShuffleVectorZeroOrUndef(
6215 Item, Idx * 2, true, Subtarget, DAG));
6219 // If we have a constant or non-constant insertion into the low element of
6220 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
6221 // the rest of the elements. This will be matched as movd/movq/movss/movsd
6222 // depending on what the source datatype is.
6225 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6227 if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
6228 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
6229 if (VT.is512BitVector()) {
6230 SDValue ZeroVec = getZeroVector(VT, Subtarget, DAG, dl);
6231 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, ZeroVec,
6232 Item, DAG.getIntPtrConstant(0, dl));
6234 assert((VT.is128BitVector() || VT.is256BitVector()) &&
6235 "Expected an SSE value type!");
6236 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6237 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
6238 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6241 // We can't directly insert an i8 or i16 into a vector, so zero extend
6243 if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
6244 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
6245 if (VT.is256BitVector()) {
6246 if (Subtarget->hasAVX()) {
6247 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v8i32, Item);
6248 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6250 // Without AVX, we need to extend to a 128-bit vector and then
6251 // insert into the 256-bit vector.
6252 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
6253 SDValue ZeroVec = getZeroVector(MVT::v8i32, Subtarget, DAG, dl);
6254 Item = Insert128BitVector(ZeroVec, Item, 0, DAG, dl);
6257 assert(VT.is128BitVector() && "Expected an SSE value type!");
6258 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
6259 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6261 return DAG.getBitcast(VT, Item);
6265 // Is it a vector logical left shift?
6266 if (NumElems == 2 && Idx == 1 &&
6267 X86::isZeroNode(Op.getOperand(0)) &&
6268 !X86::isZeroNode(Op.getOperand(1))) {
6269 unsigned NumBits = VT.getSizeInBits();
6270 return getVShift(true, VT,
6271 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6272 VT, Op.getOperand(1)),
6273 NumBits/2, DAG, *this, dl);
6276 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
6279 // Otherwise, if this is a vector with i32 or f32 elements, and the element
6280 // is a non-constant being inserted into an element other than the low one,
6281 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
6282 // movd/movss) to move this into the low element, then shuffle it into
6284 if (EVTBits == 32) {
6285 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6286 return getShuffleVectorZeroOrUndef(Item, Idx, NumZero > 0, Subtarget, DAG);
6290 // Splat is obviously ok. Let legalizer expand it to a shuffle.
6291 if (Values.size() == 1) {
6292 if (EVTBits == 32) {
6293 // Instead of a shuffle like this:
6294 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
6295 // Check if it's possible to issue this instead.
6296 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
6297 unsigned Idx = countTrailingZeros(NonZeros);
6298 SDValue Item = Op.getOperand(Idx);
6299 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
6300 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
6305 // A vector full of immediates; various special cases are already
6306 // handled, so this is best done with a single constant-pool load.
6310 // For AVX-length vectors, see if we can use a vector load to get all of the
6311 // elements, otherwise build the individual 128-bit pieces and use
6312 // shuffles to put them in place.
6313 if (VT.is256BitVector() || VT.is512BitVector()) {
6314 SmallVector<SDValue, 64> V(Op->op_begin(), Op->op_begin() + NumElems);
6316 // Check for a build vector of consecutive loads.
6317 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
6320 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
6322 // Build both the lower and upper subvector.
6323 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
6324 makeArrayRef(&V[0], NumElems/2));
6325 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
6326 makeArrayRef(&V[NumElems / 2], NumElems/2));
6328 // Recreate the wider vector with the lower and upper part.
6329 if (VT.is256BitVector())
6330 return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
6331 return Concat256BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
6334 // Let legalizer expand 2-wide build_vectors.
6335 if (EVTBits == 64) {
6336 if (NumNonZero == 1) {
6337 // One half is zero or undef.
6338 unsigned Idx = countTrailingZeros(NonZeros);
6339 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
6340 Op.getOperand(Idx));
6341 return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG);
6346 // If element VT is < 32 bits, convert it to inserts into a zero vector.
6347 if (EVTBits == 8 && NumElems == 16)
6348 if (SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
6352 if (EVTBits == 16 && NumElems == 8)
6353 if (SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
6357 // If element VT is == 32 bits and has 4 elems, try to generate an INSERTPS
6358 if (EVTBits == 32 && NumElems == 4)
6359 if (SDValue V = LowerBuildVectorv4x32(Op, DAG, Subtarget, *this))
6362 // If element VT is == 32 bits, turn it into a number of shuffles.
6363 SmallVector<SDValue, 8> V(NumElems);
6364 if (NumElems == 4 && NumZero > 0) {
6365 for (unsigned i = 0; i < 4; ++i) {
6366 bool isZero = !(NonZeros & (1 << i));
6368 V[i] = getZeroVector(VT, Subtarget, DAG, dl);
6370 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6373 for (unsigned i = 0; i < 2; ++i) {
6374 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
6377 V[i] = V[i*2]; // Must be a zero vector.
6380 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
6383 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
6386 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
6391 bool Reverse1 = (NonZeros & 0x3) == 2;
6392 bool Reverse2 = ((NonZeros & (0x3 << 2)) >> 2) == 2;
6396 static_cast<int>(Reverse2 ? NumElems+1 : NumElems),
6397 static_cast<int>(Reverse2 ? NumElems : NumElems+1)
6399 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
6402 if (Values.size() > 1 && VT.is128BitVector()) {
6403 // Check for a build vector of consecutive loads.
6404 for (unsigned i = 0; i < NumElems; ++i)
6405 V[i] = Op.getOperand(i);
6407 // Check for elements which are consecutive loads.
6408 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
6411 // Check for a build vector from mostly shuffle plus few inserting.
6412 if (SDValue Sh = buildFromShuffleMostly(Op, DAG))
6415 // For SSE 4.1, use insertps to put the high elements into the low element.
6416 if (Subtarget->hasSSE41()) {
6418 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
6419 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
6421 Result = DAG.getUNDEF(VT);
6423 for (unsigned i = 1; i < NumElems; ++i) {
6424 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
6425 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
6426 Op.getOperand(i), DAG.getIntPtrConstant(i, dl));
6431 // Otherwise, expand into a number of unpckl*, start by extending each of
6432 // our (non-undef) elements to the full vector width with the element in the
6433 // bottom slot of the vector (which generates no code for SSE).
6434 for (unsigned i = 0; i < NumElems; ++i) {
6435 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
6436 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6438 V[i] = DAG.getUNDEF(VT);
6441 // Next, we iteratively mix elements, e.g. for v4f32:
6442 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
6443 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
6444 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
6445 unsigned EltStride = NumElems >> 1;
6446 while (EltStride != 0) {
6447 for (unsigned i = 0; i < EltStride; ++i) {
6448 // If V[i+EltStride] is undef and this is the first round of mixing,
6449 // then it is safe to just drop this shuffle: V[i] is already in the
6450 // right place, the one element (since it's the first round) being
6451 // inserted as undef can be dropped. This isn't safe for successive
6452 // rounds because they will permute elements within both vectors.
6453 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
6454 EltStride == NumElems/2)
6457 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
6466 // 256-bit AVX can use the vinsertf128 instruction
6467 // to create 256-bit vectors from two other 128-bit ones.
6468 static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
6470 MVT ResVT = Op.getSimpleValueType();
6472 assert((ResVT.is256BitVector() ||
6473 ResVT.is512BitVector()) && "Value type must be 256-/512-bit wide");
6475 SDValue V1 = Op.getOperand(0);
6476 SDValue V2 = Op.getOperand(1);
6477 unsigned NumElems = ResVT.getVectorNumElements();
6478 if (ResVT.is256BitVector())
6479 return Concat128BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6481 if (Op.getNumOperands() == 4) {
6482 MVT HalfVT = MVT::getVectorVT(ResVT.getVectorElementType(),
6483 ResVT.getVectorNumElements()/2);
6484 SDValue V3 = Op.getOperand(2);
6485 SDValue V4 = Op.getOperand(3);
6486 return Concat256BitVectors(Concat128BitVectors(V1, V2, HalfVT, NumElems/2, DAG, dl),
6487 Concat128BitVectors(V3, V4, HalfVT, NumElems/2, DAG, dl), ResVT, NumElems, DAG, dl);
6489 return Concat256BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6492 static SDValue LowerCONCAT_VECTORSvXi1(SDValue Op,
6493 const X86Subtarget *Subtarget,
6494 SelectionDAG & DAG) {
6496 MVT ResVT = Op.getSimpleValueType();
6497 unsigned NumOfOperands = Op.getNumOperands();
6499 assert(isPowerOf2_32(NumOfOperands) &&
6500 "Unexpected number of operands in CONCAT_VECTORS");
6502 if (NumOfOperands > 2) {
6503 MVT HalfVT = MVT::getVectorVT(ResVT.getVectorElementType(),
6504 ResVT.getVectorNumElements()/2);
6505 SmallVector<SDValue, 2> Ops;
6506 for (unsigned i = 0; i < NumOfOperands/2; i++)
6507 Ops.push_back(Op.getOperand(i));
6508 SDValue Lo = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6510 for (unsigned i = NumOfOperands/2; i < NumOfOperands; i++)
6511 Ops.push_back(Op.getOperand(i));
6512 SDValue Hi = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6513 return DAG.getNode(ISD::CONCAT_VECTORS, dl, ResVT, Lo, Hi);
6516 SDValue V1 = Op.getOperand(0);
6517 SDValue V2 = Op.getOperand(1);
6518 bool IsZeroV1 = ISD::isBuildVectorAllZeros(V1.getNode());
6519 bool IsZeroV2 = ISD::isBuildVectorAllZeros(V2.getNode());
6521 if (IsZeroV1 && IsZeroV2)
6522 return getZeroVector(ResVT, Subtarget, DAG, dl);
6524 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
6525 SDValue Undef = DAG.getUNDEF(ResVT);
6526 unsigned NumElems = ResVT.getVectorNumElements();
6527 SDValue ShiftBits = DAG.getConstant(NumElems/2, dl, MVT::i8);
6529 V2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V2, ZeroIdx);
6530 V2 = DAG.getNode(X86ISD::VSHLI, dl, ResVT, V2, ShiftBits);
6534 V1 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V1, ZeroIdx);
6535 // Zero the upper bits of V1
6536 V1 = DAG.getNode(X86ISD::VSHLI, dl, ResVT, V1, ShiftBits);
6537 V1 = DAG.getNode(X86ISD::VSRLI, dl, ResVT, V1, ShiftBits);
6540 return DAG.getNode(ISD::OR, dl, ResVT, V1, V2);
6543 static SDValue LowerCONCAT_VECTORS(SDValue Op,
6544 const X86Subtarget *Subtarget,
6545 SelectionDAG &DAG) {
6546 MVT VT = Op.getSimpleValueType();
6547 if (VT.getVectorElementType() == MVT::i1)
6548 return LowerCONCAT_VECTORSvXi1(Op, Subtarget, DAG);
6550 assert((VT.is256BitVector() && Op.getNumOperands() == 2) ||
6551 (VT.is512BitVector() && (Op.getNumOperands() == 2 ||
6552 Op.getNumOperands() == 4)));
6554 // AVX can use the vinsertf128 instruction to create 256-bit vectors
6555 // from two other 128-bit ones.
6557 // 512-bit vector may contain 2 256-bit vectors or 4 128-bit vectors
6558 return LowerAVXCONCAT_VECTORS(Op, DAG);
6561 //===----------------------------------------------------------------------===//
6562 // Vector shuffle lowering
6564 // This is an experimental code path for lowering vector shuffles on x86. It is
6565 // designed to handle arbitrary vector shuffles and blends, gracefully
6566 // degrading performance as necessary. It works hard to recognize idiomatic
6567 // shuffles and lower them to optimal instruction patterns without leaving
6568 // a framework that allows reasonably efficient handling of all vector shuffle
6570 //===----------------------------------------------------------------------===//
6572 /// \brief Tiny helper function to identify a no-op mask.
6574 /// This is a somewhat boring predicate function. It checks whether the mask
6575 /// array input, which is assumed to be a single-input shuffle mask of the kind
6576 /// used by the X86 shuffle instructions (not a fully general
6577 /// ShuffleVectorSDNode mask) requires any shuffles to occur. Both undef and an
6578 /// in-place shuffle are 'no-op's.
6579 static bool isNoopShuffleMask(ArrayRef<int> Mask) {
6580 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6581 if (Mask[i] != -1 && Mask[i] != i)
6586 /// \brief Helper function to classify a mask as a single-input mask.
6588 /// This isn't a generic single-input test because in the vector shuffle
6589 /// lowering we canonicalize single inputs to be the first input operand. This
6590 /// means we can more quickly test for a single input by only checking whether
6591 /// an input from the second operand exists. We also assume that the size of
6592 /// mask corresponds to the size of the input vectors which isn't true in the
6593 /// fully general case.
6594 static bool isSingleInputShuffleMask(ArrayRef<int> Mask) {
6596 if (M >= (int)Mask.size())
6601 /// \brief Test whether there are elements crossing 128-bit lanes in this
6604 /// X86 divides up its shuffles into in-lane and cross-lane shuffle operations
6605 /// and we routinely test for these.
6606 static bool is128BitLaneCrossingShuffleMask(MVT VT, ArrayRef<int> Mask) {
6607 int LaneSize = 128 / VT.getScalarSizeInBits();
6608 int Size = Mask.size();
6609 for (int i = 0; i < Size; ++i)
6610 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
6615 /// \brief Test whether a shuffle mask is equivalent within each 128-bit lane.
6617 /// This checks a shuffle mask to see if it is performing the same
6618 /// 128-bit lane-relative shuffle in each 128-bit lane. This trivially implies
6619 /// that it is also not lane-crossing. It may however involve a blend from the
6620 /// same lane of a second vector.
6622 /// The specific repeated shuffle mask is populated in \p RepeatedMask, as it is
6623 /// non-trivial to compute in the face of undef lanes. The representation is
6624 /// *not* suitable for use with existing 128-bit shuffles as it will contain
6625 /// entries from both V1 and V2 inputs to the wider mask.
6627 is128BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask,
6628 SmallVectorImpl<int> &RepeatedMask) {
6629 int LaneSize = 128 / VT.getScalarSizeInBits();
6630 RepeatedMask.resize(LaneSize, -1);
6631 int Size = Mask.size();
6632 for (int i = 0; i < Size; ++i) {
6635 if ((Mask[i] % Size) / LaneSize != i / LaneSize)
6636 // This entry crosses lanes, so there is no way to model this shuffle.
6639 // Ok, handle the in-lane shuffles by detecting if and when they repeat.
6640 if (RepeatedMask[i % LaneSize] == -1)
6641 // This is the first non-undef entry in this slot of a 128-bit lane.
6642 RepeatedMask[i % LaneSize] =
6643 Mask[i] < Size ? Mask[i] % LaneSize : Mask[i] % LaneSize + Size;
6644 else if (RepeatedMask[i % LaneSize] + (i / LaneSize) * LaneSize != Mask[i])
6645 // Found a mismatch with the repeated mask.
6651 /// \brief Checks whether a shuffle mask is equivalent to an explicit list of
6654 /// This is a fast way to test a shuffle mask against a fixed pattern:
6656 /// if (isShuffleEquivalent(Mask, 3, 2, {1, 0})) { ... }
6658 /// It returns true if the mask is exactly as wide as the argument list, and
6659 /// each element of the mask is either -1 (signifying undef) or the value given
6660 /// in the argument.
6661 static bool isShuffleEquivalent(SDValue V1, SDValue V2, ArrayRef<int> Mask,
6662 ArrayRef<int> ExpectedMask) {
6663 if (Mask.size() != ExpectedMask.size())
6666 int Size = Mask.size();
6668 // If the values are build vectors, we can look through them to find
6669 // equivalent inputs that make the shuffles equivalent.
6670 auto *BV1 = dyn_cast<BuildVectorSDNode>(V1);
6671 auto *BV2 = dyn_cast<BuildVectorSDNode>(V2);
6673 for (int i = 0; i < Size; ++i)
6674 if (Mask[i] != -1 && Mask[i] != ExpectedMask[i]) {
6675 auto *MaskBV = Mask[i] < Size ? BV1 : BV2;
6676 auto *ExpectedBV = ExpectedMask[i] < Size ? BV1 : BV2;
6677 if (!MaskBV || !ExpectedBV ||
6678 MaskBV->getOperand(Mask[i] % Size) !=
6679 ExpectedBV->getOperand(ExpectedMask[i] % Size))
6686 /// \brief Get a 4-lane 8-bit shuffle immediate for a mask.
6688 /// This helper function produces an 8-bit shuffle immediate corresponding to
6689 /// the ubiquitous shuffle encoding scheme used in x86 instructions for
6690 /// shuffling 4 lanes. It can be used with most of the PSHUF instructions for
6693 /// NB: We rely heavily on "undef" masks preserving the input lane.
6694 static SDValue getV4X86ShuffleImm8ForMask(ArrayRef<int> Mask, SDLoc DL,
6695 SelectionDAG &DAG) {
6696 assert(Mask.size() == 4 && "Only 4-lane shuffle masks");
6697 assert(Mask[0] >= -1 && Mask[0] < 4 && "Out of bound mask element!");
6698 assert(Mask[1] >= -1 && Mask[1] < 4 && "Out of bound mask element!");
6699 assert(Mask[2] >= -1 && Mask[2] < 4 && "Out of bound mask element!");
6700 assert(Mask[3] >= -1 && Mask[3] < 4 && "Out of bound mask element!");
6703 Imm |= (Mask[0] == -1 ? 0 : Mask[0]) << 0;
6704 Imm |= (Mask[1] == -1 ? 1 : Mask[1]) << 2;
6705 Imm |= (Mask[2] == -1 ? 2 : Mask[2]) << 4;
6706 Imm |= (Mask[3] == -1 ? 3 : Mask[3]) << 6;
6707 return DAG.getConstant(Imm, DL, MVT::i8);
6710 /// \brief Compute whether each element of a shuffle is zeroable.
6712 /// A "zeroable" vector shuffle element is one which can be lowered to zero.
6713 /// Either it is an undef element in the shuffle mask, the element of the input
6714 /// referenced is undef, or the element of the input referenced is known to be
6715 /// zero. Many x86 shuffles can zero lanes cheaply and we often want to handle
6716 /// as many lanes with this technique as possible to simplify the remaining
6718 static SmallBitVector computeZeroableShuffleElements(ArrayRef<int> Mask,
6719 SDValue V1, SDValue V2) {
6720 SmallBitVector Zeroable(Mask.size(), false);
6722 while (V1.getOpcode() == ISD::BITCAST)
6723 V1 = V1->getOperand(0);
6724 while (V2.getOpcode() == ISD::BITCAST)
6725 V2 = V2->getOperand(0);
6727 bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
6728 bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
6730 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6732 // Handle the easy cases.
6733 if (M < 0 || (M >= 0 && M < Size && V1IsZero) || (M >= Size && V2IsZero)) {
6738 // If this is an index into a build_vector node (which has the same number
6739 // of elements), dig out the input value and use it.
6740 SDValue V = M < Size ? V1 : V2;
6741 if (V.getOpcode() != ISD::BUILD_VECTOR || Size != (int)V.getNumOperands())
6744 SDValue Input = V.getOperand(M % Size);
6745 // The UNDEF opcode check really should be dead code here, but not quite
6746 // worth asserting on (it isn't invalid, just unexpected).
6747 if (Input.getOpcode() == ISD::UNDEF || X86::isZeroNode(Input))
6754 // X86 has dedicated unpack instructions that can handle specific blend
6755 // operations: UNPCKH and UNPCKL.
6756 static SDValue lowerVectorShuffleWithUNPCK(SDLoc DL, MVT VT, ArrayRef<int> Mask,
6757 SDValue V1, SDValue V2,
6758 SelectionDAG &DAG) {
6759 int NumElts = VT.getVectorNumElements();
6760 int NumEltsInLane = 128 / VT.getScalarSizeInBits();
6761 SmallVector<int, 8> Unpckl;
6762 SmallVector<int, 8> Unpckh;
6764 for (int i = 0; i < NumElts; ++i) {
6765 unsigned LaneStart = (i / NumEltsInLane) * NumEltsInLane;
6766 int LoPos = (i % NumEltsInLane) / 2 + LaneStart + NumElts * (i % 2);
6767 int HiPos = LoPos + NumEltsInLane / 2;
6768 Unpckl.push_back(LoPos);
6769 Unpckh.push_back(HiPos);
6772 if (isShuffleEquivalent(V1, V2, Mask, Unpckl))
6773 return DAG.getNode(X86ISD::UNPCKL, DL, VT, V1, V2);
6774 if (isShuffleEquivalent(V1, V2, Mask, Unpckh))
6775 return DAG.getNode(X86ISD::UNPCKH, DL, VT, V1, V2);
6777 // Commute and try again.
6778 ShuffleVectorSDNode::commuteMask(Unpckl);
6779 if (isShuffleEquivalent(V1, V2, Mask, Unpckl))
6780 return DAG.getNode(X86ISD::UNPCKL, DL, VT, V2, V1);
6782 ShuffleVectorSDNode::commuteMask(Unpckh);
6783 if (isShuffleEquivalent(V1, V2, Mask, Unpckh))
6784 return DAG.getNode(X86ISD::UNPCKH, DL, VT, V2, V1);
6789 /// \brief Try to emit a bitmask instruction for a shuffle.
6791 /// This handles cases where we can model a blend exactly as a bitmask due to
6792 /// one of the inputs being zeroable.
6793 static SDValue lowerVectorShuffleAsBitMask(SDLoc DL, MVT VT, SDValue V1,
6794 SDValue V2, ArrayRef<int> Mask,
6795 SelectionDAG &DAG) {
6796 MVT EltVT = VT.getVectorElementType();
6797 int NumEltBits = EltVT.getSizeInBits();
6798 MVT IntEltVT = MVT::getIntegerVT(NumEltBits);
6799 SDValue Zero = DAG.getConstant(0, DL, IntEltVT);
6800 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
6802 if (EltVT.isFloatingPoint()) {
6803 Zero = DAG.getBitcast(EltVT, Zero);
6804 AllOnes = DAG.getBitcast(EltVT, AllOnes);
6806 SmallVector<SDValue, 16> VMaskOps(Mask.size(), Zero);
6807 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6809 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6812 if (Mask[i] % Size != i)
6813 return SDValue(); // Not a blend.
6815 V = Mask[i] < Size ? V1 : V2;
6816 else if (V != (Mask[i] < Size ? V1 : V2))
6817 return SDValue(); // Can only let one input through the mask.
6819 VMaskOps[i] = AllOnes;
6822 return SDValue(); // No non-zeroable elements!
6824 SDValue VMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, VMaskOps);
6825 V = DAG.getNode(VT.isFloatingPoint()
6826 ? (unsigned) X86ISD::FAND : (unsigned) ISD::AND,
6831 /// \brief Try to emit a blend instruction for a shuffle using bit math.
6833 /// This is used as a fallback approach when first class blend instructions are
6834 /// unavailable. Currently it is only suitable for integer vectors, but could
6835 /// be generalized for floating point vectors if desirable.
6836 static SDValue lowerVectorShuffleAsBitBlend(SDLoc DL, MVT VT, SDValue V1,
6837 SDValue V2, ArrayRef<int> Mask,
6838 SelectionDAG &DAG) {
6839 assert(VT.isInteger() && "Only supports integer vector types!");
6840 MVT EltVT = VT.getVectorElementType();
6841 int NumEltBits = EltVT.getSizeInBits();
6842 SDValue Zero = DAG.getConstant(0, DL, EltVT);
6843 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
6845 SmallVector<SDValue, 16> MaskOps;
6846 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6847 if (Mask[i] != -1 && Mask[i] != i && Mask[i] != i + Size)
6848 return SDValue(); // Shuffled input!
6849 MaskOps.push_back(Mask[i] < Size ? AllOnes : Zero);
6852 SDValue V1Mask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, MaskOps);
6853 V1 = DAG.getNode(ISD::AND, DL, VT, V1, V1Mask);
6854 // We have to cast V2 around.
6855 MVT MaskVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
6856 V2 = DAG.getBitcast(VT, DAG.getNode(X86ISD::ANDNP, DL, MaskVT,
6857 DAG.getBitcast(MaskVT, V1Mask),
6858 DAG.getBitcast(MaskVT, V2)));
6859 return DAG.getNode(ISD::OR, DL, VT, V1, V2);
6862 /// \brief Try to emit a blend instruction for a shuffle.
6864 /// This doesn't do any checks for the availability of instructions for blending
6865 /// these values. It relies on the availability of the X86ISD::BLENDI pattern to
6866 /// be matched in the backend with the type given. What it does check for is
6867 /// that the shuffle mask is a blend, or convertible into a blend with zero.
6868 static SDValue lowerVectorShuffleAsBlend(SDLoc DL, MVT VT, SDValue V1,
6869 SDValue V2, ArrayRef<int> Original,
6870 const X86Subtarget *Subtarget,
6871 SelectionDAG &DAG) {
6872 bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
6873 bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
6874 SmallVector<int, 8> Mask(Original.begin(), Original.end());
6875 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6876 bool ForceV1Zero = false, ForceV2Zero = false;
6878 // Attempt to generate the binary blend mask. If an input is zero then
6879 // we can use any lane.
6880 // TODO: generalize the zero matching to any scalar like isShuffleEquivalent.
6881 unsigned BlendMask = 0;
6882 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6888 if (M == i + Size) {
6889 BlendMask |= 1u << i;
6900 BlendMask |= 1u << i;
6905 return SDValue(); // Shuffled input!
6908 // Create a REAL zero vector - ISD::isBuildVectorAllZeros allows UNDEFs.
6910 V1 = getZeroVector(VT, Subtarget, DAG, DL);
6912 V2 = getZeroVector(VT, Subtarget, DAG, DL);
6914 auto ScaleBlendMask = [](unsigned BlendMask, int Size, int Scale) {
6915 unsigned ScaledMask = 0;
6916 for (int i = 0; i != Size; ++i)
6917 if (BlendMask & (1u << i))
6918 for (int j = 0; j != Scale; ++j)
6919 ScaledMask |= 1u << (i * Scale + j);
6923 switch (VT.SimpleTy) {
6928 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V2,
6929 DAG.getConstant(BlendMask, DL, MVT::i8));
6933 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6937 // If we have AVX2 it is faster to use VPBLENDD when the shuffle fits into
6938 // that instruction.
6939 if (Subtarget->hasAVX2()) {
6940 // Scale the blend by the number of 32-bit dwords per element.
6941 int Scale = VT.getScalarSizeInBits() / 32;
6942 BlendMask = ScaleBlendMask(BlendMask, Mask.size(), Scale);
6943 MVT BlendVT = VT.getSizeInBits() > 128 ? MVT::v8i32 : MVT::v4i32;
6944 V1 = DAG.getBitcast(BlendVT, V1);
6945 V2 = DAG.getBitcast(BlendVT, V2);
6946 return DAG.getBitcast(
6947 VT, DAG.getNode(X86ISD::BLENDI, DL, BlendVT, V1, V2,
6948 DAG.getConstant(BlendMask, DL, MVT::i8)));
6952 // For integer shuffles we need to expand the mask and cast the inputs to
6953 // v8i16s prior to blending.
6954 int Scale = 8 / VT.getVectorNumElements();
6955 BlendMask = ScaleBlendMask(BlendMask, Mask.size(), Scale);
6956 V1 = DAG.getBitcast(MVT::v8i16, V1);
6957 V2 = DAG.getBitcast(MVT::v8i16, V2);
6958 return DAG.getBitcast(VT,
6959 DAG.getNode(X86ISD::BLENDI, DL, MVT::v8i16, V1, V2,
6960 DAG.getConstant(BlendMask, DL, MVT::i8)));
6964 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6965 SmallVector<int, 8> RepeatedMask;
6966 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
6967 // We can lower these with PBLENDW which is mirrored across 128-bit lanes.
6968 assert(RepeatedMask.size() == 8 && "Repeated mask size doesn't match!");
6970 for (int i = 0; i < 8; ++i)
6971 if (RepeatedMask[i] >= 16)
6972 BlendMask |= 1u << i;
6973 return DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2,
6974 DAG.getConstant(BlendMask, DL, MVT::i8));
6980 assert((VT.is128BitVector() || Subtarget->hasAVX2()) &&
6981 "256-bit byte-blends require AVX2 support!");
6983 // Attempt to lower to a bitmask if we can. VPAND is faster than VPBLENDVB.
6984 if (SDValue Masked = lowerVectorShuffleAsBitMask(DL, VT, V1, V2, Mask, DAG))
6987 // Scale the blend by the number of bytes per element.
6988 int Scale = VT.getScalarSizeInBits() / 8;
6990 // This form of blend is always done on bytes. Compute the byte vector
6992 MVT BlendVT = MVT::getVectorVT(MVT::i8, VT.getSizeInBits() / 8);
6994 // Compute the VSELECT mask. Note that VSELECT is really confusing in the
6995 // mix of LLVM's code generator and the x86 backend. We tell the code
6996 // generator that boolean values in the elements of an x86 vector register
6997 // are -1 for true and 0 for false. We then use the LLVM semantics of 'true'
6998 // mapping a select to operand #1, and 'false' mapping to operand #2. The
6999 // reality in x86 is that vector masks (pre-AVX-512) use only the high bit
7000 // of the element (the remaining are ignored) and 0 in that high bit would
7001 // mean operand #1 while 1 in the high bit would mean operand #2. So while
7002 // the LLVM model for boolean values in vector elements gets the relevant
7003 // bit set, it is set backwards and over constrained relative to x86's
7005 SmallVector<SDValue, 32> VSELECTMask;
7006 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7007 for (int j = 0; j < Scale; ++j)
7008 VSELECTMask.push_back(
7009 Mask[i] < 0 ? DAG.getUNDEF(MVT::i8)
7010 : DAG.getConstant(Mask[i] < Size ? -1 : 0, DL,
7013 V1 = DAG.getBitcast(BlendVT, V1);
7014 V2 = DAG.getBitcast(BlendVT, V2);
7015 return DAG.getBitcast(VT, DAG.getNode(ISD::VSELECT, DL, BlendVT,
7016 DAG.getNode(ISD::BUILD_VECTOR, DL,
7017 BlendVT, VSELECTMask),
7022 llvm_unreachable("Not a supported integer vector type!");
7026 /// \brief Try to lower as a blend of elements from two inputs followed by
7027 /// a single-input permutation.
7029 /// This matches the pattern where we can blend elements from two inputs and
7030 /// then reduce the shuffle to a single-input permutation.
7031 static SDValue lowerVectorShuffleAsBlendAndPermute(SDLoc DL, MVT VT, SDValue V1,
7034 SelectionDAG &DAG) {
7035 // We build up the blend mask while checking whether a blend is a viable way
7036 // to reduce the shuffle.
7037 SmallVector<int, 32> BlendMask(Mask.size(), -1);
7038 SmallVector<int, 32> PermuteMask(Mask.size(), -1);
7040 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
7044 assert(Mask[i] < Size * 2 && "Shuffle input is out of bounds.");
7046 if (BlendMask[Mask[i] % Size] == -1)
7047 BlendMask[Mask[i] % Size] = Mask[i];
7048 else if (BlendMask[Mask[i] % Size] != Mask[i])
7049 return SDValue(); // Can't blend in the needed input!
7051 PermuteMask[i] = Mask[i] % Size;
7054 SDValue V = DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
7055 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), PermuteMask);
7058 /// \brief Generic routine to decompose a shuffle and blend into indepndent
7059 /// blends and permutes.
7061 /// This matches the extremely common pattern for handling combined
7062 /// shuffle+blend operations on newer X86 ISAs where we have very fast blend
7063 /// operations. It will try to pick the best arrangement of shuffles and
7065 static SDValue lowerVectorShuffleAsDecomposedShuffleBlend(SDLoc DL, MVT VT,
7069 SelectionDAG &DAG) {
7070 // Shuffle the input elements into the desired positions in V1 and V2 and
7071 // blend them together.
7072 SmallVector<int, 32> V1Mask(Mask.size(), -1);
7073 SmallVector<int, 32> V2Mask(Mask.size(), -1);
7074 SmallVector<int, 32> BlendMask(Mask.size(), -1);
7075 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7076 if (Mask[i] >= 0 && Mask[i] < Size) {
7077 V1Mask[i] = Mask[i];
7079 } else if (Mask[i] >= Size) {
7080 V2Mask[i] = Mask[i] - Size;
7081 BlendMask[i] = i + Size;
7084 // Try to lower with the simpler initial blend strategy unless one of the
7085 // input shuffles would be a no-op. We prefer to shuffle inputs as the
7086 // shuffle may be able to fold with a load or other benefit. However, when
7087 // we'll have to do 2x as many shuffles in order to achieve this, blending
7088 // first is a better strategy.
7089 if (!isNoopShuffleMask(V1Mask) && !isNoopShuffleMask(V2Mask))
7090 if (SDValue BlendPerm =
7091 lowerVectorShuffleAsBlendAndPermute(DL, VT, V1, V2, Mask, DAG))
7094 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
7095 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
7096 return DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
7099 /// \brief Try to lower a vector shuffle as a byte rotation.
7101 /// SSSE3 has a generic PALIGNR instruction in x86 that will do an arbitrary
7102 /// byte-rotation of the concatenation of two vectors; pre-SSSE3 can use
7103 /// a PSRLDQ/PSLLDQ/POR pattern to get a similar effect. This routine will
7104 /// try to generically lower a vector shuffle through such an pattern. It
7105 /// does not check for the profitability of lowering either as PALIGNR or
7106 /// PSRLDQ/PSLLDQ/POR, only whether the mask is valid to lower in that form.
7107 /// This matches shuffle vectors that look like:
7109 /// v8i16 [11, 12, 13, 14, 15, 0, 1, 2]
7111 /// Essentially it concatenates V1 and V2, shifts right by some number of
7112 /// elements, and takes the low elements as the result. Note that while this is
7113 /// specified as a *right shift* because x86 is little-endian, it is a *left
7114 /// rotate* of the vector lanes.
7115 static SDValue lowerVectorShuffleAsByteRotate(SDLoc DL, MVT VT, SDValue V1,
7118 const X86Subtarget *Subtarget,
7119 SelectionDAG &DAG) {
7120 assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!");
7122 int NumElts = Mask.size();
7123 int NumLanes = VT.getSizeInBits() / 128;
7124 int NumLaneElts = NumElts / NumLanes;
7126 // We need to detect various ways of spelling a rotation:
7127 // [11, 12, 13, 14, 15, 0, 1, 2]
7128 // [-1, 12, 13, 14, -1, -1, 1, -1]
7129 // [-1, -1, -1, -1, -1, -1, 1, 2]
7130 // [ 3, 4, 5, 6, 7, 8, 9, 10]
7131 // [-1, 4, 5, 6, -1, -1, 9, -1]
7132 // [-1, 4, 5, 6, -1, -1, -1, -1]
7135 for (int l = 0; l < NumElts; l += NumLaneElts) {
7136 for (int i = 0; i < NumLaneElts; ++i) {
7137 if (Mask[l + i] == -1)
7139 assert(Mask[l + i] >= 0 && "Only -1 is a valid negative mask element!");
7141 // Get the mod-Size index and lane correct it.
7142 int LaneIdx = (Mask[l + i] % NumElts) - l;
7143 // Make sure it was in this lane.
7144 if (LaneIdx < 0 || LaneIdx >= NumLaneElts)
7147 // Determine where a rotated vector would have started.
7148 int StartIdx = i - LaneIdx;
7150 // The identity rotation isn't interesting, stop.
7153 // If we found the tail of a vector the rotation must be the missing
7154 // front. If we found the head of a vector, it must be how much of the
7156 int CandidateRotation = StartIdx < 0 ? -StartIdx : NumLaneElts - StartIdx;
7159 Rotation = CandidateRotation;
7160 else if (Rotation != CandidateRotation)
7161 // The rotations don't match, so we can't match this mask.
7164 // Compute which value this mask is pointing at.
7165 SDValue MaskV = Mask[l + i] < NumElts ? V1 : V2;
7167 // Compute which of the two target values this index should be assigned
7168 // to. This reflects whether the high elements are remaining or the low
7169 // elements are remaining.
7170 SDValue &TargetV = StartIdx < 0 ? Hi : Lo;
7172 // Either set up this value if we've not encountered it before, or check
7173 // that it remains consistent.
7176 else if (TargetV != MaskV)
7177 // This may be a rotation, but it pulls from the inputs in some
7178 // unsupported interleaving.
7183 // Check that we successfully analyzed the mask, and normalize the results.
7184 assert(Rotation != 0 && "Failed to locate a viable rotation!");
7185 assert((Lo || Hi) && "Failed to find a rotated input vector!");
7191 // The actual rotate instruction rotates bytes, so we need to scale the
7192 // rotation based on how many bytes are in the vector lane.
7193 int Scale = 16 / NumLaneElts;
7195 // SSSE3 targets can use the palignr instruction.
7196 if (Subtarget->hasSSSE3()) {
7197 // Cast the inputs to i8 vector of correct length to match PALIGNR.
7198 MVT AlignVT = MVT::getVectorVT(MVT::i8, 16 * NumLanes);
7199 Lo = DAG.getBitcast(AlignVT, Lo);
7200 Hi = DAG.getBitcast(AlignVT, Hi);
7202 return DAG.getBitcast(
7203 VT, DAG.getNode(X86ISD::PALIGNR, DL, AlignVT, Lo, Hi,
7204 DAG.getConstant(Rotation * Scale, DL, MVT::i8)));
7207 assert(VT.is128BitVector() &&
7208 "Rotate-based lowering only supports 128-bit lowering!");
7209 assert(Mask.size() <= 16 &&
7210 "Can shuffle at most 16 bytes in a 128-bit vector!");
7212 // Default SSE2 implementation
7213 int LoByteShift = 16 - Rotation * Scale;
7214 int HiByteShift = Rotation * Scale;
7216 // Cast the inputs to v2i64 to match PSLLDQ/PSRLDQ.
7217 Lo = DAG.getBitcast(MVT::v2i64, Lo);
7218 Hi = DAG.getBitcast(MVT::v2i64, Hi);
7220 SDValue LoShift = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v2i64, Lo,
7221 DAG.getConstant(LoByteShift, DL, MVT::i8));
7222 SDValue HiShift = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v2i64, Hi,
7223 DAG.getConstant(HiByteShift, DL, MVT::i8));
7224 return DAG.getBitcast(VT,
7225 DAG.getNode(ISD::OR, DL, MVT::v2i64, LoShift, HiShift));
7228 /// \brief Try to lower a vector shuffle as a bit shift (shifts in zeros).
7230 /// Attempts to match a shuffle mask against the PSLL(W/D/Q/DQ) and
7231 /// PSRL(W/D/Q/DQ) SSE2 and AVX2 logical bit-shift instructions. The function
7232 /// matches elements from one of the input vectors shuffled to the left or
7233 /// right with zeroable elements 'shifted in'. It handles both the strictly
7234 /// bit-wise element shifts and the byte shift across an entire 128-bit double
7237 /// PSHL : (little-endian) left bit shift.
7238 /// [ zz, 0, zz, 2 ]
7239 /// [ -1, 4, zz, -1 ]
7240 /// PSRL : (little-endian) right bit shift.
7242 /// [ -1, -1, 7, zz]
7243 /// PSLLDQ : (little-endian) left byte shift
7244 /// [ zz, 0, 1, 2, 3, 4, 5, 6]
7245 /// [ zz, zz, -1, -1, 2, 3, 4, -1]
7246 /// [ zz, zz, zz, zz, zz, zz, -1, 1]
7247 /// PSRLDQ : (little-endian) right byte shift
7248 /// [ 5, 6, 7, zz, zz, zz, zz, zz]
7249 /// [ -1, 5, 6, 7, zz, zz, zz, zz]
7250 /// [ 1, 2, -1, -1, -1, -1, zz, zz]
7251 static SDValue lowerVectorShuffleAsShift(SDLoc DL, MVT VT, SDValue V1,
7252 SDValue V2, ArrayRef<int> Mask,
7253 SelectionDAG &DAG) {
7254 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7256 int Size = Mask.size();
7257 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
7259 auto CheckZeros = [&](int Shift, int Scale, bool Left) {
7260 for (int i = 0; i < Size; i += Scale)
7261 for (int j = 0; j < Shift; ++j)
7262 if (!Zeroable[i + j + (Left ? 0 : (Scale - Shift))])
7268 auto MatchShift = [&](int Shift, int Scale, bool Left, SDValue V) {
7269 for (int i = 0; i != Size; i += Scale) {
7270 unsigned Pos = Left ? i + Shift : i;
7271 unsigned Low = Left ? i : i + Shift;
7272 unsigned Len = Scale - Shift;
7273 if (!isSequentialOrUndefInRange(Mask, Pos, Len,
7274 Low + (V == V1 ? 0 : Size)))
7278 int ShiftEltBits = VT.getScalarSizeInBits() * Scale;
7279 bool ByteShift = ShiftEltBits > 64;
7280 unsigned OpCode = Left ? (ByteShift ? X86ISD::VSHLDQ : X86ISD::VSHLI)
7281 : (ByteShift ? X86ISD::VSRLDQ : X86ISD::VSRLI);
7282 int ShiftAmt = Shift * VT.getScalarSizeInBits() / (ByteShift ? 8 : 1);
7284 // Normalize the scale for byte shifts to still produce an i64 element
7286 Scale = ByteShift ? Scale / 2 : Scale;
7288 // We need to round trip through the appropriate type for the shift.
7289 MVT ShiftSVT = MVT::getIntegerVT(VT.getScalarSizeInBits() * Scale);
7290 MVT ShiftVT = MVT::getVectorVT(ShiftSVT, Size / Scale);
7291 assert(DAG.getTargetLoweringInfo().isTypeLegal(ShiftVT) &&
7292 "Illegal integer vector type");
7293 V = DAG.getBitcast(ShiftVT, V);
7295 V = DAG.getNode(OpCode, DL, ShiftVT, V,
7296 DAG.getConstant(ShiftAmt, DL, MVT::i8));
7297 return DAG.getBitcast(VT, V);
7300 // SSE/AVX supports logical shifts up to 64-bit integers - so we can just
7301 // keep doubling the size of the integer elements up to that. We can
7302 // then shift the elements of the integer vector by whole multiples of
7303 // their width within the elements of the larger integer vector. Test each
7304 // multiple to see if we can find a match with the moved element indices
7305 // and that the shifted in elements are all zeroable.
7306 for (int Scale = 2; Scale * VT.getScalarSizeInBits() <= 128; Scale *= 2)
7307 for (int Shift = 1; Shift != Scale; ++Shift)
7308 for (bool Left : {true, false})
7309 if (CheckZeros(Shift, Scale, Left))
7310 for (SDValue V : {V1, V2})
7311 if (SDValue Match = MatchShift(Shift, Scale, Left, V))
7318 /// \brief Try to lower a vector shuffle using SSE4a EXTRQ/INSERTQ.
7319 static SDValue lowerVectorShuffleWithSSE4A(SDLoc DL, MVT VT, SDValue V1,
7320 SDValue V2, ArrayRef<int> Mask,
7321 SelectionDAG &DAG) {
7322 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7323 assert(!Zeroable.all() && "Fully zeroable shuffle mask");
7325 int Size = Mask.size();
7326 int HalfSize = Size / 2;
7327 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
7329 // Upper half must be undefined.
7330 if (!isUndefInRange(Mask, HalfSize, HalfSize))
7333 // EXTRQ: Extract Len elements from lower half of source, starting at Idx.
7334 // Remainder of lower half result is zero and upper half is all undef.
7335 auto LowerAsEXTRQ = [&]() {
7336 // Determine the extraction length from the part of the
7337 // lower half that isn't zeroable.
7339 for (; Len > 0; --Len)
7340 if (!Zeroable[Len - 1])
7342 assert(Len > 0 && "Zeroable shuffle mask");
7344 // Attempt to match first Len sequential elements from the lower half.
7347 for (int i = 0; i != Len; ++i) {
7351 SDValue &V = (M < Size ? V1 : V2);
7354 // All mask elements must be in the lower half.
7358 if (Idx < 0 || (Src == V && Idx == (M - i))) {
7369 assert((Idx + Len) <= HalfSize && "Illegal extraction mask");
7370 int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
7371 int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
7372 return DAG.getNode(X86ISD::EXTRQI, DL, VT, Src,
7373 DAG.getConstant(BitLen, DL, MVT::i8),
7374 DAG.getConstant(BitIdx, DL, MVT::i8));
7377 if (SDValue ExtrQ = LowerAsEXTRQ())
7380 // INSERTQ: Extract lowest Len elements from lower half of second source and
7381 // insert over first source, starting at Idx.
7382 // { A[0], .., A[Idx-1], B[0], .., B[Len-1], A[Idx+Len], .., UNDEF, ... }
7383 auto LowerAsInsertQ = [&]() {
7384 for (int Idx = 0; Idx != HalfSize; ++Idx) {
7387 // Attempt to match first source from mask before insertion point.
7388 if (isUndefInRange(Mask, 0, Idx)) {
7390 } else if (isSequentialOrUndefInRange(Mask, 0, Idx, 0)) {
7392 } else if (isSequentialOrUndefInRange(Mask, 0, Idx, Size)) {
7398 // Extend the extraction length looking to match both the insertion of
7399 // the second source and the remaining elements of the first.
7400 for (int Hi = Idx + 1; Hi <= HalfSize; ++Hi) {
7405 if (isSequentialOrUndefInRange(Mask, Idx, Len, 0)) {
7407 } else if (isSequentialOrUndefInRange(Mask, Idx, Len, Size)) {
7413 // Match the remaining elements of the lower half.
7414 if (isUndefInRange(Mask, Hi, HalfSize - Hi)) {
7416 } else if ((!Base || (Base == V1)) &&
7417 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi, Hi)) {
7419 } else if ((!Base || (Base == V2)) &&
7420 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi,
7427 // We may not have a base (first source) - this can safely be undefined.
7429 Base = DAG.getUNDEF(VT);
7431 int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
7432 int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
7433 return DAG.getNode(X86ISD::INSERTQI, DL, VT, Base, Insert,
7434 DAG.getConstant(BitLen, DL, MVT::i8),
7435 DAG.getConstant(BitIdx, DL, MVT::i8));
7442 if (SDValue InsertQ = LowerAsInsertQ())
7448 /// \brief Lower a vector shuffle as a zero or any extension.
7450 /// Given a specific number of elements, element bit width, and extension
7451 /// stride, produce either a zero or any extension based on the available
7452 /// features of the subtarget. The extended elements are consecutive and
7453 /// begin and can start from an offseted element index in the input; to
7454 /// avoid excess shuffling the offset must either being in the bottom lane
7455 /// or at the start of a higher lane. All extended elements must be from
7457 static SDValue lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7458 SDLoc DL, MVT VT, int Scale, int Offset, bool AnyExt, SDValue InputV,
7459 ArrayRef<int> Mask, const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7460 assert(Scale > 1 && "Need a scale to extend.");
7461 int EltBits = VT.getScalarSizeInBits();
7462 int NumElements = VT.getVectorNumElements();
7463 int NumEltsPerLane = 128 / EltBits;
7464 int OffsetLane = Offset / NumEltsPerLane;
7465 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
7466 "Only 8, 16, and 32 bit elements can be extended.");
7467 assert(Scale * EltBits <= 64 && "Cannot zero extend past 64 bits.");
7468 assert(0 <= Offset && "Extension offset must be positive.");
7469 assert((Offset < NumEltsPerLane || Offset % NumEltsPerLane == 0) &&
7470 "Extension offset must be in the first lane or start an upper lane.");
7472 // Check that an index is in same lane as the base offset.
7473 auto SafeOffset = [&](int Idx) {
7474 return OffsetLane == (Idx / NumEltsPerLane);
7477 // Shift along an input so that the offset base moves to the first element.
7478 auto ShuffleOffset = [&](SDValue V) {
7482 SmallVector<int, 8> ShMask((unsigned)NumElements, -1);
7483 for (int i = 0; i * Scale < NumElements; ++i) {
7484 int SrcIdx = i + Offset;
7485 ShMask[i] = SafeOffset(SrcIdx) ? SrcIdx : -1;
7487 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), ShMask);
7490 // Found a valid zext mask! Try various lowering strategies based on the
7491 // input type and available ISA extensions.
7492 if (Subtarget->hasSSE41()) {
7493 // Not worth offseting 128-bit vectors if scale == 2, a pattern using
7494 // PUNPCK will catch this in a later shuffle match.
7495 if (Offset && Scale == 2 && VT.is128BitVector())
7497 MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits * Scale),
7498 NumElements / Scale);
7499 InputV = DAG.getNode(X86ISD::VZEXT, DL, ExtVT, ShuffleOffset(InputV));
7500 return DAG.getBitcast(VT, InputV);
7503 assert(VT.is128BitVector() && "Only 128-bit vectors can be extended.");
7505 // For any extends we can cheat for larger element sizes and use shuffle
7506 // instructions that can fold with a load and/or copy.
7507 if (AnyExt && EltBits == 32) {
7508 int PSHUFDMask[4] = {Offset, -1, SafeOffset(Offset + 1) ? Offset + 1 : -1,
7510 return DAG.getBitcast(
7511 VT, DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7512 DAG.getBitcast(MVT::v4i32, InputV),
7513 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
7515 if (AnyExt && EltBits == 16 && Scale > 2) {
7516 int PSHUFDMask[4] = {Offset / 2, -1,
7517 SafeOffset(Offset + 1) ? (Offset + 1) / 2 : -1, -1};
7518 InputV = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7519 DAG.getBitcast(MVT::v4i32, InputV),
7520 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG));
7521 int PSHUFWMask[4] = {1, -1, -1, -1};
7522 unsigned OddEvenOp = (Offset & 1 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW);
7523 return DAG.getBitcast(
7524 VT, DAG.getNode(OddEvenOp, DL, MVT::v8i16,
7525 DAG.getBitcast(MVT::v8i16, InputV),
7526 getV4X86ShuffleImm8ForMask(PSHUFWMask, DL, DAG)));
7529 // The SSE4A EXTRQ instruction can efficiently extend the first 2 lanes
7531 if ((Scale * EltBits) == 64 && EltBits < 32 && Subtarget->hasSSE4A()) {
7532 assert(NumElements == (int)Mask.size() && "Unexpected shuffle mask size!");
7533 assert(VT.is128BitVector() && "Unexpected vector width!");
7535 int LoIdx = Offset * EltBits;
7536 SDValue Lo = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7537 DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
7538 DAG.getConstant(EltBits, DL, MVT::i8),
7539 DAG.getConstant(LoIdx, DL, MVT::i8)));
7541 if (isUndefInRange(Mask, NumElements / 2, NumElements / 2) ||
7542 !SafeOffset(Offset + 1))
7543 return DAG.getNode(ISD::BITCAST, DL, VT, Lo);
7545 int HiIdx = (Offset + 1) * EltBits;
7546 SDValue Hi = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7547 DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
7548 DAG.getConstant(EltBits, DL, MVT::i8),
7549 DAG.getConstant(HiIdx, DL, MVT::i8)));
7550 return DAG.getNode(ISD::BITCAST, DL, VT,
7551 DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, Lo, Hi));
7554 // If this would require more than 2 unpack instructions to expand, use
7555 // pshufb when available. We can only use more than 2 unpack instructions
7556 // when zero extending i8 elements which also makes it easier to use pshufb.
7557 if (Scale > 4 && EltBits == 8 && Subtarget->hasSSSE3()) {
7558 assert(NumElements == 16 && "Unexpected byte vector width!");
7559 SDValue PSHUFBMask[16];
7560 for (int i = 0; i < 16; ++i) {
7561 int Idx = Offset + (i / Scale);
7562 PSHUFBMask[i] = DAG.getConstant(
7563 (i % Scale == 0 && SafeOffset(Idx)) ? Idx : 0x80, DL, MVT::i8);
7565 InputV = DAG.getBitcast(MVT::v16i8, InputV);
7566 return DAG.getBitcast(VT,
7567 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, InputV,
7568 DAG.getNode(ISD::BUILD_VECTOR, DL,
7569 MVT::v16i8, PSHUFBMask)));
7572 // If we are extending from an offset, ensure we start on a boundary that
7573 // we can unpack from.
7574 int AlignToUnpack = Offset % (NumElements / Scale);
7575 if (AlignToUnpack) {
7576 SmallVector<int, 8> ShMask((unsigned)NumElements, -1);
7577 for (int i = AlignToUnpack; i < NumElements; ++i)
7578 ShMask[i - AlignToUnpack] = i;
7579 InputV = DAG.getVectorShuffle(VT, DL, InputV, DAG.getUNDEF(VT), ShMask);
7580 Offset -= AlignToUnpack;
7583 // Otherwise emit a sequence of unpacks.
7585 unsigned UnpackLoHi = X86ISD::UNPCKL;
7586 if (Offset >= (NumElements / 2)) {
7587 UnpackLoHi = X86ISD::UNPCKH;
7588 Offset -= (NumElements / 2);
7591 MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements);
7592 SDValue Ext = AnyExt ? DAG.getUNDEF(InputVT)
7593 : getZeroVector(InputVT, Subtarget, DAG, DL);
7594 InputV = DAG.getBitcast(InputVT, InputV);
7595 InputV = DAG.getNode(UnpackLoHi, DL, InputVT, InputV, Ext);
7599 } while (Scale > 1);
7600 return DAG.getBitcast(VT, InputV);
7603 /// \brief Try to lower a vector shuffle as a zero extension on any microarch.
7605 /// This routine will try to do everything in its power to cleverly lower
7606 /// a shuffle which happens to match the pattern of a zero extend. It doesn't
7607 /// check for the profitability of this lowering, it tries to aggressively
7608 /// match this pattern. It will use all of the micro-architectural details it
7609 /// can to emit an efficient lowering. It handles both blends with all-zero
7610 /// inputs to explicitly zero-extend and undef-lanes (sometimes undef due to
7611 /// masking out later).
7613 /// The reason we have dedicated lowering for zext-style shuffles is that they
7614 /// are both incredibly common and often quite performance sensitive.
7615 static SDValue lowerVectorShuffleAsZeroOrAnyExtend(
7616 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7617 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7618 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7620 int Bits = VT.getSizeInBits();
7621 int NumLanes = Bits / 128;
7622 int NumElements = VT.getVectorNumElements();
7623 int NumEltsPerLane = NumElements / NumLanes;
7624 assert(VT.getScalarSizeInBits() <= 32 &&
7625 "Exceeds 32-bit integer zero extension limit");
7626 assert((int)Mask.size() == NumElements && "Unexpected shuffle mask size");
7628 // Define a helper function to check a particular ext-scale and lower to it if
7630 auto Lower = [&](int Scale) -> SDValue {
7635 for (int i = 0; i < NumElements; ++i) {
7638 continue; // Valid anywhere but doesn't tell us anything.
7639 if (i % Scale != 0) {
7640 // Each of the extended elements need to be zeroable.
7644 // We no longer are in the anyext case.
7649 // Each of the base elements needs to be consecutive indices into the
7650 // same input vector.
7651 SDValue V = M < NumElements ? V1 : V2;
7652 M = M % NumElements;
7655 Offset = M - (i / Scale);
7656 } else if (InputV != V)
7657 return SDValue(); // Flip-flopping inputs.
7659 // Offset must start in the lowest 128-bit lane or at the start of an
7661 // FIXME: Is it ever worth allowing a negative base offset?
7662 if (!((0 <= Offset && Offset < NumEltsPerLane) ||
7663 (Offset % NumEltsPerLane) == 0))
7666 // If we are offsetting, all referenced entries must come from the same
7668 if (Offset && (Offset / NumEltsPerLane) != (M / NumEltsPerLane))
7671 if ((M % NumElements) != (Offset + (i / Scale)))
7672 return SDValue(); // Non-consecutive strided elements.
7676 // If we fail to find an input, we have a zero-shuffle which should always
7677 // have already been handled.
7678 // FIXME: Maybe handle this here in case during blending we end up with one?
7682 // If we are offsetting, don't extend if we only match a single input, we
7683 // can always do better by using a basic PSHUF or PUNPCK.
7684 if (Offset != 0 && Matches < 2)
7687 return lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7688 DL, VT, Scale, Offset, AnyExt, InputV, Mask, Subtarget, DAG);
7691 // The widest scale possible for extending is to a 64-bit integer.
7692 assert(Bits % 64 == 0 &&
7693 "The number of bits in a vector must be divisible by 64 on x86!");
7694 int NumExtElements = Bits / 64;
7696 // Each iteration, try extending the elements half as much, but into twice as
7698 for (; NumExtElements < NumElements; NumExtElements *= 2) {
7699 assert(NumElements % NumExtElements == 0 &&
7700 "The input vector size must be divisible by the extended size.");
7701 if (SDValue V = Lower(NumElements / NumExtElements))
7705 // General extends failed, but 128-bit vectors may be able to use MOVQ.
7709 // Returns one of the source operands if the shuffle can be reduced to a
7710 // MOVQ, copying the lower 64-bits and zero-extending to the upper 64-bits.
7711 auto CanZExtLowHalf = [&]() {
7712 for (int i = NumElements / 2; i != NumElements; ++i)
7715 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, 0))
7717 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, NumElements))
7722 if (SDValue V = CanZExtLowHalf()) {
7723 V = DAG.getBitcast(MVT::v2i64, V);
7724 V = DAG.getNode(X86ISD::VZEXT_MOVL, DL, MVT::v2i64, V);
7725 return DAG.getBitcast(VT, V);
7728 // No viable ext lowering found.
7732 /// \brief Try to get a scalar value for a specific element of a vector.
7734 /// Looks through BUILD_VECTOR and SCALAR_TO_VECTOR nodes to find a scalar.
7735 static SDValue getScalarValueForVectorElement(SDValue V, int Idx,
7736 SelectionDAG &DAG) {
7737 MVT VT = V.getSimpleValueType();
7738 MVT EltVT = VT.getVectorElementType();
7739 while (V.getOpcode() == ISD::BITCAST)
7740 V = V.getOperand(0);
7741 // If the bitcasts shift the element size, we can't extract an equivalent
7743 MVT NewVT = V.getSimpleValueType();
7744 if (!NewVT.isVector() || NewVT.getScalarSizeInBits() != VT.getScalarSizeInBits())
7747 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7748 (Idx == 0 && V.getOpcode() == ISD::SCALAR_TO_VECTOR)) {
7749 // Ensure the scalar operand is the same size as the destination.
7750 // FIXME: Add support for scalar truncation where possible.
7751 SDValue S = V.getOperand(Idx);
7752 if (EltVT.getSizeInBits() == S.getSimpleValueType().getSizeInBits())
7753 return DAG.getNode(ISD::BITCAST, SDLoc(V), EltVT, S);
7759 /// \brief Helper to test for a load that can be folded with x86 shuffles.
7761 /// This is particularly important because the set of instructions varies
7762 /// significantly based on whether the operand is a load or not.
7763 static bool isShuffleFoldableLoad(SDValue V) {
7764 while (V.getOpcode() == ISD::BITCAST)
7765 V = V.getOperand(0);
7767 return ISD::isNON_EXTLoad(V.getNode());
7770 /// \brief Try to lower insertion of a single element into a zero vector.
7772 /// This is a common pattern that we have especially efficient patterns to lower
7773 /// across all subtarget feature sets.
7774 static SDValue lowerVectorShuffleAsElementInsertion(
7775 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7776 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7777 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7779 MVT EltVT = VT.getVectorElementType();
7781 int V2Index = std::find_if(Mask.begin(), Mask.end(),
7782 [&Mask](int M) { return M >= (int)Mask.size(); }) -
7784 bool IsV1Zeroable = true;
7785 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7786 if (i != V2Index && !Zeroable[i]) {
7787 IsV1Zeroable = false;
7791 // Check for a single input from a SCALAR_TO_VECTOR node.
7792 // FIXME: All of this should be canonicalized into INSERT_VECTOR_ELT and
7793 // all the smarts here sunk into that routine. However, the current
7794 // lowering of BUILD_VECTOR makes that nearly impossible until the old
7795 // vector shuffle lowering is dead.
7796 SDValue V2S = getScalarValueForVectorElement(V2, Mask[V2Index] - Mask.size(),
7798 if (V2S && DAG.getTargetLoweringInfo().isTypeLegal(V2S.getValueType())) {
7799 // We need to zext the scalar if it is smaller than an i32.
7800 V2S = DAG.getBitcast(EltVT, V2S);
7801 if (EltVT == MVT::i8 || EltVT == MVT::i16) {
7802 // Using zext to expand a narrow element won't work for non-zero
7807 // Zero-extend directly to i32.
7809 V2S = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, V2S);
7811 V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtVT, V2S);
7812 } else if (Mask[V2Index] != (int)Mask.size() || EltVT == MVT::i8 ||
7813 EltVT == MVT::i16) {
7814 // Either not inserting from the low element of the input or the input
7815 // element size is too small to use VZEXT_MOVL to clear the high bits.
7819 if (!IsV1Zeroable) {
7820 // If V1 can't be treated as a zero vector we have fewer options to lower
7821 // this. We can't support integer vectors or non-zero targets cheaply, and
7822 // the V1 elements can't be permuted in any way.
7823 assert(VT == ExtVT && "Cannot change extended type when non-zeroable!");
7824 if (!VT.isFloatingPoint() || V2Index != 0)
7826 SmallVector<int, 8> V1Mask(Mask.begin(), Mask.end());
7827 V1Mask[V2Index] = -1;
7828 if (!isNoopShuffleMask(V1Mask))
7830 // This is essentially a special case blend operation, but if we have
7831 // general purpose blend operations, they are always faster. Bail and let
7832 // the rest of the lowering handle these as blends.
7833 if (Subtarget->hasSSE41())
7836 // Otherwise, use MOVSD or MOVSS.
7837 assert((EltVT == MVT::f32 || EltVT == MVT::f64) &&
7838 "Only two types of floating point element types to handle!");
7839 return DAG.getNode(EltVT == MVT::f32 ? X86ISD::MOVSS : X86ISD::MOVSD, DL,
7843 // This lowering only works for the low element with floating point vectors.
7844 if (VT.isFloatingPoint() && V2Index != 0)
7847 V2 = DAG.getNode(X86ISD::VZEXT_MOVL, DL, ExtVT, V2);
7849 V2 = DAG.getBitcast(VT, V2);
7852 // If we have 4 or fewer lanes we can cheaply shuffle the element into
7853 // the desired position. Otherwise it is more efficient to do a vector
7854 // shift left. We know that we can do a vector shift left because all
7855 // the inputs are zero.
7856 if (VT.isFloatingPoint() || VT.getVectorNumElements() <= 4) {
7857 SmallVector<int, 4> V2Shuffle(Mask.size(), 1);
7858 V2Shuffle[V2Index] = 0;
7859 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Shuffle);
7861 V2 = DAG.getBitcast(MVT::v2i64, V2);
7863 X86ISD::VSHLDQ, DL, MVT::v2i64, V2,
7864 DAG.getConstant(V2Index * EltVT.getSizeInBits() / 8, DL,
7865 DAG.getTargetLoweringInfo().getScalarShiftAmountTy(
7866 DAG.getDataLayout(), VT)));
7867 V2 = DAG.getBitcast(VT, V2);
7873 /// \brief Try to lower broadcast of a single - truncated - integer element,
7874 /// coming from a scalar_to_vector/build_vector node \p V0 with larger elements.
7876 /// This assumes we have AVX2.
7877 static SDValue lowerVectorShuffleAsTruncBroadcast(SDLoc DL, MVT VT, SDValue V0,
7879 const X86Subtarget *Subtarget,
7880 SelectionDAG &DAG) {
7881 assert(Subtarget->hasAVX2() &&
7882 "We can only lower integer broadcasts with AVX2!");
7884 EVT EltVT = VT.getVectorElementType();
7885 EVT V0VT = V0.getValueType();
7887 assert(VT.isInteger() && "Unexpected non-integer trunc broadcast!");
7888 assert(V0VT.isVector() && "Unexpected non-vector vector-sized value!");
7890 EVT V0EltVT = V0VT.getVectorElementType();
7891 if (!V0EltVT.isInteger())
7894 const unsigned EltSize = EltVT.getSizeInBits();
7895 const unsigned V0EltSize = V0EltVT.getSizeInBits();
7897 // This is only a truncation if the original element type is larger.
7898 if (V0EltSize <= EltSize)
7901 assert(((V0EltSize % EltSize) == 0) &&
7902 "Scalar type sizes must all be powers of 2 on x86!");
7904 const unsigned V0Opc = V0.getOpcode();
7905 const unsigned Scale = V0EltSize / EltSize;
7906 const unsigned V0BroadcastIdx = BroadcastIdx / Scale;
7908 if ((V0Opc != ISD::SCALAR_TO_VECTOR || V0BroadcastIdx != 0) &&
7909 V0Opc != ISD::BUILD_VECTOR)
7912 SDValue Scalar = V0.getOperand(V0BroadcastIdx);
7914 // If we're extracting non-least-significant bits, shift so we can truncate.
7915 // Hopefully, we can fold away the trunc/srl/load into the broadcast.
7916 // Even if we can't (and !isShuffleFoldableLoad(Scalar)), prefer
7917 // vpbroadcast+vmovd+shr to vpshufb(m)+vmovd.
7918 if (const int OffsetIdx = BroadcastIdx % Scale)
7919 Scalar = DAG.getNode(ISD::SRL, DL, Scalar.getValueType(), Scalar,
7920 DAG.getConstant(OffsetIdx * EltSize, DL, Scalar.getValueType()));
7922 return DAG.getNode(X86ISD::VBROADCAST, DL, VT,
7923 DAG.getNode(ISD::TRUNCATE, DL, EltVT, Scalar));
7926 /// \brief Try to lower broadcast of a single element.
7928 /// For convenience, this code also bundles all of the subtarget feature set
7929 /// filtering. While a little annoying to re-dispatch on type here, there isn't
7930 /// a convenient way to factor it out.
7931 static SDValue lowerVectorShuffleAsBroadcast(SDLoc DL, MVT VT, SDValue V,
7933 const X86Subtarget *Subtarget,
7934 SelectionDAG &DAG) {
7935 if (!Subtarget->hasAVX())
7937 if (VT.isInteger() && !Subtarget->hasAVX2())
7940 // Check that the mask is a broadcast.
7941 int BroadcastIdx = -1;
7943 if (M >= 0 && BroadcastIdx == -1)
7945 else if (M >= 0 && M != BroadcastIdx)
7948 assert(BroadcastIdx < (int)Mask.size() && "We only expect to be called with "
7949 "a sorted mask where the broadcast "
7952 // Go up the chain of (vector) values to find a scalar load that we can
7953 // combine with the broadcast.
7955 switch (V.getOpcode()) {
7956 case ISD::CONCAT_VECTORS: {
7957 int OperandSize = Mask.size() / V.getNumOperands();
7958 V = V.getOperand(BroadcastIdx / OperandSize);
7959 BroadcastIdx %= OperandSize;
7963 case ISD::INSERT_SUBVECTOR: {
7964 SDValue VOuter = V.getOperand(0), VInner = V.getOperand(1);
7965 auto ConstantIdx = dyn_cast<ConstantSDNode>(V.getOperand(2));
7969 int BeginIdx = (int)ConstantIdx->getZExtValue();
7971 BeginIdx + (int)VInner.getSimpleValueType().getVectorNumElements();
7972 if (BroadcastIdx >= BeginIdx && BroadcastIdx < EndIdx) {
7973 BroadcastIdx -= BeginIdx;
7984 // Check if this is a broadcast of a scalar. We special case lowering
7985 // for scalars so that we can more effectively fold with loads.
7986 // First, look through bitcast: if the original value has a larger element
7987 // type than the shuffle, the broadcast element is in essence truncated.
7988 // Make that explicit to ease folding.
7989 if (V.getOpcode() == ISD::BITCAST && VT.isInteger())
7990 if (SDValue TruncBroadcast = lowerVectorShuffleAsTruncBroadcast(
7991 DL, VT, V.getOperand(0), BroadcastIdx, Subtarget, DAG))
7992 return TruncBroadcast;
7994 // Also check the simpler case, where we can directly reuse the scalar.
7995 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7996 (V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)) {
7997 V = V.getOperand(BroadcastIdx);
7999 // If the scalar isn't a load, we can't broadcast from it in AVX1.
8000 // Only AVX2 has register broadcasts.
8001 if (!Subtarget->hasAVX2() && !isShuffleFoldableLoad(V))
8003 } else if (BroadcastIdx != 0 || !Subtarget->hasAVX2()) {
8004 // We can't broadcast from a vector register without AVX2, and we can only
8005 // broadcast from the zero-element of a vector register.
8009 return DAG.getNode(X86ISD::VBROADCAST, DL, VT, V);
8012 // Check for whether we can use INSERTPS to perform the shuffle. We only use
8013 // INSERTPS when the V1 elements are already in the correct locations
8014 // because otherwise we can just always use two SHUFPS instructions which
8015 // are much smaller to encode than a SHUFPS and an INSERTPS. We can also
8016 // perform INSERTPS if a single V1 element is out of place and all V2
8017 // elements are zeroable.
8018 static SDValue lowerVectorShuffleAsInsertPS(SDValue Op, SDValue V1, SDValue V2,
8020 SelectionDAG &DAG) {
8021 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
8022 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8023 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8024 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8026 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
8029 int V1DstIndex = -1;
8030 int V2DstIndex = -1;
8031 bool V1UsedInPlace = false;
8033 for (int i = 0; i < 4; ++i) {
8034 // Synthesize a zero mask from the zeroable elements (includes undefs).
8040 // Flag if we use any V1 inputs in place.
8042 V1UsedInPlace = true;
8046 // We can only insert a single non-zeroable element.
8047 if (V1DstIndex != -1 || V2DstIndex != -1)
8051 // V1 input out of place for insertion.
8054 // V2 input for insertion.
8059 // Don't bother if we have no (non-zeroable) element for insertion.
8060 if (V1DstIndex == -1 && V2DstIndex == -1)
8063 // Determine element insertion src/dst indices. The src index is from the
8064 // start of the inserted vector, not the start of the concatenated vector.
8065 unsigned V2SrcIndex = 0;
8066 if (V1DstIndex != -1) {
8067 // If we have a V1 input out of place, we use V1 as the V2 element insertion
8068 // and don't use the original V2 at all.
8069 V2SrcIndex = Mask[V1DstIndex];
8070 V2DstIndex = V1DstIndex;
8073 V2SrcIndex = Mask[V2DstIndex] - 4;
8076 // If no V1 inputs are used in place, then the result is created only from
8077 // the zero mask and the V2 insertion - so remove V1 dependency.
8079 V1 = DAG.getUNDEF(MVT::v4f32);
8081 unsigned InsertPSMask = V2SrcIndex << 6 | V2DstIndex << 4 | ZMask;
8082 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
8084 // Insert the V2 element into the desired position.
8086 return DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
8087 DAG.getConstant(InsertPSMask, DL, MVT::i8));
8090 /// \brief Try to lower a shuffle as a permute of the inputs followed by an
8091 /// UNPCK instruction.
8093 /// This specifically targets cases where we end up with alternating between
8094 /// the two inputs, and so can permute them into something that feeds a single
8095 /// UNPCK instruction. Note that this routine only targets integer vectors
8096 /// because for floating point vectors we have a generalized SHUFPS lowering
8097 /// strategy that handles everything that doesn't *exactly* match an unpack,
8098 /// making this clever lowering unnecessary.
8099 static SDValue lowerVectorShuffleAsPermuteAndUnpack(SDLoc DL, MVT VT,
8100 SDValue V1, SDValue V2,
8102 SelectionDAG &DAG) {
8103 assert(!VT.isFloatingPoint() &&
8104 "This routine only supports integer vectors.");
8105 assert(!isSingleInputShuffleMask(Mask) &&
8106 "This routine should only be used when blending two inputs.");
8107 assert(Mask.size() >= 2 && "Single element masks are invalid.");
8109 int Size = Mask.size();
8111 int NumLoInputs = std::count_if(Mask.begin(), Mask.end(), [Size](int M) {
8112 return M >= 0 && M % Size < Size / 2;
8114 int NumHiInputs = std::count_if(
8115 Mask.begin(), Mask.end(), [Size](int M) { return M % Size >= Size / 2; });
8117 bool UnpackLo = NumLoInputs >= NumHiInputs;
8119 auto TryUnpack = [&](MVT UnpackVT, int Scale) {
8120 SmallVector<int, 32> V1Mask(Mask.size(), -1);
8121 SmallVector<int, 32> V2Mask(Mask.size(), -1);
8123 for (int i = 0; i < Size; ++i) {
8127 // Each element of the unpack contains Scale elements from this mask.
8128 int UnpackIdx = i / Scale;
8130 // We only handle the case where V1 feeds the first slots of the unpack.
8131 // We rely on canonicalization to ensure this is the case.
8132 if ((UnpackIdx % 2 == 0) != (Mask[i] < Size))
8135 // Setup the mask for this input. The indexing is tricky as we have to
8136 // handle the unpack stride.
8137 SmallVectorImpl<int> &VMask = (UnpackIdx % 2 == 0) ? V1Mask : V2Mask;
8138 VMask[(UnpackIdx / 2) * Scale + i % Scale + (UnpackLo ? 0 : Size / 2)] =
8142 // If we will have to shuffle both inputs to use the unpack, check whether
8143 // we can just unpack first and shuffle the result. If so, skip this unpack.
8144 if ((NumLoInputs == 0 || NumHiInputs == 0) && !isNoopShuffleMask(V1Mask) &&
8145 !isNoopShuffleMask(V2Mask))
8148 // Shuffle the inputs into place.
8149 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
8150 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
8152 // Cast the inputs to the type we will use to unpack them.
8153 V1 = DAG.getBitcast(UnpackVT, V1);
8154 V2 = DAG.getBitcast(UnpackVT, V2);
8156 // Unpack the inputs and cast the result back to the desired type.
8157 return DAG.getBitcast(
8158 VT, DAG.getNode(UnpackLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
8162 // We try each unpack from the largest to the smallest to try and find one
8163 // that fits this mask.
8164 int OrigNumElements = VT.getVectorNumElements();
8165 int OrigScalarSize = VT.getScalarSizeInBits();
8166 for (int ScalarSize = 64; ScalarSize >= OrigScalarSize; ScalarSize /= 2) {
8167 int Scale = ScalarSize / OrigScalarSize;
8168 int NumElements = OrigNumElements / Scale;
8169 MVT UnpackVT = MVT::getVectorVT(MVT::getIntegerVT(ScalarSize), NumElements);
8170 if (SDValue Unpack = TryUnpack(UnpackVT, Scale))
8174 // If none of the unpack-rooted lowerings worked (or were profitable) try an
8176 if (NumLoInputs == 0 || NumHiInputs == 0) {
8177 assert((NumLoInputs > 0 || NumHiInputs > 0) &&
8178 "We have to have *some* inputs!");
8179 int HalfOffset = NumLoInputs == 0 ? Size / 2 : 0;
8181 // FIXME: We could consider the total complexity of the permute of each
8182 // possible unpacking. Or at the least we should consider how many
8183 // half-crossings are created.
8184 // FIXME: We could consider commuting the unpacks.
8186 SmallVector<int, 32> PermMask;
8187 PermMask.assign(Size, -1);
8188 for (int i = 0; i < Size; ++i) {
8192 assert(Mask[i] % Size >= HalfOffset && "Found input from wrong half!");
8195 2 * ((Mask[i] % Size) - HalfOffset) + (Mask[i] < Size ? 0 : 1);
8197 return DAG.getVectorShuffle(
8198 VT, DL, DAG.getNode(NumLoInputs == 0 ? X86ISD::UNPCKH : X86ISD::UNPCKL,
8200 DAG.getUNDEF(VT), PermMask);
8206 /// \brief Handle lowering of 2-lane 64-bit floating point shuffles.
8208 /// This is the basis function for the 2-lane 64-bit shuffles as we have full
8209 /// support for floating point shuffles but not integer shuffles. These
8210 /// instructions will incur a domain crossing penalty on some chips though so
8211 /// it is better to avoid lowering through this for integer vectors where
8213 static SDValue lowerV2F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8214 const X86Subtarget *Subtarget,
8215 SelectionDAG &DAG) {
8217 assert(Op.getSimpleValueType() == MVT::v2f64 && "Bad shuffle type!");
8218 assert(V1.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
8219 assert(V2.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
8220 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8221 ArrayRef<int> Mask = SVOp->getMask();
8222 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
8224 if (isSingleInputShuffleMask(Mask)) {
8225 // Use low duplicate instructions for masks that match their pattern.
8226 if (Subtarget->hasSSE3())
8227 if (isShuffleEquivalent(V1, V2, Mask, {0, 0}))
8228 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v2f64, V1);
8230 // Straight shuffle of a single input vector. Simulate this by using the
8231 // single input as both of the "inputs" to this instruction..
8232 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1);
8234 if (Subtarget->hasAVX()) {
8235 // If we have AVX, we can use VPERMILPS which will allow folding a load
8236 // into the shuffle.
8237 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v2f64, V1,
8238 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8241 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V1,
8242 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8244 assert(Mask[0] >= 0 && Mask[0] < 2 && "Non-canonicalized blend!");
8245 assert(Mask[1] >= 2 && "Non-canonicalized blend!");
8247 // If we have a single input, insert that into V1 if we can do so cheaply.
8248 if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1) {
8249 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8250 DL, MVT::v2f64, V1, V2, Mask, Subtarget, DAG))
8252 // Try inverting the insertion since for v2 masks it is easy to do and we
8253 // can't reliably sort the mask one way or the other.
8254 int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2),
8255 Mask[1] < 0 ? -1 : (Mask[1] ^ 2)};
8256 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8257 DL, MVT::v2f64, V2, V1, InverseMask, Subtarget, DAG))
8261 // Try to use one of the special instruction patterns to handle two common
8262 // blend patterns if a zero-blend above didn't work.
8263 if (isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
8264 isShuffleEquivalent(V1, V2, Mask, {1, 3}))
8265 if (SDValue V1S = getScalarValueForVectorElement(V1, Mask[0], DAG))
8266 // We can either use a special instruction to load over the low double or
8267 // to move just the low double.
8269 isShuffleFoldableLoad(V1S) ? X86ISD::MOVLPD : X86ISD::MOVSD,
8271 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, V1S));
8273 if (Subtarget->hasSSE41())
8274 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2f64, V1, V2, Mask,
8278 // Use dedicated unpack instructions for masks that match their pattern.
8280 lowerVectorShuffleWithUNPCK(DL, MVT::v2f64, Mask, V1, V2, DAG))
8283 unsigned SHUFPDMask = (Mask[0] == 1) | (((Mask[1] - 2) == 1) << 1);
8284 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V2,
8285 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8288 /// \brief Handle lowering of 2-lane 64-bit integer shuffles.
8290 /// Tries to lower a 2-lane 64-bit shuffle using shuffle operations provided by
8291 /// the integer unit to minimize domain crossing penalties. However, for blends
8292 /// it falls back to the floating point shuffle operation with appropriate bit
8294 static SDValue lowerV2I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8295 const X86Subtarget *Subtarget,
8296 SelectionDAG &DAG) {
8298 assert(Op.getSimpleValueType() == MVT::v2i64 && "Bad shuffle type!");
8299 assert(V1.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
8300 assert(V2.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
8301 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8302 ArrayRef<int> Mask = SVOp->getMask();
8303 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
8305 if (isSingleInputShuffleMask(Mask)) {
8306 // Check for being able to broadcast a single element.
8307 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v2i64, V1,
8308 Mask, Subtarget, DAG))
8311 // Straight shuffle of a single input vector. For everything from SSE2
8312 // onward this has a single fast instruction with no scary immediates.
8313 // We have to map the mask as it is actually a v4i32 shuffle instruction.
8314 V1 = DAG.getBitcast(MVT::v4i32, V1);
8315 int WidenedMask[4] = {
8316 std::max(Mask[0], 0) * 2, std::max(Mask[0], 0) * 2 + 1,
8317 std::max(Mask[1], 0) * 2, std::max(Mask[1], 0) * 2 + 1};
8318 return DAG.getBitcast(
8320 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
8321 getV4X86ShuffleImm8ForMask(WidenedMask, DL, DAG)));
8323 assert(Mask[0] != -1 && "No undef lanes in multi-input v2 shuffles!");
8324 assert(Mask[1] != -1 && "No undef lanes in multi-input v2 shuffles!");
8325 assert(Mask[0] < 2 && "We sort V1 to be the first input.");
8326 assert(Mask[1] >= 2 && "We sort V2 to be the second input.");
8328 // If we have a blend of two PACKUS operations an the blend aligns with the
8329 // low and half halves, we can just merge the PACKUS operations. This is
8330 // particularly important as it lets us merge shuffles that this routine itself
8332 auto GetPackNode = [](SDValue V) {
8333 while (V.getOpcode() == ISD::BITCAST)
8334 V = V.getOperand(0);
8336 return V.getOpcode() == X86ISD::PACKUS ? V : SDValue();
8338 if (SDValue V1Pack = GetPackNode(V1))
8339 if (SDValue V2Pack = GetPackNode(V2))
8340 return DAG.getBitcast(MVT::v2i64,
8341 DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8,
8342 Mask[0] == 0 ? V1Pack.getOperand(0)
8343 : V1Pack.getOperand(1),
8344 Mask[1] == 2 ? V2Pack.getOperand(0)
8345 : V2Pack.getOperand(1)));
8347 // Try to use shift instructions.
8349 lowerVectorShuffleAsShift(DL, MVT::v2i64, V1, V2, Mask, DAG))
8352 // When loading a scalar and then shuffling it into a vector we can often do
8353 // the insertion cheaply.
8354 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8355 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
8357 // Try inverting the insertion since for v2 masks it is easy to do and we
8358 // can't reliably sort the mask one way or the other.
8359 int InverseMask[2] = {Mask[0] ^ 2, Mask[1] ^ 2};
8360 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8361 DL, MVT::v2i64, V2, V1, InverseMask, Subtarget, DAG))
8364 // We have different paths for blend lowering, but they all must use the
8365 // *exact* same predicate.
8366 bool IsBlendSupported = Subtarget->hasSSE41();
8367 if (IsBlendSupported)
8368 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2i64, V1, V2, Mask,
8372 // Use dedicated unpack instructions for masks that match their pattern.
8374 lowerVectorShuffleWithUNPCK(DL, MVT::v2i64, Mask, V1, V2, DAG))
8377 // Try to use byte rotation instructions.
8378 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8379 if (Subtarget->hasSSSE3())
8380 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8381 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
8384 // If we have direct support for blends, we should lower by decomposing into
8385 // a permute. That will be faster than the domain cross.
8386 if (IsBlendSupported)
8387 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v2i64, V1, V2,
8390 // We implement this with SHUFPD which is pretty lame because it will likely
8391 // incur 2 cycles of stall for integer vectors on Nehalem and older chips.
8392 // However, all the alternatives are still more cycles and newer chips don't
8393 // have this problem. It would be really nice if x86 had better shuffles here.
8394 V1 = DAG.getBitcast(MVT::v2f64, V1);
8395 V2 = DAG.getBitcast(MVT::v2f64, V2);
8396 return DAG.getBitcast(MVT::v2i64,
8397 DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask));
8400 /// \brief Test whether this can be lowered with a single SHUFPS instruction.
8402 /// This is used to disable more specialized lowerings when the shufps lowering
8403 /// will happen to be efficient.
8404 static bool isSingleSHUFPSMask(ArrayRef<int> Mask) {
8405 // This routine only handles 128-bit shufps.
8406 assert(Mask.size() == 4 && "Unsupported mask size!");
8408 // To lower with a single SHUFPS we need to have the low half and high half
8409 // each requiring a single input.
8410 if (Mask[0] != -1 && Mask[1] != -1 && (Mask[0] < 4) != (Mask[1] < 4))
8412 if (Mask[2] != -1 && Mask[3] != -1 && (Mask[2] < 4) != (Mask[3] < 4))
8418 /// \brief Lower a vector shuffle using the SHUFPS instruction.
8420 /// This is a helper routine dedicated to lowering vector shuffles using SHUFPS.
8421 /// It makes no assumptions about whether this is the *best* lowering, it simply
8423 static SDValue lowerVectorShuffleWithSHUFPS(SDLoc DL, MVT VT,
8424 ArrayRef<int> Mask, SDValue V1,
8425 SDValue V2, SelectionDAG &DAG) {
8426 SDValue LowV = V1, HighV = V2;
8427 int NewMask[4] = {Mask[0], Mask[1], Mask[2], Mask[3]};
8430 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8432 if (NumV2Elements == 1) {
8434 std::find_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }) -
8437 // Compute the index adjacent to V2Index and in the same half by toggling
8439 int V2AdjIndex = V2Index ^ 1;
8441 if (Mask[V2AdjIndex] == -1) {
8442 // Handles all the cases where we have a single V2 element and an undef.
8443 // This will only ever happen in the high lanes because we commute the
8444 // vector otherwise.
8446 std::swap(LowV, HighV);
8447 NewMask[V2Index] -= 4;
8449 // Handle the case where the V2 element ends up adjacent to a V1 element.
8450 // To make this work, blend them together as the first step.
8451 int V1Index = V2AdjIndex;
8452 int BlendMask[4] = {Mask[V2Index] - 4, 0, Mask[V1Index], 0};
8453 V2 = DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
8454 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
8456 // Now proceed to reconstruct the final blend as we have the necessary
8457 // high or low half formed.
8464 NewMask[V1Index] = 2; // We put the V1 element in V2[2].
8465 NewMask[V2Index] = 0; // We shifted the V2 element into V2[0].
8467 } else if (NumV2Elements == 2) {
8468 if (Mask[0] < 4 && Mask[1] < 4) {
8469 // Handle the easy case where we have V1 in the low lanes and V2 in the
8473 } else if (Mask[2] < 4 && Mask[3] < 4) {
8474 // We also handle the reversed case because this utility may get called
8475 // when we detect a SHUFPS pattern but can't easily commute the shuffle to
8476 // arrange things in the right direction.
8482 // We have a mixture of V1 and V2 in both low and high lanes. Rather than
8483 // trying to place elements directly, just blend them and set up the final
8484 // shuffle to place them.
8486 // The first two blend mask elements are for V1, the second two are for
8488 int BlendMask[4] = {Mask[0] < 4 ? Mask[0] : Mask[1],
8489 Mask[2] < 4 ? Mask[2] : Mask[3],
8490 (Mask[0] >= 4 ? Mask[0] : Mask[1]) - 4,
8491 (Mask[2] >= 4 ? Mask[2] : Mask[3]) - 4};
8492 V1 = DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
8493 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
8495 // Now we do a normal shuffle of V1 by giving V1 as both operands to
8498 NewMask[0] = Mask[0] < 4 ? 0 : 2;
8499 NewMask[1] = Mask[0] < 4 ? 2 : 0;
8500 NewMask[2] = Mask[2] < 4 ? 1 : 3;
8501 NewMask[3] = Mask[2] < 4 ? 3 : 1;
8504 return DAG.getNode(X86ISD::SHUFP, DL, VT, LowV, HighV,
8505 getV4X86ShuffleImm8ForMask(NewMask, DL, DAG));
8508 /// \brief Lower 4-lane 32-bit floating point shuffles.
8510 /// Uses instructions exclusively from the floating point unit to minimize
8511 /// domain crossing penalties, as these are sufficient to implement all v4f32
8513 static SDValue lowerV4F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8514 const X86Subtarget *Subtarget,
8515 SelectionDAG &DAG) {
8517 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
8518 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8519 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8520 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8521 ArrayRef<int> Mask = SVOp->getMask();
8522 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8525 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8527 if (NumV2Elements == 0) {
8528 // Check for being able to broadcast a single element.
8529 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f32, V1,
8530 Mask, Subtarget, DAG))
8533 // Use even/odd duplicate instructions for masks that match their pattern.
8534 if (Subtarget->hasSSE3()) {
8535 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
8536 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v4f32, V1);
8537 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3}))
8538 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v4f32, V1);
8541 if (Subtarget->hasAVX()) {
8542 // If we have AVX, we can use VPERMILPS which will allow folding a load
8543 // into the shuffle.
8544 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f32, V1,
8545 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8548 // Otherwise, use a straight shuffle of a single input vector. We pass the
8549 // input vector to both operands to simulate this with a SHUFPS.
8550 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1,
8551 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8554 // There are special ways we can lower some single-element blends. However, we
8555 // have custom ways we can lower more complex single-element blends below that
8556 // we defer to if both this and BLENDPS fail to match, so restrict this to
8557 // when the V2 input is targeting element 0 of the mask -- that is the fast
8559 if (NumV2Elements == 1 && Mask[0] >= 4)
8560 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4f32, V1, V2,
8561 Mask, Subtarget, DAG))
8564 if (Subtarget->hasSSE41()) {
8565 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f32, V1, V2, Mask,
8569 // Use INSERTPS if we can complete the shuffle efficiently.
8570 if (SDValue V = lowerVectorShuffleAsInsertPS(Op, V1, V2, Mask, DAG))
8573 if (!isSingleSHUFPSMask(Mask))
8574 if (SDValue BlendPerm = lowerVectorShuffleAsBlendAndPermute(
8575 DL, MVT::v4f32, V1, V2, Mask, DAG))
8579 // Use dedicated unpack instructions for masks that match their pattern.
8581 lowerVectorShuffleWithUNPCK(DL, MVT::v4f32, Mask, V1, V2, DAG))
8584 // Otherwise fall back to a SHUFPS lowering strategy.
8585 return lowerVectorShuffleWithSHUFPS(DL, MVT::v4f32, Mask, V1, V2, DAG);
8588 /// \brief Lower 4-lane i32 vector shuffles.
8590 /// We try to handle these with integer-domain shuffles where we can, but for
8591 /// blends we use the floating point domain blend instructions.
8592 static SDValue lowerV4I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8593 const X86Subtarget *Subtarget,
8594 SelectionDAG &DAG) {
8596 assert(Op.getSimpleValueType() == MVT::v4i32 && "Bad shuffle type!");
8597 assert(V1.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8598 assert(V2.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8599 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8600 ArrayRef<int> Mask = SVOp->getMask();
8601 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8603 // Whenever we can lower this as a zext, that instruction is strictly faster
8604 // than any alternative. It also allows us to fold memory operands into the
8605 // shuffle in many cases.
8606 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v4i32, V1, V2,
8607 Mask, Subtarget, DAG))
8611 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8613 if (NumV2Elements == 0) {
8614 // Check for being able to broadcast a single element.
8615 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i32, V1,
8616 Mask, Subtarget, DAG))
8619 // Straight shuffle of a single input vector. For everything from SSE2
8620 // onward this has a single fast instruction with no scary immediates.
8621 // We coerce the shuffle pattern to be compatible with UNPCK instructions
8622 // but we aren't actually going to use the UNPCK instruction because doing
8623 // so prevents folding a load into this instruction or making a copy.
8624 const int UnpackLoMask[] = {0, 0, 1, 1};
8625 const int UnpackHiMask[] = {2, 2, 3, 3};
8626 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 1, 1}))
8627 Mask = UnpackLoMask;
8628 else if (isShuffleEquivalent(V1, V2, Mask, {2, 2, 3, 3}))
8629 Mask = UnpackHiMask;
8631 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
8632 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8635 // Try to use shift instructions.
8637 lowerVectorShuffleAsShift(DL, MVT::v4i32, V1, V2, Mask, DAG))
8640 // There are special ways we can lower some single-element blends.
8641 if (NumV2Elements == 1)
8642 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4i32, V1, V2,
8643 Mask, Subtarget, DAG))
8646 // We have different paths for blend lowering, but they all must use the
8647 // *exact* same predicate.
8648 bool IsBlendSupported = Subtarget->hasSSE41();
8649 if (IsBlendSupported)
8650 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i32, V1, V2, Mask,
8654 if (SDValue Masked =
8655 lowerVectorShuffleAsBitMask(DL, MVT::v4i32, V1, V2, Mask, DAG))
8658 // Use dedicated unpack instructions for masks that match their pattern.
8660 lowerVectorShuffleWithUNPCK(DL, MVT::v4i32, Mask, V1, V2, DAG))
8663 // Try to use byte rotation instructions.
8664 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8665 if (Subtarget->hasSSSE3())
8666 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8667 DL, MVT::v4i32, V1, V2, Mask, Subtarget, DAG))
8670 // If we have direct support for blends, we should lower by decomposing into
8671 // a permute. That will be faster than the domain cross.
8672 if (IsBlendSupported)
8673 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i32, V1, V2,
8676 // Try to lower by permuting the inputs into an unpack instruction.
8677 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(DL, MVT::v4i32, V1,
8681 // We implement this with SHUFPS because it can blend from two vectors.
8682 // Because we're going to eventually use SHUFPS, we use SHUFPS even to build
8683 // up the inputs, bypassing domain shift penalties that we would encur if we
8684 // directly used PSHUFD on Nehalem and older. For newer chips, this isn't
8686 return DAG.getBitcast(
8688 DAG.getVectorShuffle(MVT::v4f32, DL, DAG.getBitcast(MVT::v4f32, V1),
8689 DAG.getBitcast(MVT::v4f32, V2), Mask));
8692 /// \brief Lowering of single-input v8i16 shuffles is the cornerstone of SSE2
8693 /// shuffle lowering, and the most complex part.
8695 /// The lowering strategy is to try to form pairs of input lanes which are
8696 /// targeted at the same half of the final vector, and then use a dword shuffle
8697 /// to place them onto the right half, and finally unpack the paired lanes into
8698 /// their final position.
8700 /// The exact breakdown of how to form these dword pairs and align them on the
8701 /// correct sides is really tricky. See the comments within the function for
8702 /// more of the details.
8704 /// This code also handles repeated 128-bit lanes of v8i16 shuffles, but each
8705 /// lane must shuffle the *exact* same way. In fact, you must pass a v8 Mask to
8706 /// this routine for it to work correctly. To shuffle a 256-bit or 512-bit i16
8707 /// vector, form the analogous 128-bit 8-element Mask.
8708 static SDValue lowerV8I16GeneralSingleInputVectorShuffle(
8709 SDLoc DL, MVT VT, SDValue V, MutableArrayRef<int> Mask,
8710 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
8711 assert(VT.getVectorElementType() == MVT::i16 && "Bad input type!");
8712 MVT PSHUFDVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
8714 assert(Mask.size() == 8 && "Shuffle mask length doen't match!");
8715 MutableArrayRef<int> LoMask = Mask.slice(0, 4);
8716 MutableArrayRef<int> HiMask = Mask.slice(4, 4);
8718 SmallVector<int, 4> LoInputs;
8719 std::copy_if(LoMask.begin(), LoMask.end(), std::back_inserter(LoInputs),
8720 [](int M) { return M >= 0; });
8721 std::sort(LoInputs.begin(), LoInputs.end());
8722 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), LoInputs.end());
8723 SmallVector<int, 4> HiInputs;
8724 std::copy_if(HiMask.begin(), HiMask.end(), std::back_inserter(HiInputs),
8725 [](int M) { return M >= 0; });
8726 std::sort(HiInputs.begin(), HiInputs.end());
8727 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), HiInputs.end());
8729 std::lower_bound(LoInputs.begin(), LoInputs.end(), 4) - LoInputs.begin();
8730 int NumHToL = LoInputs.size() - NumLToL;
8732 std::lower_bound(HiInputs.begin(), HiInputs.end(), 4) - HiInputs.begin();
8733 int NumHToH = HiInputs.size() - NumLToH;
8734 MutableArrayRef<int> LToLInputs(LoInputs.data(), NumLToL);
8735 MutableArrayRef<int> LToHInputs(HiInputs.data(), NumLToH);
8736 MutableArrayRef<int> HToLInputs(LoInputs.data() + NumLToL, NumHToL);
8737 MutableArrayRef<int> HToHInputs(HiInputs.data() + NumLToH, NumHToH);
8739 // Simplify the 1-into-3 and 3-into-1 cases with a single pshufd. For all
8740 // such inputs we can swap two of the dwords across the half mark and end up
8741 // with <=2 inputs to each half in each half. Once there, we can fall through
8742 // to the generic code below. For example:
8744 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8745 // Mask: [0, 1, 2, 7, 4, 5, 6, 3] -----------------> [0, 1, 4, 7, 2, 3, 6, 5]
8747 // However in some very rare cases we have a 1-into-3 or 3-into-1 on one half
8748 // and an existing 2-into-2 on the other half. In this case we may have to
8749 // pre-shuffle the 2-into-2 half to avoid turning it into a 3-into-1 or
8750 // 1-into-3 which could cause us to cycle endlessly fixing each side in turn.
8751 // Fortunately, we don't have to handle anything but a 2-into-2 pattern
8752 // because any other situation (including a 3-into-1 or 1-into-3 in the other
8753 // half than the one we target for fixing) will be fixed when we re-enter this
8754 // path. We will also combine away any sequence of PSHUFD instructions that
8755 // result into a single instruction. Here is an example of the tricky case:
8757 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8758 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -THIS-IS-BAD!!!!-> [5, 7, 1, 0, 4, 7, 5, 3]
8760 // This now has a 1-into-3 in the high half! Instead, we do two shuffles:
8762 // Input: [a, b, c, d, e, f, g, h] PSHUFHW[0,2,1,3]-> [a, b, c, d, e, g, f, h]
8763 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -----------------> [3, 7, 1, 0, 2, 7, 3, 6]
8765 // Input: [a, b, c, d, e, g, f, h] -PSHUFD[0,2,1,3]-> [a, b, e, g, c, d, f, h]
8766 // Mask: [3, 7, 1, 0, 2, 7, 3, 6] -----------------> [5, 7, 1, 0, 4, 7, 5, 6]
8768 // The result is fine to be handled by the generic logic.
8769 auto balanceSides = [&](ArrayRef<int> AToAInputs, ArrayRef<int> BToAInputs,
8770 ArrayRef<int> BToBInputs, ArrayRef<int> AToBInputs,
8771 int AOffset, int BOffset) {
8772 assert((AToAInputs.size() == 3 || AToAInputs.size() == 1) &&
8773 "Must call this with A having 3 or 1 inputs from the A half.");
8774 assert((BToAInputs.size() == 1 || BToAInputs.size() == 3) &&
8775 "Must call this with B having 1 or 3 inputs from the B half.");
8776 assert(AToAInputs.size() + BToAInputs.size() == 4 &&
8777 "Must call this with either 3:1 or 1:3 inputs (summing to 4).");
8779 bool ThreeAInputs = AToAInputs.size() == 3;
8781 // Compute the index of dword with only one word among the three inputs in
8782 // a half by taking the sum of the half with three inputs and subtracting
8783 // the sum of the actual three inputs. The difference is the remaining
8786 int &TripleDWord = ThreeAInputs ? ADWord : BDWord;
8787 int &OneInputDWord = ThreeAInputs ? BDWord : ADWord;
8788 int TripleInputOffset = ThreeAInputs ? AOffset : BOffset;
8789 ArrayRef<int> TripleInputs = ThreeAInputs ? AToAInputs : BToAInputs;
8790 int OneInput = ThreeAInputs ? BToAInputs[0] : AToAInputs[0];
8791 int TripleInputSum = 0 + 1 + 2 + 3 + (4 * TripleInputOffset);
8792 int TripleNonInputIdx =
8793 TripleInputSum - std::accumulate(TripleInputs.begin(), TripleInputs.end(), 0);
8794 TripleDWord = TripleNonInputIdx / 2;
8796 // We use xor with one to compute the adjacent DWord to whichever one the
8798 OneInputDWord = (OneInput / 2) ^ 1;
8800 // Check for one tricky case: We're fixing a 3<-1 or a 1<-3 shuffle for AToA
8801 // and BToA inputs. If there is also such a problem with the BToB and AToB
8802 // inputs, we don't try to fix it necessarily -- we'll recurse and see it in
8803 // the next pass. However, if we have a 2<-2 in the BToB and AToB inputs, it
8804 // is essential that we don't *create* a 3<-1 as then we might oscillate.
8805 if (BToBInputs.size() == 2 && AToBInputs.size() == 2) {
8806 // Compute how many inputs will be flipped by swapping these DWords. We
8808 // to balance this to ensure we don't form a 3-1 shuffle in the other
8810 int NumFlippedAToBInputs =
8811 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord) +
8812 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord + 1);
8813 int NumFlippedBToBInputs =
8814 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord) +
8815 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord + 1);
8816 if ((NumFlippedAToBInputs == 1 &&
8817 (NumFlippedBToBInputs == 0 || NumFlippedBToBInputs == 2)) ||
8818 (NumFlippedBToBInputs == 1 &&
8819 (NumFlippedAToBInputs == 0 || NumFlippedAToBInputs == 2))) {
8820 // We choose whether to fix the A half or B half based on whether that
8821 // half has zero flipped inputs. At zero, we may not be able to fix it
8822 // with that half. We also bias towards fixing the B half because that
8823 // will more commonly be the high half, and we have to bias one way.
8824 auto FixFlippedInputs = [&V, &DL, &Mask, &DAG](int PinnedIdx, int DWord,
8825 ArrayRef<int> Inputs) {
8826 int FixIdx = PinnedIdx ^ 1; // The adjacent slot to the pinned slot.
8827 bool IsFixIdxInput = std::find(Inputs.begin(), Inputs.end(),
8828 PinnedIdx ^ 1) != Inputs.end();
8829 // Determine whether the free index is in the flipped dword or the
8830 // unflipped dword based on where the pinned index is. We use this bit
8831 // in an xor to conditionally select the adjacent dword.
8832 int FixFreeIdx = 2 * (DWord ^ (PinnedIdx / 2 == DWord));
8833 bool IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8834 FixFreeIdx) != Inputs.end();
8835 if (IsFixIdxInput == IsFixFreeIdxInput)
8837 IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8838 FixFreeIdx) != Inputs.end();
8839 assert(IsFixIdxInput != IsFixFreeIdxInput &&
8840 "We need to be changing the number of flipped inputs!");
8841 int PSHUFHalfMask[] = {0, 1, 2, 3};
8842 std::swap(PSHUFHalfMask[FixFreeIdx % 4], PSHUFHalfMask[FixIdx % 4]);
8843 V = DAG.getNode(FixIdx < 4 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW, DL,
8845 getV4X86ShuffleImm8ForMask(PSHUFHalfMask, DL, DAG));
8848 if (M != -1 && M == FixIdx)
8850 else if (M != -1 && M == FixFreeIdx)
8853 if (NumFlippedBToBInputs != 0) {
8855 BToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
8856 FixFlippedInputs(BPinnedIdx, BDWord, BToBInputs);
8858 assert(NumFlippedAToBInputs != 0 && "Impossible given predicates!");
8859 int APinnedIdx = ThreeAInputs ? TripleNonInputIdx : OneInput;
8860 FixFlippedInputs(APinnedIdx, ADWord, AToBInputs);
8865 int PSHUFDMask[] = {0, 1, 2, 3};
8866 PSHUFDMask[ADWord] = BDWord;
8867 PSHUFDMask[BDWord] = ADWord;
8870 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
8871 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
8873 // Adjust the mask to match the new locations of A and B.
8875 if (M != -1 && M/2 == ADWord)
8876 M = 2 * BDWord + M % 2;
8877 else if (M != -1 && M/2 == BDWord)
8878 M = 2 * ADWord + M % 2;
8880 // Recurse back into this routine to re-compute state now that this isn't
8881 // a 3 and 1 problem.
8882 return lowerV8I16GeneralSingleInputVectorShuffle(DL, VT, V, Mask, Subtarget,
8885 if ((NumLToL == 3 && NumHToL == 1) || (NumLToL == 1 && NumHToL == 3))
8886 return balanceSides(LToLInputs, HToLInputs, HToHInputs, LToHInputs, 0, 4);
8887 else if ((NumHToH == 3 && NumLToH == 1) || (NumHToH == 1 && NumLToH == 3))
8888 return balanceSides(HToHInputs, LToHInputs, LToLInputs, HToLInputs, 4, 0);
8890 // At this point there are at most two inputs to the low and high halves from
8891 // each half. That means the inputs can always be grouped into dwords and
8892 // those dwords can then be moved to the correct half with a dword shuffle.
8893 // We use at most one low and one high word shuffle to collect these paired
8894 // inputs into dwords, and finally a dword shuffle to place them.
8895 int PSHUFLMask[4] = {-1, -1, -1, -1};
8896 int PSHUFHMask[4] = {-1, -1, -1, -1};
8897 int PSHUFDMask[4] = {-1, -1, -1, -1};
8899 // First fix the masks for all the inputs that are staying in their
8900 // original halves. This will then dictate the targets of the cross-half
8902 auto fixInPlaceInputs =
8903 [&PSHUFDMask](ArrayRef<int> InPlaceInputs, ArrayRef<int> IncomingInputs,
8904 MutableArrayRef<int> SourceHalfMask,
8905 MutableArrayRef<int> HalfMask, int HalfOffset) {
8906 if (InPlaceInputs.empty())
8908 if (InPlaceInputs.size() == 1) {
8909 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8910 InPlaceInputs[0] - HalfOffset;
8911 PSHUFDMask[InPlaceInputs[0] / 2] = InPlaceInputs[0] / 2;
8914 if (IncomingInputs.empty()) {
8915 // Just fix all of the in place inputs.
8916 for (int Input : InPlaceInputs) {
8917 SourceHalfMask[Input - HalfOffset] = Input - HalfOffset;
8918 PSHUFDMask[Input / 2] = Input / 2;
8923 assert(InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!");
8924 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8925 InPlaceInputs[0] - HalfOffset;
8926 // Put the second input next to the first so that they are packed into
8927 // a dword. We find the adjacent index by toggling the low bit.
8928 int AdjIndex = InPlaceInputs[0] ^ 1;
8929 SourceHalfMask[AdjIndex - HalfOffset] = InPlaceInputs[1] - HalfOffset;
8930 std::replace(HalfMask.begin(), HalfMask.end(), InPlaceInputs[1], AdjIndex);
8931 PSHUFDMask[AdjIndex / 2] = AdjIndex / 2;
8933 fixInPlaceInputs(LToLInputs, HToLInputs, PSHUFLMask, LoMask, 0);
8934 fixInPlaceInputs(HToHInputs, LToHInputs, PSHUFHMask, HiMask, 4);
8936 // Now gather the cross-half inputs and place them into a free dword of
8937 // their target half.
8938 // FIXME: This operation could almost certainly be simplified dramatically to
8939 // look more like the 3-1 fixing operation.
8940 auto moveInputsToRightHalf = [&PSHUFDMask](
8941 MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs,
8942 MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask,
8943 MutableArrayRef<int> FinalSourceHalfMask, int SourceOffset,
8945 auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) {
8946 return SourceHalfMask[Word] != -1 && SourceHalfMask[Word] != Word;
8948 auto isDWordClobbered = [&isWordClobbered](ArrayRef<int> SourceHalfMask,
8950 int LowWord = Word & ~1;
8951 int HighWord = Word | 1;
8952 return isWordClobbered(SourceHalfMask, LowWord) ||
8953 isWordClobbered(SourceHalfMask, HighWord);
8956 if (IncomingInputs.empty())
8959 if (ExistingInputs.empty()) {
8960 // Map any dwords with inputs from them into the right half.
8961 for (int Input : IncomingInputs) {
8962 // If the source half mask maps over the inputs, turn those into
8963 // swaps and use the swapped lane.
8964 if (isWordClobbered(SourceHalfMask, Input - SourceOffset)) {
8965 if (SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == -1) {
8966 SourceHalfMask[SourceHalfMask[Input - SourceOffset]] =
8967 Input - SourceOffset;
8968 // We have to swap the uses in our half mask in one sweep.
8969 for (int &M : HalfMask)
8970 if (M == SourceHalfMask[Input - SourceOffset] + SourceOffset)
8972 else if (M == Input)
8973 M = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8975 assert(SourceHalfMask[SourceHalfMask[Input - SourceOffset]] ==
8976 Input - SourceOffset &&
8977 "Previous placement doesn't match!");
8979 // Note that this correctly re-maps both when we do a swap and when
8980 // we observe the other side of the swap above. We rely on that to
8981 // avoid swapping the members of the input list directly.
8982 Input = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8985 // Map the input's dword into the correct half.
8986 if (PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == -1)
8987 PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] = Input / 2;
8989 assert(PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] ==
8991 "Previous placement doesn't match!");
8994 // And just directly shift any other-half mask elements to be same-half
8995 // as we will have mirrored the dword containing the element into the
8996 // same position within that half.
8997 for (int &M : HalfMask)
8998 if (M >= SourceOffset && M < SourceOffset + 4) {
8999 M = M - SourceOffset + DestOffset;
9000 assert(M >= 0 && "This should never wrap below zero!");
9005 // Ensure we have the input in a viable dword of its current half. This
9006 // is particularly tricky because the original position may be clobbered
9007 // by inputs being moved and *staying* in that half.
9008 if (IncomingInputs.size() == 1) {
9009 if (isWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
9010 int InputFixed = std::find(std::begin(SourceHalfMask),
9011 std::end(SourceHalfMask), -1) -
9012 std::begin(SourceHalfMask) + SourceOffset;
9013 SourceHalfMask[InputFixed - SourceOffset] =
9014 IncomingInputs[0] - SourceOffset;
9015 std::replace(HalfMask.begin(), HalfMask.end(), IncomingInputs[0],
9017 IncomingInputs[0] = InputFixed;
9019 } else if (IncomingInputs.size() == 2) {
9020 if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 ||
9021 isDWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
9022 // We have two non-adjacent or clobbered inputs we need to extract from
9023 // the source half. To do this, we need to map them into some adjacent
9024 // dword slot in the source mask.
9025 int InputsFixed[2] = {IncomingInputs[0] - SourceOffset,
9026 IncomingInputs[1] - SourceOffset};
9028 // If there is a free slot in the source half mask adjacent to one of
9029 // the inputs, place the other input in it. We use (Index XOR 1) to
9030 // compute an adjacent index.
9031 if (!isWordClobbered(SourceHalfMask, InputsFixed[0]) &&
9032 SourceHalfMask[InputsFixed[0] ^ 1] == -1) {
9033 SourceHalfMask[InputsFixed[0]] = InputsFixed[0];
9034 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
9035 InputsFixed[1] = InputsFixed[0] ^ 1;
9036 } else if (!isWordClobbered(SourceHalfMask, InputsFixed[1]) &&
9037 SourceHalfMask[InputsFixed[1] ^ 1] == -1) {
9038 SourceHalfMask[InputsFixed[1]] = InputsFixed[1];
9039 SourceHalfMask[InputsFixed[1] ^ 1] = InputsFixed[0];
9040 InputsFixed[0] = InputsFixed[1] ^ 1;
9041 } else if (SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] == -1 &&
9042 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] == -1) {
9043 // The two inputs are in the same DWord but it is clobbered and the
9044 // adjacent DWord isn't used at all. Move both inputs to the free
9046 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] = InputsFixed[0];
9047 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] = InputsFixed[1];
9048 InputsFixed[0] = 2 * ((InputsFixed[0] / 2) ^ 1);
9049 InputsFixed[1] = 2 * ((InputsFixed[0] / 2) ^ 1) + 1;
9051 // The only way we hit this point is if there is no clobbering
9052 // (because there are no off-half inputs to this half) and there is no
9053 // free slot adjacent to one of the inputs. In this case, we have to
9054 // swap an input with a non-input.
9055 for (int i = 0; i < 4; ++i)
9056 assert((SourceHalfMask[i] == -1 || SourceHalfMask[i] == i) &&
9057 "We can't handle any clobbers here!");
9058 assert(InputsFixed[1] != (InputsFixed[0] ^ 1) &&
9059 "Cannot have adjacent inputs here!");
9061 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
9062 SourceHalfMask[InputsFixed[1]] = InputsFixed[0] ^ 1;
9064 // We also have to update the final source mask in this case because
9065 // it may need to undo the above swap.
9066 for (int &M : FinalSourceHalfMask)
9067 if (M == (InputsFixed[0] ^ 1) + SourceOffset)
9068 M = InputsFixed[1] + SourceOffset;
9069 else if (M == InputsFixed[1] + SourceOffset)
9070 M = (InputsFixed[0] ^ 1) + SourceOffset;
9072 InputsFixed[1] = InputsFixed[0] ^ 1;
9075 // Point everything at the fixed inputs.
9076 for (int &M : HalfMask)
9077 if (M == IncomingInputs[0])
9078 M = InputsFixed[0] + SourceOffset;
9079 else if (M == IncomingInputs[1])
9080 M = InputsFixed[1] + SourceOffset;
9082 IncomingInputs[0] = InputsFixed[0] + SourceOffset;
9083 IncomingInputs[1] = InputsFixed[1] + SourceOffset;
9086 llvm_unreachable("Unhandled input size!");
9089 // Now hoist the DWord down to the right half.
9090 int FreeDWord = (PSHUFDMask[DestOffset / 2] == -1 ? 0 : 1) + DestOffset / 2;
9091 assert(PSHUFDMask[FreeDWord] == -1 && "DWord not free");
9092 PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2;
9093 for (int &M : HalfMask)
9094 for (int Input : IncomingInputs)
9096 M = FreeDWord * 2 + Input % 2;
9098 moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask, HiMask,
9099 /*SourceOffset*/ 4, /*DestOffset*/ 0);
9100 moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask, LoMask,
9101 /*SourceOffset*/ 0, /*DestOffset*/ 4);
9103 // Now enact all the shuffles we've computed to move the inputs into their
9105 if (!isNoopShuffleMask(PSHUFLMask))
9106 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
9107 getV4X86ShuffleImm8ForMask(PSHUFLMask, DL, DAG));
9108 if (!isNoopShuffleMask(PSHUFHMask))
9109 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
9110 getV4X86ShuffleImm8ForMask(PSHUFHMask, DL, DAG));
9111 if (!isNoopShuffleMask(PSHUFDMask))
9114 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
9115 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
9117 // At this point, each half should contain all its inputs, and we can then
9118 // just shuffle them into their final position.
9119 assert(std::count_if(LoMask.begin(), LoMask.end(),
9120 [](int M) { return M >= 4; }) == 0 &&
9121 "Failed to lift all the high half inputs to the low mask!");
9122 assert(std::count_if(HiMask.begin(), HiMask.end(),
9123 [](int M) { return M >= 0 && M < 4; }) == 0 &&
9124 "Failed to lift all the low half inputs to the high mask!");
9126 // Do a half shuffle for the low mask.
9127 if (!isNoopShuffleMask(LoMask))
9128 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
9129 getV4X86ShuffleImm8ForMask(LoMask, DL, DAG));
9131 // Do a half shuffle with the high mask after shifting its values down.
9132 for (int &M : HiMask)
9135 if (!isNoopShuffleMask(HiMask))
9136 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
9137 getV4X86ShuffleImm8ForMask(HiMask, DL, DAG));
9142 /// \brief Helper to form a PSHUFB-based shuffle+blend.
9143 static SDValue lowerVectorShuffleAsPSHUFB(SDLoc DL, MVT VT, SDValue V1,
9144 SDValue V2, ArrayRef<int> Mask,
9145 SelectionDAG &DAG, bool &V1InUse,
9147 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
9153 int Size = Mask.size();
9154 int Scale = 16 / Size;
9155 for (int i = 0; i < 16; ++i) {
9156 if (Mask[i / Scale] == -1) {
9157 V1Mask[i] = V2Mask[i] = DAG.getUNDEF(MVT::i8);
9159 const int ZeroMask = 0x80;
9160 int V1Idx = Mask[i / Scale] < Size ? Mask[i / Scale] * Scale + i % Scale
9162 int V2Idx = Mask[i / Scale] < Size
9164 : (Mask[i / Scale] - Size) * Scale + i % Scale;
9165 if (Zeroable[i / Scale])
9166 V1Idx = V2Idx = ZeroMask;
9167 V1Mask[i] = DAG.getConstant(V1Idx, DL, MVT::i8);
9168 V2Mask[i] = DAG.getConstant(V2Idx, DL, MVT::i8);
9169 V1InUse |= (ZeroMask != V1Idx);
9170 V2InUse |= (ZeroMask != V2Idx);
9175 V1 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
9176 DAG.getBitcast(MVT::v16i8, V1),
9177 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V1Mask));
9179 V2 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
9180 DAG.getBitcast(MVT::v16i8, V2),
9181 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V2Mask));
9183 // If we need shuffled inputs from both, blend the two.
9185 if (V1InUse && V2InUse)
9186 V = DAG.getNode(ISD::OR, DL, MVT::v16i8, V1, V2);
9188 V = V1InUse ? V1 : V2;
9190 // Cast the result back to the correct type.
9191 return DAG.getBitcast(VT, V);
9194 /// \brief Generic lowering of 8-lane i16 shuffles.
9196 /// This handles both single-input shuffles and combined shuffle/blends with
9197 /// two inputs. The single input shuffles are immediately delegated to
9198 /// a dedicated lowering routine.
9200 /// The blends are lowered in one of three fundamental ways. If there are few
9201 /// enough inputs, it delegates to a basic UNPCK-based strategy. If the shuffle
9202 /// of the input is significantly cheaper when lowered as an interleaving of
9203 /// the two inputs, try to interleave them. Otherwise, blend the low and high
9204 /// halves of the inputs separately (making them have relatively few inputs)
9205 /// and then concatenate them.
9206 static SDValue lowerV8I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9207 const X86Subtarget *Subtarget,
9208 SelectionDAG &DAG) {
9210 assert(Op.getSimpleValueType() == MVT::v8i16 && "Bad shuffle type!");
9211 assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
9212 assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
9213 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9214 ArrayRef<int> OrigMask = SVOp->getMask();
9215 int MaskStorage[8] = {OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
9216 OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7]};
9217 MutableArrayRef<int> Mask(MaskStorage);
9219 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9221 // Whenever we can lower this as a zext, that instruction is strictly faster
9222 // than any alternative.
9223 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
9224 DL, MVT::v8i16, V1, V2, OrigMask, Subtarget, DAG))
9227 auto isV1 = [](int M) { return M >= 0 && M < 8; };
9229 auto isV2 = [](int M) { return M >= 8; };
9231 int NumV2Inputs = std::count_if(Mask.begin(), Mask.end(), isV2);
9233 if (NumV2Inputs == 0) {
9234 // Check for being able to broadcast a single element.
9235 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i16, V1,
9236 Mask, Subtarget, DAG))
9239 // Try to use shift instructions.
9241 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V1, Mask, DAG))
9244 // Use dedicated unpack instructions for masks that match their pattern.
9246 lowerVectorShuffleWithUNPCK(DL, MVT::v8i16, Mask, V1, V2, DAG))
9249 // Try to use byte rotation instructions.
9250 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(DL, MVT::v8i16, V1, V1,
9251 Mask, Subtarget, DAG))
9254 return lowerV8I16GeneralSingleInputVectorShuffle(DL, MVT::v8i16, V1, Mask,
9258 assert(std::any_of(Mask.begin(), Mask.end(), isV1) &&
9259 "All single-input shuffles should be canonicalized to be V1-input "
9262 // Try to use shift instructions.
9264 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V2, Mask, DAG))
9267 // See if we can use SSE4A Extraction / Insertion.
9268 if (Subtarget->hasSSE4A())
9269 if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v8i16, V1, V2, Mask, DAG))
9272 // There are special ways we can lower some single-element blends.
9273 if (NumV2Inputs == 1)
9274 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v8i16, V1, V2,
9275 Mask, Subtarget, DAG))
9278 // We have different paths for blend lowering, but they all must use the
9279 // *exact* same predicate.
9280 bool IsBlendSupported = Subtarget->hasSSE41();
9281 if (IsBlendSupported)
9282 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i16, V1, V2, Mask,
9286 if (SDValue Masked =
9287 lowerVectorShuffleAsBitMask(DL, MVT::v8i16, V1, V2, Mask, DAG))
9290 // Use dedicated unpack instructions for masks that match their pattern.
9292 lowerVectorShuffleWithUNPCK(DL, MVT::v8i16, Mask, V1, V2, DAG))
9295 // Try to use byte rotation instructions.
9296 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9297 DL, MVT::v8i16, V1, V2, Mask, Subtarget, DAG))
9300 if (SDValue BitBlend =
9301 lowerVectorShuffleAsBitBlend(DL, MVT::v8i16, V1, V2, Mask, DAG))
9304 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(DL, MVT::v8i16, V1,
9308 // If we can't directly blend but can use PSHUFB, that will be better as it
9309 // can both shuffle and set up the inefficient blend.
9310 if (!IsBlendSupported && Subtarget->hasSSSE3()) {
9311 bool V1InUse, V2InUse;
9312 return lowerVectorShuffleAsPSHUFB(DL, MVT::v8i16, V1, V2, Mask, DAG,
9316 // We can always bit-blend if we have to so the fallback strategy is to
9317 // decompose into single-input permutes and blends.
9318 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i16, V1, V2,
9322 /// \brief Check whether a compaction lowering can be done by dropping even
9323 /// elements and compute how many times even elements must be dropped.
9325 /// This handles shuffles which take every Nth element where N is a power of
9326 /// two. Example shuffle masks:
9328 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 0, 2, 4, 6, 8, 10, 12, 14
9329 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30
9330 /// N = 2: 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12
9331 /// N = 2: 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28
9332 /// N = 3: 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8
9333 /// N = 3: 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24
9335 /// Any of these lanes can of course be undef.
9337 /// This routine only supports N <= 3.
9338 /// FIXME: Evaluate whether either AVX or AVX-512 have any opportunities here
9341 /// \returns N above, or the number of times even elements must be dropped if
9342 /// there is such a number. Otherwise returns zero.
9343 static int canLowerByDroppingEvenElements(ArrayRef<int> Mask) {
9344 // Figure out whether we're looping over two inputs or just one.
9345 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9347 // The modulus for the shuffle vector entries is based on whether this is
9348 // a single input or not.
9349 int ShuffleModulus = Mask.size() * (IsSingleInput ? 1 : 2);
9350 assert(isPowerOf2_32((uint32_t)ShuffleModulus) &&
9351 "We should only be called with masks with a power-of-2 size!");
9353 uint64_t ModMask = (uint64_t)ShuffleModulus - 1;
9355 // We track whether the input is viable for all power-of-2 strides 2^1, 2^2,
9356 // and 2^3 simultaneously. This is because we may have ambiguity with
9357 // partially undef inputs.
9358 bool ViableForN[3] = {true, true, true};
9360 for (int i = 0, e = Mask.size(); i < e; ++i) {
9361 // Ignore undef lanes, we'll optimistically collapse them to the pattern we
9366 bool IsAnyViable = false;
9367 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
9368 if (ViableForN[j]) {
9371 // The shuffle mask must be equal to (i * 2^N) % M.
9372 if ((uint64_t)Mask[i] == (((uint64_t)i << N) & ModMask))
9375 ViableForN[j] = false;
9377 // Early exit if we exhaust the possible powers of two.
9382 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
9386 // Return 0 as there is no viable power of two.
9390 /// \brief Generic lowering of v16i8 shuffles.
9392 /// This is a hybrid strategy to lower v16i8 vectors. It first attempts to
9393 /// detect any complexity reducing interleaving. If that doesn't help, it uses
9394 /// UNPCK to spread the i8 elements across two i16-element vectors, and uses
9395 /// the existing lowering for v8i16 blends on each half, finally PACK-ing them
9397 static SDValue lowerV16I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9398 const X86Subtarget *Subtarget,
9399 SelectionDAG &DAG) {
9401 assert(Op.getSimpleValueType() == MVT::v16i8 && "Bad shuffle type!");
9402 assert(V1.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
9403 assert(V2.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
9404 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9405 ArrayRef<int> Mask = SVOp->getMask();
9406 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
9408 // Try to use shift instructions.
9410 lowerVectorShuffleAsShift(DL, MVT::v16i8, V1, V2, Mask, DAG))
9413 // Try to use byte rotation instructions.
9414 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9415 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
9418 // Try to use a zext lowering.
9419 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
9420 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
9423 // See if we can use SSE4A Extraction / Insertion.
9424 if (Subtarget->hasSSE4A())
9425 if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v16i8, V1, V2, Mask, DAG))
9429 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 16; });
9431 // For single-input shuffles, there are some nicer lowering tricks we can use.
9432 if (NumV2Elements == 0) {
9433 // Check for being able to broadcast a single element.
9434 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i8, V1,
9435 Mask, Subtarget, DAG))
9438 // Check whether we can widen this to an i16 shuffle by duplicating bytes.
9439 // Notably, this handles splat and partial-splat shuffles more efficiently.
9440 // However, it only makes sense if the pre-duplication shuffle simplifies
9441 // things significantly. Currently, this means we need to be able to
9442 // express the pre-duplication shuffle as an i16 shuffle.
9444 // FIXME: We should check for other patterns which can be widened into an
9445 // i16 shuffle as well.
9446 auto canWidenViaDuplication = [](ArrayRef<int> Mask) {
9447 for (int i = 0; i < 16; i += 2)
9448 if (Mask[i] != -1 && Mask[i + 1] != -1 && Mask[i] != Mask[i + 1])
9453 auto tryToWidenViaDuplication = [&]() -> SDValue {
9454 if (!canWidenViaDuplication(Mask))
9456 SmallVector<int, 4> LoInputs;
9457 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(LoInputs),
9458 [](int M) { return M >= 0 && M < 8; });
9459 std::sort(LoInputs.begin(), LoInputs.end());
9460 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()),
9462 SmallVector<int, 4> HiInputs;
9463 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(HiInputs),
9464 [](int M) { return M >= 8; });
9465 std::sort(HiInputs.begin(), HiInputs.end());
9466 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()),
9469 bool TargetLo = LoInputs.size() >= HiInputs.size();
9470 ArrayRef<int> InPlaceInputs = TargetLo ? LoInputs : HiInputs;
9471 ArrayRef<int> MovingInputs = TargetLo ? HiInputs : LoInputs;
9473 int PreDupI16Shuffle[] = {-1, -1, -1, -1, -1, -1, -1, -1};
9474 SmallDenseMap<int, int, 8> LaneMap;
9475 for (int I : InPlaceInputs) {
9476 PreDupI16Shuffle[I/2] = I/2;
9479 int j = TargetLo ? 0 : 4, je = j + 4;
9480 for (int i = 0, ie = MovingInputs.size(); i < ie; ++i) {
9481 // Check if j is already a shuffle of this input. This happens when
9482 // there are two adjacent bytes after we move the low one.
9483 if (PreDupI16Shuffle[j] != MovingInputs[i] / 2) {
9484 // If we haven't yet mapped the input, search for a slot into which
9486 while (j < je && PreDupI16Shuffle[j] != -1)
9490 // We can't place the inputs into a single half with a simple i16 shuffle, so bail.
9493 // Map this input with the i16 shuffle.
9494 PreDupI16Shuffle[j] = MovingInputs[i] / 2;
9497 // Update the lane map based on the mapping we ended up with.
9498 LaneMap[MovingInputs[i]] = 2 * j + MovingInputs[i] % 2;
9500 V1 = DAG.getBitcast(
9502 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
9503 DAG.getUNDEF(MVT::v8i16), PreDupI16Shuffle));
9505 // Unpack the bytes to form the i16s that will be shuffled into place.
9506 V1 = DAG.getNode(TargetLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
9507 MVT::v16i8, V1, V1);
9509 int PostDupI16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9510 for (int i = 0; i < 16; ++i)
9511 if (Mask[i] != -1) {
9512 int MappedMask = LaneMap[Mask[i]] - (TargetLo ? 0 : 8);
9513 assert(MappedMask < 8 && "Invalid v8 shuffle mask!");
9514 if (PostDupI16Shuffle[i / 2] == -1)
9515 PostDupI16Shuffle[i / 2] = MappedMask;
9517 assert(PostDupI16Shuffle[i / 2] == MappedMask &&
9518 "Conflicting entrties in the original shuffle!");
9520 return DAG.getBitcast(
9522 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
9523 DAG.getUNDEF(MVT::v8i16), PostDupI16Shuffle));
9525 if (SDValue V = tryToWidenViaDuplication())
9529 if (SDValue Masked =
9530 lowerVectorShuffleAsBitMask(DL, MVT::v16i8, V1, V2, Mask, DAG))
9533 // Use dedicated unpack instructions for masks that match their pattern.
9535 lowerVectorShuffleWithUNPCK(DL, MVT::v16i8, Mask, V1, V2, DAG))
9538 // Check for SSSE3 which lets us lower all v16i8 shuffles much more directly
9539 // with PSHUFB. It is important to do this before we attempt to generate any
9540 // blends but after all of the single-input lowerings. If the single input
9541 // lowerings can find an instruction sequence that is faster than a PSHUFB, we
9542 // want to preserve that and we can DAG combine any longer sequences into
9543 // a PSHUFB in the end. But once we start blending from multiple inputs,
9544 // the complexity of DAG combining bad patterns back into PSHUFB is too high,
9545 // and there are *very* few patterns that would actually be faster than the
9546 // PSHUFB approach because of its ability to zero lanes.
9548 // FIXME: The only exceptions to the above are blends which are exact
9549 // interleavings with direct instructions supporting them. We currently don't
9550 // handle those well here.
9551 if (Subtarget->hasSSSE3()) {
9552 bool V1InUse = false;
9553 bool V2InUse = false;
9555 SDValue PSHUFB = lowerVectorShuffleAsPSHUFB(DL, MVT::v16i8, V1, V2, Mask,
9556 DAG, V1InUse, V2InUse);
9558 // If both V1 and V2 are in use and we can use a direct blend or an unpack,
9559 // do so. This avoids using them to handle blends-with-zero which is
9560 // important as a single pshufb is significantly faster for that.
9561 if (V1InUse && V2InUse) {
9562 if (Subtarget->hasSSE41())
9563 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i8, V1, V2,
9564 Mask, Subtarget, DAG))
9567 // We can use an unpack to do the blending rather than an or in some
9568 // cases. Even though the or may be (very minorly) more efficient, we
9569 // preference this lowering because there are common cases where part of
9570 // the complexity of the shuffles goes away when we do the final blend as
9572 // FIXME: It might be worth trying to detect if the unpack-feeding
9573 // shuffles will both be pshufb, in which case we shouldn't bother with
9575 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(
9576 DL, MVT::v16i8, V1, V2, Mask, DAG))
9583 // There are special ways we can lower some single-element blends.
9584 if (NumV2Elements == 1)
9585 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v16i8, V1, V2,
9586 Mask, Subtarget, DAG))
9589 if (SDValue BitBlend =
9590 lowerVectorShuffleAsBitBlend(DL, MVT::v16i8, V1, V2, Mask, DAG))
9593 // Check whether a compaction lowering can be done. This handles shuffles
9594 // which take every Nth element for some even N. See the helper function for
9597 // We special case these as they can be particularly efficiently handled with
9598 // the PACKUSB instruction on x86 and they show up in common patterns of
9599 // rearranging bytes to truncate wide elements.
9600 if (int NumEvenDrops = canLowerByDroppingEvenElements(Mask)) {
9601 // NumEvenDrops is the power of two stride of the elements. Another way of
9602 // thinking about it is that we need to drop the even elements this many
9603 // times to get the original input.
9604 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9606 // First we need to zero all the dropped bytes.
9607 assert(NumEvenDrops <= 3 &&
9608 "No support for dropping even elements more than 3 times.");
9609 // We use the mask type to pick which bytes are preserved based on how many
9610 // elements are dropped.
9611 MVT MaskVTs[] = { MVT::v8i16, MVT::v4i32, MVT::v2i64 };
9612 SDValue ByteClearMask = DAG.getBitcast(
9613 MVT::v16i8, DAG.getConstant(0xFF, DL, MaskVTs[NumEvenDrops - 1]));
9614 V1 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V1, ByteClearMask);
9616 V2 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V2, ByteClearMask);
9618 // Now pack things back together.
9619 V1 = DAG.getBitcast(MVT::v8i16, V1);
9620 V2 = IsSingleInput ? V1 : DAG.getBitcast(MVT::v8i16, V2);
9621 SDValue Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, V1, V2);
9622 for (int i = 1; i < NumEvenDrops; ++i) {
9623 Result = DAG.getBitcast(MVT::v8i16, Result);
9624 Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, Result, Result);
9630 // Handle multi-input cases by blending single-input shuffles.
9631 if (NumV2Elements > 0)
9632 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v16i8, V1, V2,
9635 // The fallback path for single-input shuffles widens this into two v8i16
9636 // vectors with unpacks, shuffles those, and then pulls them back together
9640 int LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9641 int HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9642 for (int i = 0; i < 16; ++i)
9644 (i < 8 ? LoBlendMask[i] : HiBlendMask[i % 8]) = Mask[i];
9646 SDValue Zero = getZeroVector(MVT::v8i16, Subtarget, DAG, DL);
9648 SDValue VLoHalf, VHiHalf;
9649 // Check if any of the odd lanes in the v16i8 are used. If not, we can mask
9650 // them out and avoid using UNPCK{L,H} to extract the elements of V as
9652 if (std::none_of(std::begin(LoBlendMask), std::end(LoBlendMask),
9653 [](int M) { return M >= 0 && M % 2 == 1; }) &&
9654 std::none_of(std::begin(HiBlendMask), std::end(HiBlendMask),
9655 [](int M) { return M >= 0 && M % 2 == 1; })) {
9656 // Use a mask to drop the high bytes.
9657 VLoHalf = DAG.getBitcast(MVT::v8i16, V);
9658 VLoHalf = DAG.getNode(ISD::AND, DL, MVT::v8i16, VLoHalf,
9659 DAG.getConstant(0x00FF, DL, MVT::v8i16));
9661 // This will be a single vector shuffle instead of a blend so nuke VHiHalf.
9662 VHiHalf = DAG.getUNDEF(MVT::v8i16);
9664 // Squash the masks to point directly into VLoHalf.
9665 for (int &M : LoBlendMask)
9668 for (int &M : HiBlendMask)
9672 // Otherwise just unpack the low half of V into VLoHalf and the high half into
9673 // VHiHalf so that we can blend them as i16s.
9674 VLoHalf = DAG.getBitcast(
9675 MVT::v8i16, DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V, Zero));
9676 VHiHalf = DAG.getBitcast(
9677 MVT::v8i16, DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V, Zero));
9680 SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, LoBlendMask);
9681 SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, HiBlendMask);
9683 return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, LoV, HiV);
9686 /// \brief Dispatching routine to lower various 128-bit x86 vector shuffles.
9688 /// This routine breaks down the specific type of 128-bit shuffle and
9689 /// dispatches to the lowering routines accordingly.
9690 static SDValue lower128BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9691 MVT VT, const X86Subtarget *Subtarget,
9692 SelectionDAG &DAG) {
9693 switch (VT.SimpleTy) {
9695 return lowerV2I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9697 return lowerV2F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9699 return lowerV4I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9701 return lowerV4F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9703 return lowerV8I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
9705 return lowerV16I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
9708 llvm_unreachable("Unimplemented!");
9712 /// \brief Helper function to test whether a shuffle mask could be
9713 /// simplified by widening the elements being shuffled.
9715 /// Appends the mask for wider elements in WidenedMask if valid. Otherwise
9716 /// leaves it in an unspecified state.
9718 /// NOTE: This must handle normal vector shuffle masks and *target* vector
9719 /// shuffle masks. The latter have the special property of a '-2' representing
9720 /// a zero-ed lane of a vector.
9721 static bool canWidenShuffleElements(ArrayRef<int> Mask,
9722 SmallVectorImpl<int> &WidenedMask) {
9723 for (int i = 0, Size = Mask.size(); i < Size; i += 2) {
9724 // If both elements are undef, its trivial.
9725 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] == SM_SentinelUndef) {
9726 WidenedMask.push_back(SM_SentinelUndef);
9730 // Check for an undef mask and a mask value properly aligned to fit with
9731 // a pair of values. If we find such a case, use the non-undef mask's value.
9732 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] >= 0 && Mask[i + 1] % 2 == 1) {
9733 WidenedMask.push_back(Mask[i + 1] / 2);
9736 if (Mask[i + 1] == SM_SentinelUndef && Mask[i] >= 0 && Mask[i] % 2 == 0) {
9737 WidenedMask.push_back(Mask[i] / 2);
9741 // When zeroing, we need to spread the zeroing across both lanes to widen.
9742 if (Mask[i] == SM_SentinelZero || Mask[i + 1] == SM_SentinelZero) {
9743 if ((Mask[i] == SM_SentinelZero || Mask[i] == SM_SentinelUndef) &&
9744 (Mask[i + 1] == SM_SentinelZero || Mask[i + 1] == SM_SentinelUndef)) {
9745 WidenedMask.push_back(SM_SentinelZero);
9751 // Finally check if the two mask values are adjacent and aligned with
9753 if (Mask[i] != SM_SentinelUndef && Mask[i] % 2 == 0 && Mask[i] + 1 == Mask[i + 1]) {
9754 WidenedMask.push_back(Mask[i] / 2);
9758 // Otherwise we can't safely widen the elements used in this shuffle.
9761 assert(WidenedMask.size() == Mask.size() / 2 &&
9762 "Incorrect size of mask after widening the elements!");
9767 /// \brief Generic routine to split vector shuffle into half-sized shuffles.
9769 /// This routine just extracts two subvectors, shuffles them independently, and
9770 /// then concatenates them back together. This should work effectively with all
9771 /// AVX vector shuffle types.
9772 static SDValue splitAndLowerVectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9773 SDValue V2, ArrayRef<int> Mask,
9774 SelectionDAG &DAG) {
9775 assert(VT.getSizeInBits() >= 256 &&
9776 "Only for 256-bit or wider vector shuffles!");
9777 assert(V1.getSimpleValueType() == VT && "Bad operand type!");
9778 assert(V2.getSimpleValueType() == VT && "Bad operand type!");
9780 ArrayRef<int> LoMask = Mask.slice(0, Mask.size() / 2);
9781 ArrayRef<int> HiMask = Mask.slice(Mask.size() / 2);
9783 int NumElements = VT.getVectorNumElements();
9784 int SplitNumElements = NumElements / 2;
9785 MVT ScalarVT = VT.getVectorElementType();
9786 MVT SplitVT = MVT::getVectorVT(ScalarVT, NumElements / 2);
9788 // Rather than splitting build-vectors, just build two narrower build
9789 // vectors. This helps shuffling with splats and zeros.
9790 auto SplitVector = [&](SDValue V) {
9791 while (V.getOpcode() == ISD::BITCAST)
9792 V = V->getOperand(0);
9794 MVT OrigVT = V.getSimpleValueType();
9795 int OrigNumElements = OrigVT.getVectorNumElements();
9796 int OrigSplitNumElements = OrigNumElements / 2;
9797 MVT OrigScalarVT = OrigVT.getVectorElementType();
9798 MVT OrigSplitVT = MVT::getVectorVT(OrigScalarVT, OrigNumElements / 2);
9802 auto *BV = dyn_cast<BuildVectorSDNode>(V);
9804 LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9805 DAG.getIntPtrConstant(0, DL));
9806 HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9807 DAG.getIntPtrConstant(OrigSplitNumElements, DL));
9810 SmallVector<SDValue, 16> LoOps, HiOps;
9811 for (int i = 0; i < OrigSplitNumElements; ++i) {
9812 LoOps.push_back(BV->getOperand(i));
9813 HiOps.push_back(BV->getOperand(i + OrigSplitNumElements));
9815 LoV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, LoOps);
9816 HiV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, HiOps);
9818 return std::make_pair(DAG.getBitcast(SplitVT, LoV),
9819 DAG.getBitcast(SplitVT, HiV));
9822 SDValue LoV1, HiV1, LoV2, HiV2;
9823 std::tie(LoV1, HiV1) = SplitVector(V1);
9824 std::tie(LoV2, HiV2) = SplitVector(V2);
9826 // Now create two 4-way blends of these half-width vectors.
9827 auto HalfBlend = [&](ArrayRef<int> HalfMask) {
9828 bool UseLoV1 = false, UseHiV1 = false, UseLoV2 = false, UseHiV2 = false;
9829 SmallVector<int, 32> V1BlendMask, V2BlendMask, BlendMask;
9830 for (int i = 0; i < SplitNumElements; ++i) {
9831 int M = HalfMask[i];
9832 if (M >= NumElements) {
9833 if (M >= NumElements + SplitNumElements)
9837 V2BlendMask.push_back(M - NumElements);
9838 V1BlendMask.push_back(-1);
9839 BlendMask.push_back(SplitNumElements + i);
9840 } else if (M >= 0) {
9841 if (M >= SplitNumElements)
9845 V2BlendMask.push_back(-1);
9846 V1BlendMask.push_back(M);
9847 BlendMask.push_back(i);
9849 V2BlendMask.push_back(-1);
9850 V1BlendMask.push_back(-1);
9851 BlendMask.push_back(-1);
9855 // Because the lowering happens after all combining takes place, we need to
9856 // manually combine these blend masks as much as possible so that we create
9857 // a minimal number of high-level vector shuffle nodes.
9859 // First try just blending the halves of V1 or V2.
9860 if (!UseLoV1 && !UseHiV1 && !UseLoV2 && !UseHiV2)
9861 return DAG.getUNDEF(SplitVT);
9862 if (!UseLoV2 && !UseHiV2)
9863 return DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9864 if (!UseLoV1 && !UseHiV1)
9865 return DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9867 SDValue V1Blend, V2Blend;
9868 if (UseLoV1 && UseHiV1) {
9870 DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9872 // We only use half of V1 so map the usage down into the final blend mask.
9873 V1Blend = UseLoV1 ? LoV1 : HiV1;
9874 for (int i = 0; i < SplitNumElements; ++i)
9875 if (BlendMask[i] >= 0 && BlendMask[i] < SplitNumElements)
9876 BlendMask[i] = V1BlendMask[i] - (UseLoV1 ? 0 : SplitNumElements);
9878 if (UseLoV2 && UseHiV2) {
9880 DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9882 // We only use half of V2 so map the usage down into the final blend mask.
9883 V2Blend = UseLoV2 ? LoV2 : HiV2;
9884 for (int i = 0; i < SplitNumElements; ++i)
9885 if (BlendMask[i] >= SplitNumElements)
9886 BlendMask[i] = V2BlendMask[i] + (UseLoV2 ? SplitNumElements : 0);
9888 return DAG.getVectorShuffle(SplitVT, DL, V1Blend, V2Blend, BlendMask);
9890 SDValue Lo = HalfBlend(LoMask);
9891 SDValue Hi = HalfBlend(HiMask);
9892 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
9895 /// \brief Either split a vector in halves or decompose the shuffles and the
9898 /// This is provided as a good fallback for many lowerings of non-single-input
9899 /// shuffles with more than one 128-bit lane. In those cases, we want to select
9900 /// between splitting the shuffle into 128-bit components and stitching those
9901 /// back together vs. extracting the single-input shuffles and blending those
9903 static SDValue lowerVectorShuffleAsSplitOrBlend(SDLoc DL, MVT VT, SDValue V1,
9904 SDValue V2, ArrayRef<int> Mask,
9905 SelectionDAG &DAG) {
9906 assert(!isSingleInputShuffleMask(Mask) && "This routine must not be used to "
9907 "lower single-input shuffles as it "
9908 "could then recurse on itself.");
9909 int Size = Mask.size();
9911 // If this can be modeled as a broadcast of two elements followed by a blend,
9912 // prefer that lowering. This is especially important because broadcasts can
9913 // often fold with memory operands.
9914 auto DoBothBroadcast = [&] {
9915 int V1BroadcastIdx = -1, V2BroadcastIdx = -1;
9918 if (V2BroadcastIdx == -1)
9919 V2BroadcastIdx = M - Size;
9920 else if (M - Size != V2BroadcastIdx)
9922 } else if (M >= 0) {
9923 if (V1BroadcastIdx == -1)
9925 else if (M != V1BroadcastIdx)
9930 if (DoBothBroadcast())
9931 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask,
9934 // If the inputs all stem from a single 128-bit lane of each input, then we
9935 // split them rather than blending because the split will decompose to
9936 // unusually few instructions.
9937 int LaneCount = VT.getSizeInBits() / 128;
9938 int LaneSize = Size / LaneCount;
9939 SmallBitVector LaneInputs[2];
9940 LaneInputs[0].resize(LaneCount, false);
9941 LaneInputs[1].resize(LaneCount, false);
9942 for (int i = 0; i < Size; ++i)
9944 LaneInputs[Mask[i] / Size][(Mask[i] % Size) / LaneSize] = true;
9945 if (LaneInputs[0].count() <= 1 && LaneInputs[1].count() <= 1)
9946 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9948 // Otherwise, just fall back to decomposed shuffles and a blend. This requires
9949 // that the decomposed single-input shuffles don't end up here.
9950 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
9953 /// \brief Lower a vector shuffle crossing multiple 128-bit lanes as
9954 /// a permutation and blend of those lanes.
9956 /// This essentially blends the out-of-lane inputs to each lane into the lane
9957 /// from a permuted copy of the vector. This lowering strategy results in four
9958 /// instructions in the worst case for a single-input cross lane shuffle which
9959 /// is lower than any other fully general cross-lane shuffle strategy I'm aware
9960 /// of. Special cases for each particular shuffle pattern should be handled
9961 /// prior to trying this lowering.
9962 static SDValue lowerVectorShuffleAsLanePermuteAndBlend(SDLoc DL, MVT VT,
9963 SDValue V1, SDValue V2,
9965 SelectionDAG &DAG) {
9966 // FIXME: This should probably be generalized for 512-bit vectors as well.
9967 assert(VT.is256BitVector() && "Only for 256-bit vector shuffles!");
9968 int LaneSize = Mask.size() / 2;
9970 // If there are only inputs from one 128-bit lane, splitting will in fact be
9971 // less expensive. The flags track whether the given lane contains an element
9972 // that crosses to another lane.
9973 bool LaneCrossing[2] = {false, false};
9974 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9975 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
9976 LaneCrossing[(Mask[i] % Size) / LaneSize] = true;
9977 if (!LaneCrossing[0] || !LaneCrossing[1])
9978 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9980 if (isSingleInputShuffleMask(Mask)) {
9981 SmallVector<int, 32> FlippedBlendMask;
9982 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9983 FlippedBlendMask.push_back(
9984 Mask[i] < 0 ? -1 : (((Mask[i] % Size) / LaneSize == i / LaneSize)
9986 : Mask[i] % LaneSize +
9987 (i / LaneSize) * LaneSize + Size));
9989 // Flip the vector, and blend the results which should now be in-lane. The
9990 // VPERM2X128 mask uses the low 2 bits for the low source and bits 4 and
9991 // 5 for the high source. The value 3 selects the high half of source 2 and
9992 // the value 2 selects the low half of source 2. We only use source 2 to
9993 // allow folding it into a memory operand.
9994 unsigned PERMMask = 3 | 2 << 4;
9995 SDValue Flipped = DAG.getNode(X86ISD::VPERM2X128, DL, VT, DAG.getUNDEF(VT),
9996 V1, DAG.getConstant(PERMMask, DL, MVT::i8));
9997 return DAG.getVectorShuffle(VT, DL, V1, Flipped, FlippedBlendMask);
10000 // This now reduces to two single-input shuffles of V1 and V2 which at worst
10001 // will be handled by the above logic and a blend of the results, much like
10002 // other patterns in AVX.
10003 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
10006 /// \brief Handle lowering 2-lane 128-bit shuffles.
10007 static SDValue lowerV2X128VectorShuffle(SDLoc DL, MVT VT, SDValue V1,
10008 SDValue V2, ArrayRef<int> Mask,
10009 const X86Subtarget *Subtarget,
10010 SelectionDAG &DAG) {
10011 // TODO: If minimizing size and one of the inputs is a zero vector and the
10012 // the zero vector has only one use, we could use a VPERM2X128 to save the
10013 // instruction bytes needed to explicitly generate the zero vector.
10015 // Blends are faster and handle all the non-lane-crossing cases.
10016 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, VT, V1, V2, Mask,
10020 bool IsV1Zero = ISD::isBuildVectorAllZeros(V1.getNode());
10021 bool IsV2Zero = ISD::isBuildVectorAllZeros(V2.getNode());
10023 // If either input operand is a zero vector, use VPERM2X128 because its mask
10024 // allows us to replace the zero input with an implicit zero.
10025 if (!IsV1Zero && !IsV2Zero) {
10026 // Check for patterns which can be matched with a single insert of a 128-bit
10028 bool OnlyUsesV1 = isShuffleEquivalent(V1, V2, Mask, {0, 1, 0, 1});
10029 if (OnlyUsesV1 || isShuffleEquivalent(V1, V2, Mask, {0, 1, 4, 5})) {
10030 MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(),
10031 VT.getVectorNumElements() / 2);
10032 SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
10033 DAG.getIntPtrConstant(0, DL));
10034 SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT,
10035 OnlyUsesV1 ? V1 : V2,
10036 DAG.getIntPtrConstant(0, DL));
10037 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
10041 // Otherwise form a 128-bit permutation. After accounting for undefs,
10042 // convert the 64-bit shuffle mask selection values into 128-bit
10043 // selection bits by dividing the indexes by 2 and shifting into positions
10044 // defined by a vperm2*128 instruction's immediate control byte.
10046 // The immediate permute control byte looks like this:
10047 // [1:0] - select 128 bits from sources for low half of destination
10049 // [3] - zero low half of destination
10050 // [5:4] - select 128 bits from sources for high half of destination
10052 // [7] - zero high half of destination
10054 int MaskLO = Mask[0];
10055 if (MaskLO == SM_SentinelUndef)
10056 MaskLO = Mask[1] == SM_SentinelUndef ? 0 : Mask[1];
10058 int MaskHI = Mask[2];
10059 if (MaskHI == SM_SentinelUndef)
10060 MaskHI = Mask[3] == SM_SentinelUndef ? 0 : Mask[3];
10062 unsigned PermMask = MaskLO / 2 | (MaskHI / 2) << 4;
10064 // If either input is a zero vector, replace it with an undef input.
10065 // Shuffle mask values < 4 are selecting elements of V1.
10066 // Shuffle mask values >= 4 are selecting elements of V2.
10067 // Adjust each half of the permute mask by clearing the half that was
10068 // selecting the zero vector and setting the zero mask bit.
10070 V1 = DAG.getUNDEF(VT);
10072 PermMask = (PermMask & 0xf0) | 0x08;
10074 PermMask = (PermMask & 0x0f) | 0x80;
10077 V2 = DAG.getUNDEF(VT);
10079 PermMask = (PermMask & 0xf0) | 0x08;
10081 PermMask = (PermMask & 0x0f) | 0x80;
10084 return DAG.getNode(X86ISD::VPERM2X128, DL, VT, V1, V2,
10085 DAG.getConstant(PermMask, DL, MVT::i8));
10088 /// \brief Lower a vector shuffle by first fixing the 128-bit lanes and then
10089 /// shuffling each lane.
10091 /// This will only succeed when the result of fixing the 128-bit lanes results
10092 /// in a single-input non-lane-crossing shuffle with a repeating shuffle mask in
10093 /// each 128-bit lanes. This handles many cases where we can quickly blend away
10094 /// the lane crosses early and then use simpler shuffles within each lane.
10096 /// FIXME: It might be worthwhile at some point to support this without
10097 /// requiring the 128-bit lane-relative shuffles to be repeating, but currently
10098 /// in x86 only floating point has interesting non-repeating shuffles, and even
10099 /// those are still *marginally* more expensive.
10100 static SDValue lowerVectorShuffleByMerging128BitLanes(
10101 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
10102 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
10103 assert(!isSingleInputShuffleMask(Mask) &&
10104 "This is only useful with multiple inputs.");
10106 int Size = Mask.size();
10107 int LaneSize = 128 / VT.getScalarSizeInBits();
10108 int NumLanes = Size / LaneSize;
10109 assert(NumLanes > 1 && "Only handles 256-bit and wider shuffles.");
10111 // See if we can build a hypothetical 128-bit lane-fixing shuffle mask. Also
10112 // check whether the in-128-bit lane shuffles share a repeating pattern.
10113 SmallVector<int, 4> Lanes;
10114 Lanes.resize(NumLanes, -1);
10115 SmallVector<int, 4> InLaneMask;
10116 InLaneMask.resize(LaneSize, -1);
10117 for (int i = 0; i < Size; ++i) {
10121 int j = i / LaneSize;
10123 if (Lanes[j] < 0) {
10124 // First entry we've seen for this lane.
10125 Lanes[j] = Mask[i] / LaneSize;
10126 } else if (Lanes[j] != Mask[i] / LaneSize) {
10127 // This doesn't match the lane selected previously!
10131 // Check that within each lane we have a consistent shuffle mask.
10132 int k = i % LaneSize;
10133 if (InLaneMask[k] < 0) {
10134 InLaneMask[k] = Mask[i] % LaneSize;
10135 } else if (InLaneMask[k] != Mask[i] % LaneSize) {
10136 // This doesn't fit a repeating in-lane mask.
10141 // First shuffle the lanes into place.
10142 MVT LaneVT = MVT::getVectorVT(VT.isFloatingPoint() ? MVT::f64 : MVT::i64,
10143 VT.getSizeInBits() / 64);
10144 SmallVector<int, 8> LaneMask;
10145 LaneMask.resize(NumLanes * 2, -1);
10146 for (int i = 0; i < NumLanes; ++i)
10147 if (Lanes[i] >= 0) {
10148 LaneMask[2 * i + 0] = 2*Lanes[i] + 0;
10149 LaneMask[2 * i + 1] = 2*Lanes[i] + 1;
10152 V1 = DAG.getBitcast(LaneVT, V1);
10153 V2 = DAG.getBitcast(LaneVT, V2);
10154 SDValue LaneShuffle = DAG.getVectorShuffle(LaneVT, DL, V1, V2, LaneMask);
10156 // Cast it back to the type we actually want.
10157 LaneShuffle = DAG.getBitcast(VT, LaneShuffle);
10159 // Now do a simple shuffle that isn't lane crossing.
10160 SmallVector<int, 8> NewMask;
10161 NewMask.resize(Size, -1);
10162 for (int i = 0; i < Size; ++i)
10164 NewMask[i] = (i / LaneSize) * LaneSize + Mask[i] % LaneSize;
10165 assert(!is128BitLaneCrossingShuffleMask(VT, NewMask) &&
10166 "Must not introduce lane crosses at this point!");
10168 return DAG.getVectorShuffle(VT, DL, LaneShuffle, DAG.getUNDEF(VT), NewMask);
10171 /// \brief Test whether the specified input (0 or 1) is in-place blended by the
10174 /// This returns true if the elements from a particular input are already in the
10175 /// slot required by the given mask and require no permutation.
10176 static bool isShuffleMaskInputInPlace(int Input, ArrayRef<int> Mask) {
10177 assert((Input == 0 || Input == 1) && "Only two inputs to shuffles.");
10178 int Size = Mask.size();
10179 for (int i = 0; i < Size; ++i)
10180 if (Mask[i] >= 0 && Mask[i] / Size == Input && Mask[i] % Size != i)
10186 static SDValue lowerVectorShuffleWithSHUFPD(SDLoc DL, MVT VT,
10187 ArrayRef<int> Mask, SDValue V1,
10188 SDValue V2, SelectionDAG &DAG) {
10190 // Mask for V8F64: 0/1, 8/9, 2/3, 10/11, 4/5, ..
10191 // Mask for V4F64; 0/1, 4/5, 2/3, 6/7..
10192 assert(VT.getScalarSizeInBits() == 64 && "Unexpected data type for VSHUFPD");
10193 int NumElts = VT.getVectorNumElements();
10194 bool ShufpdMask = true;
10195 bool CommutableMask = true;
10196 unsigned Immediate = 0;
10197 for (int i = 0; i < NumElts; ++i) {
10200 int Val = (i & 6) + NumElts * (i & 1);
10201 int CommutVal = (i & 0xe) + NumElts * ((i & 1)^1);
10202 if (Mask[i] < Val || Mask[i] > Val + 1)
10203 ShufpdMask = false;
10204 if (Mask[i] < CommutVal || Mask[i] > CommutVal + 1)
10205 CommutableMask = false;
10206 Immediate |= (Mask[i] % 2) << i;
10209 return DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
10210 DAG.getConstant(Immediate, DL, MVT::i8));
10211 if (CommutableMask)
10212 return DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
10213 DAG.getConstant(Immediate, DL, MVT::i8));
10217 /// \brief Handle lowering of 4-lane 64-bit floating point shuffles.
10219 /// Also ends up handling lowering of 4-lane 64-bit integer shuffles when AVX2
10220 /// isn't available.
10221 static SDValue lowerV4F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10222 const X86Subtarget *Subtarget,
10223 SelectionDAG &DAG) {
10225 assert(V1.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
10226 assert(V2.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
10227 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10228 ArrayRef<int> Mask = SVOp->getMask();
10229 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
10231 SmallVector<int, 4> WidenedMask;
10232 if (canWidenShuffleElements(Mask, WidenedMask))
10233 return lowerV2X128VectorShuffle(DL, MVT::v4f64, V1, V2, Mask, Subtarget,
10236 if (isSingleInputShuffleMask(Mask)) {
10237 // Check for being able to broadcast a single element.
10238 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f64, V1,
10239 Mask, Subtarget, DAG))
10242 // Use low duplicate instructions for masks that match their pattern.
10243 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
10244 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v4f64, V1);
10246 if (!is128BitLaneCrossingShuffleMask(MVT::v4f64, Mask)) {
10247 // Non-half-crossing single input shuffles can be lowerid with an
10248 // interleaved permutation.
10249 unsigned VPERMILPMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1) |
10250 ((Mask[2] == 3) << 2) | ((Mask[3] == 3) << 3);
10251 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f64, V1,
10252 DAG.getConstant(VPERMILPMask, DL, MVT::i8));
10255 // With AVX2 we have direct support for this permutation.
10256 if (Subtarget->hasAVX2())
10257 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4f64, V1,
10258 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
10260 // Otherwise, fall back.
10261 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v4f64, V1, V2, Mask,
10265 // Use dedicated unpack instructions for masks that match their pattern.
10267 lowerVectorShuffleWithUNPCK(DL, MVT::v4f64, Mask, V1, V2, DAG))
10270 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f64, V1, V2, Mask,
10274 // Check if the blend happens to exactly fit that of SHUFPD.
10276 lowerVectorShuffleWithSHUFPD(DL, MVT::v4f64, Mask, V1, V2, DAG))
10279 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10280 // shuffle. However, if we have AVX2 and either inputs are already in place,
10281 // we will be able to shuffle even across lanes the other input in a single
10282 // instruction so skip this pattern.
10283 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
10284 isShuffleMaskInputInPlace(1, Mask))))
10285 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10286 DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG))
10289 // If we have AVX2 then we always want to lower with a blend because an v4 we
10290 // can fully permute the elements.
10291 if (Subtarget->hasAVX2())
10292 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4f64, V1, V2,
10295 // Otherwise fall back on generic lowering.
10296 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v4f64, V1, V2, Mask, DAG);
10299 /// \brief Handle lowering of 4-lane 64-bit integer shuffles.
10301 /// This routine is only called when we have AVX2 and thus a reasonable
10302 /// instruction set for v4i64 shuffling..
10303 static SDValue lowerV4I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10304 const X86Subtarget *Subtarget,
10305 SelectionDAG &DAG) {
10307 assert(V1.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
10308 assert(V2.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
10309 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10310 ArrayRef<int> Mask = SVOp->getMask();
10311 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
10312 assert(Subtarget->hasAVX2() && "We can only lower v4i64 with AVX2!");
10314 SmallVector<int, 4> WidenedMask;
10315 if (canWidenShuffleElements(Mask, WidenedMask))
10316 return lowerV2X128VectorShuffle(DL, MVT::v4i64, V1, V2, Mask, Subtarget,
10319 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i64, V1, V2, Mask,
10323 // Check for being able to broadcast a single element.
10324 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i64, V1,
10325 Mask, Subtarget, DAG))
10328 // When the shuffle is mirrored between the 128-bit lanes of the unit, we can
10329 // use lower latency instructions that will operate on both 128-bit lanes.
10330 SmallVector<int, 2> RepeatedMask;
10331 if (is128BitLaneRepeatedShuffleMask(MVT::v4i64, Mask, RepeatedMask)) {
10332 if (isSingleInputShuffleMask(Mask)) {
10333 int PSHUFDMask[] = {-1, -1, -1, -1};
10334 for (int i = 0; i < 2; ++i)
10335 if (RepeatedMask[i] >= 0) {
10336 PSHUFDMask[2 * i] = 2 * RepeatedMask[i];
10337 PSHUFDMask[2 * i + 1] = 2 * RepeatedMask[i] + 1;
10339 return DAG.getBitcast(
10341 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32,
10342 DAG.getBitcast(MVT::v8i32, V1),
10343 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
10347 // AVX2 provides a direct instruction for permuting a single input across
10349 if (isSingleInputShuffleMask(Mask))
10350 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4i64, V1,
10351 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
10353 // Try to use shift instructions.
10354 if (SDValue Shift =
10355 lowerVectorShuffleAsShift(DL, MVT::v4i64, V1, V2, Mask, DAG))
10358 // Use dedicated unpack instructions for masks that match their pattern.
10360 lowerVectorShuffleWithUNPCK(DL, MVT::v4i64, Mask, V1, V2, DAG))
10363 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10364 // shuffle. However, if we have AVX2 and either inputs are already in place,
10365 // we will be able to shuffle even across lanes the other input in a single
10366 // instruction so skip this pattern.
10367 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
10368 isShuffleMaskInputInPlace(1, Mask))))
10369 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10370 DL, MVT::v4i64, V1, V2, Mask, Subtarget, DAG))
10373 // Otherwise fall back on generic blend lowering.
10374 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i64, V1, V2,
10378 /// \brief Handle lowering of 8-lane 32-bit floating point shuffles.
10380 /// Also ends up handling lowering of 8-lane 32-bit integer shuffles when AVX2
10381 /// isn't available.
10382 static SDValue lowerV8F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10383 const X86Subtarget *Subtarget,
10384 SelectionDAG &DAG) {
10386 assert(V1.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
10387 assert(V2.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
10388 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10389 ArrayRef<int> Mask = SVOp->getMask();
10390 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10392 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8f32, V1, V2, Mask,
10396 // Check for being able to broadcast a single element.
10397 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8f32, V1,
10398 Mask, Subtarget, DAG))
10401 // If the shuffle mask is repeated in each 128-bit lane, we have many more
10402 // options to efficiently lower the shuffle.
10403 SmallVector<int, 4> RepeatedMask;
10404 if (is128BitLaneRepeatedShuffleMask(MVT::v8f32, Mask, RepeatedMask)) {
10405 assert(RepeatedMask.size() == 4 &&
10406 "Repeated masks must be half the mask width!");
10408 // Use even/odd duplicate instructions for masks that match their pattern.
10409 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2, 4, 4, 6, 6}))
10410 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v8f32, V1);
10411 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3, 5, 5, 7, 7}))
10412 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v8f32, V1);
10414 if (isSingleInputShuffleMask(Mask))
10415 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f32, V1,
10416 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
10418 // Use dedicated unpack instructions for masks that match their pattern.
10420 lowerVectorShuffleWithUNPCK(DL, MVT::v8f32, Mask, V1, V2, DAG))
10423 // Otherwise, fall back to a SHUFPS sequence. Here it is important that we
10424 // have already handled any direct blends. We also need to squash the
10425 // repeated mask into a simulated v4f32 mask.
10426 for (int i = 0; i < 4; ++i)
10427 if (RepeatedMask[i] >= 8)
10428 RepeatedMask[i] -= 4;
10429 return lowerVectorShuffleWithSHUFPS(DL, MVT::v8f32, RepeatedMask, V1, V2, DAG);
10432 // If we have a single input shuffle with different shuffle patterns in the
10433 // two 128-bit lanes use the variable mask to VPERMILPS.
10434 if (isSingleInputShuffleMask(Mask)) {
10435 SDValue VPermMask[8];
10436 for (int i = 0; i < 8; ++i)
10437 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10438 : DAG.getConstant(Mask[i], DL, MVT::i32);
10439 if (!is128BitLaneCrossingShuffleMask(MVT::v8f32, Mask))
10440 return DAG.getNode(
10441 X86ISD::VPERMILPV, DL, MVT::v8f32, V1,
10442 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask));
10444 if (Subtarget->hasAVX2())
10445 return DAG.getNode(
10446 X86ISD::VPERMV, DL, MVT::v8f32,
10447 DAG.getBitcast(MVT::v8f32, DAG.getNode(ISD::BUILD_VECTOR, DL,
10448 MVT::v8i32, VPermMask)),
10451 // Otherwise, fall back.
10452 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v8f32, V1, V2, Mask,
10456 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10458 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10459 DL, MVT::v8f32, V1, V2, Mask, Subtarget, DAG))
10462 // If we have AVX2 then we always want to lower with a blend because at v8 we
10463 // can fully permute the elements.
10464 if (Subtarget->hasAVX2())
10465 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8f32, V1, V2,
10468 // Otherwise fall back on generic lowering.
10469 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v8f32, V1, V2, Mask, DAG);
10472 /// \brief Handle lowering of 8-lane 32-bit integer shuffles.
10474 /// This routine is only called when we have AVX2 and thus a reasonable
10475 /// instruction set for v8i32 shuffling..
10476 static SDValue lowerV8I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10477 const X86Subtarget *Subtarget,
10478 SelectionDAG &DAG) {
10480 assert(V1.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10481 assert(V2.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10482 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10483 ArrayRef<int> Mask = SVOp->getMask();
10484 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10485 assert(Subtarget->hasAVX2() && "We can only lower v8i32 with AVX2!");
10487 // Whenever we can lower this as a zext, that instruction is strictly faster
10488 // than any alternative. It also allows us to fold memory operands into the
10489 // shuffle in many cases.
10490 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v8i32, V1, V2,
10491 Mask, Subtarget, DAG))
10494 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i32, V1, V2, Mask,
10498 // Check for being able to broadcast a single element.
10499 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i32, V1,
10500 Mask, Subtarget, DAG))
10503 // If the shuffle mask is repeated in each 128-bit lane we can use more
10504 // efficient instructions that mirror the shuffles across the two 128-bit
10506 SmallVector<int, 4> RepeatedMask;
10507 if (is128BitLaneRepeatedShuffleMask(MVT::v8i32, Mask, RepeatedMask)) {
10508 assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!");
10509 if (isSingleInputShuffleMask(Mask))
10510 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, V1,
10511 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
10513 // Use dedicated unpack instructions for masks that match their pattern.
10515 lowerVectorShuffleWithUNPCK(DL, MVT::v8i32, Mask, V1, V2, DAG))
10519 // Try to use shift instructions.
10520 if (SDValue Shift =
10521 lowerVectorShuffleAsShift(DL, MVT::v8i32, V1, V2, Mask, DAG))
10524 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10525 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10528 // If the shuffle patterns aren't repeated but it is a single input, directly
10529 // generate a cross-lane VPERMD instruction.
10530 if (isSingleInputShuffleMask(Mask)) {
10531 SDValue VPermMask[8];
10532 for (int i = 0; i < 8; ++i)
10533 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10534 : DAG.getConstant(Mask[i], DL, MVT::i32);
10535 return DAG.getNode(
10536 X86ISD::VPERMV, DL, MVT::v8i32,
10537 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1);
10540 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10542 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10543 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10546 // Otherwise fall back on generic blend lowering.
10547 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i32, V1, V2,
10551 /// \brief Handle lowering of 16-lane 16-bit integer shuffles.
10553 /// This routine is only called when we have AVX2 and thus a reasonable
10554 /// instruction set for v16i16 shuffling..
10555 static SDValue lowerV16I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10556 const X86Subtarget *Subtarget,
10557 SelectionDAG &DAG) {
10559 assert(V1.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10560 assert(V2.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10561 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10562 ArrayRef<int> Mask = SVOp->getMask();
10563 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10564 assert(Subtarget->hasAVX2() && "We can only lower v16i16 with AVX2!");
10566 // Whenever we can lower this as a zext, that instruction is strictly faster
10567 // than any alternative. It also allows us to fold memory operands into the
10568 // shuffle in many cases.
10569 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v16i16, V1, V2,
10570 Mask, Subtarget, DAG))
10573 // Check for being able to broadcast a single element.
10574 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i16, V1,
10575 Mask, Subtarget, DAG))
10578 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i16, V1, V2, Mask,
10582 // Use dedicated unpack instructions for masks that match their pattern.
10584 lowerVectorShuffleWithUNPCK(DL, MVT::v16i16, Mask, V1, V2, DAG))
10587 // Try to use shift instructions.
10588 if (SDValue Shift =
10589 lowerVectorShuffleAsShift(DL, MVT::v16i16, V1, V2, Mask, DAG))
10592 // Try to use byte rotation instructions.
10593 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10594 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10597 if (isSingleInputShuffleMask(Mask)) {
10598 // There are no generalized cross-lane shuffle operations available on i16
10600 if (is128BitLaneCrossingShuffleMask(MVT::v16i16, Mask))
10601 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v16i16, V1, V2,
10604 SmallVector<int, 8> RepeatedMask;
10605 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
10606 // As this is a single-input shuffle, the repeated mask should be
10607 // a strictly valid v8i16 mask that we can pass through to the v8i16
10608 // lowering to handle even the v16 case.
10609 return lowerV8I16GeneralSingleInputVectorShuffle(
10610 DL, MVT::v16i16, V1, RepeatedMask, Subtarget, DAG);
10613 SDValue PSHUFBMask[32];
10614 for (int i = 0; i < 16; ++i) {
10615 if (Mask[i] == -1) {
10616 PSHUFBMask[2 * i] = PSHUFBMask[2 * i + 1] = DAG.getUNDEF(MVT::i8);
10620 int M = i < 8 ? Mask[i] : Mask[i] - 8;
10621 assert(M >= 0 && M < 8 && "Invalid single-input mask!");
10622 PSHUFBMask[2 * i] = DAG.getConstant(2 * M, DL, MVT::i8);
10623 PSHUFBMask[2 * i + 1] = DAG.getConstant(2 * M + 1, DL, MVT::i8);
10625 return DAG.getBitcast(MVT::v16i16,
10626 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8,
10627 DAG.getBitcast(MVT::v32i8, V1),
10628 DAG.getNode(ISD::BUILD_VECTOR, DL,
10629 MVT::v32i8, PSHUFBMask)));
10632 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10634 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10635 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10638 // Otherwise fall back on generic lowering.
10639 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v16i16, V1, V2, Mask, DAG);
10642 /// \brief Handle lowering of 32-lane 8-bit integer shuffles.
10644 /// This routine is only called when we have AVX2 and thus a reasonable
10645 /// instruction set for v32i8 shuffling..
10646 static SDValue lowerV32I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10647 const X86Subtarget *Subtarget,
10648 SelectionDAG &DAG) {
10650 assert(V1.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10651 assert(V2.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10652 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10653 ArrayRef<int> Mask = SVOp->getMask();
10654 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10655 assert(Subtarget->hasAVX2() && "We can only lower v32i8 with AVX2!");
10657 // Whenever we can lower this as a zext, that instruction is strictly faster
10658 // than any alternative. It also allows us to fold memory operands into the
10659 // shuffle in many cases.
10660 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v32i8, V1, V2,
10661 Mask, Subtarget, DAG))
10664 // Check for being able to broadcast a single element.
10665 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v32i8, V1,
10666 Mask, Subtarget, DAG))
10669 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v32i8, V1, V2, Mask,
10673 // Use dedicated unpack instructions for masks that match their pattern.
10675 lowerVectorShuffleWithUNPCK(DL, MVT::v32i8, Mask, V1, V2, DAG))
10678 // Try to use shift instructions.
10679 if (SDValue Shift =
10680 lowerVectorShuffleAsShift(DL, MVT::v32i8, V1, V2, Mask, DAG))
10683 // Try to use byte rotation instructions.
10684 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10685 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10688 if (isSingleInputShuffleMask(Mask)) {
10689 // There are no generalized cross-lane shuffle operations available on i8
10691 if (is128BitLaneCrossingShuffleMask(MVT::v32i8, Mask))
10692 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v32i8, V1, V2,
10695 SDValue PSHUFBMask[32];
10696 for (int i = 0; i < 32; ++i)
10699 ? DAG.getUNDEF(MVT::i8)
10700 : DAG.getConstant(Mask[i] < 16 ? Mask[i] : Mask[i] - 16, DL,
10703 return DAG.getNode(
10704 X86ISD::PSHUFB, DL, MVT::v32i8, V1,
10705 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask));
10708 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10710 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10711 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10714 // Otherwise fall back on generic lowering.
10715 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v32i8, V1, V2, Mask, DAG);
10718 /// \brief High-level routine to lower various 256-bit x86 vector shuffles.
10720 /// This routine either breaks down the specific type of a 256-bit x86 vector
10721 /// shuffle or splits it into two 128-bit shuffles and fuses the results back
10722 /// together based on the available instructions.
10723 static SDValue lower256BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10724 MVT VT, const X86Subtarget *Subtarget,
10725 SelectionDAG &DAG) {
10727 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10728 ArrayRef<int> Mask = SVOp->getMask();
10730 // If we have a single input to the zero element, insert that into V1 if we
10731 // can do so cheaply.
10732 int NumElts = VT.getVectorNumElements();
10733 int NumV2Elements = std::count_if(Mask.begin(), Mask.end(), [NumElts](int M) {
10734 return M >= NumElts;
10737 if (NumV2Elements == 1 && Mask[0] >= NumElts)
10738 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
10739 DL, VT, V1, V2, Mask, Subtarget, DAG))
10742 // There is a really nice hard cut-over between AVX1 and AVX2 that means we
10743 // can check for those subtargets here and avoid much of the subtarget
10744 // querying in the per-vector-type lowering routines. With AVX1 we have
10745 // essentially *zero* ability to manipulate a 256-bit vector with integer
10746 // types. Since we'll use floating point types there eventually, just
10747 // immediately cast everything to a float and operate entirely in that domain.
10748 if (VT.isInteger() && !Subtarget->hasAVX2()) {
10749 int ElementBits = VT.getScalarSizeInBits();
10750 if (ElementBits < 32)
10751 // No floating point type available, decompose into 128-bit vectors.
10752 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10754 MVT FpVT = MVT::getVectorVT(MVT::getFloatingPointVT(ElementBits),
10755 VT.getVectorNumElements());
10756 V1 = DAG.getBitcast(FpVT, V1);
10757 V2 = DAG.getBitcast(FpVT, V2);
10758 return DAG.getBitcast(VT, DAG.getVectorShuffle(FpVT, DL, V1, V2, Mask));
10761 switch (VT.SimpleTy) {
10763 return lowerV4F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10765 return lowerV4I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10767 return lowerV8F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10769 return lowerV8I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10771 return lowerV16I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10773 return lowerV32I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10776 llvm_unreachable("Not a valid 256-bit x86 vector type!");
10780 /// \brief Try to lower a vector shuffle as a 128-bit shuffles.
10781 static SDValue lowerV4X128VectorShuffle(SDLoc DL, MVT VT,
10782 ArrayRef<int> Mask,
10783 SDValue V1, SDValue V2,
10784 SelectionDAG &DAG) {
10785 assert(VT.getScalarSizeInBits() == 64 &&
10786 "Unexpected element type size for 128bit shuffle.");
10788 // To handle 256 bit vector requires VLX and most probably
10789 // function lowerV2X128VectorShuffle() is better solution.
10790 assert(VT.is512BitVector() && "Unexpected vector size for 128bit shuffle.");
10792 SmallVector<int, 4> WidenedMask;
10793 if (!canWidenShuffleElements(Mask, WidenedMask))
10796 // Form a 128-bit permutation.
10797 // Convert the 64-bit shuffle mask selection values into 128-bit selection
10798 // bits defined by a vshuf64x2 instruction's immediate control byte.
10799 unsigned PermMask = 0, Imm = 0;
10800 unsigned ControlBitsNum = WidenedMask.size() / 2;
10802 for (int i = 0, Size = WidenedMask.size(); i < Size; ++i) {
10803 if (WidenedMask[i] == SM_SentinelZero)
10806 // Use first element in place of undef mask.
10807 Imm = (WidenedMask[i] == SM_SentinelUndef) ? 0 : WidenedMask[i];
10808 PermMask |= (Imm % WidenedMask.size()) << (i * ControlBitsNum);
10811 return DAG.getNode(X86ISD::SHUF128, DL, VT, V1, V2,
10812 DAG.getConstant(PermMask, DL, MVT::i8));
10815 static SDValue lowerVectorShuffleWithPERMV(SDLoc DL, MVT VT,
10816 ArrayRef<int> Mask, SDValue V1,
10817 SDValue V2, SelectionDAG &DAG) {
10819 assert(VT.getScalarSizeInBits() >= 16 && "Unexpected data type for PERMV");
10821 MVT MaskEltVT = MVT::getIntegerVT(VT.getScalarSizeInBits());
10822 MVT MaskVecVT = MVT::getVectorVT(MaskEltVT, VT.getVectorNumElements());
10824 SDValue MaskNode = getConstVector(Mask, MaskVecVT, DAG, DL, true);
10825 if (isSingleInputShuffleMask(Mask))
10826 return DAG.getNode(X86ISD::VPERMV, DL, VT, MaskNode, V1);
10828 return DAG.getNode(X86ISD::VPERMV3, DL, VT, V1, MaskNode, V2);
10831 /// \brief Handle lowering of 8-lane 64-bit floating point shuffles.
10832 static SDValue lowerV8F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10833 const X86Subtarget *Subtarget,
10834 SelectionDAG &DAG) {
10836 assert(V1.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10837 assert(V2.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10838 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10839 ArrayRef<int> Mask = SVOp->getMask();
10840 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10842 if (SDValue Shuf128 =
10843 lowerV4X128VectorShuffle(DL, MVT::v8f64, Mask, V1, V2, DAG))
10846 if (SDValue Unpck =
10847 lowerVectorShuffleWithUNPCK(DL, MVT::v8f64, Mask, V1, V2, DAG))
10850 return lowerVectorShuffleWithPERMV(DL, MVT::v8f64, Mask, V1, V2, DAG);
10853 /// \brief Handle lowering of 16-lane 32-bit floating point shuffles.
10854 static SDValue lowerV16F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10855 const X86Subtarget *Subtarget,
10856 SelectionDAG &DAG) {
10858 assert(V1.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10859 assert(V2.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10860 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10861 ArrayRef<int> Mask = SVOp->getMask();
10862 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10864 if (SDValue Unpck =
10865 lowerVectorShuffleWithUNPCK(DL, MVT::v16f32, Mask, V1, V2, DAG))
10868 return lowerVectorShuffleWithPERMV(DL, MVT::v16f32, Mask, V1, V2, DAG);
10871 /// \brief Handle lowering of 8-lane 64-bit integer shuffles.
10872 static SDValue lowerV8I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10873 const X86Subtarget *Subtarget,
10874 SelectionDAG &DAG) {
10876 assert(V1.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
10877 assert(V2.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
10878 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10879 ArrayRef<int> Mask = SVOp->getMask();
10880 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10882 if (SDValue Shuf128 =
10883 lowerV4X128VectorShuffle(DL, MVT::v8i64, Mask, V1, V2, DAG))
10886 if (SDValue Unpck =
10887 lowerVectorShuffleWithUNPCK(DL, MVT::v8i64, Mask, V1, V2, DAG))
10890 return lowerVectorShuffleWithPERMV(DL, MVT::v8i64, Mask, V1, V2, DAG);
10893 /// \brief Handle lowering of 16-lane 32-bit integer shuffles.
10894 static SDValue lowerV16I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10895 const X86Subtarget *Subtarget,
10896 SelectionDAG &DAG) {
10898 assert(V1.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
10899 assert(V2.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
10900 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10901 ArrayRef<int> Mask = SVOp->getMask();
10902 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10904 if (SDValue Unpck =
10905 lowerVectorShuffleWithUNPCK(DL, MVT::v16i32, Mask, V1, V2, DAG))
10908 return lowerVectorShuffleWithPERMV(DL, MVT::v16i32, Mask, V1, V2, DAG);
10911 /// \brief Handle lowering of 32-lane 16-bit integer shuffles.
10912 static SDValue lowerV32I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10913 const X86Subtarget *Subtarget,
10914 SelectionDAG &DAG) {
10916 assert(V1.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10917 assert(V2.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10918 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10919 ArrayRef<int> Mask = SVOp->getMask();
10920 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10921 assert(Subtarget->hasBWI() && "We can only lower v32i16 with AVX-512-BWI!");
10923 return lowerVectorShuffleWithPERMV(DL, MVT::v32i16, Mask, V1, V2, DAG);
10926 /// \brief Handle lowering of 64-lane 8-bit integer shuffles.
10927 static SDValue lowerV64I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10928 const X86Subtarget *Subtarget,
10929 SelectionDAG &DAG) {
10931 assert(V1.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10932 assert(V2.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10933 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10934 ArrayRef<int> Mask = SVOp->getMask();
10935 assert(Mask.size() == 64 && "Unexpected mask size for v64 shuffle!");
10936 assert(Subtarget->hasBWI() && "We can only lower v64i8 with AVX-512-BWI!");
10938 // FIXME: Implement direct support for this type!
10939 return splitAndLowerVectorShuffle(DL, MVT::v64i8, V1, V2, Mask, DAG);
10942 /// \brief High-level routine to lower various 512-bit x86 vector shuffles.
10944 /// This routine either breaks down the specific type of a 512-bit x86 vector
10945 /// shuffle or splits it into two 256-bit shuffles and fuses the results back
10946 /// together based on the available instructions.
10947 static SDValue lower512BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10948 MVT VT, const X86Subtarget *Subtarget,
10949 SelectionDAG &DAG) {
10951 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10952 ArrayRef<int> Mask = SVOp->getMask();
10953 assert(Subtarget->hasAVX512() &&
10954 "Cannot lower 512-bit vectors w/ basic ISA!");
10956 // Check for being able to broadcast a single element.
10957 if (SDValue Broadcast =
10958 lowerVectorShuffleAsBroadcast(DL, VT, V1, Mask, Subtarget, DAG))
10961 // Dispatch to each element type for lowering. If we don't have supprot for
10962 // specific element type shuffles at 512 bits, immediately split them and
10963 // lower them. Each lowering routine of a given type is allowed to assume that
10964 // the requisite ISA extensions for that element type are available.
10965 switch (VT.SimpleTy) {
10967 return lowerV8F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10969 return lowerV16F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10971 return lowerV8I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10973 return lowerV16I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10975 if (Subtarget->hasBWI())
10976 return lowerV32I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10979 if (Subtarget->hasBWI())
10980 return lowerV64I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10984 llvm_unreachable("Not a valid 512-bit x86 vector type!");
10987 // Otherwise fall back on splitting.
10988 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10991 // Lower vXi1 vector shuffles.
10992 // There is no a dedicated instruction on AVX-512 that shuffles the masks.
10993 // The only way to shuffle bits is to sign-extend the mask vector to SIMD
10994 // vector, shuffle and then truncate it back.
10995 static SDValue lower1BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10996 MVT VT, const X86Subtarget *Subtarget,
10997 SelectionDAG &DAG) {
10999 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11000 ArrayRef<int> Mask = SVOp->getMask();
11001 assert(Subtarget->hasAVX512() &&
11002 "Cannot lower 512-bit vectors w/o basic ISA!");
11004 switch (VT.SimpleTy) {
11006 llvm_unreachable("Expected a vector of i1 elements");
11008 ExtVT = MVT::v2i64;
11011 ExtVT = MVT::v4i32;
11014 ExtVT = MVT::v8i64; // Take 512-bit type, more shuffles on KNL
11017 ExtVT = MVT::v16i32;
11020 ExtVT = MVT::v32i16;
11023 ExtVT = MVT::v64i8;
11027 if (ISD::isBuildVectorAllZeros(V1.getNode()))
11028 V1 = getZeroVector(ExtVT, Subtarget, DAG, DL);
11029 else if (ISD::isBuildVectorAllOnes(V1.getNode()))
11030 V1 = getOnesVector(ExtVT, Subtarget, DAG, DL);
11032 V1 = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, V1);
11035 V2 = DAG.getUNDEF(ExtVT);
11036 else if (ISD::isBuildVectorAllZeros(V2.getNode()))
11037 V2 = getZeroVector(ExtVT, Subtarget, DAG, DL);
11038 else if (ISD::isBuildVectorAllOnes(V2.getNode()))
11039 V2 = getOnesVector(ExtVT, Subtarget, DAG, DL);
11041 V2 = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, V2);
11042 return DAG.getNode(ISD::TRUNCATE, DL, VT,
11043 DAG.getVectorShuffle(ExtVT, DL, V1, V2, Mask));
11045 /// \brief Top-level lowering for x86 vector shuffles.
11047 /// This handles decomposition, canonicalization, and lowering of all x86
11048 /// vector shuffles. Most of the specific lowering strategies are encapsulated
11049 /// above in helper routines. The canonicalization attempts to widen shuffles
11050 /// to involve fewer lanes of wider elements, consolidate symmetric patterns
11051 /// s.t. only one of the two inputs needs to be tested, etc.
11052 static SDValue lowerVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
11053 SelectionDAG &DAG) {
11054 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11055 ArrayRef<int> Mask = SVOp->getMask();
11056 SDValue V1 = Op.getOperand(0);
11057 SDValue V2 = Op.getOperand(1);
11058 MVT VT = Op.getSimpleValueType();
11059 int NumElements = VT.getVectorNumElements();
11061 bool Is1BitVector = (VT.getVectorElementType() == MVT::i1);
11063 assert((VT.getSizeInBits() != 64 || Is1BitVector) &&
11064 "Can't lower MMX shuffles");
11066 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
11067 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
11068 if (V1IsUndef && V2IsUndef)
11069 return DAG.getUNDEF(VT);
11071 // When we create a shuffle node we put the UNDEF node to second operand,
11072 // but in some cases the first operand may be transformed to UNDEF.
11073 // In this case we should just commute the node.
11075 return DAG.getCommutedVectorShuffle(*SVOp);
11077 // Check for non-undef masks pointing at an undef vector and make the masks
11078 // undef as well. This makes it easier to match the shuffle based solely on
11082 if (M >= NumElements) {
11083 SmallVector<int, 8> NewMask(Mask.begin(), Mask.end());
11084 for (int &M : NewMask)
11085 if (M >= NumElements)
11087 return DAG.getVectorShuffle(VT, dl, V1, V2, NewMask);
11090 // We actually see shuffles that are entirely re-arrangements of a set of
11091 // zero inputs. This mostly happens while decomposing complex shuffles into
11092 // simple ones. Directly lower these as a buildvector of zeros.
11093 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
11094 if (Zeroable.all())
11095 return getZeroVector(VT, Subtarget, DAG, dl);
11097 // Try to collapse shuffles into using a vector type with fewer elements but
11098 // wider element types. We cap this to not form integers or floating point
11099 // elements wider than 64 bits, but it might be interesting to form i128
11100 // integers to handle flipping the low and high halves of AVX 256-bit vectors.
11101 SmallVector<int, 16> WidenedMask;
11102 if (VT.getScalarSizeInBits() < 64 && !Is1BitVector &&
11103 canWidenShuffleElements(Mask, WidenedMask)) {
11104 MVT NewEltVT = VT.isFloatingPoint()
11105 ? MVT::getFloatingPointVT(VT.getScalarSizeInBits() * 2)
11106 : MVT::getIntegerVT(VT.getScalarSizeInBits() * 2);
11107 MVT NewVT = MVT::getVectorVT(NewEltVT, VT.getVectorNumElements() / 2);
11108 // Make sure that the new vector type is legal. For example, v2f64 isn't
11110 if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) {
11111 V1 = DAG.getBitcast(NewVT, V1);
11112 V2 = DAG.getBitcast(NewVT, V2);
11113 return DAG.getBitcast(
11114 VT, DAG.getVectorShuffle(NewVT, dl, V1, V2, WidenedMask));
11118 int NumV1Elements = 0, NumUndefElements = 0, NumV2Elements = 0;
11119 for (int M : SVOp->getMask())
11121 ++NumUndefElements;
11122 else if (M < NumElements)
11127 // Commute the shuffle as needed such that more elements come from V1 than
11128 // V2. This allows us to match the shuffle pattern strictly on how many
11129 // elements come from V1 without handling the symmetric cases.
11130 if (NumV2Elements > NumV1Elements)
11131 return DAG.getCommutedVectorShuffle(*SVOp);
11133 // When the number of V1 and V2 elements are the same, try to minimize the
11134 // number of uses of V2 in the low half of the vector. When that is tied,
11135 // ensure that the sum of indices for V1 is equal to or lower than the sum
11136 // indices for V2. When those are equal, try to ensure that the number of odd
11137 // indices for V1 is lower than the number of odd indices for V2.
11138 if (NumV1Elements == NumV2Elements) {
11139 int LowV1Elements = 0, LowV2Elements = 0;
11140 for (int M : SVOp->getMask().slice(0, NumElements / 2))
11141 if (M >= NumElements)
11145 if (LowV2Elements > LowV1Elements) {
11146 return DAG.getCommutedVectorShuffle(*SVOp);
11147 } else if (LowV2Elements == LowV1Elements) {
11148 int SumV1Indices = 0, SumV2Indices = 0;
11149 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
11150 if (SVOp->getMask()[i] >= NumElements)
11152 else if (SVOp->getMask()[i] >= 0)
11154 if (SumV2Indices < SumV1Indices) {
11155 return DAG.getCommutedVectorShuffle(*SVOp);
11156 } else if (SumV2Indices == SumV1Indices) {
11157 int NumV1OddIndices = 0, NumV2OddIndices = 0;
11158 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
11159 if (SVOp->getMask()[i] >= NumElements)
11160 NumV2OddIndices += i % 2;
11161 else if (SVOp->getMask()[i] >= 0)
11162 NumV1OddIndices += i % 2;
11163 if (NumV2OddIndices < NumV1OddIndices)
11164 return DAG.getCommutedVectorShuffle(*SVOp);
11169 // For each vector width, delegate to a specialized lowering routine.
11170 if (VT.is128BitVector())
11171 return lower128BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11173 if (VT.is256BitVector())
11174 return lower256BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11176 if (VT.is512BitVector())
11177 return lower512BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11180 return lower1BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11181 llvm_unreachable("Unimplemented!");
11184 // This function assumes its argument is a BUILD_VECTOR of constants or
11185 // undef SDNodes. i.e: ISD::isBuildVectorOfConstantSDNodes(BuildVector) is
11187 static bool BUILD_VECTORtoBlendMask(BuildVectorSDNode *BuildVector,
11188 unsigned &MaskValue) {
11190 unsigned NumElems = BuildVector->getNumOperands();
11192 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
11193 // We don't handle the >2 lanes case right now.
11194 unsigned NumLanes = (NumElems - 1) / 8 + 1;
11198 unsigned NumElemsInLane = NumElems / NumLanes;
11200 // Blend for v16i16 should be symmetric for the both lanes.
11201 for (unsigned i = 0; i < NumElemsInLane; ++i) {
11202 SDValue EltCond = BuildVector->getOperand(i);
11203 SDValue SndLaneEltCond =
11204 (NumLanes == 2) ? BuildVector->getOperand(i + NumElemsInLane) : EltCond;
11206 int Lane1Cond = -1, Lane2Cond = -1;
11207 if (isa<ConstantSDNode>(EltCond))
11208 Lane1Cond = !isZero(EltCond);
11209 if (isa<ConstantSDNode>(SndLaneEltCond))
11210 Lane2Cond = !isZero(SndLaneEltCond);
11212 unsigned LaneMask = 0;
11213 if (Lane1Cond == Lane2Cond || Lane2Cond < 0)
11214 // Lane1Cond != 0, means we want the first argument.
11215 // Lane1Cond == 0, means we want the second argument.
11216 // The encoding of this argument is 0 for the first argument, 1
11217 // for the second. Therefore, invert the condition.
11218 LaneMask = !Lane1Cond << i;
11219 else if (Lane1Cond < 0)
11220 LaneMask = !Lane2Cond << i;
11224 MaskValue |= LaneMask;
11226 MaskValue |= LaneMask << NumElemsInLane;
11231 /// \brief Try to lower a VSELECT instruction to a vector shuffle.
11232 static SDValue lowerVSELECTtoVectorShuffle(SDValue Op,
11233 const X86Subtarget *Subtarget,
11234 SelectionDAG &DAG) {
11235 SDValue Cond = Op.getOperand(0);
11236 SDValue LHS = Op.getOperand(1);
11237 SDValue RHS = Op.getOperand(2);
11239 MVT VT = Op.getSimpleValueType();
11241 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
11243 auto *CondBV = cast<BuildVectorSDNode>(Cond);
11245 // Only non-legal VSELECTs reach this lowering, convert those into generic
11246 // shuffles and re-use the shuffle lowering path for blends.
11247 SmallVector<int, 32> Mask;
11248 for (int i = 0, Size = VT.getVectorNumElements(); i < Size; ++i) {
11249 SDValue CondElt = CondBV->getOperand(i);
11251 isa<ConstantSDNode>(CondElt) ? i + (isZero(CondElt) ? Size : 0) : -1);
11253 return DAG.getVectorShuffle(VT, dl, LHS, RHS, Mask);
11256 SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const {
11257 // A vselect where all conditions and data are constants can be optimized into
11258 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
11259 if (ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(0).getNode()) &&
11260 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(1).getNode()) &&
11261 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(2).getNode()))
11264 // Try to lower this to a blend-style vector shuffle. This can handle all
11265 // constant condition cases.
11266 if (SDValue BlendOp = lowerVSELECTtoVectorShuffle(Op, Subtarget, DAG))
11269 // Variable blends are only legal from SSE4.1 onward.
11270 if (!Subtarget->hasSSE41())
11273 // Only some types will be legal on some subtargets. If we can emit a legal
11274 // VSELECT-matching blend, return Op, and but if we need to expand, return
11276 switch (Op.getSimpleValueType().SimpleTy) {
11278 // Most of the vector types have blends past SSE4.1.
11282 // The byte blends for AVX vectors were introduced only in AVX2.
11283 if (Subtarget->hasAVX2())
11290 // AVX-512 BWI and VLX features support VSELECT with i16 elements.
11291 if (Subtarget->hasBWI() && Subtarget->hasVLX())
11294 // FIXME: We should custom lower this by fixing the condition and using i8
11300 static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
11301 MVT VT = Op.getSimpleValueType();
11304 if (!Op.getOperand(0).getSimpleValueType().is128BitVector())
11307 if (VT.getSizeInBits() == 8) {
11308 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
11309 Op.getOperand(0), Op.getOperand(1));
11310 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
11311 DAG.getValueType(VT));
11312 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11315 if (VT.getSizeInBits() == 16) {
11316 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11317 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
11319 return DAG.getNode(
11320 ISD::TRUNCATE, dl, MVT::i16,
11321 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11322 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
11323 Op.getOperand(1)));
11324 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
11325 Op.getOperand(0), Op.getOperand(1));
11326 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
11327 DAG.getValueType(VT));
11328 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11331 if (VT == MVT::f32) {
11332 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
11333 // the result back to FR32 register. It's only worth matching if the
11334 // result has a single use which is a store or a bitcast to i32. And in
11335 // the case of a store, it's not worth it if the index is a constant 0,
11336 // because a MOVSSmr can be used instead, which is smaller and faster.
11337 if (!Op.hasOneUse())
11339 SDNode *User = *Op.getNode()->use_begin();
11340 if ((User->getOpcode() != ISD::STORE ||
11341 (isa<ConstantSDNode>(Op.getOperand(1)) &&
11342 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
11343 (User->getOpcode() != ISD::BITCAST ||
11344 User->getValueType(0) != MVT::i32))
11346 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11347 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
11349 return DAG.getBitcast(MVT::f32, Extract);
11352 if (VT == MVT::i32 || VT == MVT::i64) {
11353 // ExtractPS/pextrq works with constant index.
11354 if (isa<ConstantSDNode>(Op.getOperand(1)))
11360 /// Extract one bit from mask vector, like v16i1 or v8i1.
11361 /// AVX-512 feature.
11363 X86TargetLowering::ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const {
11364 SDValue Vec = Op.getOperand(0);
11366 MVT VecVT = Vec.getSimpleValueType();
11367 SDValue Idx = Op.getOperand(1);
11368 MVT EltVT = Op.getSimpleValueType();
11370 assert((EltVT == MVT::i1) && "Unexpected operands in ExtractBitFromMaskVector");
11371 assert((VecVT.getVectorNumElements() <= 16 || Subtarget->hasBWI()) &&
11372 "Unexpected vector type in ExtractBitFromMaskVector");
11374 // variable index can't be handled in mask registers,
11375 // extend vector to VR512
11376 if (!isa<ConstantSDNode>(Idx)) {
11377 MVT ExtVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
11378 SDValue Ext = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVT, Vec);
11379 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
11380 ExtVT.getVectorElementType(), Ext, Idx);
11381 return DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
11384 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11385 const TargetRegisterClass* rc = getRegClassFor(VecVT);
11386 if (!Subtarget->hasDQI() && (VecVT.getVectorNumElements() <= 8))
11387 rc = getRegClassFor(MVT::v16i1);
11388 unsigned MaxSift = rc->getSize()*8 - 1;
11389 Vec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, Vec,
11390 DAG.getConstant(MaxSift - IdxVal, dl, MVT::i8));
11391 Vec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, Vec,
11392 DAG.getConstant(MaxSift, dl, MVT::i8));
11393 return DAG.getNode(X86ISD::VEXTRACT, dl, MVT::i1, Vec,
11394 DAG.getIntPtrConstant(0, dl));
11398 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
11399 SelectionDAG &DAG) const {
11401 SDValue Vec = Op.getOperand(0);
11402 MVT VecVT = Vec.getSimpleValueType();
11403 SDValue Idx = Op.getOperand(1);
11405 if (Op.getSimpleValueType() == MVT::i1)
11406 return ExtractBitFromMaskVector(Op, DAG);
11408 if (!isa<ConstantSDNode>(Idx)) {
11409 if (VecVT.is512BitVector() ||
11410 (VecVT.is256BitVector() && Subtarget->hasInt256() &&
11411 VecVT.getVectorElementType().getSizeInBits() == 32)) {
11414 MVT::getIntegerVT(VecVT.getVectorElementType().getSizeInBits());
11415 MVT MaskVT = MVT::getVectorVT(MaskEltVT, VecVT.getSizeInBits() /
11416 MaskEltVT.getSizeInBits());
11418 Idx = DAG.getZExtOrTrunc(Idx, dl, MaskEltVT);
11419 auto PtrVT = getPointerTy(DAG.getDataLayout());
11420 SDValue Mask = DAG.getNode(X86ISD::VINSERT, dl, MaskVT,
11421 getZeroVector(MaskVT, Subtarget, DAG, dl), Idx,
11422 DAG.getConstant(0, dl, PtrVT));
11423 SDValue Perm = DAG.getNode(X86ISD::VPERMV, dl, VecVT, Mask, Vec);
11424 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Perm,
11425 DAG.getConstant(0, dl, PtrVT));
11430 // If this is a 256-bit vector result, first extract the 128-bit vector and
11431 // then extract the element from the 128-bit vector.
11432 if (VecVT.is256BitVector() || VecVT.is512BitVector()) {
11434 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11435 // Get the 128-bit vector.
11436 Vec = Extract128BitVector(Vec, IdxVal, DAG, dl);
11437 MVT EltVT = VecVT.getVectorElementType();
11439 unsigned ElemsPerChunk = 128 / EltVT.getSizeInBits();
11440 assert(isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2");
11442 // Find IdxVal modulo ElemsPerChunk. Since ElemsPerChunk is a power of 2
11443 // this can be done with a mask.
11444 IdxVal &= ElemsPerChunk - 1;
11445 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
11446 DAG.getConstant(IdxVal, dl, MVT::i32));
11449 assert(VecVT.is128BitVector() && "Unexpected vector length");
11451 if (Subtarget->hasSSE41())
11452 if (SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG))
11455 MVT VT = Op.getSimpleValueType();
11456 // TODO: handle v16i8.
11457 if (VT.getSizeInBits() == 16) {
11458 SDValue Vec = Op.getOperand(0);
11459 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11461 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
11462 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11463 DAG.getBitcast(MVT::v4i32, Vec),
11464 Op.getOperand(1)));
11465 // Transform it so it match pextrw which produces a 32-bit result.
11466 MVT EltVT = MVT::i32;
11467 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
11468 Op.getOperand(0), Op.getOperand(1));
11469 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
11470 DAG.getValueType(VT));
11471 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11474 if (VT.getSizeInBits() == 32) {
11475 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11479 // SHUFPS the element to the lowest double word, then movss.
11480 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
11481 MVT VVT = Op.getOperand(0).getSimpleValueType();
11482 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
11483 DAG.getUNDEF(VVT), Mask);
11484 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
11485 DAG.getIntPtrConstant(0, dl));
11488 if (VT.getSizeInBits() == 64) {
11489 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
11490 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
11491 // to match extract_elt for f64.
11492 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11496 // UNPCKHPD the element to the lowest double word, then movsd.
11497 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
11498 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
11499 int Mask[2] = { 1, -1 };
11500 MVT VVT = Op.getOperand(0).getSimpleValueType();
11501 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
11502 DAG.getUNDEF(VVT), Mask);
11503 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
11504 DAG.getIntPtrConstant(0, dl));
11510 /// Insert one bit to mask vector, like v16i1 or v8i1.
11511 /// AVX-512 feature.
11513 X86TargetLowering::InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const {
11515 SDValue Vec = Op.getOperand(0);
11516 SDValue Elt = Op.getOperand(1);
11517 SDValue Idx = Op.getOperand(2);
11518 MVT VecVT = Vec.getSimpleValueType();
11520 if (!isa<ConstantSDNode>(Idx)) {
11521 // Non constant index. Extend source and destination,
11522 // insert element and then truncate the result.
11523 MVT ExtVecVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
11524 MVT ExtEltVT = (VecVT == MVT::v8i1 ? MVT::i64 : MVT::i32);
11525 SDValue ExtOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ExtVecVT,
11526 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVecVT, Vec),
11527 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtEltVT, Elt), Idx);
11528 return DAG.getNode(ISD::TRUNCATE, dl, VecVT, ExtOp);
11531 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11532 SDValue EltInVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Elt);
11534 EltInVec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
11535 DAG.getConstant(IdxVal, dl, MVT::i8));
11536 if (Vec.getOpcode() == ISD::UNDEF)
11538 return DAG.getNode(ISD::OR, dl, VecVT, Vec, EltInVec);
11541 SDValue X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
11542 SelectionDAG &DAG) const {
11543 MVT VT = Op.getSimpleValueType();
11544 MVT EltVT = VT.getVectorElementType();
11546 if (EltVT == MVT::i1)
11547 return InsertBitToMaskVector(Op, DAG);
11550 SDValue N0 = Op.getOperand(0);
11551 SDValue N1 = Op.getOperand(1);
11552 SDValue N2 = Op.getOperand(2);
11553 if (!isa<ConstantSDNode>(N2))
11555 auto *N2C = cast<ConstantSDNode>(N2);
11556 unsigned IdxVal = N2C->getZExtValue();
11558 // If the vector is wider than 128 bits, extract the 128-bit subvector, insert
11559 // into that, and then insert the subvector back into the result.
11560 if (VT.is256BitVector() || VT.is512BitVector()) {
11561 // With a 256-bit vector, we can insert into the zero element efficiently
11562 // using a blend if we have AVX or AVX2 and the right data type.
11563 if (VT.is256BitVector() && IdxVal == 0) {
11564 // TODO: It is worthwhile to cast integer to floating point and back
11565 // and incur a domain crossing penalty if that's what we'll end up
11566 // doing anyway after extracting to a 128-bit vector.
11567 if ((Subtarget->hasAVX() && (EltVT == MVT::f64 || EltVT == MVT::f32)) ||
11568 (Subtarget->hasAVX2() && EltVT == MVT::i32)) {
11569 SDValue N1Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, N1);
11570 N2 = DAG.getIntPtrConstant(1, dl);
11571 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1Vec, N2);
11575 // Get the desired 128-bit vector chunk.
11576 SDValue V = Extract128BitVector(N0, IdxVal, DAG, dl);
11578 // Insert the element into the desired chunk.
11579 unsigned NumEltsIn128 = 128 / EltVT.getSizeInBits();
11580 assert(isPowerOf2_32(NumEltsIn128));
11581 // Since NumEltsIn128 is a power of 2 we can use mask instead of modulo.
11582 unsigned IdxIn128 = IdxVal & (NumEltsIn128 - 1);
11584 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1,
11585 DAG.getConstant(IdxIn128, dl, MVT::i32));
11587 // Insert the changed part back into the bigger vector
11588 return Insert128BitVector(N0, V, IdxVal, DAG, dl);
11590 assert(VT.is128BitVector() && "Only 128-bit vector types should be left!");
11592 if (Subtarget->hasSSE41()) {
11593 if (EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) {
11595 if (VT == MVT::v8i16) {
11596 Opc = X86ISD::PINSRW;
11598 assert(VT == MVT::v16i8);
11599 Opc = X86ISD::PINSRB;
11602 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
11604 if (N1.getValueType() != MVT::i32)
11605 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11606 if (N2.getValueType() != MVT::i32)
11607 N2 = DAG.getIntPtrConstant(IdxVal, dl);
11608 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
11611 if (EltVT == MVT::f32) {
11612 // Bits [7:6] of the constant are the source select. This will always be
11613 // zero here. The DAG Combiner may combine an extract_elt index into
11614 // these bits. For example (insert (extract, 3), 2) could be matched by
11615 // putting the '3' into bits [7:6] of X86ISD::INSERTPS.
11616 // Bits [5:4] of the constant are the destination select. This is the
11617 // value of the incoming immediate.
11618 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
11619 // combine either bitwise AND or insert of float 0.0 to set these bits.
11621 bool MinSize = DAG.getMachineFunction().getFunction()->optForMinSize();
11622 if (IdxVal == 0 && (!MinSize || !MayFoldLoad(N1))) {
11623 // If this is an insertion of 32-bits into the low 32-bits of
11624 // a vector, we prefer to generate a blend with immediate rather
11625 // than an insertps. Blends are simpler operations in hardware and so
11626 // will always have equal or better performance than insertps.
11627 // But if optimizing for size and there's a load folding opportunity,
11628 // generate insertps because blendps does not have a 32-bit memory
11630 N2 = DAG.getIntPtrConstant(1, dl);
11631 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
11632 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1, N2);
11634 N2 = DAG.getIntPtrConstant(IdxVal << 4, dl);
11635 // Create this as a scalar to vector..
11636 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
11637 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
11640 if (EltVT == MVT::i32 || EltVT == MVT::i64) {
11641 // PINSR* works with constant index.
11646 if (EltVT == MVT::i8)
11649 if (EltVT.getSizeInBits() == 16) {
11650 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
11651 // as its second argument.
11652 if (N1.getValueType() != MVT::i32)
11653 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11654 if (N2.getValueType() != MVT::i32)
11655 N2 = DAG.getIntPtrConstant(IdxVal, dl);
11656 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
11661 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
11663 MVT OpVT = Op.getSimpleValueType();
11665 // If this is a 256-bit vector result, first insert into a 128-bit
11666 // vector and then insert into the 256-bit vector.
11667 if (!OpVT.is128BitVector()) {
11668 // Insert into a 128-bit vector.
11669 unsigned SizeFactor = OpVT.getSizeInBits()/128;
11670 MVT VT128 = MVT::getVectorVT(OpVT.getVectorElementType(),
11671 OpVT.getVectorNumElements() / SizeFactor);
11673 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
11675 // Insert the 128-bit vector.
11676 return Insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl);
11679 if (OpVT == MVT::v1i64 &&
11680 Op.getOperand(0).getValueType() == MVT::i64)
11681 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
11683 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
11684 assert(OpVT.is128BitVector() && "Expected an SSE type!");
11685 return DAG.getBitcast(
11686 OpVT, DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, AnyExt));
11689 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
11690 // a simple subregister reference or explicit instructions to grab
11691 // upper bits of a vector.
11692 static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
11693 SelectionDAG &DAG) {
11695 SDValue In = Op.getOperand(0);
11696 SDValue Idx = Op.getOperand(1);
11697 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11698 MVT ResVT = Op.getSimpleValueType();
11699 MVT InVT = In.getSimpleValueType();
11701 if (Subtarget->hasFp256()) {
11702 if (ResVT.is128BitVector() &&
11703 (InVT.is256BitVector() || InVT.is512BitVector()) &&
11704 isa<ConstantSDNode>(Idx)) {
11705 return Extract128BitVector(In, IdxVal, DAG, dl);
11707 if (ResVT.is256BitVector() && InVT.is512BitVector() &&
11708 isa<ConstantSDNode>(Idx)) {
11709 return Extract256BitVector(In, IdxVal, DAG, dl);
11715 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
11716 // simple superregister reference or explicit instructions to insert
11717 // the upper bits of a vector.
11718 static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
11719 SelectionDAG &DAG) {
11720 if (!Subtarget->hasAVX())
11724 SDValue Vec = Op.getOperand(0);
11725 SDValue SubVec = Op.getOperand(1);
11726 SDValue Idx = Op.getOperand(2);
11728 if (!isa<ConstantSDNode>(Idx))
11731 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11732 MVT OpVT = Op.getSimpleValueType();
11733 MVT SubVecVT = SubVec.getSimpleValueType();
11735 // Fold two 16-byte subvector loads into one 32-byte load:
11736 // (insert_subvector (insert_subvector undef, (load addr), 0),
11737 // (load addr + 16), Elts/2)
11739 if ((IdxVal == OpVT.getVectorNumElements() / 2) &&
11740 Vec.getOpcode() == ISD::INSERT_SUBVECTOR &&
11741 OpVT.is256BitVector() && SubVecVT.is128BitVector()) {
11742 auto *Idx2 = dyn_cast<ConstantSDNode>(Vec.getOperand(2));
11743 if (Idx2 && Idx2->getZExtValue() == 0) {
11744 SDValue SubVec2 = Vec.getOperand(1);
11745 // If needed, look through a bitcast to get to the load.
11746 if (SubVec2.getNode() && SubVec2.getOpcode() == ISD::BITCAST)
11747 SubVec2 = SubVec2.getOperand(0);
11749 if (auto *FirstLd = dyn_cast<LoadSDNode>(SubVec2)) {
11751 unsigned Alignment = FirstLd->getAlignment();
11752 unsigned AS = FirstLd->getAddressSpace();
11753 const X86TargetLowering *TLI = Subtarget->getTargetLowering();
11754 if (TLI->allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(),
11755 OpVT, AS, Alignment, &Fast) && Fast) {
11756 SDValue Ops[] = { SubVec2, SubVec };
11757 if (SDValue Ld = EltsFromConsecutiveLoads(OpVT, Ops, dl, DAG, false))
11764 if ((OpVT.is256BitVector() || OpVT.is512BitVector()) &&
11765 SubVecVT.is128BitVector())
11766 return Insert128BitVector(Vec, SubVec, IdxVal, DAG, dl);
11768 if (OpVT.is512BitVector() && SubVecVT.is256BitVector())
11769 return Insert256BitVector(Vec, SubVec, IdxVal, DAG, dl);
11771 if (OpVT.getVectorElementType() == MVT::i1) {
11772 if (IdxVal == 0 && Vec.getOpcode() == ISD::UNDEF) // the operation is legal
11774 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
11775 SDValue Undef = DAG.getUNDEF(OpVT);
11776 unsigned NumElems = OpVT.getVectorNumElements();
11777 SDValue ShiftBits = DAG.getConstant(NumElems/2, dl, MVT::i8);
11779 if (IdxVal == OpVT.getVectorNumElements() / 2) {
11780 // Zero upper bits of the Vec
11781 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
11782 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
11784 SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
11786 Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
11787 return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
11790 SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
11792 // Zero upper bits of the Vec2
11793 Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
11794 Vec2 = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec2, ShiftBits);
11795 // Zero lower bits of the Vec
11796 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
11797 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
11798 // Merge them together
11799 return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
11805 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
11806 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
11807 // one of the above mentioned nodes. It has to be wrapped because otherwise
11808 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
11809 // be used to form addressing mode. These wrapped nodes will be selected
11812 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
11813 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
11815 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11816 // global base reg.
11817 unsigned char OpFlag = 0;
11818 unsigned WrapperKind = X86ISD::Wrapper;
11819 CodeModel::Model M = DAG.getTarget().getCodeModel();
11821 if (Subtarget->isPICStyleRIPRel() &&
11822 (M == CodeModel::Small || M == CodeModel::Kernel))
11823 WrapperKind = X86ISD::WrapperRIP;
11824 else if (Subtarget->isPICStyleGOT())
11825 OpFlag = X86II::MO_GOTOFF;
11826 else if (Subtarget->isPICStyleStubPIC())
11827 OpFlag = X86II::MO_PIC_BASE_OFFSET;
11829 auto PtrVT = getPointerTy(DAG.getDataLayout());
11830 SDValue Result = DAG.getTargetConstantPool(
11831 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(), OpFlag);
11833 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11834 // With PIC, the address is actually $g + Offset.
11837 DAG.getNode(ISD::ADD, DL, PtrVT,
11838 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11844 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
11845 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
11847 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11848 // global base reg.
11849 unsigned char OpFlag = 0;
11850 unsigned WrapperKind = X86ISD::Wrapper;
11851 CodeModel::Model M = DAG.getTarget().getCodeModel();
11853 if (Subtarget->isPICStyleRIPRel() &&
11854 (M == CodeModel::Small || M == CodeModel::Kernel))
11855 WrapperKind = X86ISD::WrapperRIP;
11856 else if (Subtarget->isPICStyleGOT())
11857 OpFlag = X86II::MO_GOTOFF;
11858 else if (Subtarget->isPICStyleStubPIC())
11859 OpFlag = X86II::MO_PIC_BASE_OFFSET;
11861 auto PtrVT = getPointerTy(DAG.getDataLayout());
11862 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, OpFlag);
11864 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11866 // With PIC, the address is actually $g + Offset.
11869 DAG.getNode(ISD::ADD, DL, PtrVT,
11870 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11876 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
11877 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
11879 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11880 // global base reg.
11881 unsigned char OpFlag = 0;
11882 unsigned WrapperKind = X86ISD::Wrapper;
11883 CodeModel::Model M = DAG.getTarget().getCodeModel();
11885 if (Subtarget->isPICStyleRIPRel() &&
11886 (M == CodeModel::Small || M == CodeModel::Kernel)) {
11887 if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF())
11888 OpFlag = X86II::MO_GOTPCREL;
11889 WrapperKind = X86ISD::WrapperRIP;
11890 } else if (Subtarget->isPICStyleGOT()) {
11891 OpFlag = X86II::MO_GOT;
11892 } else if (Subtarget->isPICStyleStubPIC()) {
11893 OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE;
11894 } else if (Subtarget->isPICStyleStubNoDynamic()) {
11895 OpFlag = X86II::MO_DARWIN_NONLAZY;
11898 auto PtrVT = getPointerTy(DAG.getDataLayout());
11899 SDValue Result = DAG.getTargetExternalSymbol(Sym, PtrVT, OpFlag);
11902 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11904 // With PIC, the address is actually $g + Offset.
11905 if (DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
11906 !Subtarget->is64Bit()) {
11908 DAG.getNode(ISD::ADD, DL, PtrVT,
11909 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11912 // For symbols that require a load from a stub to get the address, emit the
11914 if (isGlobalStubReference(OpFlag))
11915 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
11916 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
11917 false, false, false, 0);
11923 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
11924 // Create the TargetBlockAddressAddress node.
11925 unsigned char OpFlags =
11926 Subtarget->ClassifyBlockAddressReference();
11927 CodeModel::Model M = DAG.getTarget().getCodeModel();
11928 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
11929 int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset();
11931 auto PtrVT = getPointerTy(DAG.getDataLayout());
11932 SDValue Result = DAG.getTargetBlockAddress(BA, PtrVT, Offset, OpFlags);
11934 if (Subtarget->isPICStyleRIPRel() &&
11935 (M == CodeModel::Small || M == CodeModel::Kernel))
11936 Result = DAG.getNode(X86ISD::WrapperRIP, dl, PtrVT, Result);
11938 Result = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, Result);
11940 // With PIC, the address is actually $g + Offset.
11941 if (isGlobalRelativeToPICBase(OpFlags)) {
11942 Result = DAG.getNode(ISD::ADD, dl, PtrVT,
11943 DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
11950 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
11951 int64_t Offset, SelectionDAG &DAG) const {
11952 // Create the TargetGlobalAddress node, folding in the constant
11953 // offset if it is legal.
11954 unsigned char OpFlags =
11955 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget());
11956 CodeModel::Model M = DAG.getTarget().getCodeModel();
11957 auto PtrVT = getPointerTy(DAG.getDataLayout());
11959 if (OpFlags == X86II::MO_NO_FLAG &&
11960 X86::isOffsetSuitableForCodeModel(Offset, M)) {
11961 // A direct static reference to a global.
11962 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, Offset);
11965 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, OpFlags);
11968 if (Subtarget->isPICStyleRIPRel() &&
11969 (M == CodeModel::Small || M == CodeModel::Kernel))
11970 Result = DAG.getNode(X86ISD::WrapperRIP, dl, PtrVT, Result);
11972 Result = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, Result);
11974 // With PIC, the address is actually $g + Offset.
11975 if (isGlobalRelativeToPICBase(OpFlags)) {
11976 Result = DAG.getNode(ISD::ADD, dl, PtrVT,
11977 DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
11980 // For globals that require a load from a stub to get the address, emit the
11982 if (isGlobalStubReference(OpFlags))
11983 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
11984 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
11985 false, false, false, 0);
11987 // If there was a non-zero offset that we didn't fold, create an explicit
11988 // addition for it.
11990 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result,
11991 DAG.getConstant(Offset, dl, PtrVT));
11997 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
11998 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
11999 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
12000 return LowerGlobalAddress(GV, SDLoc(Op), Offset, DAG);
12004 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
12005 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
12006 unsigned char OperandFlags, bool LocalDynamic = false) {
12007 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
12008 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
12010 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
12011 GA->getValueType(0),
12015 X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR
12019 SDValue Ops[] = { Chain, TGA, *InFlag };
12020 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
12022 SDValue Ops[] = { Chain, TGA };
12023 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
12026 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
12027 MFI->setAdjustsStack(true);
12028 MFI->setHasCalls(true);
12030 SDValue Flag = Chain.getValue(1);
12031 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
12034 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
12036 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
12039 SDLoc dl(GA); // ? function entry point might be better
12040 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
12041 DAG.getNode(X86ISD::GlobalBaseReg,
12042 SDLoc(), PtrVT), InFlag);
12043 InFlag = Chain.getValue(1);
12045 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
12048 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
12050 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
12052 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT,
12053 X86::RAX, X86II::MO_TLSGD);
12056 static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA,
12062 // Get the start address of the TLS block for this module.
12063 X86MachineFunctionInfo* MFI = DAG.getMachineFunction()
12064 .getInfo<X86MachineFunctionInfo>();
12065 MFI->incNumLocalDynamicTLSAccesses();
12069 Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, X86::RAX,
12070 X86II::MO_TLSLD, /*LocalDynamic=*/true);
12073 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
12074 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), InFlag);
12075 InFlag = Chain.getValue(1);
12076 Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX,
12077 X86II::MO_TLSLDM, /*LocalDynamic=*/true);
12080 // Note: the CleanupLocalDynamicTLSPass will remove redundant computations
12084 unsigned char OperandFlags = X86II::MO_DTPOFF;
12085 unsigned WrapperKind = X86ISD::Wrapper;
12086 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
12087 GA->getValueType(0),
12088 GA->getOffset(), OperandFlags);
12089 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
12091 // Add x@dtpoff with the base.
12092 return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base);
12095 // Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model.
12096 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
12097 const EVT PtrVT, TLSModel::Model model,
12098 bool is64Bit, bool isPIC) {
12101 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
12102 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
12103 is64Bit ? 257 : 256));
12105 SDValue ThreadPointer =
12106 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), DAG.getIntPtrConstant(0, dl),
12107 MachinePointerInfo(Ptr), false, false, false, 0);
12109 unsigned char OperandFlags = 0;
12110 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
12112 unsigned WrapperKind = X86ISD::Wrapper;
12113 if (model == TLSModel::LocalExec) {
12114 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
12115 } else if (model == TLSModel::InitialExec) {
12117 OperandFlags = X86II::MO_GOTTPOFF;
12118 WrapperKind = X86ISD::WrapperRIP;
12120 OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF;
12123 llvm_unreachable("Unexpected model");
12126 // emit "addl x@ntpoff,%eax" (local exec)
12127 // or "addl x@indntpoff,%eax" (initial exec)
12128 // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic)
12130 DAG.getTargetGlobalAddress(GA->getGlobal(), dl, GA->getValueType(0),
12131 GA->getOffset(), OperandFlags);
12132 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
12134 if (model == TLSModel::InitialExec) {
12135 if (isPIC && !is64Bit) {
12136 Offset = DAG.getNode(ISD::ADD, dl, PtrVT,
12137 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
12141 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
12142 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
12143 false, false, false, 0);
12146 // The address of the thread local variable is the add of the thread
12147 // pointer with the offset of the variable.
12148 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
12152 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
12154 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
12155 const GlobalValue *GV = GA->getGlobal();
12156 auto PtrVT = getPointerTy(DAG.getDataLayout());
12158 if (Subtarget->isTargetELF()) {
12159 if (DAG.getTarget().Options.EmulatedTLS)
12160 return LowerToTLSEmulatedModel(GA, DAG);
12161 TLSModel::Model model = DAG.getTarget().getTLSModel(GV);
12163 case TLSModel::GeneralDynamic:
12164 if (Subtarget->is64Bit())
12165 return LowerToTLSGeneralDynamicModel64(GA, DAG, PtrVT);
12166 return LowerToTLSGeneralDynamicModel32(GA, DAG, PtrVT);
12167 case TLSModel::LocalDynamic:
12168 return LowerToTLSLocalDynamicModel(GA, DAG, PtrVT,
12169 Subtarget->is64Bit());
12170 case TLSModel::InitialExec:
12171 case TLSModel::LocalExec:
12172 return LowerToTLSExecModel(GA, DAG, PtrVT, model, Subtarget->is64Bit(),
12173 DAG.getTarget().getRelocationModel() ==
12176 llvm_unreachable("Unknown TLS model.");
12179 if (Subtarget->isTargetDarwin()) {
12180 // Darwin only has one model of TLS. Lower to that.
12181 unsigned char OpFlag = 0;
12182 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
12183 X86ISD::WrapperRIP : X86ISD::Wrapper;
12185 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
12186 // global base reg.
12187 bool PIC32 = (DAG.getTarget().getRelocationModel() == Reloc::PIC_) &&
12188 !Subtarget->is64Bit();
12190 OpFlag = X86II::MO_TLVP_PIC_BASE;
12192 OpFlag = X86II::MO_TLVP;
12194 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
12195 GA->getValueType(0),
12196 GA->getOffset(), OpFlag);
12197 SDValue Offset = DAG.getNode(WrapperKind, DL, PtrVT, Result);
12199 // With PIC32, the address is actually $g + Offset.
12201 Offset = DAG.getNode(ISD::ADD, DL, PtrVT,
12202 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
12205 // Lowering the machine isd will make sure everything is in the right
12207 SDValue Chain = DAG.getEntryNode();
12208 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
12209 SDValue Args[] = { Chain, Offset };
12210 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args);
12212 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
12213 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
12214 MFI->setAdjustsStack(true);
12216 // And our return value (tls address) is in the standard call return value
12218 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
12219 return DAG.getCopyFromReg(Chain, DL, Reg, PtrVT, Chain.getValue(1));
12222 if (Subtarget->isTargetKnownWindowsMSVC() ||
12223 Subtarget->isTargetWindowsGNU()) {
12224 // Just use the implicit TLS architecture
12225 // Need to generate someting similar to:
12226 // mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage
12228 // mov ecx, dword [rel _tls_index]: Load index (from C runtime)
12229 // mov rcx, qword [rdx+rcx*8]
12230 // mov eax, .tls$:tlsvar
12231 // [rax+rcx] contains the address
12232 // Windows 64bit: gs:0x58
12233 // Windows 32bit: fs:__tls_array
12236 SDValue Chain = DAG.getEntryNode();
12238 // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or
12239 // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly
12240 // use its literal value of 0x2C.
12241 Value *Ptr = Constant::getNullValue(Subtarget->is64Bit()
12242 ? Type::getInt8PtrTy(*DAG.getContext(),
12244 : Type::getInt32PtrTy(*DAG.getContext(),
12247 SDValue TlsArray = Subtarget->is64Bit()
12248 ? DAG.getIntPtrConstant(0x58, dl)
12249 : (Subtarget->isTargetWindowsGNU()
12250 ? DAG.getIntPtrConstant(0x2C, dl)
12251 : DAG.getExternalSymbol("_tls_array", PtrVT));
12253 SDValue ThreadPointer =
12254 DAG.getLoad(PtrVT, dl, Chain, TlsArray, MachinePointerInfo(Ptr), false,
12258 if (GV->getThreadLocalMode() == GlobalVariable::LocalExecTLSModel) {
12259 res = ThreadPointer;
12261 // Load the _tls_index variable
12262 SDValue IDX = DAG.getExternalSymbol("_tls_index", PtrVT);
12263 if (Subtarget->is64Bit())
12264 IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, PtrVT, Chain, IDX,
12265 MachinePointerInfo(), MVT::i32, false, false,
12268 IDX = DAG.getLoad(PtrVT, dl, Chain, IDX, MachinePointerInfo(), false,
12271 auto &DL = DAG.getDataLayout();
12273 DAG.getConstant(Log2_64_Ceil(DL.getPointerSize()), dl, PtrVT);
12274 IDX = DAG.getNode(ISD::SHL, dl, PtrVT, IDX, Scale);
12276 res = DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, IDX);
12279 res = DAG.getLoad(PtrVT, dl, Chain, res, MachinePointerInfo(), false, false,
12282 // Get the offset of start of .tls section
12283 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
12284 GA->getValueType(0),
12285 GA->getOffset(), X86II::MO_SECREL);
12286 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, TGA);
12288 // The address of the thread local variable is the add of the thread
12289 // pointer with the offset of the variable.
12290 return DAG.getNode(ISD::ADD, dl, PtrVT, res, Offset);
12293 llvm_unreachable("TLS not implemented for this target.");
12296 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values
12297 /// and take a 2 x i32 value to shift plus a shift amount.
12298 static SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) {
12299 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
12300 MVT VT = Op.getSimpleValueType();
12301 unsigned VTBits = VT.getSizeInBits();
12303 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
12304 SDValue ShOpLo = Op.getOperand(0);
12305 SDValue ShOpHi = Op.getOperand(1);
12306 SDValue ShAmt = Op.getOperand(2);
12307 // X86ISD::SHLD and X86ISD::SHRD have defined overflow behavior but the
12308 // generic ISD nodes haven't. Insert an AND to be safe, it's optimized away
12310 SDValue SafeShAmt = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
12311 DAG.getConstant(VTBits - 1, dl, MVT::i8));
12312 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
12313 DAG.getConstant(VTBits - 1, dl, MVT::i8))
12314 : DAG.getConstant(0, dl, VT);
12316 SDValue Tmp2, Tmp3;
12317 if (Op.getOpcode() == ISD::SHL_PARTS) {
12318 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
12319 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, SafeShAmt);
12321 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
12322 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, SafeShAmt);
12325 // If the shift amount is larger or equal than the width of a part we can't
12326 // rely on the results of shld/shrd. Insert a test and select the appropriate
12327 // values for large shift amounts.
12328 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
12329 DAG.getConstant(VTBits, dl, MVT::i8));
12330 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
12331 AndNode, DAG.getConstant(0, dl, MVT::i8));
12334 SDValue CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
12335 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
12336 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
12338 if (Op.getOpcode() == ISD::SHL_PARTS) {
12339 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
12340 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
12342 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
12343 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
12346 SDValue Ops[2] = { Lo, Hi };
12347 return DAG.getMergeValues(Ops, dl);
12350 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
12351 SelectionDAG &DAG) const {
12352 SDValue Src = Op.getOperand(0);
12353 MVT SrcVT = Src.getSimpleValueType();
12354 MVT VT = Op.getSimpleValueType();
12357 if (SrcVT.isVector()) {
12358 if (SrcVT == MVT::v2i32 && VT == MVT::v2f64) {
12359 return DAG.getNode(X86ISD::CVTDQ2PD, dl, VT,
12360 DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4i32, Src,
12361 DAG.getUNDEF(SrcVT)));
12363 if (SrcVT.getVectorElementType() == MVT::i1) {
12364 MVT IntegerVT = MVT::getVectorVT(MVT::i32, SrcVT.getVectorNumElements());
12365 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
12366 DAG.getNode(ISD::SIGN_EXTEND, dl, IntegerVT, Src));
12371 assert(SrcVT <= MVT::i64 && SrcVT >= MVT::i16 &&
12372 "Unknown SINT_TO_FP to lower!");
12374 // These are really Legal; return the operand so the caller accepts it as
12376 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
12378 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
12379 Subtarget->is64Bit()) {
12383 unsigned Size = SrcVT.getSizeInBits()/8;
12384 MachineFunction &MF = DAG.getMachineFunction();
12385 auto PtrVT = getPointerTy(MF.getDataLayout());
12386 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
12387 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12388 SDValue Chain = DAG.getStore(
12389 DAG.getEntryNode(), dl, Op.getOperand(0), StackSlot,
12390 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI), false,
12392 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
12395 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
12397 SelectionDAG &DAG) const {
12401 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
12403 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
12405 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
12407 unsigned ByteSize = SrcVT.getSizeInBits()/8;
12409 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
12410 MachineMemOperand *MMO;
12412 int SSFI = FI->getIndex();
12413 MMO = DAG.getMachineFunction().getMachineMemOperand(
12414 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12415 MachineMemOperand::MOLoad, ByteSize, ByteSize);
12417 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
12418 StackSlot = StackSlot.getOperand(1);
12420 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
12421 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
12423 Tys, Ops, SrcVT, MMO);
12426 Chain = Result.getValue(1);
12427 SDValue InFlag = Result.getValue(2);
12429 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
12430 // shouldn't be necessary except that RFP cannot be live across
12431 // multiple blocks. When stackifier is fixed, they can be uncoupled.
12432 MachineFunction &MF = DAG.getMachineFunction();
12433 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
12434 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
12435 auto PtrVT = getPointerTy(MF.getDataLayout());
12436 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12437 Tys = DAG.getVTList(MVT::Other);
12439 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
12441 MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand(
12442 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12443 MachineMemOperand::MOStore, SSFISize, SSFISize);
12445 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
12446 Ops, Op.getValueType(), MMO);
12447 Result = DAG.getLoad(
12448 Op.getValueType(), DL, Chain, StackSlot,
12449 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12450 false, false, false, 0);
12456 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
12457 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
12458 SelectionDAG &DAG) const {
12459 // This algorithm is not obvious. Here it is what we're trying to output:
12462 punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U }
12463 subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 }
12465 haddpd %xmm0, %xmm0
12467 pshufd $0x4e, %xmm0, %xmm1
12473 LLVMContext *Context = DAG.getContext();
12475 // Build some magic constants.
12476 static const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 };
12477 Constant *C0 = ConstantDataVector::get(*Context, CV0);
12478 auto PtrVT = getPointerTy(DAG.getDataLayout());
12479 SDValue CPIdx0 = DAG.getConstantPool(C0, PtrVT, 16);
12481 SmallVector<Constant*,2> CV1;
12483 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
12484 APInt(64, 0x4330000000000000ULL))));
12486 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
12487 APInt(64, 0x4530000000000000ULL))));
12488 Constant *C1 = ConstantVector::get(CV1);
12489 SDValue CPIdx1 = DAG.getConstantPool(C1, PtrVT, 16);
12491 // Load the 64-bit value into an XMM register.
12492 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
12495 DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
12496 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12497 false, false, false, 16);
12499 getUnpackl(DAG, dl, MVT::v4i32, DAG.getBitcast(MVT::v4i32, XR1), CLod0);
12502 DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
12503 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12504 false, false, false, 16);
12505 SDValue XR2F = DAG.getBitcast(MVT::v2f64, Unpck1);
12506 // TODO: Are there any fast-math-flags to propagate here?
12507 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
12510 if (Subtarget->hasSSE3()) {
12511 // FIXME: The 'haddpd' instruction may be slower than 'movhlps + addsd'.
12512 Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub);
12514 SDValue S2F = DAG.getBitcast(MVT::v4i32, Sub);
12515 SDValue Shuffle = getTargetShuffleNode(X86ISD::PSHUFD, dl, MVT::v4i32,
12517 Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64,
12518 DAG.getBitcast(MVT::v2f64, Shuffle), Sub);
12521 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result,
12522 DAG.getIntPtrConstant(0, dl));
12525 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
12526 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
12527 SelectionDAG &DAG) const {
12529 // FP constant to bias correct the final result.
12530 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
12533 // Load the 32-bit value into an XMM register.
12534 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
12537 // Zero out the upper parts of the register.
12538 Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG);
12540 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12541 DAG.getBitcast(MVT::v2f64, Load),
12542 DAG.getIntPtrConstant(0, dl));
12544 // Or the load with the bias.
12545 SDValue Or = DAG.getNode(
12546 ISD::OR, dl, MVT::v2i64,
12547 DAG.getBitcast(MVT::v2i64,
12548 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Load)),
12549 DAG.getBitcast(MVT::v2i64,
12550 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Bias)));
12552 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12553 DAG.getBitcast(MVT::v2f64, Or), DAG.getIntPtrConstant(0, dl));
12555 // Subtract the bias.
12556 // TODO: Are there any fast-math-flags to propagate here?
12557 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
12559 // Handle final rounding.
12560 MVT DestVT = Op.getSimpleValueType();
12562 if (DestVT.bitsLT(MVT::f64))
12563 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
12564 DAG.getIntPtrConstant(0, dl));
12565 if (DestVT.bitsGT(MVT::f64))
12566 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
12568 // Handle final rounding.
12572 static SDValue lowerUINT_TO_FP_vXi32(SDValue Op, SelectionDAG &DAG,
12573 const X86Subtarget &Subtarget) {
12574 // The algorithm is the following:
12575 // #ifdef __SSE4_1__
12576 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
12577 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
12578 // (uint4) 0x53000000, 0xaa);
12580 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
12581 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
12583 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
12584 // return (float4) lo + fhi;
12586 // We shouldn't use it when unsafe-fp-math is enabled though: we might later
12587 // reassociate the two FADDs, and if we do that, the algorithm fails
12588 // spectacularly (PR24512).
12589 // FIXME: If we ever have some kind of Machine FMF, this should be marked
12590 // as non-fast and always be enabled. Why isn't SDAG FMF enough? Because
12591 // there's also the MachineCombiner reassociations happening on Machine IR.
12592 if (DAG.getTarget().Options.UnsafeFPMath)
12596 SDValue V = Op->getOperand(0);
12597 MVT VecIntVT = V.getSimpleValueType();
12598 bool Is128 = VecIntVT == MVT::v4i32;
12599 MVT VecFloatVT = Is128 ? MVT::v4f32 : MVT::v8f32;
12600 // If we convert to something else than the supported type, e.g., to v4f64,
12602 if (VecFloatVT != Op->getSimpleValueType(0))
12605 unsigned NumElts = VecIntVT.getVectorNumElements();
12606 assert((VecIntVT == MVT::v4i32 || VecIntVT == MVT::v8i32) &&
12607 "Unsupported custom type");
12608 assert(NumElts <= 8 && "The size of the constant array must be fixed");
12610 // In the #idef/#else code, we have in common:
12611 // - The vector of constants:
12617 // Create the splat vector for 0x4b000000.
12618 SDValue CstLow = DAG.getConstant(0x4b000000, DL, MVT::i32);
12619 SDValue CstLowArray[] = {CstLow, CstLow, CstLow, CstLow,
12620 CstLow, CstLow, CstLow, CstLow};
12621 SDValue VecCstLow = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12622 makeArrayRef(&CstLowArray[0], NumElts));
12623 // Create the splat vector for 0x53000000.
12624 SDValue CstHigh = DAG.getConstant(0x53000000, DL, MVT::i32);
12625 SDValue CstHighArray[] = {CstHigh, CstHigh, CstHigh, CstHigh,
12626 CstHigh, CstHigh, CstHigh, CstHigh};
12627 SDValue VecCstHigh = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12628 makeArrayRef(&CstHighArray[0], NumElts));
12630 // Create the right shift.
12631 SDValue CstShift = DAG.getConstant(16, DL, MVT::i32);
12632 SDValue CstShiftArray[] = {CstShift, CstShift, CstShift, CstShift,
12633 CstShift, CstShift, CstShift, CstShift};
12634 SDValue VecCstShift = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12635 makeArrayRef(&CstShiftArray[0], NumElts));
12636 SDValue HighShift = DAG.getNode(ISD::SRL, DL, VecIntVT, V, VecCstShift);
12639 if (Subtarget.hasSSE41()) {
12640 MVT VecI16VT = Is128 ? MVT::v8i16 : MVT::v16i16;
12641 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
12642 SDValue VecCstLowBitcast = DAG.getBitcast(VecI16VT, VecCstLow);
12643 SDValue VecBitcast = DAG.getBitcast(VecI16VT, V);
12644 // Low will be bitcasted right away, so do not bother bitcasting back to its
12646 Low = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecBitcast,
12647 VecCstLowBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
12648 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
12649 // (uint4) 0x53000000, 0xaa);
12650 SDValue VecCstHighBitcast = DAG.getBitcast(VecI16VT, VecCstHigh);
12651 SDValue VecShiftBitcast = DAG.getBitcast(VecI16VT, HighShift);
12652 // High will be bitcasted right away, so do not bother bitcasting back to
12653 // its original type.
12654 High = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecShiftBitcast,
12655 VecCstHighBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
12657 SDValue CstMask = DAG.getConstant(0xffff, DL, MVT::i32);
12658 SDValue VecCstMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT, CstMask,
12659 CstMask, CstMask, CstMask);
12660 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
12661 SDValue LowAnd = DAG.getNode(ISD::AND, DL, VecIntVT, V, VecCstMask);
12662 Low = DAG.getNode(ISD::OR, DL, VecIntVT, LowAnd, VecCstLow);
12664 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
12665 High = DAG.getNode(ISD::OR, DL, VecIntVT, HighShift, VecCstHigh);
12668 // Create the vector constant for -(0x1.0p39f + 0x1.0p23f).
12669 SDValue CstFAdd = DAG.getConstantFP(
12670 APFloat(APFloat::IEEEsingle, APInt(32, 0xD3000080)), DL, MVT::f32);
12671 SDValue CstFAddArray[] = {CstFAdd, CstFAdd, CstFAdd, CstFAdd,
12672 CstFAdd, CstFAdd, CstFAdd, CstFAdd};
12673 SDValue VecCstFAdd = DAG.getNode(ISD::BUILD_VECTOR, DL, VecFloatVT,
12674 makeArrayRef(&CstFAddArray[0], NumElts));
12676 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
12677 SDValue HighBitcast = DAG.getBitcast(VecFloatVT, High);
12678 // TODO: Are there any fast-math-flags to propagate here?
12680 DAG.getNode(ISD::FADD, DL, VecFloatVT, HighBitcast, VecCstFAdd);
12681 // return (float4) lo + fhi;
12682 SDValue LowBitcast = DAG.getBitcast(VecFloatVT, Low);
12683 return DAG.getNode(ISD::FADD, DL, VecFloatVT, LowBitcast, FHigh);
12686 SDValue X86TargetLowering::lowerUINT_TO_FP_vec(SDValue Op,
12687 SelectionDAG &DAG) const {
12688 SDValue N0 = Op.getOperand(0);
12689 MVT SVT = N0.getSimpleValueType();
12692 switch (SVT.SimpleTy) {
12694 llvm_unreachable("Custom UINT_TO_FP is not supported!");
12699 MVT NVT = MVT::getVectorVT(MVT::i32, SVT.getVectorNumElements());
12700 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
12701 DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0));
12705 return lowerUINT_TO_FP_vXi32(Op, DAG, *Subtarget);
12708 assert(Subtarget->hasAVX512());
12709 return DAG.getNode(ISD::UINT_TO_FP, dl, Op.getValueType(),
12710 DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v16i32, N0));
12714 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
12715 SelectionDAG &DAG) const {
12716 SDValue N0 = Op.getOperand(0);
12718 auto PtrVT = getPointerTy(DAG.getDataLayout());
12720 if (Op.getSimpleValueType().isVector())
12721 return lowerUINT_TO_FP_vec(Op, DAG);
12723 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
12724 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
12725 // the optimization here.
12726 if (DAG.SignBitIsZero(N0))
12727 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
12729 MVT SrcVT = N0.getSimpleValueType();
12730 MVT DstVT = Op.getSimpleValueType();
12732 if (Subtarget->hasAVX512() && isScalarFPTypeInSSEReg(DstVT) &&
12733 (SrcVT == MVT::i32 || (SrcVT == MVT::i64 && Subtarget->is64Bit()))) {
12734 // Conversions from unsigned i32 to f32/f64 are legal,
12735 // using VCVTUSI2SS/SD. Same for i64 in 64-bit mode.
12739 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
12740 return LowerUINT_TO_FP_i64(Op, DAG);
12741 if (SrcVT == MVT::i32 && X86ScalarSSEf64)
12742 return LowerUINT_TO_FP_i32(Op, DAG);
12743 if (Subtarget->is64Bit() && SrcVT == MVT::i64 && DstVT == MVT::f32)
12746 // Make a 64-bit buffer, and use it to build an FILD.
12747 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
12748 if (SrcVT == MVT::i32) {
12749 SDValue WordOff = DAG.getConstant(4, dl, PtrVT);
12750 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, WordOff);
12751 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
12752 StackSlot, MachinePointerInfo(),
12754 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, dl, MVT::i32),
12755 OffsetSlot, MachinePointerInfo(),
12757 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
12761 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
12762 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
12763 StackSlot, MachinePointerInfo(),
12765 // For i64 source, we need to add the appropriate power of 2 if the input
12766 // was negative. This is the same as the optimization in
12767 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
12768 // we must be careful to do the computation in x87 extended precision, not
12769 // in SSE. (The generic code can't know it's OK to do this, or how to.)
12770 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
12771 MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand(
12772 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12773 MachineMemOperand::MOLoad, 8, 8);
12775 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
12776 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
12777 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops,
12780 APInt FF(32, 0x5F800000ULL);
12782 // Check whether the sign bit is set.
12783 SDValue SignSet = DAG.getSetCC(
12784 dl, getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i64),
12785 Op.getOperand(0), DAG.getConstant(0, dl, MVT::i64), ISD::SETLT);
12787 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
12788 SDValue FudgePtr = DAG.getConstantPool(
12789 ConstantInt::get(*DAG.getContext(), FF.zext(64)), PtrVT);
12791 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
12792 SDValue Zero = DAG.getIntPtrConstant(0, dl);
12793 SDValue Four = DAG.getIntPtrConstant(4, dl);
12794 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
12796 FudgePtr = DAG.getNode(ISD::ADD, dl, PtrVT, FudgePtr, Offset);
12798 // Load the value out, extending it from f32 to f80.
12799 // FIXME: Avoid the extend by constructing the right constant pool?
12800 SDValue Fudge = DAG.getExtLoad(
12801 ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(), FudgePtr,
12802 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), MVT::f32,
12803 false, false, false, 4);
12804 // Extend everything to 80 bits to force it to be done on x87.
12805 // TODO: Are there any fast-math-flags to propagate here?
12806 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
12807 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add,
12808 DAG.getIntPtrConstant(0, dl));
12811 // If the given FP_TO_SINT (IsSigned) or FP_TO_UINT (!IsSigned) operation
12812 // is legal, or has an fp128 or f16 source (which needs to be promoted to f32),
12813 // just return an <SDValue(), SDValue()> pair.
12814 // Otherwise it is assumed to be a conversion from one of f32, f64 or f80
12815 // to i16, i32 or i64, and we lower it to a legal sequence.
12816 // If lowered to the final integer result we return a <result, SDValue()> pair.
12817 // Otherwise we lower it to a sequence ending with a FIST, return a
12818 // <FIST, StackSlot> pair, and the caller is responsible for loading
12819 // the final integer result from StackSlot.
12820 std::pair<SDValue,SDValue>
12821 X86TargetLowering::FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
12822 bool IsSigned, bool IsReplace) const {
12825 EVT DstTy = Op.getValueType();
12826 EVT TheVT = Op.getOperand(0).getValueType();
12827 auto PtrVT = getPointerTy(DAG.getDataLayout());
12829 if (TheVT != MVT::f32 && TheVT != MVT::f64 && TheVT != MVT::f80) {
12830 // f16 must be promoted before using the lowering in this routine.
12831 // fp128 does not use this lowering.
12832 return std::make_pair(SDValue(), SDValue());
12835 // If using FIST to compute an unsigned i64, we'll need some fixup
12836 // to handle values above the maximum signed i64. A FIST is always
12837 // used for the 32-bit subtarget, but also for f80 on a 64-bit target.
12838 bool UnsignedFixup = !IsSigned &&
12839 DstTy == MVT::i64 &&
12840 (!Subtarget->is64Bit() ||
12841 !isScalarFPTypeInSSEReg(TheVT));
12843 if (!IsSigned && DstTy != MVT::i64 && !Subtarget->hasAVX512()) {
12844 // Replace the fp-to-uint32 operation with an fp-to-sint64 FIST.
12845 // The low 32 bits of the fist result will have the correct uint32 result.
12846 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
12850 assert(DstTy.getSimpleVT() <= MVT::i64 &&
12851 DstTy.getSimpleVT() >= MVT::i16 &&
12852 "Unknown FP_TO_INT to lower!");
12854 // These are really Legal.
12855 if (DstTy == MVT::i32 &&
12856 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
12857 return std::make_pair(SDValue(), SDValue());
12858 if (Subtarget->is64Bit() &&
12859 DstTy == MVT::i64 &&
12860 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
12861 return std::make_pair(SDValue(), SDValue());
12863 // We lower FP->int64 into FISTP64 followed by a load from a temporary
12865 MachineFunction &MF = DAG.getMachineFunction();
12866 unsigned MemSize = DstTy.getSizeInBits()/8;
12867 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
12868 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12871 switch (DstTy.getSimpleVT().SimpleTy) {
12872 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
12873 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
12874 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
12875 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
12878 SDValue Chain = DAG.getEntryNode();
12879 SDValue Value = Op.getOperand(0);
12880 SDValue Adjust; // 0x0 or 0x80000000, for result sign bit adjustment.
12882 if (UnsignedFixup) {
12884 // Conversion to unsigned i64 is implemented with a select,
12885 // depending on whether the source value fits in the range
12886 // of a signed i64. Let Thresh be the FP equivalent of
12887 // 0x8000000000000000ULL.
12889 // Adjust i32 = (Value < Thresh) ? 0 : 0x80000000;
12890 // FistSrc = (Value < Thresh) ? Value : (Value - Thresh);
12891 // Fist-to-mem64 FistSrc
12892 // Add 0 or 0x800...0ULL to the 64-bit result, which is equivalent
12893 // to XOR'ing the high 32 bits with Adjust.
12895 // Being a power of 2, Thresh is exactly representable in all FP formats.
12896 // For X87 we'd like to use the smallest FP type for this constant, but
12897 // for DAG type consistency we have to match the FP operand type.
12899 APFloat Thresh(APFloat::IEEEsingle, APInt(32, 0x5f000000));
12900 LLVM_ATTRIBUTE_UNUSED APFloat::opStatus Status = APFloat::opOK;
12901 bool LosesInfo = false;
12902 if (TheVT == MVT::f64)
12903 // The rounding mode is irrelevant as the conversion should be exact.
12904 Status = Thresh.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven,
12906 else if (TheVT == MVT::f80)
12907 Status = Thresh.convert(APFloat::x87DoubleExtended,
12908 APFloat::rmNearestTiesToEven, &LosesInfo);
12910 assert(Status == APFloat::opOK && !LosesInfo &&
12911 "FP conversion should have been exact");
12913 SDValue ThreshVal = DAG.getConstantFP(Thresh, DL, TheVT);
12915 SDValue Cmp = DAG.getSetCC(DL,
12916 getSetCCResultType(DAG.getDataLayout(),
12917 *DAG.getContext(), TheVT),
12918 Value, ThreshVal, ISD::SETLT);
12919 Adjust = DAG.getSelect(DL, MVT::i32, Cmp,
12920 DAG.getConstant(0, DL, MVT::i32),
12921 DAG.getConstant(0x80000000, DL, MVT::i32));
12922 SDValue Sub = DAG.getNode(ISD::FSUB, DL, TheVT, Value, ThreshVal);
12923 Cmp = DAG.getSetCC(DL, getSetCCResultType(DAG.getDataLayout(),
12924 *DAG.getContext(), TheVT),
12925 Value, ThreshVal, ISD::SETLT);
12926 Value = DAG.getSelect(DL, TheVT, Cmp, Value, Sub);
12929 // FIXME This causes a redundant load/store if the SSE-class value is already
12930 // in memory, such as if it is on the callstack.
12931 if (isScalarFPTypeInSSEReg(TheVT)) {
12932 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
12933 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
12934 MachinePointerInfo::getFixedStack(MF, SSFI), false,
12936 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
12938 Chain, StackSlot, DAG.getValueType(TheVT)
12941 MachineMemOperand *MMO =
12942 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
12943 MachineMemOperand::MOLoad, MemSize, MemSize);
12944 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, DstTy, MMO);
12945 Chain = Value.getValue(1);
12946 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
12947 StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12950 MachineMemOperand *MMO =
12951 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
12952 MachineMemOperand::MOStore, MemSize, MemSize);
12954 if (UnsignedFixup) {
12956 // Insert the FIST, load its result as two i32's,
12957 // and XOR the high i32 with Adjust.
12959 SDValue FistOps[] = { Chain, Value, StackSlot };
12960 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
12961 FistOps, DstTy, MMO);
12963 SDValue Low32 = DAG.getLoad(MVT::i32, DL, FIST, StackSlot,
12964 MachinePointerInfo(),
12965 false, false, false, 0);
12966 SDValue HighAddr = DAG.getNode(ISD::ADD, DL, PtrVT, StackSlot,
12967 DAG.getConstant(4, DL, PtrVT));
12969 SDValue High32 = DAG.getLoad(MVT::i32, DL, FIST, HighAddr,
12970 MachinePointerInfo(),
12971 false, false, false, 0);
12972 High32 = DAG.getNode(ISD::XOR, DL, MVT::i32, High32, Adjust);
12974 if (Subtarget->is64Bit()) {
12975 // Join High32 and Low32 into a 64-bit result.
12976 // (High32 << 32) | Low32
12977 Low32 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, Low32);
12978 High32 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, High32);
12979 High32 = DAG.getNode(ISD::SHL, DL, MVT::i64, High32,
12980 DAG.getConstant(32, DL, MVT::i8));
12981 SDValue Result = DAG.getNode(ISD::OR, DL, MVT::i64, High32, Low32);
12982 return std::make_pair(Result, SDValue());
12985 SDValue ResultOps[] = { Low32, High32 };
12987 SDValue pair = IsReplace
12988 ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, ResultOps)
12989 : DAG.getMergeValues(ResultOps, DL);
12990 return std::make_pair(pair, SDValue());
12992 // Build the FP_TO_INT*_IN_MEM
12993 SDValue Ops[] = { Chain, Value, StackSlot };
12994 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
12996 return std::make_pair(FIST, StackSlot);
13000 static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
13001 const X86Subtarget *Subtarget) {
13002 MVT VT = Op->getSimpleValueType(0);
13003 SDValue In = Op->getOperand(0);
13004 MVT InVT = In.getSimpleValueType();
13007 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
13008 return DAG.getNode(ISD::ZERO_EXTEND, dl, VT, In);
13010 // Optimize vectors in AVX mode:
13013 // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32.
13014 // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32.
13015 // Concat upper and lower parts.
13018 // Use vpunpckldq for 4 lower elements v4i32 -> v2i64.
13019 // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64.
13020 // Concat upper and lower parts.
13023 if (((VT != MVT::v16i16) || (InVT != MVT::v16i8)) &&
13024 ((VT != MVT::v8i32) || (InVT != MVT::v8i16)) &&
13025 ((VT != MVT::v4i64) || (InVT != MVT::v4i32)))
13028 if (Subtarget->hasInt256())
13029 return DAG.getNode(X86ISD::VZEXT, dl, VT, In);
13031 SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl);
13032 SDValue Undef = DAG.getUNDEF(InVT);
13033 bool NeedZero = Op.getOpcode() == ISD::ZERO_EXTEND;
13034 SDValue OpLo = getUnpackl(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
13035 SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
13037 MVT HVT = MVT::getVectorVT(VT.getVectorElementType(),
13038 VT.getVectorNumElements()/2);
13040 OpLo = DAG.getBitcast(HVT, OpLo);
13041 OpHi = DAG.getBitcast(HVT, OpHi);
13043 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
13046 static SDValue LowerZERO_EXTEND_AVX512(SDValue Op,
13047 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
13048 MVT VT = Op->getSimpleValueType(0);
13049 SDValue In = Op->getOperand(0);
13050 MVT InVT = In.getSimpleValueType();
13052 unsigned int NumElts = VT.getVectorNumElements();
13053 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
13056 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1)
13057 return DAG.getNode(X86ISD::VZEXT, DL, VT, In);
13059 assert(InVT.getVectorElementType() == MVT::i1);
13060 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
13062 DAG.getConstant(APInt(ExtVT.getScalarSizeInBits(), 1), DL, ExtVT);
13064 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), DL, ExtVT);
13066 SDValue V = DAG.getNode(ISD::VSELECT, DL, ExtVT, In, One, Zero);
13067 if (VT.is512BitVector())
13069 return DAG.getNode(X86ISD::VTRUNC, DL, VT, V);
13072 static SDValue LowerANY_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
13073 SelectionDAG &DAG) {
13074 if (Subtarget->hasFp256())
13075 if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget))
13081 static SDValue LowerZERO_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
13082 SelectionDAG &DAG) {
13084 MVT VT = Op.getSimpleValueType();
13085 SDValue In = Op.getOperand(0);
13086 MVT SVT = In.getSimpleValueType();
13088 if (VT.is512BitVector() || SVT.getVectorElementType() == MVT::i1)
13089 return LowerZERO_EXTEND_AVX512(Op, Subtarget, DAG);
13091 if (Subtarget->hasFp256())
13092 if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget))
13095 assert(!VT.is256BitVector() || !SVT.is128BitVector() ||
13096 VT.getVectorNumElements() != SVT.getVectorNumElements());
13100 SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
13102 MVT VT = Op.getSimpleValueType();
13103 SDValue In = Op.getOperand(0);
13104 MVT InVT = In.getSimpleValueType();
13106 if (VT == MVT::i1) {
13107 assert((InVT.isInteger() && (InVT.getSizeInBits() <= 64)) &&
13108 "Invalid scalar TRUNCATE operation");
13109 if (InVT.getSizeInBits() >= 32)
13111 In = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, In);
13112 return DAG.getNode(ISD::TRUNCATE, DL, VT, In);
13114 assert(VT.getVectorNumElements() == InVT.getVectorNumElements() &&
13115 "Invalid TRUNCATE operation");
13117 // move vector to mask - truncate solution for SKX
13118 if (VT.getVectorElementType() == MVT::i1) {
13119 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() <= 16 &&
13120 Subtarget->hasBWI())
13121 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
13122 if ((InVT.is256BitVector() || InVT.is128BitVector())
13123 && InVT.getScalarSizeInBits() <= 16 &&
13124 Subtarget->hasBWI() && Subtarget->hasVLX())
13125 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
13126 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() >= 32 &&
13127 Subtarget->hasDQI())
13128 return Op; // legal, will go to VPMOVD2M, VPMOVQ2M
13129 if ((InVT.is256BitVector() || InVT.is128BitVector())
13130 && InVT.getScalarSizeInBits() >= 32 &&
13131 Subtarget->hasDQI() && Subtarget->hasVLX())
13132 return Op; // legal, will go to VPMOVB2M, VPMOVQ2M
13135 if (VT.getVectorElementType() == MVT::i1) {
13136 assert(VT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
13137 unsigned NumElts = InVT.getVectorNumElements();
13138 assert ((NumElts == 8 || NumElts == 16) && "Unexpected vector type");
13139 if (InVT.getSizeInBits() < 512) {
13140 MVT ExtVT = (NumElts == 16)? MVT::v16i32 : MVT::v8i64;
13141 In = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, In);
13146 DAG.getConstant(APInt::getSignBit(InVT.getScalarSizeInBits()), DL, InVT);
13147 SDValue And = DAG.getNode(ISD::AND, DL, InVT, OneV, In);
13148 return DAG.getNode(X86ISD::TESTM, DL, VT, And, And);
13151 // vpmovqb/w/d, vpmovdb/w, vpmovwb
13152 if (Subtarget->hasAVX512()) {
13153 // word to byte only under BWI
13154 if (InVT == MVT::v16i16 && !Subtarget->hasBWI()) // v16i16 -> v16i8
13155 return DAG.getNode(X86ISD::VTRUNC, DL, VT,
13156 DAG.getNode(X86ISD::VSEXT, DL, MVT::v16i32, In));
13157 return DAG.getNode(X86ISD::VTRUNC, DL, VT, In);
13159 if ((VT == MVT::v4i32) && (InVT == MVT::v4i64)) {
13160 // On AVX2, v4i64 -> v4i32 becomes VPERMD.
13161 if (Subtarget->hasInt256()) {
13162 static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1};
13163 In = DAG.getBitcast(MVT::v8i32, In);
13164 In = DAG.getVectorShuffle(MVT::v8i32, DL, In, DAG.getUNDEF(MVT::v8i32),
13166 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, In,
13167 DAG.getIntPtrConstant(0, DL));
13170 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13171 DAG.getIntPtrConstant(0, DL));
13172 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13173 DAG.getIntPtrConstant(2, DL));
13174 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
13175 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
13176 static const int ShufMask[] = {0, 2, 4, 6};
13177 return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask);
13180 if ((VT == MVT::v8i16) && (InVT == MVT::v8i32)) {
13181 // On AVX2, v8i32 -> v8i16 becomed PSHUFB.
13182 if (Subtarget->hasInt256()) {
13183 In = DAG.getBitcast(MVT::v32i8, In);
13185 SmallVector<SDValue,32> pshufbMask;
13186 for (unsigned i = 0; i < 2; ++i) {
13187 pshufbMask.push_back(DAG.getConstant(0x0, DL, MVT::i8));
13188 pshufbMask.push_back(DAG.getConstant(0x1, DL, MVT::i8));
13189 pshufbMask.push_back(DAG.getConstant(0x4, DL, MVT::i8));
13190 pshufbMask.push_back(DAG.getConstant(0x5, DL, MVT::i8));
13191 pshufbMask.push_back(DAG.getConstant(0x8, DL, MVT::i8));
13192 pshufbMask.push_back(DAG.getConstant(0x9, DL, MVT::i8));
13193 pshufbMask.push_back(DAG.getConstant(0xc, DL, MVT::i8));
13194 pshufbMask.push_back(DAG.getConstant(0xd, DL, MVT::i8));
13195 for (unsigned j = 0; j < 8; ++j)
13196 pshufbMask.push_back(DAG.getConstant(0x80, DL, MVT::i8));
13198 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, pshufbMask);
13199 In = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8, In, BV);
13200 In = DAG.getBitcast(MVT::v4i64, In);
13202 static const int ShufMask[] = {0, 2, -1, -1};
13203 In = DAG.getVectorShuffle(MVT::v4i64, DL, In, DAG.getUNDEF(MVT::v4i64),
13205 In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13206 DAG.getIntPtrConstant(0, DL));
13207 return DAG.getBitcast(VT, In);
13210 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
13211 DAG.getIntPtrConstant(0, DL));
13213 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
13214 DAG.getIntPtrConstant(4, DL));
13216 OpLo = DAG.getBitcast(MVT::v16i8, OpLo);
13217 OpHi = DAG.getBitcast(MVT::v16i8, OpHi);
13219 // The PSHUFB mask:
13220 static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13,
13221 -1, -1, -1, -1, -1, -1, -1, -1};
13223 SDValue Undef = DAG.getUNDEF(MVT::v16i8);
13224 OpLo = DAG.getVectorShuffle(MVT::v16i8, DL, OpLo, Undef, ShufMask1);
13225 OpHi = DAG.getVectorShuffle(MVT::v16i8, DL, OpHi, Undef, ShufMask1);
13227 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
13228 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
13230 // The MOVLHPS Mask:
13231 static const int ShufMask2[] = {0, 1, 4, 5};
13232 SDValue res = DAG.getVectorShuffle(MVT::v4i32, DL, OpLo, OpHi, ShufMask2);
13233 return DAG.getBitcast(MVT::v8i16, res);
13236 // Handle truncation of V256 to V128 using shuffles.
13237 if (!VT.is128BitVector() || !InVT.is256BitVector())
13240 assert(Subtarget->hasFp256() && "256-bit vector without AVX!");
13242 unsigned NumElems = VT.getVectorNumElements();
13243 MVT NVT = MVT::getVectorVT(VT.getVectorElementType(), NumElems * 2);
13245 SmallVector<int, 16> MaskVec(NumElems * 2, -1);
13246 // Prepare truncation shuffle mask
13247 for (unsigned i = 0; i != NumElems; ++i)
13248 MaskVec[i] = i * 2;
13249 SDValue V = DAG.getVectorShuffle(NVT, DL, DAG.getBitcast(NVT, In),
13250 DAG.getUNDEF(NVT), &MaskVec[0]);
13251 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V,
13252 DAG.getIntPtrConstant(0, DL));
13255 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
13256 SelectionDAG &DAG) const {
13257 assert(!Op.getSimpleValueType().isVector());
13259 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
13260 /*IsSigned=*/ true, /*IsReplace=*/ false);
13261 SDValue FIST = Vals.first, StackSlot = Vals.second;
13262 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
13263 if (!FIST.getNode())
13266 if (StackSlot.getNode())
13267 // Load the result.
13268 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
13269 FIST, StackSlot, MachinePointerInfo(),
13270 false, false, false, 0);
13272 // The node is the result.
13276 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
13277 SelectionDAG &DAG) const {
13278 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
13279 /*IsSigned=*/ false, /*IsReplace=*/ false);
13280 SDValue FIST = Vals.first, StackSlot = Vals.second;
13281 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
13282 if (!FIST.getNode())
13285 if (StackSlot.getNode())
13286 // Load the result.
13287 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
13288 FIST, StackSlot, MachinePointerInfo(),
13289 false, false, false, 0);
13291 // The node is the result.
13295 static SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) {
13297 MVT VT = Op.getSimpleValueType();
13298 SDValue In = Op.getOperand(0);
13299 MVT SVT = In.getSimpleValueType();
13301 assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!");
13303 return DAG.getNode(X86ISD::VFPEXT, DL, VT,
13304 DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32,
13305 In, DAG.getUNDEF(SVT)));
13308 /// The only differences between FABS and FNEG are the mask and the logic op.
13309 /// FNEG also has a folding opportunity for FNEG(FABS(x)).
13310 static SDValue LowerFABSorFNEG(SDValue Op, SelectionDAG &DAG) {
13311 assert((Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG) &&
13312 "Wrong opcode for lowering FABS or FNEG.");
13314 bool IsFABS = (Op.getOpcode() == ISD::FABS);
13316 // If this is a FABS and it has an FNEG user, bail out to fold the combination
13317 // into an FNABS. We'll lower the FABS after that if it is still in use.
13319 for (SDNode *User : Op->uses())
13320 if (User->getOpcode() == ISD::FNEG)
13324 MVT VT = Op.getSimpleValueType();
13326 // FIXME: Use function attribute "OptimizeForSize" and/or CodeGenOpt::Level to
13327 // decide if we should generate a 16-byte constant mask when we only need 4 or
13328 // 8 bytes for the scalar case.
13334 if (VT.isVector()) {
13336 EltVT = VT.getVectorElementType();
13337 NumElts = VT.getVectorNumElements();
13339 // There are no scalar bitwise logical SSE/AVX instructions, so we
13340 // generate a 16-byte vector constant and logic op even for the scalar case.
13341 // Using a 16-byte mask allows folding the load of the mask with
13342 // the logic op, so it can save (~4 bytes) on code size.
13343 LogicVT = (VT == MVT::f64) ? MVT::v2f64 : MVT::v4f32;
13345 NumElts = (VT == MVT::f64) ? 2 : 4;
13348 unsigned EltBits = EltVT.getSizeInBits();
13349 LLVMContext *Context = DAG.getContext();
13350 // For FABS, mask is 0x7f...; for FNEG, mask is 0x80...
13352 IsFABS ? APInt::getSignedMaxValue(EltBits) : APInt::getSignBit(EltBits);
13353 Constant *C = ConstantInt::get(*Context, MaskElt);
13354 C = ConstantVector::getSplat(NumElts, C);
13355 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13356 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout()));
13357 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
13359 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13360 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13361 false, false, false, Alignment);
13363 SDValue Op0 = Op.getOperand(0);
13364 bool IsFNABS = !IsFABS && (Op0.getOpcode() == ISD::FABS);
13366 IsFABS ? X86ISD::FAND : IsFNABS ? X86ISD::FOR : X86ISD::FXOR;
13367 SDValue Operand = IsFNABS ? Op0.getOperand(0) : Op0;
13370 return DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask);
13372 // For the scalar case extend to a 128-bit vector, perform the logic op,
13373 // and extract the scalar result back out.
13374 Operand = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Operand);
13375 SDValue LogicNode = DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask);
13376 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, LogicNode,
13377 DAG.getIntPtrConstant(0, dl));
13380 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
13381 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13382 LLVMContext *Context = DAG.getContext();
13383 SDValue Op0 = Op.getOperand(0);
13384 SDValue Op1 = Op.getOperand(1);
13386 MVT VT = Op.getSimpleValueType();
13387 MVT SrcVT = Op1.getSimpleValueType();
13389 // If second operand is smaller, extend it first.
13390 if (SrcVT.bitsLT(VT)) {
13391 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
13394 // And if it is bigger, shrink it first.
13395 if (SrcVT.bitsGT(VT)) {
13396 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1, dl));
13400 // At this point the operands and the result should have the same
13401 // type, and that won't be f80 since that is not custom lowered.
13403 const fltSemantics &Sem =
13404 VT == MVT::f64 ? APFloat::IEEEdouble : APFloat::IEEEsingle;
13405 const unsigned SizeInBits = VT.getSizeInBits();
13407 SmallVector<Constant *, 4> CV(
13408 VT == MVT::f64 ? 2 : 4,
13409 ConstantFP::get(*Context, APFloat(Sem, APInt(SizeInBits, 0))));
13411 // First, clear all bits but the sign bit from the second operand (sign).
13412 CV[0] = ConstantFP::get(*Context,
13413 APFloat(Sem, APInt::getHighBitsSet(SizeInBits, 1)));
13414 Constant *C = ConstantVector::get(CV);
13415 auto PtrVT = TLI.getPointerTy(DAG.getDataLayout());
13416 SDValue CPIdx = DAG.getConstantPool(C, PtrVT, 16);
13418 // Perform all logic operations as 16-byte vectors because there are no
13419 // scalar FP logic instructions in SSE. This allows load folding of the
13420 // constants into the logic instructions.
13421 MVT LogicVT = (VT == MVT::f64) ? MVT::v2f64 : MVT::v4f32;
13423 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13424 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13425 false, false, false, 16);
13426 Op1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Op1);
13427 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, LogicVT, Op1, Mask1);
13429 // Next, clear the sign bit from the first operand (magnitude).
13430 // If it's a constant, we can clear it here.
13431 if (ConstantFPSDNode *Op0CN = dyn_cast<ConstantFPSDNode>(Op0)) {
13432 APFloat APF = Op0CN->getValueAPF();
13433 // If the magnitude is a positive zero, the sign bit alone is enough.
13434 if (APF.isPosZero())
13435 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcVT, SignBit,
13436 DAG.getIntPtrConstant(0, dl));
13438 CV[0] = ConstantFP::get(*Context, APF);
13440 CV[0] = ConstantFP::get(
13442 APFloat(Sem, APInt::getLowBitsSet(SizeInBits, SizeInBits - 1)));
13444 C = ConstantVector::get(CV);
13445 CPIdx = DAG.getConstantPool(C, PtrVT, 16);
13447 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13448 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13449 false, false, false, 16);
13450 // If the magnitude operand wasn't a constant, we need to AND out the sign.
13451 if (!isa<ConstantFPSDNode>(Op0)) {
13452 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Op0);
13453 Val = DAG.getNode(X86ISD::FAND, dl, LogicVT, Op0, Val);
13455 // OR the magnitude value with the sign bit.
13456 Val = DAG.getNode(X86ISD::FOR, dl, LogicVT, Val, SignBit);
13457 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcVT, Val,
13458 DAG.getIntPtrConstant(0, dl));
13461 static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) {
13462 SDValue N0 = Op.getOperand(0);
13464 MVT VT = Op.getSimpleValueType();
13466 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
13467 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
13468 DAG.getConstant(1, dl, VT));
13469 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, dl, VT));
13472 // Check whether an OR'd tree is PTEST-able.
13473 static SDValue LowerVectorAllZeroTest(SDValue Op, const X86Subtarget *Subtarget,
13474 SelectionDAG &DAG) {
13475 assert(Op.getOpcode() == ISD::OR && "Only check OR'd tree.");
13477 if (!Subtarget->hasSSE41())
13480 if (!Op->hasOneUse())
13483 SDNode *N = Op.getNode();
13486 SmallVector<SDValue, 8> Opnds;
13487 DenseMap<SDValue, unsigned> VecInMap;
13488 SmallVector<SDValue, 8> VecIns;
13489 EVT VT = MVT::Other;
13491 // Recognize a special case where a vector is casted into wide integer to
13493 Opnds.push_back(N->getOperand(0));
13494 Opnds.push_back(N->getOperand(1));
13496 for (unsigned Slot = 0, e = Opnds.size(); Slot < e; ++Slot) {
13497 SmallVectorImpl<SDValue>::const_iterator I = Opnds.begin() + Slot;
13498 // BFS traverse all OR'd operands.
13499 if (I->getOpcode() == ISD::OR) {
13500 Opnds.push_back(I->getOperand(0));
13501 Opnds.push_back(I->getOperand(1));
13502 // Re-evaluate the number of nodes to be traversed.
13503 e += 2; // 2 more nodes (LHS and RHS) are pushed.
13507 // Quit if a non-EXTRACT_VECTOR_ELT
13508 if (I->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
13511 // Quit if without a constant index.
13512 SDValue Idx = I->getOperand(1);
13513 if (!isa<ConstantSDNode>(Idx))
13516 SDValue ExtractedFromVec = I->getOperand(0);
13517 DenseMap<SDValue, unsigned>::iterator M = VecInMap.find(ExtractedFromVec);
13518 if (M == VecInMap.end()) {
13519 VT = ExtractedFromVec.getValueType();
13520 // Quit if not 128/256-bit vector.
13521 if (!VT.is128BitVector() && !VT.is256BitVector())
13523 // Quit if not the same type.
13524 if (VecInMap.begin() != VecInMap.end() &&
13525 VT != VecInMap.begin()->first.getValueType())
13527 M = VecInMap.insert(std::make_pair(ExtractedFromVec, 0)).first;
13528 VecIns.push_back(ExtractedFromVec);
13530 M->second |= 1U << cast<ConstantSDNode>(Idx)->getZExtValue();
13533 assert((VT.is128BitVector() || VT.is256BitVector()) &&
13534 "Not extracted from 128-/256-bit vector.");
13536 unsigned FullMask = (1U << VT.getVectorNumElements()) - 1U;
13538 for (DenseMap<SDValue, unsigned>::const_iterator
13539 I = VecInMap.begin(), E = VecInMap.end(); I != E; ++I) {
13540 // Quit if not all elements are used.
13541 if (I->second != FullMask)
13545 MVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
13547 // Cast all vectors into TestVT for PTEST.
13548 for (unsigned i = 0, e = VecIns.size(); i < e; ++i)
13549 VecIns[i] = DAG.getBitcast(TestVT, VecIns[i]);
13551 // If more than one full vectors are evaluated, OR them first before PTEST.
13552 for (unsigned Slot = 0, e = VecIns.size(); e - Slot > 1; Slot += 2, e += 1) {
13553 // Each iteration will OR 2 nodes and append the result until there is only
13554 // 1 node left, i.e. the final OR'd value of all vectors.
13555 SDValue LHS = VecIns[Slot];
13556 SDValue RHS = VecIns[Slot + 1];
13557 VecIns.push_back(DAG.getNode(ISD::OR, DL, TestVT, LHS, RHS));
13560 return DAG.getNode(X86ISD::PTEST, DL, MVT::i32,
13561 VecIns.back(), VecIns.back());
13564 /// \brief return true if \c Op has a use that doesn't just read flags.
13565 static bool hasNonFlagsUse(SDValue Op) {
13566 for (SDNode::use_iterator UI = Op->use_begin(), UE = Op->use_end(); UI != UE;
13568 SDNode *User = *UI;
13569 unsigned UOpNo = UI.getOperandNo();
13570 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
13571 // Look pass truncate.
13572 UOpNo = User->use_begin().getOperandNo();
13573 User = *User->use_begin();
13576 if (User->getOpcode() != ISD::BRCOND && User->getOpcode() != ISD::SETCC &&
13577 !(User->getOpcode() == ISD::SELECT && UOpNo == 0))
13583 /// Emit nodes that will be selected as "test Op0,Op0", or something
13585 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC, SDLoc dl,
13586 SelectionDAG &DAG) const {
13587 if (Op.getValueType() == MVT::i1) {
13588 SDValue ExtOp = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i8, Op);
13589 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, ExtOp,
13590 DAG.getConstant(0, dl, MVT::i8));
13592 // CF and OF aren't always set the way we want. Determine which
13593 // of these we need.
13594 bool NeedCF = false;
13595 bool NeedOF = false;
13598 case X86::COND_A: case X86::COND_AE:
13599 case X86::COND_B: case X86::COND_BE:
13602 case X86::COND_G: case X86::COND_GE:
13603 case X86::COND_L: case X86::COND_LE:
13604 case X86::COND_O: case X86::COND_NO: {
13605 // Check if we really need to set the
13606 // Overflow flag. If NoSignedWrap is present
13607 // that is not actually needed.
13608 switch (Op->getOpcode()) {
13613 const auto *BinNode = cast<BinaryWithFlagsSDNode>(Op.getNode());
13614 if (BinNode->Flags.hasNoSignedWrap())
13624 // See if we can use the EFLAGS value from the operand instead of
13625 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
13626 // we prove that the arithmetic won't overflow, we can't use OF or CF.
13627 if (Op.getResNo() != 0 || NeedOF || NeedCF) {
13628 // Emit a CMP with 0, which is the TEST pattern.
13629 //if (Op.getValueType() == MVT::i1)
13630 // return DAG.getNode(X86ISD::CMP, dl, MVT::i1, Op,
13631 // DAG.getConstant(0, MVT::i1));
13632 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
13633 DAG.getConstant(0, dl, Op.getValueType()));
13635 unsigned Opcode = 0;
13636 unsigned NumOperands = 0;
13638 // Truncate operations may prevent the merge of the SETCC instruction
13639 // and the arithmetic instruction before it. Attempt to truncate the operands
13640 // of the arithmetic instruction and use a reduced bit-width instruction.
13641 bool NeedTruncation = false;
13642 SDValue ArithOp = Op;
13643 if (Op->getOpcode() == ISD::TRUNCATE && Op->hasOneUse()) {
13644 SDValue Arith = Op->getOperand(0);
13645 // Both the trunc and the arithmetic op need to have one user each.
13646 if (Arith->hasOneUse())
13647 switch (Arith.getOpcode()) {
13654 NeedTruncation = true;
13660 // NOTICE: In the code below we use ArithOp to hold the arithmetic operation
13661 // which may be the result of a CAST. We use the variable 'Op', which is the
13662 // non-casted variable when we check for possible users.
13663 switch (ArithOp.getOpcode()) {
13665 // Due to an isel shortcoming, be conservative if this add is likely to be
13666 // selected as part of a load-modify-store instruction. When the root node
13667 // in a match is a store, isel doesn't know how to remap non-chain non-flag
13668 // uses of other nodes in the match, such as the ADD in this case. This
13669 // leads to the ADD being left around and reselected, with the result being
13670 // two adds in the output. Alas, even if none our users are stores, that
13671 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
13672 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
13673 // climbing the DAG back to the root, and it doesn't seem to be worth the
13675 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13676 UE = Op.getNode()->use_end(); UI != UE; ++UI)
13677 if (UI->getOpcode() != ISD::CopyToReg &&
13678 UI->getOpcode() != ISD::SETCC &&
13679 UI->getOpcode() != ISD::STORE)
13682 if (ConstantSDNode *C =
13683 dyn_cast<ConstantSDNode>(ArithOp.getNode()->getOperand(1))) {
13684 // An add of one will be selected as an INC.
13685 if (C->getAPIntValue() == 1 && !Subtarget->slowIncDec()) {
13686 Opcode = X86ISD::INC;
13691 // An add of negative one (subtract of one) will be selected as a DEC.
13692 if (C->getAPIntValue().isAllOnesValue() && !Subtarget->slowIncDec()) {
13693 Opcode = X86ISD::DEC;
13699 // Otherwise use a regular EFLAGS-setting add.
13700 Opcode = X86ISD::ADD;
13705 // If we have a constant logical shift that's only used in a comparison
13706 // against zero turn it into an equivalent AND. This allows turning it into
13707 // a TEST instruction later.
13708 if ((X86CC == X86::COND_E || X86CC == X86::COND_NE) && Op->hasOneUse() &&
13709 isa<ConstantSDNode>(Op->getOperand(1)) && !hasNonFlagsUse(Op)) {
13710 EVT VT = Op.getValueType();
13711 unsigned BitWidth = VT.getSizeInBits();
13712 unsigned ShAmt = Op->getConstantOperandVal(1);
13713 if (ShAmt >= BitWidth) // Avoid undefined shifts.
13715 APInt Mask = ArithOp.getOpcode() == ISD::SRL
13716 ? APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt)
13717 : APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt);
13718 if (!Mask.isSignedIntN(32)) // Avoid large immediates.
13720 SDValue New = DAG.getNode(ISD::AND, dl, VT, Op->getOperand(0),
13721 DAG.getConstant(Mask, dl, VT));
13722 DAG.ReplaceAllUsesWith(Op, New);
13728 // If the primary and result isn't used, don't bother using X86ISD::AND,
13729 // because a TEST instruction will be better.
13730 if (!hasNonFlagsUse(Op))
13736 // Due to the ISEL shortcoming noted above, be conservative if this op is
13737 // likely to be selected as part of a load-modify-store instruction.
13738 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13739 UE = Op.getNode()->use_end(); UI != UE; ++UI)
13740 if (UI->getOpcode() == ISD::STORE)
13743 // Otherwise use a regular EFLAGS-setting instruction.
13744 switch (ArithOp.getOpcode()) {
13745 default: llvm_unreachable("unexpected operator!");
13746 case ISD::SUB: Opcode = X86ISD::SUB; break;
13747 case ISD::XOR: Opcode = X86ISD::XOR; break;
13748 case ISD::AND: Opcode = X86ISD::AND; break;
13750 if (!NeedTruncation && (X86CC == X86::COND_E || X86CC == X86::COND_NE)) {
13751 SDValue EFLAGS = LowerVectorAllZeroTest(Op, Subtarget, DAG);
13752 if (EFLAGS.getNode())
13755 Opcode = X86ISD::OR;
13769 return SDValue(Op.getNode(), 1);
13775 // If we found that truncation is beneficial, perform the truncation and
13777 if (NeedTruncation) {
13778 EVT VT = Op.getValueType();
13779 SDValue WideVal = Op->getOperand(0);
13780 EVT WideVT = WideVal.getValueType();
13781 unsigned ConvertedOp = 0;
13782 // Use a target machine opcode to prevent further DAGCombine
13783 // optimizations that may separate the arithmetic operations
13784 // from the setcc node.
13785 switch (WideVal.getOpcode()) {
13787 case ISD::ADD: ConvertedOp = X86ISD::ADD; break;
13788 case ISD::SUB: ConvertedOp = X86ISD::SUB; break;
13789 case ISD::AND: ConvertedOp = X86ISD::AND; break;
13790 case ISD::OR: ConvertedOp = X86ISD::OR; break;
13791 case ISD::XOR: ConvertedOp = X86ISD::XOR; break;
13795 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13796 if (TLI.isOperationLegal(WideVal.getOpcode(), WideVT)) {
13797 SDValue V0 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(0));
13798 SDValue V1 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(1));
13799 Op = DAG.getNode(ConvertedOp, dl, VT, V0, V1);
13805 // Emit a CMP with 0, which is the TEST pattern.
13806 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
13807 DAG.getConstant(0, dl, Op.getValueType()));
13809 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
13810 SmallVector<SDValue, 4> Ops(Op->op_begin(), Op->op_begin() + NumOperands);
13812 SDValue New = DAG.getNode(Opcode, dl, VTs, Ops);
13813 DAG.ReplaceAllUsesWith(Op, New);
13814 return SDValue(New.getNode(), 1);
13817 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
13819 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
13820 SDLoc dl, SelectionDAG &DAG) const {
13821 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1)) {
13822 if (C->getAPIntValue() == 0)
13823 return EmitTest(Op0, X86CC, dl, DAG);
13825 assert(Op0.getValueType() != MVT::i1 &&
13826 "Unexpected comparison operation for MVT::i1 operands");
13829 if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 ||
13830 Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) {
13831 // Do the comparison at i32 if it's smaller, besides the Atom case.
13832 // This avoids subregister aliasing issues. Keep the smaller reference
13833 // if we're optimizing for size, however, as that'll allow better folding
13834 // of memory operations.
13835 if (Op0.getValueType() != MVT::i32 && Op0.getValueType() != MVT::i64 &&
13836 !DAG.getMachineFunction().getFunction()->optForMinSize() &&
13837 !Subtarget->isAtom()) {
13838 unsigned ExtendOp =
13839 isX86CCUnsigned(X86CC) ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
13840 Op0 = DAG.getNode(ExtendOp, dl, MVT::i32, Op0);
13841 Op1 = DAG.getNode(ExtendOp, dl, MVT::i32, Op1);
13843 // Use SUB instead of CMP to enable CSE between SUB and CMP.
13844 SDVTList VTs = DAG.getVTList(Op0.getValueType(), MVT::i32);
13845 SDValue Sub = DAG.getNode(X86ISD::SUB, dl, VTs,
13847 return SDValue(Sub.getNode(), 1);
13849 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
13852 /// Convert a comparison if required by the subtarget.
13853 SDValue X86TargetLowering::ConvertCmpIfNecessary(SDValue Cmp,
13854 SelectionDAG &DAG) const {
13855 // If the subtarget does not support the FUCOMI instruction, floating-point
13856 // comparisons have to be converted.
13857 if (Subtarget->hasCMov() ||
13858 Cmp.getOpcode() != X86ISD::CMP ||
13859 !Cmp.getOperand(0).getValueType().isFloatingPoint() ||
13860 !Cmp.getOperand(1).getValueType().isFloatingPoint())
13863 // The instruction selector will select an FUCOM instruction instead of
13864 // FUCOMI, which writes the comparison result to FPSW instead of EFLAGS. Hence
13865 // build an SDNode sequence that transfers the result from FPSW into EFLAGS:
13866 // (X86sahf (trunc (srl (X86fp_stsw (trunc (X86cmp ...)), 8))))
13868 SDValue TruncFPSW = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, Cmp);
13869 SDValue FNStSW = DAG.getNode(X86ISD::FNSTSW16r, dl, MVT::i16, TruncFPSW);
13870 SDValue Srl = DAG.getNode(ISD::SRL, dl, MVT::i16, FNStSW,
13871 DAG.getConstant(8, dl, MVT::i8));
13872 SDValue TruncSrl = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Srl);
13873 return DAG.getNode(X86ISD::SAHF, dl, MVT::i32, TruncSrl);
13876 /// The minimum architected relative accuracy is 2^-12. We need one
13877 /// Newton-Raphson step to have a good float result (24 bits of precision).
13878 SDValue X86TargetLowering::getRsqrtEstimate(SDValue Op,
13879 DAGCombinerInfo &DCI,
13880 unsigned &RefinementSteps,
13881 bool &UseOneConstNR) const {
13882 EVT VT = Op.getValueType();
13883 const char *RecipOp;
13885 // SSE1 has rsqrtss and rsqrtps. AVX adds a 256-bit variant for rsqrtps.
13886 // TODO: Add support for AVX512 (v16f32).
13887 // It is likely not profitable to do this for f64 because a double-precision
13888 // rsqrt estimate with refinement on x86 prior to FMA requires at least 16
13889 // instructions: convert to single, rsqrtss, convert back to double, refine
13890 // (3 steps = at least 13 insts). If an 'rsqrtsd' variant was added to the ISA
13891 // along with FMA, this could be a throughput win.
13892 if (VT == MVT::f32 && Subtarget->hasSSE1())
13894 else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
13895 (VT == MVT::v8f32 && Subtarget->hasAVX()))
13896 RecipOp = "vec-sqrtf";
13900 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
13901 if (!Recips.isEnabled(RecipOp))
13904 RefinementSteps = Recips.getRefinementSteps(RecipOp);
13905 UseOneConstNR = false;
13906 return DCI.DAG.getNode(X86ISD::FRSQRT, SDLoc(Op), VT, Op);
13909 /// The minimum architected relative accuracy is 2^-12. We need one
13910 /// Newton-Raphson step to have a good float result (24 bits of precision).
13911 SDValue X86TargetLowering::getRecipEstimate(SDValue Op,
13912 DAGCombinerInfo &DCI,
13913 unsigned &RefinementSteps) const {
13914 EVT VT = Op.getValueType();
13915 const char *RecipOp;
13917 // SSE1 has rcpss and rcpps. AVX adds a 256-bit variant for rcpps.
13918 // TODO: Add support for AVX512 (v16f32).
13919 // It is likely not profitable to do this for f64 because a double-precision
13920 // reciprocal estimate with refinement on x86 prior to FMA requires
13921 // 15 instructions: convert to single, rcpss, convert back to double, refine
13922 // (3 steps = 12 insts). If an 'rcpsd' variant was added to the ISA
13923 // along with FMA, this could be a throughput win.
13924 if (VT == MVT::f32 && Subtarget->hasSSE1())
13926 else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
13927 (VT == MVT::v8f32 && Subtarget->hasAVX()))
13928 RecipOp = "vec-divf";
13932 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
13933 if (!Recips.isEnabled(RecipOp))
13936 RefinementSteps = Recips.getRefinementSteps(RecipOp);
13937 return DCI.DAG.getNode(X86ISD::FRCP, SDLoc(Op), VT, Op);
13940 /// If we have at least two divisions that use the same divisor, convert to
13941 /// multplication by a reciprocal. This may need to be adjusted for a given
13942 /// CPU if a division's cost is not at least twice the cost of a multiplication.
13943 /// This is because we still need one division to calculate the reciprocal and
13944 /// then we need two multiplies by that reciprocal as replacements for the
13945 /// original divisions.
13946 unsigned X86TargetLowering::combineRepeatedFPDivisors() const {
13950 static bool isAllOnes(SDValue V) {
13951 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
13952 return C && C->isAllOnesValue();
13955 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
13956 /// if it's possible.
13957 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
13958 SDLoc dl, SelectionDAG &DAG) const {
13959 SDValue Op0 = And.getOperand(0);
13960 SDValue Op1 = And.getOperand(1);
13961 if (Op0.getOpcode() == ISD::TRUNCATE)
13962 Op0 = Op0.getOperand(0);
13963 if (Op1.getOpcode() == ISD::TRUNCATE)
13964 Op1 = Op1.getOperand(0);
13967 if (Op1.getOpcode() == ISD::SHL)
13968 std::swap(Op0, Op1);
13969 if (Op0.getOpcode() == ISD::SHL) {
13970 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
13971 if (And00C->getZExtValue() == 1) {
13972 // If we looked past a truncate, check that it's only truncating away
13974 unsigned BitWidth = Op0.getValueSizeInBits();
13975 unsigned AndBitWidth = And.getValueSizeInBits();
13976 if (BitWidth > AndBitWidth) {
13978 DAG.computeKnownBits(Op0, Zeros, Ones);
13979 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
13983 RHS = Op0.getOperand(1);
13985 } else if (Op1.getOpcode() == ISD::Constant) {
13986 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
13987 uint64_t AndRHSVal = AndRHS->getZExtValue();
13988 SDValue AndLHS = Op0;
13990 if (AndRHSVal == 1 && AndLHS.getOpcode() == ISD::SRL) {
13991 LHS = AndLHS.getOperand(0);
13992 RHS = AndLHS.getOperand(1);
13995 // Use BT if the immediate can't be encoded in a TEST instruction.
13996 if (!isUInt<32>(AndRHSVal) && isPowerOf2_64(AndRHSVal)) {
13998 RHS = DAG.getConstant(Log2_64_Ceil(AndRHSVal), dl, LHS.getValueType());
14002 if (LHS.getNode()) {
14003 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
14004 // instruction. Since the shift amount is in-range-or-undefined, we know
14005 // that doing a bittest on the i32 value is ok. We extend to i32 because
14006 // the encoding for the i16 version is larger than the i32 version.
14007 // Also promote i16 to i32 for performance / code size reason.
14008 if (LHS.getValueType() == MVT::i8 ||
14009 LHS.getValueType() == MVT::i16)
14010 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
14012 // If the operand types disagree, extend the shift amount to match. Since
14013 // BT ignores high bits (like shifts) we can use anyextend.
14014 if (LHS.getValueType() != RHS.getValueType())
14015 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
14017 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
14018 X86::CondCode Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
14019 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14020 DAG.getConstant(Cond, dl, MVT::i8), BT);
14026 /// \brief - Turns an ISD::CondCode into a value suitable for SSE floating point
14028 static int translateX86FSETCC(ISD::CondCode SetCCOpcode, SDValue &Op0,
14033 // SSE Condition code mapping:
14042 switch (SetCCOpcode) {
14043 default: llvm_unreachable("Unexpected SETCC condition");
14045 case ISD::SETEQ: SSECC = 0; break;
14047 case ISD::SETGT: Swap = true; // Fallthrough
14049 case ISD::SETOLT: SSECC = 1; break;
14051 case ISD::SETGE: Swap = true; // Fallthrough
14053 case ISD::SETOLE: SSECC = 2; break;
14054 case ISD::SETUO: SSECC = 3; break;
14056 case ISD::SETNE: SSECC = 4; break;
14057 case ISD::SETULE: Swap = true; // Fallthrough
14058 case ISD::SETUGE: SSECC = 5; break;
14059 case ISD::SETULT: Swap = true; // Fallthrough
14060 case ISD::SETUGT: SSECC = 6; break;
14061 case ISD::SETO: SSECC = 7; break;
14063 case ISD::SETONE: SSECC = 8; break;
14066 std::swap(Op0, Op1);
14071 // Lower256IntVSETCC - Break a VSETCC 256-bit integer VSETCC into two new 128
14072 // ones, and then concatenate the result back.
14073 static SDValue Lower256IntVSETCC(SDValue Op, SelectionDAG &DAG) {
14074 MVT VT = Op.getSimpleValueType();
14076 assert(VT.is256BitVector() && Op.getOpcode() == ISD::SETCC &&
14077 "Unsupported value type for operation");
14079 unsigned NumElems = VT.getVectorNumElements();
14081 SDValue CC = Op.getOperand(2);
14083 // Extract the LHS vectors
14084 SDValue LHS = Op.getOperand(0);
14085 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
14086 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
14088 // Extract the RHS vectors
14089 SDValue RHS = Op.getOperand(1);
14090 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
14091 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
14093 // Issue the operation on the smaller types and concatenate the result back
14094 MVT EltVT = VT.getVectorElementType();
14095 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
14096 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
14097 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC),
14098 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
14101 static SDValue LowerBoolVSETCC_AVX512(SDValue Op, SelectionDAG &DAG) {
14102 SDValue Op0 = Op.getOperand(0);
14103 SDValue Op1 = Op.getOperand(1);
14104 SDValue CC = Op.getOperand(2);
14105 MVT VT = Op.getSimpleValueType();
14108 assert(Op0.getSimpleValueType().getVectorElementType() == MVT::i1 &&
14109 "Unexpected type for boolean compare operation");
14110 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14111 SDValue NotOp0 = DAG.getNode(ISD::XOR, dl, VT, Op0,
14112 DAG.getConstant(-1, dl, VT));
14113 SDValue NotOp1 = DAG.getNode(ISD::XOR, dl, VT, Op1,
14114 DAG.getConstant(-1, dl, VT));
14115 switch (SetCCOpcode) {
14116 default: llvm_unreachable("Unexpected SETCC condition");
14118 // (x == y) -> ~(x ^ y)
14119 return DAG.getNode(ISD::XOR, dl, VT,
14120 DAG.getNode(ISD::XOR, dl, VT, Op0, Op1),
14121 DAG.getConstant(-1, dl, VT));
14123 // (x != y) -> (x ^ y)
14124 return DAG.getNode(ISD::XOR, dl, VT, Op0, Op1);
14127 // (x > y) -> (x & ~y)
14128 return DAG.getNode(ISD::AND, dl, VT, Op0, NotOp1);
14131 // (x < y) -> (~x & y)
14132 return DAG.getNode(ISD::AND, dl, VT, NotOp0, Op1);
14135 // (x <= y) -> (~x | y)
14136 return DAG.getNode(ISD::OR, dl, VT, NotOp0, Op1);
14139 // (x >=y) -> (x | ~y)
14140 return DAG.getNode(ISD::OR, dl, VT, Op0, NotOp1);
14144 static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG,
14145 const X86Subtarget *Subtarget) {
14146 SDValue Op0 = Op.getOperand(0);
14147 SDValue Op1 = Op.getOperand(1);
14148 SDValue CC = Op.getOperand(2);
14149 MVT VT = Op.getSimpleValueType();
14152 assert(Op0.getSimpleValueType().getVectorElementType().getSizeInBits() >= 8 &&
14153 Op.getSimpleValueType().getVectorElementType() == MVT::i1 &&
14154 "Cannot set masked compare for this operation");
14156 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14158 bool Unsigned = false;
14161 switch (SetCCOpcode) {
14162 default: llvm_unreachable("Unexpected SETCC condition");
14163 case ISD::SETNE: SSECC = 4; break;
14164 case ISD::SETEQ: Opc = X86ISD::PCMPEQM; break;
14165 case ISD::SETUGT: SSECC = 6; Unsigned = true; break;
14166 case ISD::SETLT: Swap = true; //fall-through
14167 case ISD::SETGT: Opc = X86ISD::PCMPGTM; break;
14168 case ISD::SETULT: SSECC = 1; Unsigned = true; break;
14169 case ISD::SETUGE: SSECC = 5; Unsigned = true; break; //NLT
14170 case ISD::SETGE: Swap = true; SSECC = 2; break; // LE + swap
14171 case ISD::SETULE: Unsigned = true; //fall-through
14172 case ISD::SETLE: SSECC = 2; break;
14176 std::swap(Op0, Op1);
14178 return DAG.getNode(Opc, dl, VT, Op0, Op1);
14179 Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM;
14180 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14181 DAG.getConstant(SSECC, dl, MVT::i8));
14184 /// \brief Try to turn a VSETULT into a VSETULE by modifying its second
14185 /// operand \p Op1. If non-trivial (for example because it's not constant)
14186 /// return an empty value.
14187 static SDValue ChangeVSETULTtoVSETULE(SDLoc dl, SDValue Op1, SelectionDAG &DAG)
14189 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Op1.getNode());
14193 MVT VT = Op1.getSimpleValueType();
14194 MVT EVT = VT.getVectorElementType();
14195 unsigned n = VT.getVectorNumElements();
14196 SmallVector<SDValue, 8> ULTOp1;
14198 for (unsigned i = 0; i < n; ++i) {
14199 ConstantSDNode *Elt = dyn_cast<ConstantSDNode>(BV->getOperand(i));
14200 if (!Elt || Elt->isOpaque() || Elt->getSimpleValueType(0) != EVT)
14203 // Avoid underflow.
14204 APInt Val = Elt->getAPIntValue();
14208 ULTOp1.push_back(DAG.getConstant(Val - 1, dl, EVT));
14211 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, ULTOp1);
14214 static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
14215 SelectionDAG &DAG) {
14216 SDValue Op0 = Op.getOperand(0);
14217 SDValue Op1 = Op.getOperand(1);
14218 SDValue CC = Op.getOperand(2);
14219 MVT VT = Op.getSimpleValueType();
14220 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14221 bool isFP = Op.getOperand(1).getSimpleValueType().isFloatingPoint();
14226 MVT EltVT = Op0.getSimpleValueType().getVectorElementType();
14227 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
14230 unsigned SSECC = translateX86FSETCC(SetCCOpcode, Op0, Op1);
14231 unsigned Opc = X86ISD::CMPP;
14232 if (Subtarget->hasAVX512() && VT.getVectorElementType() == MVT::i1) {
14233 assert(VT.getVectorNumElements() <= 16);
14234 Opc = X86ISD::CMPM;
14236 // In the two special cases we can't handle, emit two comparisons.
14239 unsigned CombineOpc;
14240 if (SetCCOpcode == ISD::SETUEQ) {
14241 CC0 = 3; CC1 = 0; CombineOpc = ISD::OR;
14243 assert(SetCCOpcode == ISD::SETONE);
14244 CC0 = 7; CC1 = 4; CombineOpc = ISD::AND;
14247 SDValue Cmp0 = DAG.getNode(Opc, dl, VT, Op0, Op1,
14248 DAG.getConstant(CC0, dl, MVT::i8));
14249 SDValue Cmp1 = DAG.getNode(Opc, dl, VT, Op0, Op1,
14250 DAG.getConstant(CC1, dl, MVT::i8));
14251 return DAG.getNode(CombineOpc, dl, VT, Cmp0, Cmp1);
14253 // Handle all other FP comparisons here.
14254 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14255 DAG.getConstant(SSECC, dl, MVT::i8));
14258 MVT VTOp0 = Op0.getSimpleValueType();
14259 assert(VTOp0 == Op1.getSimpleValueType() &&
14260 "Expected operands with same type!");
14261 assert(VT.getVectorNumElements() == VTOp0.getVectorNumElements() &&
14262 "Invalid number of packed elements for source and destination!");
14264 if (VT.is128BitVector() && VTOp0.is256BitVector()) {
14265 // On non-AVX512 targets, a vector of MVT::i1 is promoted by the type
14266 // legalizer to a wider vector type. In the case of 'vsetcc' nodes, the
14267 // legalizer firstly checks if the first operand in input to the setcc has
14268 // a legal type. If so, then it promotes the return type to that same type.
14269 // Otherwise, the return type is promoted to the 'next legal type' which,
14270 // for a vector of MVT::i1 is always a 128-bit integer vector type.
14272 // We reach this code only if the following two conditions are met:
14273 // 1. Both return type and operand type have been promoted to wider types
14274 // by the type legalizer.
14275 // 2. The original operand type has been promoted to a 256-bit vector.
14277 // Note that condition 2. only applies for AVX targets.
14278 SDValue NewOp = DAG.getSetCC(dl, VTOp0, Op0, Op1, SetCCOpcode);
14279 return DAG.getZExtOrTrunc(NewOp, dl, VT);
14282 // The non-AVX512 code below works under the assumption that source and
14283 // destination types are the same.
14284 assert((Subtarget->hasAVX512() || (VT == VTOp0)) &&
14285 "Value types for source and destination must be the same!");
14287 // Break 256-bit integer vector compare into smaller ones.
14288 if (VT.is256BitVector() && !Subtarget->hasInt256())
14289 return Lower256IntVSETCC(Op, DAG);
14291 MVT OpVT = Op1.getSimpleValueType();
14292 if (OpVT.getVectorElementType() == MVT::i1)
14293 return LowerBoolVSETCC_AVX512(Op, DAG);
14295 bool MaskResult = (VT.getVectorElementType() == MVT::i1);
14296 if (Subtarget->hasAVX512()) {
14297 if (Op1.getSimpleValueType().is512BitVector() ||
14298 (Subtarget->hasBWI() && Subtarget->hasVLX()) ||
14299 (MaskResult && OpVT.getVectorElementType().getSizeInBits() >= 32))
14300 return LowerIntVSETCC_AVX512(Op, DAG, Subtarget);
14302 // In AVX-512 architecture setcc returns mask with i1 elements,
14303 // But there is no compare instruction for i8 and i16 elements in KNL.
14304 // We are not talking about 512-bit operands in this case, these
14305 // types are illegal.
14307 (OpVT.getVectorElementType().getSizeInBits() < 32 &&
14308 OpVT.getVectorElementType().getSizeInBits() >= 8))
14309 return DAG.getNode(ISD::TRUNCATE, dl, VT,
14310 DAG.getNode(ISD::SETCC, dl, OpVT, Op0, Op1, CC));
14313 // Lower using XOP integer comparisons.
14314 if ((VT == MVT::v16i8 || VT == MVT::v8i16 ||
14315 VT == MVT::v4i32 || VT == MVT::v2i64) && Subtarget->hasXOP()) {
14316 // Translate compare code to XOP PCOM compare mode.
14317 unsigned CmpMode = 0;
14318 switch (SetCCOpcode) {
14319 default: llvm_unreachable("Unexpected SETCC condition");
14321 case ISD::SETLT: CmpMode = 0x00; break;
14323 case ISD::SETLE: CmpMode = 0x01; break;
14325 case ISD::SETGT: CmpMode = 0x02; break;
14327 case ISD::SETGE: CmpMode = 0x03; break;
14328 case ISD::SETEQ: CmpMode = 0x04; break;
14329 case ISD::SETNE: CmpMode = 0x05; break;
14332 // Are we comparing unsigned or signed integers?
14333 unsigned Opc = ISD::isUnsignedIntSetCC(SetCCOpcode)
14334 ? X86ISD::VPCOMU : X86ISD::VPCOM;
14336 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14337 DAG.getConstant(CmpMode, dl, MVT::i8));
14340 // We are handling one of the integer comparisons here. Since SSE only has
14341 // GT and EQ comparisons for integer, swapping operands and multiple
14342 // operations may be required for some comparisons.
14344 bool Swap = false, Invert = false, FlipSigns = false, MinMax = false;
14345 bool Subus = false;
14347 switch (SetCCOpcode) {
14348 default: llvm_unreachable("Unexpected SETCC condition");
14349 case ISD::SETNE: Invert = true;
14350 case ISD::SETEQ: Opc = X86ISD::PCMPEQ; break;
14351 case ISD::SETLT: Swap = true;
14352 case ISD::SETGT: Opc = X86ISD::PCMPGT; break;
14353 case ISD::SETGE: Swap = true;
14354 case ISD::SETLE: Opc = X86ISD::PCMPGT;
14355 Invert = true; break;
14356 case ISD::SETULT: Swap = true;
14357 case ISD::SETUGT: Opc = X86ISD::PCMPGT;
14358 FlipSigns = true; break;
14359 case ISD::SETUGE: Swap = true;
14360 case ISD::SETULE: Opc = X86ISD::PCMPGT;
14361 FlipSigns = true; Invert = true; break;
14364 // Special case: Use min/max operations for SETULE/SETUGE
14365 MVT VET = VT.getVectorElementType();
14367 (Subtarget->hasSSE41() && (VET >= MVT::i8 && VET <= MVT::i32))
14368 || (Subtarget->hasSSE2() && (VET == MVT::i8));
14371 switch (SetCCOpcode) {
14373 case ISD::SETULE: Opc = ISD::UMIN; MinMax = true; break;
14374 case ISD::SETUGE: Opc = ISD::UMAX; MinMax = true; break;
14377 if (MinMax) { Swap = false; Invert = false; FlipSigns = false; }
14380 bool hasSubus = Subtarget->hasSSE2() && (VET == MVT::i8 || VET == MVT::i16);
14381 if (!MinMax && hasSubus) {
14382 // As another special case, use PSUBUS[BW] when it's profitable. E.g. for
14384 // t = psubus Op0, Op1
14385 // pcmpeq t, <0..0>
14386 switch (SetCCOpcode) {
14388 case ISD::SETULT: {
14389 // If the comparison is against a constant we can turn this into a
14390 // setule. With psubus, setule does not require a swap. This is
14391 // beneficial because the constant in the register is no longer
14392 // destructed as the destination so it can be hoisted out of a loop.
14393 // Only do this pre-AVX since vpcmp* is no longer destructive.
14394 if (Subtarget->hasAVX())
14396 SDValue ULEOp1 = ChangeVSETULTtoVSETULE(dl, Op1, DAG);
14397 if (ULEOp1.getNode()) {
14399 Subus = true; Invert = false; Swap = false;
14403 // Psubus is better than flip-sign because it requires no inversion.
14404 case ISD::SETUGE: Subus = true; Invert = false; Swap = true; break;
14405 case ISD::SETULE: Subus = true; Invert = false; Swap = false; break;
14409 Opc = X86ISD::SUBUS;
14415 std::swap(Op0, Op1);
14417 // Check that the operation in question is available (most are plain SSE2,
14418 // but PCMPGTQ and PCMPEQQ have different requirements).
14419 if (VT == MVT::v2i64) {
14420 if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42()) {
14421 assert(Subtarget->hasSSE2() && "Don't know how to lower!");
14423 // First cast everything to the right type.
14424 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
14425 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
14427 // Since SSE has no unsigned integer comparisons, we need to flip the sign
14428 // bits of the inputs before performing those operations. The lower
14429 // compare is always unsigned.
14432 SB = DAG.getConstant(0x80000000U, dl, MVT::v4i32);
14434 SDValue Sign = DAG.getConstant(0x80000000U, dl, MVT::i32);
14435 SDValue Zero = DAG.getConstant(0x00000000U, dl, MVT::i32);
14436 SB = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
14437 Sign, Zero, Sign, Zero);
14439 Op0 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op0, SB);
14440 Op1 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op1, SB);
14442 // Emulate PCMPGTQ with (hi1 > hi2) | ((hi1 == hi2) & (lo1 > lo2))
14443 SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1);
14444 SDValue EQ = DAG.getNode(X86ISD::PCMPEQ, dl, MVT::v4i32, Op0, Op1);
14446 // Create masks for only the low parts/high parts of the 64 bit integers.
14447 static const int MaskHi[] = { 1, 1, 3, 3 };
14448 static const int MaskLo[] = { 0, 0, 2, 2 };
14449 SDValue EQHi = DAG.getVectorShuffle(MVT::v4i32, dl, EQ, EQ, MaskHi);
14450 SDValue GTLo = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskLo);
14451 SDValue GTHi = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi);
14453 SDValue Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, EQHi, GTLo);
14454 Result = DAG.getNode(ISD::OR, dl, MVT::v4i32, Result, GTHi);
14457 Result = DAG.getNOT(dl, Result, MVT::v4i32);
14459 return DAG.getBitcast(VT, Result);
14462 if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) {
14463 // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with
14464 // pcmpeqd + pshufd + pand.
14465 assert(Subtarget->hasSSE2() && !FlipSigns && "Don't know how to lower!");
14467 // First cast everything to the right type.
14468 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
14469 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
14472 SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1);
14474 // Make sure the lower and upper halves are both all-ones.
14475 static const int Mask[] = { 1, 0, 3, 2 };
14476 SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask);
14477 Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf);
14480 Result = DAG.getNOT(dl, Result, MVT::v4i32);
14482 return DAG.getBitcast(VT, Result);
14486 // Since SSE has no unsigned integer comparisons, we need to flip the sign
14487 // bits of the inputs before performing those operations.
14489 MVT EltVT = VT.getVectorElementType();
14490 SDValue SB = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()), dl,
14492 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SB);
14493 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SB);
14496 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
14498 // If the logical-not of the result is required, perform that now.
14500 Result = DAG.getNOT(dl, Result, VT);
14503 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Op0, Result);
14506 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Result,
14507 getZeroVector(VT, Subtarget, DAG, dl));
14512 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
14514 MVT VT = Op.getSimpleValueType();
14516 if (VT.isVector()) return LowerVSETCC(Op, Subtarget, DAG);
14518 assert(((!Subtarget->hasAVX512() && VT == MVT::i8) || (VT == MVT::i1))
14519 && "SetCC type must be 8-bit or 1-bit integer");
14520 SDValue Op0 = Op.getOperand(0);
14521 SDValue Op1 = Op.getOperand(1);
14523 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
14525 // Optimize to BT if possible.
14526 // Lower (X & (1 << N)) == 0 to BT(X, N).
14527 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
14528 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
14529 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
14530 Op1.getOpcode() == ISD::Constant &&
14531 cast<ConstantSDNode>(Op1)->isNullValue() &&
14532 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14533 if (SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG)) {
14535 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewSetCC);
14540 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
14542 if (Op1.getOpcode() == ISD::Constant &&
14543 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
14544 cast<ConstantSDNode>(Op1)->isNullValue()) &&
14545 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14547 // If the input is a setcc, then reuse the input setcc or use a new one with
14548 // the inverted condition.
14549 if (Op0.getOpcode() == X86ISD::SETCC) {
14550 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
14551 bool Invert = (CC == ISD::SETNE) ^
14552 cast<ConstantSDNode>(Op1)->isNullValue();
14556 CCode = X86::GetOppositeBranchCondition(CCode);
14557 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14558 DAG.getConstant(CCode, dl, MVT::i8),
14559 Op0.getOperand(1));
14561 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
14565 if ((Op0.getValueType() == MVT::i1) && (Op1.getOpcode() == ISD::Constant) &&
14566 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1) &&
14567 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14569 ISD::CondCode NewCC = ISD::getSetCCInverse(CC, true);
14570 return DAG.getSetCC(dl, VT, Op0, DAG.getConstant(0, dl, MVT::i1), NewCC);
14573 bool isFP = Op1.getSimpleValueType().isFloatingPoint();
14574 unsigned X86CC = TranslateX86CC(CC, dl, isFP, Op0, Op1, DAG);
14575 if (X86CC == X86::COND_INVALID)
14578 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, dl, DAG);
14579 EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
14580 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14581 DAG.getConstant(X86CC, dl, MVT::i8), EFLAGS);
14583 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
14587 SDValue X86TargetLowering::LowerSETCCE(SDValue Op, SelectionDAG &DAG) const {
14588 SDValue LHS = Op.getOperand(0);
14589 SDValue RHS = Op.getOperand(1);
14590 SDValue Carry = Op.getOperand(2);
14591 SDValue Cond = Op.getOperand(3);
14594 assert(LHS.getSimpleValueType().isInteger() && "SETCCE is integer only.");
14595 X86::CondCode CC = TranslateIntegerX86CC(cast<CondCodeSDNode>(Cond)->get());
14597 assert(Carry.getOpcode() != ISD::CARRY_FALSE);
14598 SDVTList VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
14599 SDValue Cmp = DAG.getNode(X86ISD::SBB, DL, VTs, LHS, RHS, Carry);
14600 return DAG.getNode(X86ISD::SETCC, DL, Op.getValueType(),
14601 DAG.getConstant(CC, DL, MVT::i8), Cmp.getValue(1));
14604 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
14605 static bool isX86LogicalCmp(SDValue Op) {
14606 unsigned Opc = Op.getNode()->getOpcode();
14607 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI ||
14608 Opc == X86ISD::SAHF)
14610 if (Op.getResNo() == 1 &&
14611 (Opc == X86ISD::ADD ||
14612 Opc == X86ISD::SUB ||
14613 Opc == X86ISD::ADC ||
14614 Opc == X86ISD::SBB ||
14615 Opc == X86ISD::SMUL ||
14616 Opc == X86ISD::UMUL ||
14617 Opc == X86ISD::INC ||
14618 Opc == X86ISD::DEC ||
14619 Opc == X86ISD::OR ||
14620 Opc == X86ISD::XOR ||
14621 Opc == X86ISD::AND))
14624 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
14630 static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) {
14631 if (V.getOpcode() != ISD::TRUNCATE)
14634 SDValue VOp0 = V.getOperand(0);
14635 unsigned InBits = VOp0.getValueSizeInBits();
14636 unsigned Bits = V.getValueSizeInBits();
14637 return DAG.MaskedValueIsZero(VOp0, APInt::getHighBitsSet(InBits,InBits-Bits));
14640 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
14641 bool addTest = true;
14642 SDValue Cond = Op.getOperand(0);
14643 SDValue Op1 = Op.getOperand(1);
14644 SDValue Op2 = Op.getOperand(2);
14646 MVT VT = Op1.getSimpleValueType();
14649 // Lower FP selects into a CMP/AND/ANDN/OR sequence when the necessary SSE ops
14650 // are available or VBLENDV if AVX is available.
14651 // Otherwise FP cmovs get lowered into a less efficient branch sequence later.
14652 if (Cond.getOpcode() == ISD::SETCC &&
14653 ((Subtarget->hasSSE2() && (VT == MVT::f32 || VT == MVT::f64)) ||
14654 (Subtarget->hasSSE1() && VT == MVT::f32)) &&
14655 VT == Cond.getOperand(0).getSimpleValueType() && Cond->hasOneUse()) {
14656 SDValue CondOp0 = Cond.getOperand(0), CondOp1 = Cond.getOperand(1);
14657 int SSECC = translateX86FSETCC(
14658 cast<CondCodeSDNode>(Cond.getOperand(2))->get(), CondOp0, CondOp1);
14661 if (Subtarget->hasAVX512()) {
14662 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CondOp0, CondOp1,
14663 DAG.getConstant(SSECC, DL, MVT::i8));
14664 return DAG.getNode(X86ISD::SELECT, DL, VT, Cmp, Op1, Op2);
14667 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, VT, CondOp0, CondOp1,
14668 DAG.getConstant(SSECC, DL, MVT::i8));
14670 // If we have AVX, we can use a variable vector select (VBLENDV) instead
14671 // of 3 logic instructions for size savings and potentially speed.
14672 // Unfortunately, there is no scalar form of VBLENDV.
14674 // If either operand is a constant, don't try this. We can expect to
14675 // optimize away at least one of the logic instructions later in that
14676 // case, so that sequence would be faster than a variable blend.
14678 // BLENDV was introduced with SSE 4.1, but the 2 register form implicitly
14679 // uses XMM0 as the selection register. That may need just as many
14680 // instructions as the AND/ANDN/OR sequence due to register moves, so
14683 if (Subtarget->hasAVX() &&
14684 !isa<ConstantFPSDNode>(Op1) && !isa<ConstantFPSDNode>(Op2)) {
14686 // Convert to vectors, do a VSELECT, and convert back to scalar.
14687 // All of the conversions should be optimized away.
14689 MVT VecVT = VT == MVT::f32 ? MVT::v4f32 : MVT::v2f64;
14690 SDValue VOp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op1);
14691 SDValue VOp2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op2);
14692 SDValue VCmp = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Cmp);
14694 MVT VCmpVT = VT == MVT::f32 ? MVT::v4i32 : MVT::v2i64;
14695 VCmp = DAG.getBitcast(VCmpVT, VCmp);
14697 SDValue VSel = DAG.getNode(ISD::VSELECT, DL, VecVT, VCmp, VOp1, VOp2);
14699 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT,
14700 VSel, DAG.getIntPtrConstant(0, DL));
14702 SDValue AndN = DAG.getNode(X86ISD::FANDN, DL, VT, Cmp, Op2);
14703 SDValue And = DAG.getNode(X86ISD::FAND, DL, VT, Cmp, Op1);
14704 return DAG.getNode(X86ISD::FOR, DL, VT, AndN, And);
14708 if (VT.isVector() && VT.getVectorElementType() == MVT::i1) {
14710 if (ISD::isBuildVectorOfConstantSDNodes(Op1.getNode()))
14711 Op1Scalar = ConvertI1VectorToInteger(Op1, DAG);
14712 else if (Op1.getOpcode() == ISD::BITCAST && Op1.getOperand(0))
14713 Op1Scalar = Op1.getOperand(0);
14715 if (ISD::isBuildVectorOfConstantSDNodes(Op2.getNode()))
14716 Op2Scalar = ConvertI1VectorToInteger(Op2, DAG);
14717 else if (Op2.getOpcode() == ISD::BITCAST && Op2.getOperand(0))
14718 Op2Scalar = Op2.getOperand(0);
14719 if (Op1Scalar.getNode() && Op2Scalar.getNode()) {
14720 SDValue newSelect = DAG.getNode(ISD::SELECT, DL,
14721 Op1Scalar.getValueType(),
14722 Cond, Op1Scalar, Op2Scalar);
14723 if (newSelect.getValueSizeInBits() == VT.getSizeInBits())
14724 return DAG.getBitcast(VT, newSelect);
14725 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, newSelect);
14726 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, ExtVec,
14727 DAG.getIntPtrConstant(0, DL));
14731 if (VT == MVT::v4i1 || VT == MVT::v2i1) {
14732 SDValue zeroConst = DAG.getIntPtrConstant(0, DL);
14733 Op1 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
14734 DAG.getUNDEF(MVT::v8i1), Op1, zeroConst);
14735 Op2 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
14736 DAG.getUNDEF(MVT::v8i1), Op2, zeroConst);
14737 SDValue newSelect = DAG.getNode(ISD::SELECT, DL, MVT::v8i1,
14739 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, newSelect, zeroConst);
14742 if (Cond.getOpcode() == ISD::SETCC) {
14743 SDValue NewCond = LowerSETCC(Cond, DAG);
14744 if (NewCond.getNode())
14748 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
14749 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
14750 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
14751 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
14752 if (Cond.getOpcode() == X86ISD::SETCC &&
14753 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
14754 isZero(Cond.getOperand(1).getOperand(1))) {
14755 SDValue Cmp = Cond.getOperand(1);
14757 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
14759 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
14760 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
14761 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
14763 SDValue CmpOp0 = Cmp.getOperand(0);
14764 // Apply further optimizations for special cases
14765 // (select (x != 0), -1, 0) -> neg & sbb
14766 // (select (x == 0), 0, -1) -> neg & sbb
14767 if (ConstantSDNode *YC = dyn_cast<ConstantSDNode>(Y))
14768 if (YC->isNullValue() &&
14769 (isAllOnes(Op1) == (CondCode == X86::COND_NE))) {
14770 SDVTList VTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32);
14771 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, VTs,
14772 DAG.getConstant(0, DL,
14773 CmpOp0.getValueType()),
14775 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14776 DAG.getConstant(X86::COND_B, DL, MVT::i8),
14777 SDValue(Neg.getNode(), 1));
14781 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
14782 CmpOp0, DAG.getConstant(1, DL, CmpOp0.getValueType()));
14783 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
14785 SDValue Res = // Res = 0 or -1.
14786 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14787 DAG.getConstant(X86::COND_B, DL, MVT::i8), Cmp);
14789 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
14790 Res = DAG.getNOT(DL, Res, Res.getValueType());
14792 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
14793 if (!N2C || !N2C->isNullValue())
14794 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
14799 // Look past (and (setcc_carry (cmp ...)), 1).
14800 if (Cond.getOpcode() == ISD::AND &&
14801 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
14802 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
14803 if (C && C->getAPIntValue() == 1)
14804 Cond = Cond.getOperand(0);
14807 // If condition flag is set by a X86ISD::CMP, then use it as the condition
14808 // setting operand in place of the X86ISD::SETCC.
14809 unsigned CondOpcode = Cond.getOpcode();
14810 if (CondOpcode == X86ISD::SETCC ||
14811 CondOpcode == X86ISD::SETCC_CARRY) {
14812 CC = Cond.getOperand(0);
14814 SDValue Cmp = Cond.getOperand(1);
14815 unsigned Opc = Cmp.getOpcode();
14816 MVT VT = Op.getSimpleValueType();
14818 bool IllegalFPCMov = false;
14819 if (VT.isFloatingPoint() && !VT.isVector() &&
14820 !isScalarFPTypeInSSEReg(VT)) // FPStack?
14821 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
14823 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
14824 Opc == X86ISD::BT) { // FIXME
14828 } else if (CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
14829 CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
14830 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
14831 Cond.getOperand(0).getValueType() != MVT::i8)) {
14832 SDValue LHS = Cond.getOperand(0);
14833 SDValue RHS = Cond.getOperand(1);
14834 unsigned X86Opcode;
14837 switch (CondOpcode) {
14838 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
14839 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
14840 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
14841 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
14842 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
14843 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
14844 default: llvm_unreachable("unexpected overflowing operator");
14846 if (CondOpcode == ISD::UMULO)
14847 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
14850 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
14852 SDValue X86Op = DAG.getNode(X86Opcode, DL, VTs, LHS, RHS);
14854 if (CondOpcode == ISD::UMULO)
14855 Cond = X86Op.getValue(2);
14857 Cond = X86Op.getValue(1);
14859 CC = DAG.getConstant(X86Cond, DL, MVT::i8);
14864 // Look past the truncate if the high bits are known zero.
14865 if (isTruncWithZeroHighBitsInput(Cond, DAG))
14866 Cond = Cond.getOperand(0);
14868 // We know the result of AND is compared against zero. Try to match
14870 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
14871 if (SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG)) {
14872 CC = NewSetCC.getOperand(0);
14873 Cond = NewSetCC.getOperand(1);
14880 CC = DAG.getConstant(X86::COND_NE, DL, MVT::i8);
14881 Cond = EmitTest(Cond, X86::COND_NE, DL, DAG);
14884 // a < b ? -1 : 0 -> RES = ~setcc_carry
14885 // a < b ? 0 : -1 -> RES = setcc_carry
14886 // a >= b ? -1 : 0 -> RES = setcc_carry
14887 // a >= b ? 0 : -1 -> RES = ~setcc_carry
14888 if (Cond.getOpcode() == X86ISD::SUB) {
14889 Cond = ConvertCmpIfNecessary(Cond, DAG);
14890 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
14892 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
14893 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
14894 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14895 DAG.getConstant(X86::COND_B, DL, MVT::i8),
14897 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
14898 return DAG.getNOT(DL, Res, Res.getValueType());
14903 // X86 doesn't have an i8 cmov. If both operands are the result of a truncate
14904 // widen the cmov and push the truncate through. This avoids introducing a new
14905 // branch during isel and doesn't add any extensions.
14906 if (Op.getValueType() == MVT::i8 &&
14907 Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) {
14908 SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0);
14909 if (T1.getValueType() == T2.getValueType() &&
14910 // Blacklist CopyFromReg to avoid partial register stalls.
14911 T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){
14912 SDVTList VTs = DAG.getVTList(T1.getValueType(), MVT::Glue);
14913 SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, VTs, T2, T1, CC, Cond);
14914 return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov);
14918 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
14919 // condition is true.
14920 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
14921 SDValue Ops[] = { Op2, Op1, CC, Cond };
14922 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops);
14925 static SDValue LowerSIGN_EXTEND_AVX512(SDValue Op,
14926 const X86Subtarget *Subtarget,
14927 SelectionDAG &DAG) {
14928 MVT VT = Op->getSimpleValueType(0);
14929 SDValue In = Op->getOperand(0);
14930 MVT InVT = In.getSimpleValueType();
14931 MVT VTElt = VT.getVectorElementType();
14932 MVT InVTElt = InVT.getVectorElementType();
14936 if ((InVTElt == MVT::i1) &&
14937 (((Subtarget->hasBWI() && Subtarget->hasVLX() &&
14938 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() <= 16)) ||
14940 ((Subtarget->hasBWI() && VT.is512BitVector() &&
14941 VTElt.getSizeInBits() <= 16)) ||
14943 ((Subtarget->hasDQI() && Subtarget->hasVLX() &&
14944 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() >= 32)) ||
14946 ((Subtarget->hasDQI() && VT.is512BitVector() &&
14947 VTElt.getSizeInBits() >= 32))))
14948 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14950 unsigned int NumElts = VT.getVectorNumElements();
14952 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
14955 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1) {
14956 if (In.getOpcode() == X86ISD::VSEXT || In.getOpcode() == X86ISD::VZEXT)
14957 return DAG.getNode(In.getOpcode(), dl, VT, In.getOperand(0));
14958 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14961 assert (InVT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
14962 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
14964 DAG.getConstant(APInt::getAllOnesValue(ExtVT.getScalarSizeInBits()), dl,
14967 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), dl, ExtVT);
14969 SDValue V = DAG.getNode(ISD::VSELECT, dl, ExtVT, In, NegOne, Zero);
14970 if (VT.is512BitVector())
14972 return DAG.getNode(X86ISD::VTRUNC, dl, VT, V);
14975 static SDValue LowerSIGN_EXTEND_VECTOR_INREG(SDValue Op,
14976 const X86Subtarget *Subtarget,
14977 SelectionDAG &DAG) {
14978 SDValue In = Op->getOperand(0);
14979 MVT VT = Op->getSimpleValueType(0);
14980 MVT InVT = In.getSimpleValueType();
14981 assert(VT.getSizeInBits() == InVT.getSizeInBits());
14983 MVT InSVT = InVT.getVectorElementType();
14984 assert(VT.getVectorElementType().getSizeInBits() > InSVT.getSizeInBits());
14986 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
14988 if (InSVT != MVT::i32 && InSVT != MVT::i16 && InSVT != MVT::i8)
14993 // SSE41 targets can use the pmovsx* instructions directly.
14994 if (Subtarget->hasSSE41())
14995 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14997 // pre-SSE41 targets unpack lower lanes and then sign-extend using SRAI.
15001 // As SRAI is only available on i16/i32 types, we expand only up to i32
15002 // and handle i64 separately.
15003 while (CurrVT != VT && CurrVT.getVectorElementType() != MVT::i32) {
15004 Curr = DAG.getNode(X86ISD::UNPCKL, dl, CurrVT, DAG.getUNDEF(CurrVT), Curr);
15005 MVT CurrSVT = MVT::getIntegerVT(CurrVT.getScalarSizeInBits() * 2);
15006 CurrVT = MVT::getVectorVT(CurrSVT, CurrVT.getVectorNumElements() / 2);
15007 Curr = DAG.getBitcast(CurrVT, Curr);
15010 SDValue SignExt = Curr;
15011 if (CurrVT != InVT) {
15012 unsigned SignExtShift =
15013 CurrVT.getVectorElementType().getSizeInBits() - InSVT.getSizeInBits();
15014 SignExt = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
15015 DAG.getConstant(SignExtShift, dl, MVT::i8));
15021 if (VT == MVT::v2i64 && CurrVT == MVT::v4i32) {
15022 SDValue Sign = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
15023 DAG.getConstant(31, dl, MVT::i8));
15024 SDValue Ext = DAG.getVectorShuffle(CurrVT, dl, SignExt, Sign, {0, 4, 1, 5});
15025 return DAG.getBitcast(VT, Ext);
15031 static SDValue LowerSIGN_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
15032 SelectionDAG &DAG) {
15033 MVT VT = Op->getSimpleValueType(0);
15034 SDValue In = Op->getOperand(0);
15035 MVT InVT = In.getSimpleValueType();
15038 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
15039 return LowerSIGN_EXTEND_AVX512(Op, Subtarget, DAG);
15041 if ((VT != MVT::v4i64 || InVT != MVT::v4i32) &&
15042 (VT != MVT::v8i32 || InVT != MVT::v8i16) &&
15043 (VT != MVT::v16i16 || InVT != MVT::v16i8))
15046 if (Subtarget->hasInt256())
15047 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
15049 // Optimize vectors in AVX mode
15050 // Sign extend v8i16 to v8i32 and
15053 // Divide input vector into two parts
15054 // for v4i32 the shuffle mask will be { 0, 1, -1, -1} {2, 3, -1, -1}
15055 // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32
15056 // concat the vectors to original VT
15058 unsigned NumElems = InVT.getVectorNumElements();
15059 SDValue Undef = DAG.getUNDEF(InVT);
15061 SmallVector<int,8> ShufMask1(NumElems, -1);
15062 for (unsigned i = 0; i != NumElems/2; ++i)
15065 SDValue OpLo = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask1[0]);
15067 SmallVector<int,8> ShufMask2(NumElems, -1);
15068 for (unsigned i = 0; i != NumElems/2; ++i)
15069 ShufMask2[i] = i + NumElems/2;
15071 SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask2[0]);
15073 MVT HalfVT = MVT::getVectorVT(VT.getVectorElementType(),
15074 VT.getVectorNumElements()/2);
15076 OpLo = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpLo);
15077 OpHi = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpHi);
15079 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
15082 // Lower vector extended loads using a shuffle. If SSSE3 is not available we
15083 // may emit an illegal shuffle but the expansion is still better than scalar
15084 // code. We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise
15085 // we'll emit a shuffle and a arithmetic shift.
15086 // FIXME: Is the expansion actually better than scalar code? It doesn't seem so.
15087 // TODO: It is possible to support ZExt by zeroing the undef values during
15088 // the shuffle phase or after the shuffle.
15089 static SDValue LowerExtendedLoad(SDValue Op, const X86Subtarget *Subtarget,
15090 SelectionDAG &DAG) {
15091 MVT RegVT = Op.getSimpleValueType();
15092 assert(RegVT.isVector() && "We only custom lower vector sext loads.");
15093 assert(RegVT.isInteger() &&
15094 "We only custom lower integer vector sext loads.");
15096 // Nothing useful we can do without SSE2 shuffles.
15097 assert(Subtarget->hasSSE2() && "We only custom lower sext loads with SSE2.");
15099 LoadSDNode *Ld = cast<LoadSDNode>(Op.getNode());
15101 EVT MemVT = Ld->getMemoryVT();
15102 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15103 unsigned RegSz = RegVT.getSizeInBits();
15105 ISD::LoadExtType Ext = Ld->getExtensionType();
15107 assert((Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)
15108 && "Only anyext and sext are currently implemented.");
15109 assert(MemVT != RegVT && "Cannot extend to the same type");
15110 assert(MemVT.isVector() && "Must load a vector from memory");
15112 unsigned NumElems = RegVT.getVectorNumElements();
15113 unsigned MemSz = MemVT.getSizeInBits();
15114 assert(RegSz > MemSz && "Register size must be greater than the mem size");
15116 if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256()) {
15117 // The only way in which we have a legal 256-bit vector result but not the
15118 // integer 256-bit operations needed to directly lower a sextload is if we
15119 // have AVX1 but not AVX2. In that case, we can always emit a sextload to
15120 // a 128-bit vector and a normal sign_extend to 256-bits that should get
15121 // correctly legalized. We do this late to allow the canonical form of
15122 // sextload to persist throughout the rest of the DAG combiner -- it wants
15123 // to fold together any extensions it can, and so will fuse a sign_extend
15124 // of an sextload into a sextload targeting a wider value.
15126 if (MemSz == 128) {
15127 // Just switch this to a normal load.
15128 assert(TLI.isTypeLegal(MemVT) && "If the memory type is a 128-bit type, "
15129 "it must be a legal 128-bit vector "
15131 Load = DAG.getLoad(MemVT, dl, Ld->getChain(), Ld->getBasePtr(),
15132 Ld->getPointerInfo(), Ld->isVolatile(), Ld->isNonTemporal(),
15133 Ld->isInvariant(), Ld->getAlignment());
15135 assert(MemSz < 128 &&
15136 "Can't extend a type wider than 128 bits to a 256 bit vector!");
15137 // Do an sext load to a 128-bit vector type. We want to use the same
15138 // number of elements, but elements half as wide. This will end up being
15139 // recursively lowered by this routine, but will succeed as we definitely
15140 // have all the necessary features if we're using AVX1.
15142 EVT::getIntegerVT(*DAG.getContext(), RegVT.getScalarSizeInBits() / 2);
15143 EVT HalfVecVT = EVT::getVectorVT(*DAG.getContext(), HalfEltVT, NumElems);
15145 DAG.getExtLoad(Ext, dl, HalfVecVT, Ld->getChain(), Ld->getBasePtr(),
15146 Ld->getPointerInfo(), MemVT, Ld->isVolatile(),
15147 Ld->isNonTemporal(), Ld->isInvariant(),
15148 Ld->getAlignment());
15151 // Replace chain users with the new chain.
15152 assert(Load->getNumValues() == 2 && "Loads must carry a chain!");
15153 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Load.getValue(1));
15155 // Finally, do a normal sign-extend to the desired register.
15156 return DAG.getSExtOrTrunc(Load, dl, RegVT);
15159 // All sizes must be a power of two.
15160 assert(isPowerOf2_32(RegSz * MemSz * NumElems) &&
15161 "Non-power-of-two elements are not custom lowered!");
15163 // Attempt to load the original value using scalar loads.
15164 // Find the largest scalar type that divides the total loaded size.
15165 MVT SclrLoadTy = MVT::i8;
15166 for (MVT Tp : MVT::integer_valuetypes()) {
15167 if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
15172 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
15173 if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
15175 SclrLoadTy = MVT::f64;
15177 // Calculate the number of scalar loads that we need to perform
15178 // in order to load our vector from memory.
15179 unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
15181 assert((Ext != ISD::SEXTLOAD || NumLoads == 1) &&
15182 "Can only lower sext loads with a single scalar load!");
15184 unsigned loadRegZize = RegSz;
15185 if (Ext == ISD::SEXTLOAD && RegSz >= 256)
15188 // Represent our vector as a sequence of elements which are the
15189 // largest scalar that we can load.
15190 EVT LoadUnitVecVT = EVT::getVectorVT(
15191 *DAG.getContext(), SclrLoadTy, loadRegZize / SclrLoadTy.getSizeInBits());
15193 // Represent the data using the same element type that is stored in
15194 // memory. In practice, we ''widen'' MemVT.
15196 EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
15197 loadRegZize / MemVT.getScalarSizeInBits());
15199 assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
15200 "Invalid vector type");
15202 // We can't shuffle using an illegal type.
15203 assert(TLI.isTypeLegal(WideVecVT) &&
15204 "We only lower types that form legal widened vector types");
15206 SmallVector<SDValue, 8> Chains;
15207 SDValue Ptr = Ld->getBasePtr();
15208 SDValue Increment = DAG.getConstant(SclrLoadTy.getSizeInBits() / 8, dl,
15209 TLI.getPointerTy(DAG.getDataLayout()));
15210 SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
15212 for (unsigned i = 0; i < NumLoads; ++i) {
15213 // Perform a single load.
15214 SDValue ScalarLoad =
15215 DAG.getLoad(SclrLoadTy, dl, Ld->getChain(), Ptr, Ld->getPointerInfo(),
15216 Ld->isVolatile(), Ld->isNonTemporal(), Ld->isInvariant(),
15217 Ld->getAlignment());
15218 Chains.push_back(ScalarLoad.getValue(1));
15219 // Create the first element type using SCALAR_TO_VECTOR in order to avoid
15220 // another round of DAGCombining.
15222 Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
15224 Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
15225 ScalarLoad, DAG.getIntPtrConstant(i, dl));
15227 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
15230 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
15232 // Bitcast the loaded value to a vector of the original element type, in
15233 // the size of the target vector type.
15234 SDValue SlicedVec = DAG.getBitcast(WideVecVT, Res);
15235 unsigned SizeRatio = RegSz / MemSz;
15237 if (Ext == ISD::SEXTLOAD) {
15238 // If we have SSE4.1, we can directly emit a VSEXT node.
15239 if (Subtarget->hasSSE41()) {
15240 SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
15241 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
15245 // Otherwise we'll use SIGN_EXTEND_VECTOR_INREG to sign extend the lowest
15247 assert(TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND_VECTOR_INREG, RegVT) &&
15248 "We can't implement a sext load without SIGN_EXTEND_VECTOR_INREG!");
15250 SDValue Shuff = DAG.getSignExtendVectorInReg(SlicedVec, dl, RegVT);
15251 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
15255 // Redistribute the loaded elements into the different locations.
15256 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
15257 for (unsigned i = 0; i != NumElems; ++i)
15258 ShuffleVec[i * SizeRatio] = i;
15260 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
15261 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
15263 // Bitcast to the requested type.
15264 Shuff = DAG.getBitcast(RegVT, Shuff);
15265 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
15269 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
15270 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
15271 // from the AND / OR.
15272 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
15273 Opc = Op.getOpcode();
15274 if (Opc != ISD::OR && Opc != ISD::AND)
15276 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
15277 Op.getOperand(0).hasOneUse() &&
15278 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
15279 Op.getOperand(1).hasOneUse());
15282 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
15283 // 1 and that the SETCC node has a single use.
15284 static bool isXor1OfSetCC(SDValue Op) {
15285 if (Op.getOpcode() != ISD::XOR)
15287 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
15288 if (N1C && N1C->getAPIntValue() == 1) {
15289 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
15290 Op.getOperand(0).hasOneUse();
15295 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
15296 bool addTest = true;
15297 SDValue Chain = Op.getOperand(0);
15298 SDValue Cond = Op.getOperand(1);
15299 SDValue Dest = Op.getOperand(2);
15302 bool Inverted = false;
15304 if (Cond.getOpcode() == ISD::SETCC) {
15305 // Check for setcc([su]{add,sub,mul}o == 0).
15306 if (cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETEQ &&
15307 isa<ConstantSDNode>(Cond.getOperand(1)) &&
15308 cast<ConstantSDNode>(Cond.getOperand(1))->isNullValue() &&
15309 Cond.getOperand(0).getResNo() == 1 &&
15310 (Cond.getOperand(0).getOpcode() == ISD::SADDO ||
15311 Cond.getOperand(0).getOpcode() == ISD::UADDO ||
15312 Cond.getOperand(0).getOpcode() == ISD::SSUBO ||
15313 Cond.getOperand(0).getOpcode() == ISD::USUBO ||
15314 Cond.getOperand(0).getOpcode() == ISD::SMULO ||
15315 Cond.getOperand(0).getOpcode() == ISD::UMULO)) {
15317 Cond = Cond.getOperand(0);
15319 SDValue NewCond = LowerSETCC(Cond, DAG);
15320 if (NewCond.getNode())
15325 // FIXME: LowerXALUO doesn't handle these!!
15326 else if (Cond.getOpcode() == X86ISD::ADD ||
15327 Cond.getOpcode() == X86ISD::SUB ||
15328 Cond.getOpcode() == X86ISD::SMUL ||
15329 Cond.getOpcode() == X86ISD::UMUL)
15330 Cond = LowerXALUO(Cond, DAG);
15333 // Look pass (and (setcc_carry (cmp ...)), 1).
15334 if (Cond.getOpcode() == ISD::AND &&
15335 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
15336 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
15337 if (C && C->getAPIntValue() == 1)
15338 Cond = Cond.getOperand(0);
15341 // If condition flag is set by a X86ISD::CMP, then use it as the condition
15342 // setting operand in place of the X86ISD::SETCC.
15343 unsigned CondOpcode = Cond.getOpcode();
15344 if (CondOpcode == X86ISD::SETCC ||
15345 CondOpcode == X86ISD::SETCC_CARRY) {
15346 CC = Cond.getOperand(0);
15348 SDValue Cmp = Cond.getOperand(1);
15349 unsigned Opc = Cmp.getOpcode();
15350 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
15351 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
15355 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
15359 // These can only come from an arithmetic instruction with overflow,
15360 // e.g. SADDO, UADDO.
15361 Cond = Cond.getNode()->getOperand(1);
15367 CondOpcode = Cond.getOpcode();
15368 if (CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
15369 CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
15370 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
15371 Cond.getOperand(0).getValueType() != MVT::i8)) {
15372 SDValue LHS = Cond.getOperand(0);
15373 SDValue RHS = Cond.getOperand(1);
15374 unsigned X86Opcode;
15377 // Keep this in sync with LowerXALUO, otherwise we might create redundant
15378 // instructions that can't be removed afterwards (i.e. X86ISD::ADD and
15380 switch (CondOpcode) {
15381 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
15383 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
15385 X86Opcode = X86ISD::INC; X86Cond = X86::COND_O;
15388 X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
15389 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
15391 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
15393 X86Opcode = X86ISD::DEC; X86Cond = X86::COND_O;
15396 X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
15397 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
15398 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
15399 default: llvm_unreachable("unexpected overflowing operator");
15402 X86Cond = X86::GetOppositeBranchCondition((X86::CondCode)X86Cond);
15403 if (CondOpcode == ISD::UMULO)
15404 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
15407 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
15409 SDValue X86Op = DAG.getNode(X86Opcode, dl, VTs, LHS, RHS);
15411 if (CondOpcode == ISD::UMULO)
15412 Cond = X86Op.getValue(2);
15414 Cond = X86Op.getValue(1);
15416 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
15420 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
15421 SDValue Cmp = Cond.getOperand(0).getOperand(1);
15422 if (CondOpc == ISD::OR) {
15423 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
15424 // two branches instead of an explicit OR instruction with a
15426 if (Cmp == Cond.getOperand(1).getOperand(1) &&
15427 isX86LogicalCmp(Cmp)) {
15428 CC = Cond.getOperand(0).getOperand(0);
15429 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15430 Chain, Dest, CC, Cmp);
15431 CC = Cond.getOperand(1).getOperand(0);
15435 } else { // ISD::AND
15436 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
15437 // two branches instead of an explicit AND instruction with a
15438 // separate test. However, we only do this if this block doesn't
15439 // have a fall-through edge, because this requires an explicit
15440 // jmp when the condition is false.
15441 if (Cmp == Cond.getOperand(1).getOperand(1) &&
15442 isX86LogicalCmp(Cmp) &&
15443 Op.getNode()->hasOneUse()) {
15444 X86::CondCode CCode =
15445 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
15446 CCode = X86::GetOppositeBranchCondition(CCode);
15447 CC = DAG.getConstant(CCode, dl, MVT::i8);
15448 SDNode *User = *Op.getNode()->use_begin();
15449 // Look for an unconditional branch following this conditional branch.
15450 // We need this because we need to reverse the successors in order
15451 // to implement FCMP_OEQ.
15452 if (User->getOpcode() == ISD::BR) {
15453 SDValue FalseBB = User->getOperand(1);
15455 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15456 assert(NewBR == User);
15460 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15461 Chain, Dest, CC, Cmp);
15462 X86::CondCode CCode =
15463 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
15464 CCode = X86::GetOppositeBranchCondition(CCode);
15465 CC = DAG.getConstant(CCode, dl, MVT::i8);
15471 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
15472 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
15473 // It should be transformed during dag combiner except when the condition
15474 // is set by a arithmetics with overflow node.
15475 X86::CondCode CCode =
15476 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
15477 CCode = X86::GetOppositeBranchCondition(CCode);
15478 CC = DAG.getConstant(CCode, dl, MVT::i8);
15479 Cond = Cond.getOperand(0).getOperand(1);
15481 } else if (Cond.getOpcode() == ISD::SETCC &&
15482 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETOEQ) {
15483 // For FCMP_OEQ, we can emit
15484 // two branches instead of an explicit AND instruction with a
15485 // separate test. However, we only do this if this block doesn't
15486 // have a fall-through edge, because this requires an explicit
15487 // jmp when the condition is false.
15488 if (Op.getNode()->hasOneUse()) {
15489 SDNode *User = *Op.getNode()->use_begin();
15490 // Look for an unconditional branch following this conditional branch.
15491 // We need this because we need to reverse the successors in order
15492 // to implement FCMP_OEQ.
15493 if (User->getOpcode() == ISD::BR) {
15494 SDValue FalseBB = User->getOperand(1);
15496 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15497 assert(NewBR == User);
15501 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
15502 Cond.getOperand(0), Cond.getOperand(1));
15503 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
15504 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
15505 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15506 Chain, Dest, CC, Cmp);
15507 CC = DAG.getConstant(X86::COND_P, dl, MVT::i8);
15512 } else if (Cond.getOpcode() == ISD::SETCC &&
15513 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETUNE) {
15514 // For FCMP_UNE, we can emit
15515 // two branches instead of an explicit AND instruction with a
15516 // separate test. However, we only do this if this block doesn't
15517 // have a fall-through edge, because this requires an explicit
15518 // jmp when the condition is false.
15519 if (Op.getNode()->hasOneUse()) {
15520 SDNode *User = *Op.getNode()->use_begin();
15521 // Look for an unconditional branch following this conditional branch.
15522 // We need this because we need to reverse the successors in order
15523 // to implement FCMP_UNE.
15524 if (User->getOpcode() == ISD::BR) {
15525 SDValue FalseBB = User->getOperand(1);
15527 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15528 assert(NewBR == User);
15531 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
15532 Cond.getOperand(0), Cond.getOperand(1));
15533 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
15534 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
15535 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15536 Chain, Dest, CC, Cmp);
15537 CC = DAG.getConstant(X86::COND_NP, dl, MVT::i8);
15547 // Look pass the truncate if the high bits are known zero.
15548 if (isTruncWithZeroHighBitsInput(Cond, DAG))
15549 Cond = Cond.getOperand(0);
15551 // We know the result of AND is compared against zero. Try to match
15553 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
15554 if (SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG)) {
15555 CC = NewSetCC.getOperand(0);
15556 Cond = NewSetCC.getOperand(1);
15563 X86::CondCode X86Cond = Inverted ? X86::COND_E : X86::COND_NE;
15564 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
15565 Cond = EmitTest(Cond, X86Cond, dl, DAG);
15567 Cond = ConvertCmpIfNecessary(Cond, DAG);
15568 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15569 Chain, Dest, CC, Cond);
15572 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
15573 // Calls to _alloca are needed to probe the stack when allocating more than 4k
15574 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
15575 // that the guard pages used by the OS virtual memory manager are allocated in
15576 // correct sequence.
15578 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
15579 SelectionDAG &DAG) const {
15580 MachineFunction &MF = DAG.getMachineFunction();
15581 bool SplitStack = MF.shouldSplitStack();
15582 bool Lower = (Subtarget->isOSWindows() && !Subtarget->isTargetMachO()) ||
15587 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15588 SDNode* Node = Op.getNode();
15590 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
15591 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
15592 " not tell us which reg is the stack pointer!");
15593 EVT VT = Node->getValueType(0);
15594 SDValue Tmp1 = SDValue(Node, 0);
15595 SDValue Tmp2 = SDValue(Node, 1);
15596 SDValue Tmp3 = Node->getOperand(2);
15597 SDValue Chain = Tmp1.getOperand(0);
15599 // Chain the dynamic stack allocation so that it doesn't modify the stack
15600 // pointer when other instructions are using the stack.
15601 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, dl, true),
15604 SDValue Size = Tmp2.getOperand(1);
15605 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
15606 Chain = SP.getValue(1);
15607 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
15608 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
15609 unsigned StackAlign = TFI.getStackAlignment();
15610 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
15611 if (Align > StackAlign)
15612 Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1,
15613 DAG.getConstant(-(uint64_t)Align, dl, VT));
15614 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
15616 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, dl, true),
15617 DAG.getIntPtrConstant(0, dl, true), SDValue(),
15620 SDValue Ops[2] = { Tmp1, Tmp2 };
15621 return DAG.getMergeValues(Ops, dl);
15625 SDValue Chain = Op.getOperand(0);
15626 SDValue Size = Op.getOperand(1);
15627 unsigned Align = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
15628 EVT VT = Op.getNode()->getValueType(0);
15630 bool Is64Bit = Subtarget->is64Bit();
15631 MVT SPTy = getPointerTy(DAG.getDataLayout());
15634 MachineRegisterInfo &MRI = MF.getRegInfo();
15637 // The 64 bit implementation of segmented stacks needs to clobber both r10
15638 // r11. This makes it impossible to use it along with nested parameters.
15639 const Function *F = MF.getFunction();
15641 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
15643 if (I->hasNestAttr())
15644 report_fatal_error("Cannot use segmented stacks with functions that "
15645 "have nested arguments.");
15648 const TargetRegisterClass *AddrRegClass = getRegClassFor(SPTy);
15649 unsigned Vreg = MRI.createVirtualRegister(AddrRegClass);
15650 Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size);
15651 SDValue Value = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain,
15652 DAG.getRegister(Vreg, SPTy));
15653 SDValue Ops1[2] = { Value, Chain };
15654 return DAG.getMergeValues(Ops1, dl);
15657 const unsigned Reg = (Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX);
15659 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
15660 Flag = Chain.getValue(1);
15661 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
15663 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
15665 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15666 unsigned SPReg = RegInfo->getStackRegister();
15667 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, SPTy);
15668 Chain = SP.getValue(1);
15671 SP = DAG.getNode(ISD::AND, dl, VT, SP.getValue(0),
15672 DAG.getConstant(-(uint64_t)Align, dl, VT));
15673 Chain = DAG.getCopyToReg(Chain, dl, SPReg, SP);
15676 SDValue Ops1[2] = { SP, Chain };
15677 return DAG.getMergeValues(Ops1, dl);
15681 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
15682 MachineFunction &MF = DAG.getMachineFunction();
15683 auto PtrVT = getPointerTy(MF.getDataLayout());
15684 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
15686 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
15689 if (!Subtarget->is64Bit() ||
15690 Subtarget->isCallingConvWin64(MF.getFunction()->getCallingConv())) {
15691 // vastart just stores the address of the VarArgsFrameIndex slot into the
15692 // memory location argument.
15693 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
15694 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
15695 MachinePointerInfo(SV), false, false, 0);
15699 // gp_offset (0 - 6 * 8)
15700 // fp_offset (48 - 48 + 8 * 16)
15701 // overflow_arg_area (point to parameters coming in memory).
15703 SmallVector<SDValue, 8> MemOps;
15704 SDValue FIN = Op.getOperand(1);
15706 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
15707 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
15709 FIN, MachinePointerInfo(SV), false, false, 0);
15710 MemOps.push_back(Store);
15713 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL));
15714 Store = DAG.getStore(Op.getOperand(0), DL,
15715 DAG.getConstant(FuncInfo->getVarArgsFPOffset(), DL,
15717 FIN, MachinePointerInfo(SV, 4), false, false, 0);
15718 MemOps.push_back(Store);
15720 // Store ptr to overflow_arg_area
15721 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL));
15722 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
15723 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
15724 MachinePointerInfo(SV, 8),
15726 MemOps.push_back(Store);
15728 // Store ptr to reg_save_area.
15729 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(
15730 Subtarget->isTarget64BitLP64() ? 8 : 4, DL));
15731 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(), PtrVT);
15732 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN, MachinePointerInfo(
15733 SV, Subtarget->isTarget64BitLP64() ? 16 : 12), false, false, 0);
15734 MemOps.push_back(Store);
15735 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
15738 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
15739 assert(Subtarget->is64Bit() &&
15740 "LowerVAARG only handles 64-bit va_arg!");
15741 assert(Op.getNode()->getNumOperands() == 4);
15743 MachineFunction &MF = DAG.getMachineFunction();
15744 if (Subtarget->isCallingConvWin64(MF.getFunction()->getCallingConv()))
15745 // The Win64 ABI uses char* instead of a structure.
15746 return DAG.expandVAArg(Op.getNode());
15748 SDValue Chain = Op.getOperand(0);
15749 SDValue SrcPtr = Op.getOperand(1);
15750 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
15751 unsigned Align = Op.getConstantOperandVal(3);
15754 EVT ArgVT = Op.getNode()->getValueType(0);
15755 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
15756 uint32_t ArgSize = DAG.getDataLayout().getTypeAllocSize(ArgTy);
15759 // Decide which area this value should be read from.
15760 // TODO: Implement the AMD64 ABI in its entirety. This simple
15761 // selection mechanism works only for the basic types.
15762 if (ArgVT == MVT::f80) {
15763 llvm_unreachable("va_arg for f80 not yet implemented");
15764 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
15765 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
15766 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
15767 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
15769 llvm_unreachable("Unhandled argument type in LowerVAARG");
15772 if (ArgMode == 2) {
15773 // Sanity Check: Make sure using fp_offset makes sense.
15774 assert(!Subtarget->useSoftFloat() &&
15775 !(MF.getFunction()->hasFnAttribute(Attribute::NoImplicitFloat)) &&
15776 Subtarget->hasSSE1());
15779 // Insert VAARG_64 node into the DAG
15780 // VAARG_64 returns two values: Variable Argument Address, Chain
15781 SDValue InstOps[] = {Chain, SrcPtr, DAG.getConstant(ArgSize, dl, MVT::i32),
15782 DAG.getConstant(ArgMode, dl, MVT::i8),
15783 DAG.getConstant(Align, dl, MVT::i32)};
15784 SDVTList VTs = DAG.getVTList(getPointerTy(DAG.getDataLayout()), MVT::Other);
15785 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
15786 VTs, InstOps, MVT::i64,
15787 MachinePointerInfo(SV),
15789 /*Volatile=*/false,
15791 /*WriteMem=*/true);
15792 Chain = VAARG.getValue(1);
15794 // Load the next argument and return it
15795 return DAG.getLoad(ArgVT, dl,
15798 MachinePointerInfo(),
15799 false, false, false, 0);
15802 static SDValue LowerVACOPY(SDValue Op, const X86Subtarget *Subtarget,
15803 SelectionDAG &DAG) {
15804 // X86-64 va_list is a struct { i32, i32, i8*, i8* }, except on Windows,
15805 // where a va_list is still an i8*.
15806 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
15807 if (Subtarget->isCallingConvWin64(
15808 DAG.getMachineFunction().getFunction()->getCallingConv()))
15809 // Probably a Win64 va_copy.
15810 return DAG.expandVACopy(Op.getNode());
15812 SDValue Chain = Op.getOperand(0);
15813 SDValue DstPtr = Op.getOperand(1);
15814 SDValue SrcPtr = Op.getOperand(2);
15815 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
15816 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
15819 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
15820 DAG.getIntPtrConstant(24, DL), 8, /*isVolatile*/false,
15822 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
15825 // getTargetVShiftByConstNode - Handle vector element shifts where the shift
15826 // amount is a constant. Takes immediate version of shift as input.
15827 static SDValue getTargetVShiftByConstNode(unsigned Opc, SDLoc dl, MVT VT,
15828 SDValue SrcOp, uint64_t ShiftAmt,
15829 SelectionDAG &DAG) {
15830 MVT ElementType = VT.getVectorElementType();
15832 // Fold this packed shift into its first operand if ShiftAmt is 0.
15836 // Check for ShiftAmt >= element width
15837 if (ShiftAmt >= ElementType.getSizeInBits()) {
15838 if (Opc == X86ISD::VSRAI)
15839 ShiftAmt = ElementType.getSizeInBits() - 1;
15841 return DAG.getConstant(0, dl, VT);
15844 assert((Opc == X86ISD::VSHLI || Opc == X86ISD::VSRLI || Opc == X86ISD::VSRAI)
15845 && "Unknown target vector shift-by-constant node");
15847 // Fold this packed vector shift into a build vector if SrcOp is a
15848 // vector of Constants or UNDEFs, and SrcOp valuetype is the same as VT.
15849 if (VT == SrcOp.getSimpleValueType() &&
15850 ISD::isBuildVectorOfConstantSDNodes(SrcOp.getNode())) {
15851 SmallVector<SDValue, 8> Elts;
15852 unsigned NumElts = SrcOp->getNumOperands();
15853 ConstantSDNode *ND;
15856 default: llvm_unreachable(nullptr);
15857 case X86ISD::VSHLI:
15858 for (unsigned i=0; i!=NumElts; ++i) {
15859 SDValue CurrentOp = SrcOp->getOperand(i);
15860 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15861 Elts.push_back(CurrentOp);
15864 ND = cast<ConstantSDNode>(CurrentOp);
15865 const APInt &C = ND->getAPIntValue();
15866 Elts.push_back(DAG.getConstant(C.shl(ShiftAmt), dl, ElementType));
15869 case X86ISD::VSRLI:
15870 for (unsigned i=0; i!=NumElts; ++i) {
15871 SDValue CurrentOp = SrcOp->getOperand(i);
15872 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15873 Elts.push_back(CurrentOp);
15876 ND = cast<ConstantSDNode>(CurrentOp);
15877 const APInt &C = ND->getAPIntValue();
15878 Elts.push_back(DAG.getConstant(C.lshr(ShiftAmt), dl, ElementType));
15881 case X86ISD::VSRAI:
15882 for (unsigned i=0; i!=NumElts; ++i) {
15883 SDValue CurrentOp = SrcOp->getOperand(i);
15884 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15885 Elts.push_back(CurrentOp);
15888 ND = cast<ConstantSDNode>(CurrentOp);
15889 const APInt &C = ND->getAPIntValue();
15890 Elts.push_back(DAG.getConstant(C.ashr(ShiftAmt), dl, ElementType));
15895 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
15898 return DAG.getNode(Opc, dl, VT, SrcOp,
15899 DAG.getConstant(ShiftAmt, dl, MVT::i8));
15902 // getTargetVShiftNode - Handle vector element shifts where the shift amount
15903 // may or may not be a constant. Takes immediate version of shift as input.
15904 static SDValue getTargetVShiftNode(unsigned Opc, SDLoc dl, MVT VT,
15905 SDValue SrcOp, SDValue ShAmt,
15906 SelectionDAG &DAG) {
15907 MVT SVT = ShAmt.getSimpleValueType();
15908 assert((SVT == MVT::i32 || SVT == MVT::i64) && "Unexpected value type!");
15910 // Catch shift-by-constant.
15911 if (ConstantSDNode *CShAmt = dyn_cast<ConstantSDNode>(ShAmt))
15912 return getTargetVShiftByConstNode(Opc, dl, VT, SrcOp,
15913 CShAmt->getZExtValue(), DAG);
15915 // Change opcode to non-immediate version
15917 default: llvm_unreachable("Unknown target vector shift node");
15918 case X86ISD::VSHLI: Opc = X86ISD::VSHL; break;
15919 case X86ISD::VSRLI: Opc = X86ISD::VSRL; break;
15920 case X86ISD::VSRAI: Opc = X86ISD::VSRA; break;
15923 const X86Subtarget &Subtarget =
15924 static_cast<const X86Subtarget &>(DAG.getSubtarget());
15925 if (Subtarget.hasSSE41() && ShAmt.getOpcode() == ISD::ZERO_EXTEND &&
15926 ShAmt.getOperand(0).getSimpleValueType() == MVT::i16) {
15927 // Let the shuffle legalizer expand this shift amount node.
15928 SDValue Op0 = ShAmt.getOperand(0);
15929 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(Op0), MVT::v8i16, Op0);
15930 ShAmt = getShuffleVectorZeroOrUndef(Op0, 0, true, &Subtarget, DAG);
15932 // Need to build a vector containing shift amount.
15933 // SSE/AVX packed shifts only use the lower 64-bit of the shift count.
15934 SmallVector<SDValue, 4> ShOps;
15935 ShOps.push_back(ShAmt);
15936 if (SVT == MVT::i32) {
15937 ShOps.push_back(DAG.getConstant(0, dl, SVT));
15938 ShOps.push_back(DAG.getUNDEF(SVT));
15940 ShOps.push_back(DAG.getUNDEF(SVT));
15942 MVT BVT = SVT == MVT::i32 ? MVT::v4i32 : MVT::v2i64;
15943 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, BVT, ShOps);
15946 // The return type has to be a 128-bit type with the same element
15947 // type as the input type.
15948 MVT EltVT = VT.getVectorElementType();
15949 MVT ShVT = MVT::getVectorVT(EltVT, 128/EltVT.getSizeInBits());
15951 ShAmt = DAG.getBitcast(ShVT, ShAmt);
15952 return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt);
15955 /// \brief Return (and \p Op, \p Mask) for compare instructions or
15956 /// (vselect \p Mask, \p Op, \p PreservedSrc) for others along with the
15957 /// necessary casting or extending for \p Mask when lowering masking intrinsics
15958 static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask,
15959 SDValue PreservedSrc,
15960 const X86Subtarget *Subtarget,
15961 SelectionDAG &DAG) {
15962 MVT VT = Op.getSimpleValueType();
15963 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
15965 unsigned OpcodeSelect = ISD::VSELECT;
15968 if (isAllOnes(Mask))
15971 if (MaskVT.bitsGT(Mask.getSimpleValueType())) {
15972 MVT newMaskVT = MVT::getIntegerVT(MaskVT.getSizeInBits());
15973 VMask = DAG.getBitcast(MaskVT,
15974 DAG.getNode(ISD::ANY_EXTEND, dl, newMaskVT, Mask));
15976 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
15977 Mask.getSimpleValueType().getSizeInBits());
15978 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
15979 // are extracted by EXTRACT_SUBVECTOR.
15980 VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15981 DAG.getBitcast(BitcastVT, Mask),
15982 DAG.getIntPtrConstant(0, dl));
15985 switch (Op.getOpcode()) {
15987 case X86ISD::PCMPEQM:
15988 case X86ISD::PCMPGTM:
15990 case X86ISD::CMPMU:
15991 return DAG.getNode(ISD::AND, dl, VT, Op, VMask);
15992 case X86ISD::VFPCLASS:
15993 return DAG.getNode(ISD::OR, dl, VT, Op, VMask);
15994 case X86ISD::VTRUNC:
15995 case X86ISD::VTRUNCS:
15996 case X86ISD::VTRUNCUS:
15997 // We can't use ISD::VSELECT here because it is not always "Legal"
15998 // for the destination type. For example vpmovqb require only AVX512
15999 // and vselect that can operate on byte element type require BWI
16000 OpcodeSelect = X86ISD::SELECT;
16003 if (PreservedSrc.getOpcode() == ISD::UNDEF)
16004 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
16005 return DAG.getNode(OpcodeSelect, dl, VT, VMask, Op, PreservedSrc);
16008 /// \brief Creates an SDNode for a predicated scalar operation.
16009 /// \returns (X86vselect \p Mask, \p Op, \p PreservedSrc).
16010 /// The mask is coming as MVT::i8 and it should be truncated
16011 /// to MVT::i1 while lowering masking intrinsics.
16012 /// The main difference between ScalarMaskingNode and VectorMaskingNode is using
16013 /// "X86select" instead of "vselect". We just can't create the "vselect" node
16014 /// for a scalar instruction.
16015 static SDValue getScalarMaskingNode(SDValue Op, SDValue Mask,
16016 SDValue PreservedSrc,
16017 const X86Subtarget *Subtarget,
16018 SelectionDAG &DAG) {
16019 if (isAllOnes(Mask))
16022 MVT VT = Op.getSimpleValueType();
16024 // The mask should be of type MVT::i1
16025 SDValue IMask = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Mask);
16027 if (Op.getOpcode() == X86ISD::FSETCC)
16028 return DAG.getNode(ISD::AND, dl, VT, Op, IMask);
16029 if (Op.getOpcode() == X86ISD::VFPCLASS)
16030 return DAG.getNode(ISD::OR, dl, VT, Op, IMask);
16032 if (PreservedSrc.getOpcode() == ISD::UNDEF)
16033 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
16034 return DAG.getNode(X86ISD::SELECT, dl, VT, IMask, Op, PreservedSrc);
16037 static int getSEHRegistrationNodeSize(const Function *Fn) {
16038 if (!Fn->hasPersonalityFn())
16039 report_fatal_error(
16040 "querying registration node size for function without personality");
16041 // The RegNodeSize is 6 32-bit words for SEH and 4 for C++ EH. See
16042 // WinEHStatePass for the full struct definition.
16043 switch (classifyEHPersonality(Fn->getPersonalityFn())) {
16044 case EHPersonality::MSVC_X86SEH: return 24;
16045 case EHPersonality::MSVC_CXX: return 16;
16048 report_fatal_error("can only recover FP for MSVC EH personality functions");
16051 /// When the 32-bit MSVC runtime transfers control to us, either to an outlined
16052 /// function or when returning to a parent frame after catching an exception, we
16053 /// recover the parent frame pointer by doing arithmetic on the incoming EBP.
16054 /// Here's the math:
16055 /// RegNodeBase = EntryEBP - RegNodeSize
16056 /// ParentFP = RegNodeBase - RegNodeFrameOffset
16057 /// Subtracting RegNodeSize takes us to the offset of the registration node, and
16058 /// subtracting the offset (negative on x86) takes us back to the parent FP.
16059 static SDValue recoverFramePointer(SelectionDAG &DAG, const Function *Fn,
16060 SDValue EntryEBP) {
16061 MachineFunction &MF = DAG.getMachineFunction();
16064 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16065 MVT PtrVT = TLI.getPointerTy(DAG.getDataLayout());
16067 // It's possible that the parent function no longer has a personality function
16068 // if the exceptional code was optimized away, in which case we just return
16069 // the incoming EBP.
16070 if (!Fn->hasPersonalityFn())
16073 int RegNodeSize = getSEHRegistrationNodeSize(Fn);
16075 // Get an MCSymbol that will ultimately resolve to the frame offset of the EH
16077 MCSymbol *OffsetSym =
16078 MF.getMMI().getContext().getOrCreateParentFrameOffsetSymbol(
16079 GlobalValue::getRealLinkageName(Fn->getName()));
16080 SDValue OffsetSymVal = DAG.getMCSymbol(OffsetSym, PtrVT);
16081 SDValue RegNodeFrameOffset =
16082 DAG.getNode(ISD::LOCAL_RECOVER, dl, PtrVT, OffsetSymVal);
16084 // RegNodeBase = EntryEBP - RegNodeSize
16085 // ParentFP = RegNodeBase - RegNodeFrameOffset
16086 SDValue RegNodeBase = DAG.getNode(ISD::SUB, dl, PtrVT, EntryEBP,
16087 DAG.getConstant(RegNodeSize, dl, PtrVT));
16088 return DAG.getNode(ISD::SUB, dl, PtrVT, RegNodeBase, RegNodeFrameOffset);
16091 static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
16092 SelectionDAG &DAG) {
16094 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
16095 MVT VT = Op.getSimpleValueType();
16096 const IntrinsicData* IntrData = getIntrinsicWithoutChain(IntNo);
16098 switch(IntrData->Type) {
16099 case INTR_TYPE_1OP:
16100 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1));
16101 case INTR_TYPE_2OP:
16102 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16104 case INTR_TYPE_2OP_IMM8:
16105 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16106 DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(2)));
16107 case INTR_TYPE_3OP:
16108 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16109 Op.getOperand(2), Op.getOperand(3));
16110 case INTR_TYPE_4OP:
16111 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16112 Op.getOperand(2), Op.getOperand(3), Op.getOperand(4));
16113 case INTR_TYPE_1OP_MASK_RM: {
16114 SDValue Src = Op.getOperand(1);
16115 SDValue PassThru = Op.getOperand(2);
16116 SDValue Mask = Op.getOperand(3);
16117 SDValue RoundingMode;
16118 // We allways add rounding mode to the Node.
16119 // If the rounding mode is not specified, we add the
16120 // "current direction" mode.
16121 if (Op.getNumOperands() == 4)
16123 DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
16125 RoundingMode = Op.getOperand(4);
16126 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16127 if (IntrWithRoundingModeOpcode != 0)
16128 if (cast<ConstantSDNode>(RoundingMode)->getZExtValue() !=
16129 X86::STATIC_ROUNDING::CUR_DIRECTION)
16130 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16131 dl, Op.getValueType(), Src, RoundingMode),
16132 Mask, PassThru, Subtarget, DAG);
16133 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src,
16135 Mask, PassThru, Subtarget, DAG);
16137 case INTR_TYPE_1OP_MASK: {
16138 SDValue Src = Op.getOperand(1);
16139 SDValue PassThru = Op.getOperand(2);
16140 SDValue Mask = Op.getOperand(3);
16141 // We add rounding mode to the Node when
16142 // - RM Opcode is specified and
16143 // - RM is not "current direction".
16144 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16145 if (IntrWithRoundingModeOpcode != 0) {
16146 SDValue Rnd = Op.getOperand(4);
16147 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
16148 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
16149 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16150 dl, Op.getValueType(),
16152 Mask, PassThru, Subtarget, DAG);
16155 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src),
16156 Mask, PassThru, Subtarget, DAG);
16158 case INTR_TYPE_SCALAR_MASK: {
16159 SDValue Src1 = Op.getOperand(1);
16160 SDValue Src2 = Op.getOperand(2);
16161 SDValue passThru = Op.getOperand(3);
16162 SDValue Mask = Op.getOperand(4);
16163 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2),
16164 Mask, passThru, Subtarget, DAG);
16166 case INTR_TYPE_SCALAR_MASK_RM: {
16167 SDValue Src1 = Op.getOperand(1);
16168 SDValue Src2 = Op.getOperand(2);
16169 SDValue Src0 = Op.getOperand(3);
16170 SDValue Mask = Op.getOperand(4);
16171 // There are 2 kinds of intrinsics in this group:
16172 // (1) With suppress-all-exceptions (sae) or rounding mode- 6 operands
16173 // (2) With rounding mode and sae - 7 operands.
16174 if (Op.getNumOperands() == 6) {
16175 SDValue Sae = Op.getOperand(5);
16176 unsigned Opc = IntrData->Opc1 ? IntrData->Opc1 : IntrData->Opc0;
16177 return getScalarMaskingNode(DAG.getNode(Opc, dl, VT, Src1, Src2,
16179 Mask, Src0, Subtarget, DAG);
16181 assert(Op.getNumOperands() == 7 && "Unexpected intrinsic form");
16182 SDValue RoundingMode = Op.getOperand(5);
16183 SDValue Sae = Op.getOperand(6);
16184 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2,
16185 RoundingMode, Sae),
16186 Mask, Src0, Subtarget, DAG);
16188 case INTR_TYPE_2OP_MASK:
16189 case INTR_TYPE_2OP_IMM8_MASK: {
16190 SDValue Src1 = Op.getOperand(1);
16191 SDValue Src2 = Op.getOperand(2);
16192 SDValue PassThru = Op.getOperand(3);
16193 SDValue Mask = Op.getOperand(4);
16195 if (IntrData->Type == INTR_TYPE_2OP_IMM8_MASK)
16196 Src2 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Src2);
16198 // We specify 2 possible opcodes for intrinsics with rounding modes.
16199 // First, we check if the intrinsic may have non-default rounding mode,
16200 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16201 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16202 if (IntrWithRoundingModeOpcode != 0) {
16203 SDValue Rnd = Op.getOperand(5);
16204 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
16205 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
16206 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16207 dl, Op.getValueType(),
16209 Mask, PassThru, Subtarget, DAG);
16212 // TODO: Intrinsics should have fast-math-flags to propagate.
16213 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,Src1,Src2),
16214 Mask, PassThru, Subtarget, DAG);
16216 case INTR_TYPE_2OP_MASK_RM: {
16217 SDValue Src1 = Op.getOperand(1);
16218 SDValue Src2 = Op.getOperand(2);
16219 SDValue PassThru = Op.getOperand(3);
16220 SDValue Mask = Op.getOperand(4);
16221 // We specify 2 possible modes for intrinsics, with/without rounding
16223 // First, we check if the intrinsic have rounding mode (6 operands),
16224 // if not, we set rounding mode to "current".
16226 if (Op.getNumOperands() == 6)
16227 Rnd = Op.getOperand(5);
16229 Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
16230 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16232 Mask, PassThru, Subtarget, DAG);
16234 case INTR_TYPE_3OP_SCALAR_MASK_RM: {
16235 SDValue Src1 = Op.getOperand(1);
16236 SDValue Src2 = Op.getOperand(2);
16237 SDValue Src3 = Op.getOperand(3);
16238 SDValue PassThru = Op.getOperand(4);
16239 SDValue Mask = Op.getOperand(5);
16240 SDValue Sae = Op.getOperand(6);
16242 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1,
16244 Mask, PassThru, Subtarget, DAG);
16246 case INTR_TYPE_3OP_MASK_RM: {
16247 SDValue Src1 = Op.getOperand(1);
16248 SDValue Src2 = Op.getOperand(2);
16249 SDValue Imm = Op.getOperand(3);
16250 SDValue PassThru = Op.getOperand(4);
16251 SDValue Mask = Op.getOperand(5);
16252 // We specify 2 possible modes for intrinsics, with/without rounding
16254 // First, we check if the intrinsic have rounding mode (7 operands),
16255 // if not, we set rounding mode to "current".
16257 if (Op.getNumOperands() == 7)
16258 Rnd = Op.getOperand(6);
16260 Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
16261 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16262 Src1, Src2, Imm, Rnd),
16263 Mask, PassThru, Subtarget, DAG);
16265 case INTR_TYPE_3OP_IMM8_MASK:
16266 case INTR_TYPE_3OP_MASK:
16267 case INSERT_SUBVEC: {
16268 SDValue Src1 = Op.getOperand(1);
16269 SDValue Src2 = Op.getOperand(2);
16270 SDValue Src3 = Op.getOperand(3);
16271 SDValue PassThru = Op.getOperand(4);
16272 SDValue Mask = Op.getOperand(5);
16274 if (IntrData->Type == INTR_TYPE_3OP_IMM8_MASK)
16275 Src3 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Src3);
16276 else if (IntrData->Type == INSERT_SUBVEC) {
16277 // imm should be adapted to ISD::INSERT_SUBVECTOR behavior
16278 assert(isa<ConstantSDNode>(Src3) && "Expected a ConstantSDNode here!");
16279 unsigned Imm = cast<ConstantSDNode>(Src3)->getZExtValue();
16280 Imm *= Src2.getSimpleValueType().getVectorNumElements();
16281 Src3 = DAG.getTargetConstant(Imm, dl, MVT::i32);
16284 // We specify 2 possible opcodes for intrinsics with rounding modes.
16285 // First, we check if the intrinsic may have non-default rounding mode,
16286 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16287 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16288 if (IntrWithRoundingModeOpcode != 0) {
16289 SDValue Rnd = Op.getOperand(6);
16290 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
16291 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
16292 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16293 dl, Op.getValueType(),
16294 Src1, Src2, Src3, Rnd),
16295 Mask, PassThru, Subtarget, DAG);
16298 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16300 Mask, PassThru, Subtarget, DAG);
16302 case VPERM_3OP_MASKZ:
16303 case VPERM_3OP_MASK:
16306 case FMA_OP_MASK: {
16307 SDValue Src1 = Op.getOperand(1);
16308 SDValue Src2 = Op.getOperand(2);
16309 SDValue Src3 = Op.getOperand(3);
16310 SDValue Mask = Op.getOperand(4);
16311 MVT VT = Op.getSimpleValueType();
16312 SDValue PassThru = SDValue();
16314 // set PassThru element
16315 if (IntrData->Type == VPERM_3OP_MASKZ || IntrData->Type == FMA_OP_MASKZ)
16316 PassThru = getZeroVector(VT, Subtarget, DAG, dl);
16317 else if (IntrData->Type == FMA_OP_MASK3)
16322 // We specify 2 possible opcodes for intrinsics with rounding modes.
16323 // First, we check if the intrinsic may have non-default rounding mode,
16324 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16325 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16326 if (IntrWithRoundingModeOpcode != 0) {
16327 SDValue Rnd = Op.getOperand(5);
16328 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16329 X86::STATIC_ROUNDING::CUR_DIRECTION)
16330 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16331 dl, Op.getValueType(),
16332 Src1, Src2, Src3, Rnd),
16333 Mask, PassThru, Subtarget, DAG);
16335 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0,
16336 dl, Op.getValueType(),
16338 Mask, PassThru, Subtarget, DAG);
16340 case TERLOG_OP_MASK:
16341 case TERLOG_OP_MASKZ: {
16342 SDValue Src1 = Op.getOperand(1);
16343 SDValue Src2 = Op.getOperand(2);
16344 SDValue Src3 = Op.getOperand(3);
16345 SDValue Src4 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(4));
16346 SDValue Mask = Op.getOperand(5);
16347 MVT VT = Op.getSimpleValueType();
16348 SDValue PassThru = Src1;
16349 // Set PassThru element.
16350 if (IntrData->Type == TERLOG_OP_MASKZ)
16351 PassThru = getZeroVector(VT, Subtarget, DAG, dl);
16353 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16354 Src1, Src2, Src3, Src4),
16355 Mask, PassThru, Subtarget, DAG);
16358 // FPclass intrinsics with mask
16359 SDValue Src1 = Op.getOperand(1);
16360 MVT VT = Src1.getSimpleValueType();
16361 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
16362 SDValue Imm = Op.getOperand(2);
16363 SDValue Mask = Op.getOperand(3);
16364 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
16365 Mask.getSimpleValueType().getSizeInBits());
16366 SDValue FPclass = DAG.getNode(IntrData->Opc0, dl, MaskVT, Src1, Imm);
16367 SDValue FPclassMask = getVectorMaskingNode(FPclass, Mask,
16368 DAG.getTargetConstant(0, dl, MaskVT),
16370 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
16371 DAG.getUNDEF(BitcastVT), FPclassMask,
16372 DAG.getIntPtrConstant(0, dl));
16373 return DAG.getBitcast(Op.getValueType(), Res);
16376 SDValue Src1 = Op.getOperand(1);
16377 SDValue Imm = Op.getOperand(2);
16378 SDValue Mask = Op.getOperand(3);
16379 SDValue FPclass = DAG.getNode(IntrData->Opc0, dl, MVT::i1, Src1, Imm);
16380 SDValue FPclassMask = getScalarMaskingNode(FPclass, Mask,
16381 DAG.getTargetConstant(0, dl, MVT::i1), Subtarget, DAG);
16382 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i8, FPclassMask);
16385 case CMP_MASK_CC: {
16386 // Comparison intrinsics with masks.
16387 // Example of transformation:
16388 // (i8 (int_x86_avx512_mask_pcmpeq_q_128
16389 // (v2i64 %a), (v2i64 %b), (i8 %mask))) ->
16391 // (v8i1 (insert_subvector undef,
16392 // (v2i1 (and (PCMPEQM %a, %b),
16393 // (extract_subvector
16394 // (v8i1 (bitcast %mask)), 0))), 0))))
16395 MVT VT = Op.getOperand(1).getSimpleValueType();
16396 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
16397 SDValue Mask = Op.getOperand((IntrData->Type == CMP_MASK_CC) ? 4 : 3);
16398 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
16399 Mask.getSimpleValueType().getSizeInBits());
16401 if (IntrData->Type == CMP_MASK_CC) {
16402 SDValue CC = Op.getOperand(3);
16403 CC = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, CC);
16404 // We specify 2 possible opcodes for intrinsics with rounding modes.
16405 // First, we check if the intrinsic may have non-default rounding mode,
16406 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16407 if (IntrData->Opc1 != 0) {
16408 SDValue Rnd = Op.getOperand(5);
16409 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16410 X86::STATIC_ROUNDING::CUR_DIRECTION)
16411 Cmp = DAG.getNode(IntrData->Opc1, dl, MaskVT, Op.getOperand(1),
16412 Op.getOperand(2), CC, Rnd);
16414 //default rounding mode
16416 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
16417 Op.getOperand(2), CC);
16420 assert(IntrData->Type == CMP_MASK && "Unexpected intrinsic type!");
16421 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
16424 SDValue CmpMask = getVectorMaskingNode(Cmp, Mask,
16425 DAG.getTargetConstant(0, dl,
16428 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
16429 DAG.getUNDEF(BitcastVT), CmpMask,
16430 DAG.getIntPtrConstant(0, dl));
16431 return DAG.getBitcast(Op.getValueType(), Res);
16433 case CMP_MASK_SCALAR_CC: {
16434 SDValue Src1 = Op.getOperand(1);
16435 SDValue Src2 = Op.getOperand(2);
16436 SDValue CC = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(3));
16437 SDValue Mask = Op.getOperand(4);
16440 if (IntrData->Opc1 != 0) {
16441 SDValue Rnd = Op.getOperand(5);
16442 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16443 X86::STATIC_ROUNDING::CUR_DIRECTION)
16444 Cmp = DAG.getNode(IntrData->Opc1, dl, MVT::i1, Src1, Src2, CC, Rnd);
16446 //default rounding mode
16448 Cmp = DAG.getNode(IntrData->Opc0, dl, MVT::i1, Src1, Src2, CC);
16450 SDValue CmpMask = getScalarMaskingNode(Cmp, Mask,
16451 DAG.getTargetConstant(0, dl,
16455 return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::i8,
16456 DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i8, CmpMask),
16457 DAG.getValueType(MVT::i1));
16459 case COMI: { // Comparison intrinsics
16460 ISD::CondCode CC = (ISD::CondCode)IntrData->Opc1;
16461 SDValue LHS = Op.getOperand(1);
16462 SDValue RHS = Op.getOperand(2);
16463 unsigned X86CC = TranslateX86CC(CC, dl, true, LHS, RHS, DAG);
16464 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
16465 SDValue Cond = DAG.getNode(IntrData->Opc0, dl, MVT::i32, LHS, RHS);
16466 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16467 DAG.getConstant(X86CC, dl, MVT::i8), Cond);
16468 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16471 return getTargetVShiftNode(IntrData->Opc0, dl, Op.getSimpleValueType(),
16472 Op.getOperand(1), Op.getOperand(2), DAG);
16474 return getVectorMaskingNode(getTargetVShiftNode(IntrData->Opc0, dl,
16475 Op.getSimpleValueType(),
16477 Op.getOperand(2), DAG),
16478 Op.getOperand(4), Op.getOperand(3), Subtarget,
16480 case COMPRESS_EXPAND_IN_REG: {
16481 SDValue Mask = Op.getOperand(3);
16482 SDValue DataToCompress = Op.getOperand(1);
16483 SDValue PassThru = Op.getOperand(2);
16484 if (isAllOnes(Mask)) // return data as is
16485 return Op.getOperand(1);
16487 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16489 Mask, PassThru, Subtarget, DAG);
16492 SDValue Mask = Op.getOperand(1);
16493 MVT MaskVT = MVT::getVectorVT(MVT::i1, Mask.getSimpleValueType().getSizeInBits());
16494 Mask = DAG.getBitcast(MaskVT, Mask);
16495 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Mask);
16498 SDValue Mask = Op.getOperand(3);
16499 MVT VT = Op.getSimpleValueType();
16500 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
16501 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
16502 Mask.getSimpleValueType().getSizeInBits());
16504 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16505 DAG.getBitcast(BitcastVT, Mask),
16506 DAG.getIntPtrConstant(0, dl));
16507 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, Op.getOperand(1),
16516 default: return SDValue(); // Don't custom lower most intrinsics.
16518 case Intrinsic::x86_avx2_permd:
16519 case Intrinsic::x86_avx2_permps:
16520 // Operands intentionally swapped. Mask is last operand to intrinsic,
16521 // but second operand for node/instruction.
16522 return DAG.getNode(X86ISD::VPERMV, dl, Op.getValueType(),
16523 Op.getOperand(2), Op.getOperand(1));
16525 // ptest and testp intrinsics. The intrinsic these come from are designed to
16526 // return an integer value, not just an instruction so lower it to the ptest
16527 // or testp pattern and a setcc for the result.
16528 case Intrinsic::x86_sse41_ptestz:
16529 case Intrinsic::x86_sse41_ptestc:
16530 case Intrinsic::x86_sse41_ptestnzc:
16531 case Intrinsic::x86_avx_ptestz_256:
16532 case Intrinsic::x86_avx_ptestc_256:
16533 case Intrinsic::x86_avx_ptestnzc_256:
16534 case Intrinsic::x86_avx_vtestz_ps:
16535 case Intrinsic::x86_avx_vtestc_ps:
16536 case Intrinsic::x86_avx_vtestnzc_ps:
16537 case Intrinsic::x86_avx_vtestz_pd:
16538 case Intrinsic::x86_avx_vtestc_pd:
16539 case Intrinsic::x86_avx_vtestnzc_pd:
16540 case Intrinsic::x86_avx_vtestz_ps_256:
16541 case Intrinsic::x86_avx_vtestc_ps_256:
16542 case Intrinsic::x86_avx_vtestnzc_ps_256:
16543 case Intrinsic::x86_avx_vtestz_pd_256:
16544 case Intrinsic::x86_avx_vtestc_pd_256:
16545 case Intrinsic::x86_avx_vtestnzc_pd_256: {
16546 bool IsTestPacked = false;
16549 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
16550 case Intrinsic::x86_avx_vtestz_ps:
16551 case Intrinsic::x86_avx_vtestz_pd:
16552 case Intrinsic::x86_avx_vtestz_ps_256:
16553 case Intrinsic::x86_avx_vtestz_pd_256:
16554 IsTestPacked = true; // Fallthrough
16555 case Intrinsic::x86_sse41_ptestz:
16556 case Intrinsic::x86_avx_ptestz_256:
16558 X86CC = X86::COND_E;
16560 case Intrinsic::x86_avx_vtestc_ps:
16561 case Intrinsic::x86_avx_vtestc_pd:
16562 case Intrinsic::x86_avx_vtestc_ps_256:
16563 case Intrinsic::x86_avx_vtestc_pd_256:
16564 IsTestPacked = true; // Fallthrough
16565 case Intrinsic::x86_sse41_ptestc:
16566 case Intrinsic::x86_avx_ptestc_256:
16568 X86CC = X86::COND_B;
16570 case Intrinsic::x86_avx_vtestnzc_ps:
16571 case Intrinsic::x86_avx_vtestnzc_pd:
16572 case Intrinsic::x86_avx_vtestnzc_ps_256:
16573 case Intrinsic::x86_avx_vtestnzc_pd_256:
16574 IsTestPacked = true; // Fallthrough
16575 case Intrinsic::x86_sse41_ptestnzc:
16576 case Intrinsic::x86_avx_ptestnzc_256:
16578 X86CC = X86::COND_A;
16582 SDValue LHS = Op.getOperand(1);
16583 SDValue RHS = Op.getOperand(2);
16584 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
16585 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
16586 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
16587 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
16588 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16590 case Intrinsic::x86_avx512_kortestz_w:
16591 case Intrinsic::x86_avx512_kortestc_w: {
16592 unsigned X86CC = (IntNo == Intrinsic::x86_avx512_kortestz_w)? X86::COND_E: X86::COND_B;
16593 SDValue LHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(1));
16594 SDValue RHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(2));
16595 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
16596 SDValue Test = DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, LHS, RHS);
16597 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i1, CC, Test);
16598 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16601 case Intrinsic::x86_sse42_pcmpistria128:
16602 case Intrinsic::x86_sse42_pcmpestria128:
16603 case Intrinsic::x86_sse42_pcmpistric128:
16604 case Intrinsic::x86_sse42_pcmpestric128:
16605 case Intrinsic::x86_sse42_pcmpistrio128:
16606 case Intrinsic::x86_sse42_pcmpestrio128:
16607 case Intrinsic::x86_sse42_pcmpistris128:
16608 case Intrinsic::x86_sse42_pcmpestris128:
16609 case Intrinsic::x86_sse42_pcmpistriz128:
16610 case Intrinsic::x86_sse42_pcmpestriz128: {
16614 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
16615 case Intrinsic::x86_sse42_pcmpistria128:
16616 Opcode = X86ISD::PCMPISTRI;
16617 X86CC = X86::COND_A;
16619 case Intrinsic::x86_sse42_pcmpestria128:
16620 Opcode = X86ISD::PCMPESTRI;
16621 X86CC = X86::COND_A;
16623 case Intrinsic::x86_sse42_pcmpistric128:
16624 Opcode = X86ISD::PCMPISTRI;
16625 X86CC = X86::COND_B;
16627 case Intrinsic::x86_sse42_pcmpestric128:
16628 Opcode = X86ISD::PCMPESTRI;
16629 X86CC = X86::COND_B;
16631 case Intrinsic::x86_sse42_pcmpistrio128:
16632 Opcode = X86ISD::PCMPISTRI;
16633 X86CC = X86::COND_O;
16635 case Intrinsic::x86_sse42_pcmpestrio128:
16636 Opcode = X86ISD::PCMPESTRI;
16637 X86CC = X86::COND_O;
16639 case Intrinsic::x86_sse42_pcmpistris128:
16640 Opcode = X86ISD::PCMPISTRI;
16641 X86CC = X86::COND_S;
16643 case Intrinsic::x86_sse42_pcmpestris128:
16644 Opcode = X86ISD::PCMPESTRI;
16645 X86CC = X86::COND_S;
16647 case Intrinsic::x86_sse42_pcmpistriz128:
16648 Opcode = X86ISD::PCMPISTRI;
16649 X86CC = X86::COND_E;
16651 case Intrinsic::x86_sse42_pcmpestriz128:
16652 Opcode = X86ISD::PCMPESTRI;
16653 X86CC = X86::COND_E;
16656 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
16657 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
16658 SDValue PCMP = DAG.getNode(Opcode, dl, VTs, NewOps);
16659 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16660 DAG.getConstant(X86CC, dl, MVT::i8),
16661 SDValue(PCMP.getNode(), 1));
16662 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16665 case Intrinsic::x86_sse42_pcmpistri128:
16666 case Intrinsic::x86_sse42_pcmpestri128: {
16668 if (IntNo == Intrinsic::x86_sse42_pcmpistri128)
16669 Opcode = X86ISD::PCMPISTRI;
16671 Opcode = X86ISD::PCMPESTRI;
16673 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
16674 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
16675 return DAG.getNode(Opcode, dl, VTs, NewOps);
16678 case Intrinsic::x86_seh_lsda: {
16679 // Compute the symbol for the LSDA. We know it'll get emitted later.
16680 MachineFunction &MF = DAG.getMachineFunction();
16681 SDValue Op1 = Op.getOperand(1);
16682 auto *Fn = cast<Function>(cast<GlobalAddressSDNode>(Op1)->getGlobal());
16683 MCSymbol *LSDASym = MF.getMMI().getContext().getOrCreateLSDASymbol(
16684 GlobalValue::getRealLinkageName(Fn->getName()));
16686 // Generate a simple absolute symbol reference. This intrinsic is only
16687 // supported on 32-bit Windows, which isn't PIC.
16688 SDValue Result = DAG.getMCSymbol(LSDASym, VT);
16689 return DAG.getNode(X86ISD::Wrapper, dl, VT, Result);
16692 case Intrinsic::x86_seh_recoverfp: {
16693 SDValue FnOp = Op.getOperand(1);
16694 SDValue IncomingFPOp = Op.getOperand(2);
16695 GlobalAddressSDNode *GSD = dyn_cast<GlobalAddressSDNode>(FnOp);
16696 auto *Fn = dyn_cast_or_null<Function>(GSD ? GSD->getGlobal() : nullptr);
16698 report_fatal_error(
16699 "llvm.x86.seh.recoverfp must take a function as the first argument");
16700 return recoverFramePointer(DAG, Fn, IncomingFPOp);
16703 case Intrinsic::localaddress: {
16704 // Returns one of the stack, base, or frame pointer registers, depending on
16705 // which is used to reference local variables.
16706 MachineFunction &MF = DAG.getMachineFunction();
16707 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16709 if (RegInfo->hasBasePointer(MF))
16710 Reg = RegInfo->getBaseRegister();
16711 else // This function handles the SP or FP case.
16712 Reg = RegInfo->getPtrSizedFrameRegister(MF);
16713 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
16718 static SDValue getGatherNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16719 SDValue Src, SDValue Mask, SDValue Base,
16720 SDValue Index, SDValue ScaleOp, SDValue Chain,
16721 const X86Subtarget * Subtarget) {
16723 auto *C = cast<ConstantSDNode>(ScaleOp);
16724 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16725 MVT MaskVT = MVT::getVectorVT(MVT::i1,
16726 Index.getSimpleValueType().getVectorNumElements());
16728 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16730 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16732 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
16733 Mask.getSimpleValueType().getSizeInBits());
16735 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16736 // are extracted by EXTRACT_SUBVECTOR.
16737 MaskInReg = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16738 DAG.getBitcast(BitcastVT, Mask),
16739 DAG.getIntPtrConstant(0, dl));
16741 SDVTList VTs = DAG.getVTList(Op.getValueType(), MaskVT, MVT::Other);
16742 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16743 SDValue Segment = DAG.getRegister(0, MVT::i32);
16744 if (Src.getOpcode() == ISD::UNDEF)
16745 Src = getZeroVector(Op.getValueType(), Subtarget, DAG, dl);
16746 SDValue Ops[] = {Src, MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
16747 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
16748 SDValue RetOps[] = { SDValue(Res, 0), SDValue(Res, 2) };
16749 return DAG.getMergeValues(RetOps, dl);
16752 static SDValue getScatterNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16753 SDValue Src, SDValue Mask, SDValue Base,
16754 SDValue Index, SDValue ScaleOp, SDValue Chain) {
16756 auto *C = cast<ConstantSDNode>(ScaleOp);
16757 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16758 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16759 SDValue Segment = DAG.getRegister(0, MVT::i32);
16760 MVT MaskVT = MVT::getVectorVT(MVT::i1,
16761 Index.getSimpleValueType().getVectorNumElements());
16763 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16765 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16767 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
16768 Mask.getSimpleValueType().getSizeInBits());
16770 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16771 // are extracted by EXTRACT_SUBVECTOR.
16772 MaskInReg = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16773 DAG.getBitcast(BitcastVT, Mask),
16774 DAG.getIntPtrConstant(0, dl));
16776 SDVTList VTs = DAG.getVTList(MaskVT, MVT::Other);
16777 SDValue Ops[] = {Base, Scale, Index, Disp, Segment, MaskInReg, Src, Chain};
16778 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
16779 return SDValue(Res, 1);
16782 static SDValue getPrefetchNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16783 SDValue Mask, SDValue Base, SDValue Index,
16784 SDValue ScaleOp, SDValue Chain) {
16786 auto *C = cast<ConstantSDNode>(ScaleOp);
16787 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16788 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16789 SDValue Segment = DAG.getRegister(0, MVT::i32);
16791 MVT::getVectorVT(MVT::i1, Index.getSimpleValueType().getVectorNumElements());
16793 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16795 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16797 MaskInReg = DAG.getBitcast(MaskVT, Mask);
16798 //SDVTList VTs = DAG.getVTList(MVT::Other);
16799 SDValue Ops[] = {MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
16800 SDNode *Res = DAG.getMachineNode(Opc, dl, MVT::Other, Ops);
16801 return SDValue(Res, 0);
16804 // getReadPerformanceCounter - Handles the lowering of builtin intrinsics that
16805 // read performance monitor counters (x86_rdpmc).
16806 static void getReadPerformanceCounter(SDNode *N, SDLoc DL,
16807 SelectionDAG &DAG, const X86Subtarget *Subtarget,
16808 SmallVectorImpl<SDValue> &Results) {
16809 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
16810 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
16813 // The ECX register is used to select the index of the performance counter
16815 SDValue Chain = DAG.getCopyToReg(N->getOperand(0), DL, X86::ECX,
16817 SDValue rd = DAG.getNode(X86ISD::RDPMC_DAG, DL, Tys, Chain);
16819 // Reads the content of a 64-bit performance counter and returns it in the
16820 // registers EDX:EAX.
16821 if (Subtarget->is64Bit()) {
16822 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
16823 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
16826 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
16827 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
16830 Chain = HI.getValue(1);
16832 if (Subtarget->is64Bit()) {
16833 // The EAX register is loaded with the low-order 32 bits. The EDX register
16834 // is loaded with the supported high-order bits of the counter.
16835 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
16836 DAG.getConstant(32, DL, MVT::i8));
16837 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
16838 Results.push_back(Chain);
16842 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
16843 SDValue Ops[] = { LO, HI };
16844 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
16845 Results.push_back(Pair);
16846 Results.push_back(Chain);
16849 // getReadTimeStampCounter - Handles the lowering of builtin intrinsics that
16850 // read the time stamp counter (x86_rdtsc and x86_rdtscp). This function is
16851 // also used to custom lower READCYCLECOUNTER nodes.
16852 static void getReadTimeStampCounter(SDNode *N, SDLoc DL, unsigned Opcode,
16853 SelectionDAG &DAG, const X86Subtarget *Subtarget,
16854 SmallVectorImpl<SDValue> &Results) {
16855 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
16856 SDValue rd = DAG.getNode(Opcode, DL, Tys, N->getOperand(0));
16859 // The processor's time-stamp counter (a 64-bit MSR) is stored into the
16860 // EDX:EAX registers. EDX is loaded with the high-order 32 bits of the MSR
16861 // and the EAX register is loaded with the low-order 32 bits.
16862 if (Subtarget->is64Bit()) {
16863 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
16864 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
16867 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
16868 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
16871 SDValue Chain = HI.getValue(1);
16873 if (Opcode == X86ISD::RDTSCP_DAG) {
16874 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
16876 // Instruction RDTSCP loads the IA32:TSC_AUX_MSR (address C000_0103H) into
16877 // the ECX register. Add 'ecx' explicitly to the chain.
16878 SDValue ecx = DAG.getCopyFromReg(Chain, DL, X86::ECX, MVT::i32,
16880 // Explicitly store the content of ECX at the location passed in input
16881 // to the 'rdtscp' intrinsic.
16882 Chain = DAG.getStore(ecx.getValue(1), DL, ecx, N->getOperand(2),
16883 MachinePointerInfo(), false, false, 0);
16886 if (Subtarget->is64Bit()) {
16887 // The EDX register is loaded with the high-order 32 bits of the MSR, and
16888 // the EAX register is loaded with the low-order 32 bits.
16889 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
16890 DAG.getConstant(32, DL, MVT::i8));
16891 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
16892 Results.push_back(Chain);
16896 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
16897 SDValue Ops[] = { LO, HI };
16898 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
16899 Results.push_back(Pair);
16900 Results.push_back(Chain);
16903 static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget *Subtarget,
16904 SelectionDAG &DAG) {
16905 SmallVector<SDValue, 2> Results;
16907 getReadTimeStampCounter(Op.getNode(), DL, X86ISD::RDTSC_DAG, DAG, Subtarget,
16909 return DAG.getMergeValues(Results, DL);
16912 static SDValue LowerSEHRESTOREFRAME(SDValue Op, const X86Subtarget *Subtarget,
16913 SelectionDAG &DAG) {
16914 MachineFunction &MF = DAG.getMachineFunction();
16915 const Function *Fn = MF.getFunction();
16917 SDValue Chain = Op.getOperand(0);
16919 assert(Subtarget->getFrameLowering()->hasFP(MF) &&
16920 "using llvm.x86.seh.restoreframe requires a frame pointer");
16922 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16923 MVT VT = TLI.getPointerTy(DAG.getDataLayout());
16925 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16926 unsigned FrameReg =
16927 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
16928 unsigned SPReg = RegInfo->getStackRegister();
16929 unsigned SlotSize = RegInfo->getSlotSize();
16931 // Get incoming EBP.
16932 SDValue IncomingEBP =
16933 DAG.getCopyFromReg(Chain, dl, FrameReg, VT);
16935 // SP is saved in the first field of every registration node, so load
16936 // [EBP-RegNodeSize] into SP.
16937 int RegNodeSize = getSEHRegistrationNodeSize(Fn);
16938 SDValue SPAddr = DAG.getNode(ISD::ADD, dl, VT, IncomingEBP,
16939 DAG.getConstant(-RegNodeSize, dl, VT));
16941 DAG.getLoad(VT, dl, Chain, SPAddr, MachinePointerInfo(), false, false,
16942 false, VT.getScalarSizeInBits() / 8);
16943 Chain = DAG.getCopyToReg(Chain, dl, SPReg, NewSP);
16945 if (!RegInfo->needsStackRealignment(MF)) {
16946 // Adjust EBP to point back to the original frame position.
16947 SDValue NewFP = recoverFramePointer(DAG, Fn, IncomingEBP);
16948 Chain = DAG.getCopyToReg(Chain, dl, FrameReg, NewFP);
16950 assert(RegInfo->hasBasePointer(MF) &&
16951 "functions with Win32 EH must use frame or base pointer register");
16953 // Reload the base pointer (ESI) with the adjusted incoming EBP.
16954 SDValue NewBP = recoverFramePointer(DAG, Fn, IncomingEBP);
16955 Chain = DAG.getCopyToReg(Chain, dl, RegInfo->getBaseRegister(), NewBP);
16957 // Reload the spilled EBP value, now that the stack and base pointers are
16959 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
16960 X86FI->setHasSEHFramePtrSave(true);
16961 int FI = MF.getFrameInfo()->CreateSpillStackObject(SlotSize, SlotSize);
16962 X86FI->setSEHFramePtrSaveIndex(FI);
16963 SDValue NewFP = DAG.getLoad(VT, dl, Chain, DAG.getFrameIndex(FI, VT),
16964 MachinePointerInfo(), false, false, false,
16965 VT.getScalarSizeInBits() / 8);
16966 Chain = DAG.getCopyToReg(NewFP, dl, FrameReg, NewFP);
16972 static SDValue MarkEHRegistrationNode(SDValue Op, SelectionDAG &DAG) {
16973 MachineFunction &MF = DAG.getMachineFunction();
16974 SDValue Chain = Op.getOperand(0);
16975 SDValue RegNode = Op.getOperand(2);
16976 WinEHFuncInfo *EHInfo = MF.getWinEHFuncInfo();
16978 report_fatal_error("EH registrations only live in functions using WinEH");
16980 // Cast the operand to an alloca, and remember the frame index.
16981 auto *FINode = dyn_cast<FrameIndexSDNode>(RegNode);
16983 report_fatal_error("llvm.x86.seh.ehregnode expects a static alloca");
16984 EHInfo->EHRegNodeFrameIndex = FINode->getIndex();
16986 // Return the chain operand without making any DAG nodes.
16990 /// \brief Lower intrinsics for TRUNCATE_TO_MEM case
16991 /// return truncate Store/MaskedStore Node
16992 static SDValue LowerINTRINSIC_TRUNCATE_TO_MEM(const SDValue & Op,
16996 SDValue Mask = Op.getOperand(4);
16997 SDValue DataToTruncate = Op.getOperand(3);
16998 SDValue Addr = Op.getOperand(2);
16999 SDValue Chain = Op.getOperand(0);
17001 MVT VT = DataToTruncate.getSimpleValueType();
17002 MVT SVT = MVT::getVectorVT(ElementType, VT.getVectorNumElements());
17004 if (isAllOnes(Mask)) // return just a truncate store
17005 return DAG.getTruncStore(Chain, dl, DataToTruncate, Addr,
17006 MachinePointerInfo(), SVT, false, false,
17007 SVT.getScalarSizeInBits()/8);
17009 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
17010 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
17011 Mask.getSimpleValueType().getSizeInBits());
17012 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
17013 // are extracted by EXTRACT_SUBVECTOR.
17014 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
17015 DAG.getBitcast(BitcastVT, Mask),
17016 DAG.getIntPtrConstant(0, dl));
17018 MachineMemOperand *MMO = DAG.getMachineFunction().
17019 getMachineMemOperand(MachinePointerInfo(),
17020 MachineMemOperand::MOStore, SVT.getStoreSize(),
17021 SVT.getScalarSizeInBits()/8);
17023 return DAG.getMaskedStore(Chain, dl, DataToTruncate, Addr,
17024 VMask, SVT, MMO, true);
17027 static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
17028 SelectionDAG &DAG) {
17029 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
17031 const IntrinsicData* IntrData = getIntrinsicWithChain(IntNo);
17033 if (IntNo == llvm::Intrinsic::x86_seh_restoreframe)
17034 return LowerSEHRESTOREFRAME(Op, Subtarget, DAG);
17035 else if (IntNo == llvm::Intrinsic::x86_seh_ehregnode)
17036 return MarkEHRegistrationNode(Op, DAG);
17041 switch(IntrData->Type) {
17042 default: llvm_unreachable("Unknown Intrinsic Type");
17045 // Emit the node with the right value type.
17046 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other);
17047 SDValue Result = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
17049 // If the value returned by RDRAND/RDSEED was valid (CF=1), return 1.
17050 // Otherwise return the value from Rand, which is always 0, casted to i32.
17051 SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)),
17052 DAG.getConstant(1, dl, Op->getValueType(1)),
17053 DAG.getConstant(X86::COND_B, dl, MVT::i32),
17054 SDValue(Result.getNode(), 1) };
17055 SDValue isValid = DAG.getNode(X86ISD::CMOV, dl,
17056 DAG.getVTList(Op->getValueType(1), MVT::Glue),
17059 // Return { result, isValid, chain }.
17060 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid,
17061 SDValue(Result.getNode(), 2));
17064 //gather(v1, mask, index, base, scale);
17065 SDValue Chain = Op.getOperand(0);
17066 SDValue Src = Op.getOperand(2);
17067 SDValue Base = Op.getOperand(3);
17068 SDValue Index = Op.getOperand(4);
17069 SDValue Mask = Op.getOperand(5);
17070 SDValue Scale = Op.getOperand(6);
17071 return getGatherNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale,
17075 //scatter(base, mask, index, v1, scale);
17076 SDValue Chain = Op.getOperand(0);
17077 SDValue Base = Op.getOperand(2);
17078 SDValue Mask = Op.getOperand(3);
17079 SDValue Index = Op.getOperand(4);
17080 SDValue Src = Op.getOperand(5);
17081 SDValue Scale = Op.getOperand(6);
17082 return getScatterNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index,
17086 SDValue Hint = Op.getOperand(6);
17087 unsigned HintVal = cast<ConstantSDNode>(Hint)->getZExtValue();
17088 assert(HintVal < 2 && "Wrong prefetch hint in intrinsic: should be 0 or 1");
17089 unsigned Opcode = (HintVal ? IntrData->Opc1 : IntrData->Opc0);
17090 SDValue Chain = Op.getOperand(0);
17091 SDValue Mask = Op.getOperand(2);
17092 SDValue Index = Op.getOperand(3);
17093 SDValue Base = Op.getOperand(4);
17094 SDValue Scale = Op.getOperand(5);
17095 return getPrefetchNode(Opcode, Op, DAG, Mask, Base, Index, Scale, Chain);
17097 // Read Time Stamp Counter (RDTSC) and Processor ID (RDTSCP).
17099 SmallVector<SDValue, 2> Results;
17100 getReadTimeStampCounter(Op.getNode(), dl, IntrData->Opc0, DAG, Subtarget,
17102 return DAG.getMergeValues(Results, dl);
17104 // Read Performance Monitoring Counters.
17106 SmallVector<SDValue, 2> Results;
17107 getReadPerformanceCounter(Op.getNode(), dl, DAG, Subtarget, Results);
17108 return DAG.getMergeValues(Results, dl);
17110 // XTEST intrinsics.
17112 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
17113 SDValue InTrans = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
17114 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17115 DAG.getConstant(X86::COND_NE, dl, MVT::i8),
17117 SDValue Ret = DAG.getNode(ISD::ZERO_EXTEND, dl, Op->getValueType(0), SetCC);
17118 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(),
17119 Ret, SDValue(InTrans.getNode(), 1));
17123 SmallVector<SDValue, 2> Results;
17124 SDVTList CFVTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
17125 SDVTList VTs = DAG.getVTList(Op.getOperand(3)->getValueType(0), MVT::Other);
17126 SDValue GenCF = DAG.getNode(X86ISD::ADD, dl, CFVTs, Op.getOperand(2),
17127 DAG.getConstant(-1, dl, MVT::i8));
17128 SDValue Res = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(3),
17129 Op.getOperand(4), GenCF.getValue(1));
17130 SDValue Store = DAG.getStore(Op.getOperand(0), dl, Res.getValue(0),
17131 Op.getOperand(5), MachinePointerInfo(),
17133 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17134 DAG.getConstant(X86::COND_B, dl, MVT::i8),
17136 Results.push_back(SetCC);
17137 Results.push_back(Store);
17138 return DAG.getMergeValues(Results, dl);
17140 case COMPRESS_TO_MEM: {
17142 SDValue Mask = Op.getOperand(4);
17143 SDValue DataToCompress = Op.getOperand(3);
17144 SDValue Addr = Op.getOperand(2);
17145 SDValue Chain = Op.getOperand(0);
17147 MVT VT = DataToCompress.getSimpleValueType();
17148 if (isAllOnes(Mask)) // return just a store
17149 return DAG.getStore(Chain, dl, DataToCompress, Addr,
17150 MachinePointerInfo(), false, false,
17151 VT.getScalarSizeInBits()/8);
17153 SDValue Compressed =
17154 getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, DataToCompress),
17155 Mask, DAG.getUNDEF(VT), Subtarget, DAG);
17156 return DAG.getStore(Chain, dl, Compressed, Addr,
17157 MachinePointerInfo(), false, false,
17158 VT.getScalarSizeInBits()/8);
17160 case TRUNCATE_TO_MEM_VI8:
17161 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i8);
17162 case TRUNCATE_TO_MEM_VI16:
17163 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i16);
17164 case TRUNCATE_TO_MEM_VI32:
17165 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i32);
17166 case EXPAND_FROM_MEM: {
17168 SDValue Mask = Op.getOperand(4);
17169 SDValue PassThru = Op.getOperand(3);
17170 SDValue Addr = Op.getOperand(2);
17171 SDValue Chain = Op.getOperand(0);
17172 MVT VT = Op.getSimpleValueType();
17174 if (isAllOnes(Mask)) // return just a load
17175 return DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(), false, false,
17176 false, VT.getScalarSizeInBits()/8);
17178 SDValue DataToExpand = DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(),
17179 false, false, false,
17180 VT.getScalarSizeInBits()/8);
17182 SDValue Results[] = {
17183 getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, DataToExpand),
17184 Mask, PassThru, Subtarget, DAG), Chain};
17185 return DAG.getMergeValues(Results, dl);
17190 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
17191 SelectionDAG &DAG) const {
17192 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
17193 MFI->setReturnAddressIsTaken(true);
17195 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
17198 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
17200 EVT PtrVT = getPointerTy(DAG.getDataLayout());
17203 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
17204 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17205 SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), dl, PtrVT);
17206 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
17207 DAG.getNode(ISD::ADD, dl, PtrVT,
17208 FrameAddr, Offset),
17209 MachinePointerInfo(), false, false, false, 0);
17212 // Just load the return address.
17213 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
17214 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
17215 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
17218 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
17219 MachineFunction &MF = DAG.getMachineFunction();
17220 MachineFrameInfo *MFI = MF.getFrameInfo();
17221 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
17222 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17223 EVT VT = Op.getValueType();
17225 MFI->setFrameAddressIsTaken(true);
17227 if (MF.getTarget().getMCAsmInfo()->usesWindowsCFI()) {
17228 // Depth > 0 makes no sense on targets which use Windows unwind codes. It
17229 // is not possible to crawl up the stack without looking at the unwind codes
17231 int FrameAddrIndex = FuncInfo->getFAIndex();
17232 if (!FrameAddrIndex) {
17233 // Set up a frame object for the return address.
17234 unsigned SlotSize = RegInfo->getSlotSize();
17235 FrameAddrIndex = MF.getFrameInfo()->CreateFixedObject(
17236 SlotSize, /*Offset=*/0, /*IsImmutable=*/false);
17237 FuncInfo->setFAIndex(FrameAddrIndex);
17239 return DAG.getFrameIndex(FrameAddrIndex, VT);
17242 unsigned FrameReg =
17243 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
17244 SDLoc dl(Op); // FIXME probably not meaningful
17245 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
17246 assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
17247 (FrameReg == X86::EBP && VT == MVT::i32)) &&
17248 "Invalid Frame Register!");
17249 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
17251 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
17252 MachinePointerInfo(),
17253 false, false, false, 0);
17257 // FIXME? Maybe this could be a TableGen attribute on some registers and
17258 // this table could be generated automatically from RegInfo.
17259 unsigned X86TargetLowering::getRegisterByName(const char* RegName, EVT VT,
17260 SelectionDAG &DAG) const {
17261 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
17262 const MachineFunction &MF = DAG.getMachineFunction();
17264 unsigned Reg = StringSwitch<unsigned>(RegName)
17265 .Case("esp", X86::ESP)
17266 .Case("rsp", X86::RSP)
17267 .Case("ebp", X86::EBP)
17268 .Case("rbp", X86::RBP)
17271 if (Reg == X86::EBP || Reg == X86::RBP) {
17272 if (!TFI.hasFP(MF))
17273 report_fatal_error("register " + StringRef(RegName) +
17274 " is allocatable: function has no frame pointer");
17277 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17278 unsigned FrameReg =
17279 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
17280 assert((FrameReg == X86::EBP || FrameReg == X86::RBP) &&
17281 "Invalid Frame Register!");
17289 report_fatal_error("Invalid register name global variable");
17292 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
17293 SelectionDAG &DAG) const {
17294 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17295 return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize(), SDLoc(Op));
17298 unsigned X86TargetLowering::getExceptionPointerRegister(
17299 const Constant *PersonalityFn) const {
17300 if (classifyEHPersonality(PersonalityFn) == EHPersonality::CoreCLR)
17301 return Subtarget->isTarget64BitLP64() ? X86::RDX : X86::EDX;
17303 return Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX;
17306 unsigned X86TargetLowering::getExceptionSelectorRegister(
17307 const Constant *PersonalityFn) const {
17308 // Funclet personalities don't use selectors (the runtime does the selection).
17309 assert(!isFuncletEHPersonality(classifyEHPersonality(PersonalityFn)));
17310 return Subtarget->isTarget64BitLP64() ? X86::RDX : X86::EDX;
17313 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
17314 SDValue Chain = Op.getOperand(0);
17315 SDValue Offset = Op.getOperand(1);
17316 SDValue Handler = Op.getOperand(2);
17319 EVT PtrVT = getPointerTy(DAG.getDataLayout());
17320 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17321 unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
17322 assert(((FrameReg == X86::RBP && PtrVT == MVT::i64) ||
17323 (FrameReg == X86::EBP && PtrVT == MVT::i32)) &&
17324 "Invalid Frame Register!");
17325 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, PtrVT);
17326 unsigned StoreAddrReg = (PtrVT == MVT::i64) ? X86::RCX : X86::ECX;
17328 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Frame,
17329 DAG.getIntPtrConstant(RegInfo->getSlotSize(),
17331 StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, StoreAddr, Offset);
17332 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
17334 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
17336 return DAG.getNode(X86ISD::EH_RETURN, dl, MVT::Other, Chain,
17337 DAG.getRegister(StoreAddrReg, PtrVT));
17340 SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
17341 SelectionDAG &DAG) const {
17343 return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL,
17344 DAG.getVTList(MVT::i32, MVT::Other),
17345 Op.getOperand(0), Op.getOperand(1));
17348 SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
17349 SelectionDAG &DAG) const {
17351 return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
17352 Op.getOperand(0), Op.getOperand(1));
17355 static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
17356 return Op.getOperand(0);
17359 SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
17360 SelectionDAG &DAG) const {
17361 SDValue Root = Op.getOperand(0);
17362 SDValue Trmp = Op.getOperand(1); // trampoline
17363 SDValue FPtr = Op.getOperand(2); // nested function
17364 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
17367 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
17368 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
17370 if (Subtarget->is64Bit()) {
17371 SDValue OutChains[6];
17373 // Large code-model.
17374 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
17375 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
17377 const unsigned char N86R10 = TRI->getEncodingValue(X86::R10) & 0x7;
17378 const unsigned char N86R11 = TRI->getEncodingValue(X86::R11) & 0x7;
17380 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
17382 // Load the pointer to the nested function into R11.
17383 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
17384 SDValue Addr = Trmp;
17385 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17386 Addr, MachinePointerInfo(TrmpAddr),
17389 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17390 DAG.getConstant(2, dl, MVT::i64));
17391 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
17392 MachinePointerInfo(TrmpAddr, 2),
17395 // Load the 'nest' parameter value into R10.
17396 // R10 is specified in X86CallingConv.td
17397 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
17398 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17399 DAG.getConstant(10, dl, MVT::i64));
17400 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17401 Addr, MachinePointerInfo(TrmpAddr, 10),
17404 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17405 DAG.getConstant(12, dl, MVT::i64));
17406 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
17407 MachinePointerInfo(TrmpAddr, 12),
17410 // Jump to the nested function.
17411 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
17412 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17413 DAG.getConstant(20, dl, MVT::i64));
17414 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17415 Addr, MachinePointerInfo(TrmpAddr, 20),
17418 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
17419 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17420 DAG.getConstant(22, dl, MVT::i64));
17421 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, dl, MVT::i8),
17422 Addr, MachinePointerInfo(TrmpAddr, 22),
17425 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
17427 const Function *Func =
17428 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
17429 CallingConv::ID CC = Func->getCallingConv();
17434 llvm_unreachable("Unsupported calling convention");
17435 case CallingConv::C:
17436 case CallingConv::X86_StdCall: {
17437 // Pass 'nest' parameter in ECX.
17438 // Must be kept in sync with X86CallingConv.td
17439 NestReg = X86::ECX;
17441 // Check that ECX wasn't needed by an 'inreg' parameter.
17442 FunctionType *FTy = Func->getFunctionType();
17443 const AttributeSet &Attrs = Func->getAttributes();
17445 if (!Attrs.isEmpty() && !Func->isVarArg()) {
17446 unsigned InRegCount = 0;
17449 for (FunctionType::param_iterator I = FTy->param_begin(),
17450 E = FTy->param_end(); I != E; ++I, ++Idx)
17451 if (Attrs.hasAttribute(Idx, Attribute::InReg)) {
17452 auto &DL = DAG.getDataLayout();
17453 // FIXME: should only count parameters that are lowered to integers.
17454 InRegCount += (DL.getTypeSizeInBits(*I) + 31) / 32;
17457 if (InRegCount > 2) {
17458 report_fatal_error("Nest register in use - reduce number of inreg"
17464 case CallingConv::X86_FastCall:
17465 case CallingConv::X86_ThisCall:
17466 case CallingConv::Fast:
17467 // Pass 'nest' parameter in EAX.
17468 // Must be kept in sync with X86CallingConv.td
17469 NestReg = X86::EAX;
17473 SDValue OutChains[4];
17474 SDValue Addr, Disp;
17476 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17477 DAG.getConstant(10, dl, MVT::i32));
17478 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
17480 // This is storing the opcode for MOV32ri.
17481 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
17482 const unsigned char N86Reg = TRI->getEncodingValue(NestReg) & 0x7;
17483 OutChains[0] = DAG.getStore(Root, dl,
17484 DAG.getConstant(MOV32ri|N86Reg, dl, MVT::i8),
17485 Trmp, MachinePointerInfo(TrmpAddr),
17488 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17489 DAG.getConstant(1, dl, MVT::i32));
17490 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
17491 MachinePointerInfo(TrmpAddr, 1),
17494 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
17495 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17496 DAG.getConstant(5, dl, MVT::i32));
17497 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, dl, MVT::i8),
17498 Addr, MachinePointerInfo(TrmpAddr, 5),
17501 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17502 DAG.getConstant(6, dl, MVT::i32));
17503 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
17504 MachinePointerInfo(TrmpAddr, 6),
17507 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
17511 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
17512 SelectionDAG &DAG) const {
17514 The rounding mode is in bits 11:10 of FPSR, and has the following
17516 00 Round to nearest
17521 FLT_ROUNDS, on the other hand, expects the following:
17528 To perform the conversion, we do:
17529 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
17532 MachineFunction &MF = DAG.getMachineFunction();
17533 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
17534 unsigned StackAlignment = TFI.getStackAlignment();
17535 MVT VT = Op.getSimpleValueType();
17538 // Save FP Control Word to stack slot
17539 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
17540 SDValue StackSlot =
17541 DAG.getFrameIndex(SSFI, getPointerTy(DAG.getDataLayout()));
17543 MachineMemOperand *MMO =
17544 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
17545 MachineMemOperand::MOStore, 2, 2);
17547 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
17548 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
17549 DAG.getVTList(MVT::Other),
17550 Ops, MVT::i16, MMO);
17552 // Load FP Control Word from stack slot
17553 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
17554 MachinePointerInfo(), false, false, false, 0);
17556 // Transform as necessary
17558 DAG.getNode(ISD::SRL, DL, MVT::i16,
17559 DAG.getNode(ISD::AND, DL, MVT::i16,
17560 CWD, DAG.getConstant(0x800, DL, MVT::i16)),
17561 DAG.getConstant(11, DL, MVT::i8));
17563 DAG.getNode(ISD::SRL, DL, MVT::i16,
17564 DAG.getNode(ISD::AND, DL, MVT::i16,
17565 CWD, DAG.getConstant(0x400, DL, MVT::i16)),
17566 DAG.getConstant(9, DL, MVT::i8));
17569 DAG.getNode(ISD::AND, DL, MVT::i16,
17570 DAG.getNode(ISD::ADD, DL, MVT::i16,
17571 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
17572 DAG.getConstant(1, DL, MVT::i16)),
17573 DAG.getConstant(3, DL, MVT::i16));
17575 return DAG.getNode((VT.getSizeInBits() < 16 ?
17576 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
17579 /// \brief Lower a vector CTLZ using native supported vector CTLZ instruction.
17581 // 1. i32/i64 128/256-bit vector (native support require VLX) are expended
17582 // to 512-bit vector.
17583 // 2. i8/i16 vector implemented using dword LZCNT vector instruction
17584 // ( sub(trunc(lzcnt(zext32(x)))) ). In case zext32(x) is illegal,
17585 // split the vector, perform operation on it's Lo a Hi part and
17586 // concatenate the results.
17587 static SDValue LowerVectorCTLZ_AVX512(SDValue Op, SelectionDAG &DAG) {
17589 MVT VT = Op.getSimpleValueType();
17590 MVT EltVT = VT.getVectorElementType();
17591 unsigned NumElems = VT.getVectorNumElements();
17593 if (EltVT == MVT::i64 || EltVT == MVT::i32) {
17594 // Extend to 512 bit vector.
17595 assert((VT.is256BitVector() || VT.is128BitVector()) &&
17596 "Unsupported value type for operation");
17598 MVT NewVT = MVT::getVectorVT(EltVT, 512 / VT.getScalarSizeInBits());
17599 SDValue Vec512 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, NewVT,
17600 DAG.getUNDEF(NewVT),
17602 DAG.getIntPtrConstant(0, dl));
17603 SDValue CtlzNode = DAG.getNode(ISD::CTLZ, dl, NewVT, Vec512);
17605 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, CtlzNode,
17606 DAG.getIntPtrConstant(0, dl));
17609 assert((EltVT == MVT::i8 || EltVT == MVT::i16) &&
17610 "Unsupported element type");
17612 if (16 < NumElems) {
17613 // Split vector, it's Lo and Hi parts will be handled in next iteration.
17615 std::tie(Lo, Hi) = DAG.SplitVector(Op.getOperand(0), dl);
17616 MVT OutVT = MVT::getVectorVT(EltVT, NumElems/2);
17618 Lo = DAG.getNode(Op.getOpcode(), dl, OutVT, Lo);
17619 Hi = DAG.getNode(Op.getOpcode(), dl, OutVT, Hi);
17621 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Lo, Hi);
17624 MVT NewVT = MVT::getVectorVT(MVT::i32, NumElems);
17626 assert((NewVT.is256BitVector() || NewVT.is512BitVector()) &&
17627 "Unsupported value type for operation");
17629 // Use native supported vector instruction vplzcntd.
17630 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, NewVT, Op.getOperand(0));
17631 SDValue CtlzNode = DAG.getNode(ISD::CTLZ, dl, NewVT, Op);
17632 SDValue TruncNode = DAG.getNode(ISD::TRUNCATE, dl, VT, CtlzNode);
17633 SDValue Delta = DAG.getConstant(32 - EltVT.getSizeInBits(), dl, VT);
17635 return DAG.getNode(ISD::SUB, dl, VT, TruncNode, Delta);
17638 static SDValue LowerCTLZ(SDValue Op, const X86Subtarget *Subtarget,
17639 SelectionDAG &DAG) {
17640 MVT VT = Op.getSimpleValueType();
17642 unsigned NumBits = VT.getSizeInBits();
17645 if (VT.isVector() && Subtarget->hasAVX512())
17646 return LowerVectorCTLZ_AVX512(Op, DAG);
17648 Op = Op.getOperand(0);
17649 if (VT == MVT::i8) {
17650 // Zero extend to i32 since there is not an i8 bsr.
17652 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
17655 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
17656 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
17657 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
17659 // If src is zero (i.e. bsr sets ZF), returns NumBits.
17662 DAG.getConstant(NumBits + NumBits - 1, dl, OpVT),
17663 DAG.getConstant(X86::COND_E, dl, MVT::i8),
17666 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops);
17668 // Finally xor with NumBits-1.
17669 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
17670 DAG.getConstant(NumBits - 1, dl, OpVT));
17673 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
17677 static SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, const X86Subtarget *Subtarget,
17678 SelectionDAG &DAG) {
17679 MVT VT = Op.getSimpleValueType();
17681 unsigned NumBits = VT.getSizeInBits();
17684 if (VT.isVector() && Subtarget->hasAVX512())
17685 return LowerVectorCTLZ_AVX512(Op, DAG);
17687 Op = Op.getOperand(0);
17688 if (VT == MVT::i8) {
17689 // Zero extend to i32 since there is not an i8 bsr.
17691 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
17694 // Issue a bsr (scan bits in reverse).
17695 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
17696 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
17698 // And xor with NumBits-1.
17699 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
17700 DAG.getConstant(NumBits - 1, dl, OpVT));
17703 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
17707 static SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
17708 MVT VT = Op.getSimpleValueType();
17709 unsigned NumBits = VT.getScalarSizeInBits();
17712 if (VT.isVector()) {
17713 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
17715 SDValue N0 = Op.getOperand(0);
17716 SDValue Zero = DAG.getConstant(0, dl, VT);
17718 // lsb(x) = (x & -x)
17719 SDValue LSB = DAG.getNode(ISD::AND, dl, VT, N0,
17720 DAG.getNode(ISD::SUB, dl, VT, Zero, N0));
17722 // cttz_undef(x) = (width - 1) - ctlz(lsb)
17723 if (Op.getOpcode() == ISD::CTTZ_ZERO_UNDEF &&
17724 TLI.isOperationLegal(ISD::CTLZ, VT)) {
17725 SDValue WidthMinusOne = DAG.getConstant(NumBits - 1, dl, VT);
17726 return DAG.getNode(ISD::SUB, dl, VT, WidthMinusOne,
17727 DAG.getNode(ISD::CTLZ, dl, VT, LSB));
17730 // cttz(x) = ctpop(lsb - 1)
17731 SDValue One = DAG.getConstant(1, dl, VT);
17732 return DAG.getNode(ISD::CTPOP, dl, VT,
17733 DAG.getNode(ISD::SUB, dl, VT, LSB, One));
17736 assert(Op.getOpcode() == ISD::CTTZ &&
17737 "Only scalar CTTZ requires custom lowering");
17739 // Issue a bsf (scan bits forward) which also sets EFLAGS.
17740 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
17741 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op.getOperand(0));
17743 // If src is zero (i.e. bsf sets ZF), returns NumBits.
17746 DAG.getConstant(NumBits, dl, VT),
17747 DAG.getConstant(X86::COND_E, dl, MVT::i8),
17750 return DAG.getNode(X86ISD::CMOV, dl, VT, Ops);
17753 // Lower256IntArith - Break a 256-bit integer operation into two new 128-bit
17754 // ones, and then concatenate the result back.
17755 static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) {
17756 MVT VT = Op.getSimpleValueType();
17758 assert(VT.is256BitVector() && VT.isInteger() &&
17759 "Unsupported value type for operation");
17761 unsigned NumElems = VT.getVectorNumElements();
17764 // Extract the LHS vectors
17765 SDValue LHS = Op.getOperand(0);
17766 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
17767 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
17769 // Extract the RHS vectors
17770 SDValue RHS = Op.getOperand(1);
17771 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
17772 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
17774 MVT EltVT = VT.getVectorElementType();
17775 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
17777 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
17778 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1),
17779 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
17782 static SDValue LowerADD(SDValue Op, SelectionDAG &DAG) {
17783 if (Op.getValueType() == MVT::i1)
17784 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
17785 Op.getOperand(0), Op.getOperand(1));
17786 assert(Op.getSimpleValueType().is256BitVector() &&
17787 Op.getSimpleValueType().isInteger() &&
17788 "Only handle AVX 256-bit vector integer operation");
17789 return Lower256IntArith(Op, DAG);
17792 static SDValue LowerSUB(SDValue Op, SelectionDAG &DAG) {
17793 if (Op.getValueType() == MVT::i1)
17794 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
17795 Op.getOperand(0), Op.getOperand(1));
17796 assert(Op.getSimpleValueType().is256BitVector() &&
17797 Op.getSimpleValueType().isInteger() &&
17798 "Only handle AVX 256-bit vector integer operation");
17799 return Lower256IntArith(Op, DAG);
17802 static SDValue LowerMINMAX(SDValue Op, SelectionDAG &DAG) {
17803 assert(Op.getSimpleValueType().is256BitVector() &&
17804 Op.getSimpleValueType().isInteger() &&
17805 "Only handle AVX 256-bit vector integer operation");
17806 return Lower256IntArith(Op, DAG);
17809 static SDValue LowerMUL(SDValue Op, const X86Subtarget *Subtarget,
17810 SelectionDAG &DAG) {
17812 MVT VT = Op.getSimpleValueType();
17815 return DAG.getNode(ISD::AND, dl, VT, Op.getOperand(0), Op.getOperand(1));
17817 // Decompose 256-bit ops into smaller 128-bit ops.
17818 if (VT.is256BitVector() && !Subtarget->hasInt256())
17819 return Lower256IntArith(Op, DAG);
17821 SDValue A = Op.getOperand(0);
17822 SDValue B = Op.getOperand(1);
17824 // Lower v16i8/v32i8 mul as promotion to v8i16/v16i16 vector
17825 // pairs, multiply and truncate.
17826 if (VT == MVT::v16i8 || VT == MVT::v32i8) {
17827 if (Subtarget->hasInt256()) {
17828 if (VT == MVT::v32i8) {
17829 MVT SubVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() / 2);
17830 SDValue Lo = DAG.getIntPtrConstant(0, dl);
17831 SDValue Hi = DAG.getIntPtrConstant(VT.getVectorNumElements() / 2, dl);
17832 SDValue ALo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Lo);
17833 SDValue BLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Lo);
17834 SDValue AHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Hi);
17835 SDValue BHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Hi);
17836 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
17837 DAG.getNode(ISD::MUL, dl, SubVT, ALo, BLo),
17838 DAG.getNode(ISD::MUL, dl, SubVT, AHi, BHi));
17841 MVT ExVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements());
17842 return DAG.getNode(
17843 ISD::TRUNCATE, dl, VT,
17844 DAG.getNode(ISD::MUL, dl, ExVT,
17845 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, A),
17846 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, B)));
17849 assert(VT == MVT::v16i8 &&
17850 "Pre-AVX2 support only supports v16i8 multiplication");
17851 MVT ExVT = MVT::v8i16;
17853 // Extract the lo parts and sign extend to i16
17855 if (Subtarget->hasSSE41()) {
17856 ALo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, A);
17857 BLo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, B);
17859 const int ShufMask[] = {-1, 0, -1, 1, -1, 2, -1, 3,
17860 -1, 4, -1, 5, -1, 6, -1, 7};
17861 ALo = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
17862 BLo = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
17863 ALo = DAG.getBitcast(ExVT, ALo);
17864 BLo = DAG.getBitcast(ExVT, BLo);
17865 ALo = DAG.getNode(ISD::SRA, dl, ExVT, ALo, DAG.getConstant(8, dl, ExVT));
17866 BLo = DAG.getNode(ISD::SRA, dl, ExVT, BLo, DAG.getConstant(8, dl, ExVT));
17869 // Extract the hi parts and sign extend to i16
17871 if (Subtarget->hasSSE41()) {
17872 const int ShufMask[] = {8, 9, 10, 11, 12, 13, 14, 15,
17873 -1, -1, -1, -1, -1, -1, -1, -1};
17874 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
17875 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
17876 AHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, AHi);
17877 BHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, BHi);
17879 const int ShufMask[] = {-1, 8, -1, 9, -1, 10, -1, 11,
17880 -1, 12, -1, 13, -1, 14, -1, 15};
17881 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
17882 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
17883 AHi = DAG.getBitcast(ExVT, AHi);
17884 BHi = DAG.getBitcast(ExVT, BHi);
17885 AHi = DAG.getNode(ISD::SRA, dl, ExVT, AHi, DAG.getConstant(8, dl, ExVT));
17886 BHi = DAG.getNode(ISD::SRA, dl, ExVT, BHi, DAG.getConstant(8, dl, ExVT));
17889 // Multiply, mask the lower 8bits of the lo/hi results and pack
17890 SDValue RLo = DAG.getNode(ISD::MUL, dl, ExVT, ALo, BLo);
17891 SDValue RHi = DAG.getNode(ISD::MUL, dl, ExVT, AHi, BHi);
17892 RLo = DAG.getNode(ISD::AND, dl, ExVT, RLo, DAG.getConstant(255, dl, ExVT));
17893 RHi = DAG.getNode(ISD::AND, dl, ExVT, RHi, DAG.getConstant(255, dl, ExVT));
17894 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
17897 // Lower v4i32 mul as 2x shuffle, 2x pmuludq, 2x shuffle.
17898 if (VT == MVT::v4i32) {
17899 assert(Subtarget->hasSSE2() && !Subtarget->hasSSE41() &&
17900 "Should not custom lower when pmuldq is available!");
17902 // Extract the odd parts.
17903 static const int UnpackMask[] = { 1, -1, 3, -1 };
17904 SDValue Aodds = DAG.getVectorShuffle(VT, dl, A, A, UnpackMask);
17905 SDValue Bodds = DAG.getVectorShuffle(VT, dl, B, B, UnpackMask);
17907 // Multiply the even parts.
17908 SDValue Evens = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, A, B);
17909 // Now multiply odd parts.
17910 SDValue Odds = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, Aodds, Bodds);
17912 Evens = DAG.getBitcast(VT, Evens);
17913 Odds = DAG.getBitcast(VT, Odds);
17915 // Merge the two vectors back together with a shuffle. This expands into 2
17917 static const int ShufMask[] = { 0, 4, 2, 6 };
17918 return DAG.getVectorShuffle(VT, dl, Evens, Odds, ShufMask);
17921 assert((VT == MVT::v2i64 || VT == MVT::v4i64 || VT == MVT::v8i64) &&
17922 "Only know how to lower V2I64/V4I64/V8I64 multiply");
17924 // Ahi = psrlqi(a, 32);
17925 // Bhi = psrlqi(b, 32);
17927 // AloBlo = pmuludq(a, b);
17928 // AloBhi = pmuludq(a, Bhi);
17929 // AhiBlo = pmuludq(Ahi, b);
17931 // AloBhi = psllqi(AloBhi, 32);
17932 // AhiBlo = psllqi(AhiBlo, 32);
17933 // return AloBlo + AloBhi + AhiBlo;
17935 SDValue Ahi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, A, 32, DAG);
17936 SDValue Bhi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, B, 32, DAG);
17938 SDValue AhiBlo = Ahi;
17939 SDValue AloBhi = Bhi;
17940 // Bit cast to 32-bit vectors for MULUDQ
17941 MVT MulVT = (VT == MVT::v2i64) ? MVT::v4i32 :
17942 (VT == MVT::v4i64) ? MVT::v8i32 : MVT::v16i32;
17943 A = DAG.getBitcast(MulVT, A);
17944 B = DAG.getBitcast(MulVT, B);
17945 Ahi = DAG.getBitcast(MulVT, Ahi);
17946 Bhi = DAG.getBitcast(MulVT, Bhi);
17948 SDValue AloBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, B);
17949 // After shifting right const values the result may be all-zero.
17950 if (!ISD::isBuildVectorAllZeros(Ahi.getNode())) {
17951 AhiBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Ahi, B);
17952 AhiBlo = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AhiBlo, 32, DAG);
17954 if (!ISD::isBuildVectorAllZeros(Bhi.getNode())) {
17955 AloBhi = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, Bhi);
17956 AloBhi = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AloBhi, 32, DAG);
17959 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
17960 return DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
17963 SDValue X86TargetLowering::LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const {
17964 assert(Subtarget->isTargetWin64() && "Unexpected target");
17965 EVT VT = Op.getValueType();
17966 assert(VT.isInteger() && VT.getSizeInBits() == 128 &&
17967 "Unexpected return type for lowering");
17971 switch (Op->getOpcode()) {
17972 default: llvm_unreachable("Unexpected request for libcall!");
17973 case ISD::SDIV: isSigned = true; LC = RTLIB::SDIV_I128; break;
17974 case ISD::UDIV: isSigned = false; LC = RTLIB::UDIV_I128; break;
17975 case ISD::SREM: isSigned = true; LC = RTLIB::SREM_I128; break;
17976 case ISD::UREM: isSigned = false; LC = RTLIB::UREM_I128; break;
17977 case ISD::SDIVREM: isSigned = true; LC = RTLIB::SDIVREM_I128; break;
17978 case ISD::UDIVREM: isSigned = false; LC = RTLIB::UDIVREM_I128; break;
17982 SDValue InChain = DAG.getEntryNode();
17984 TargetLowering::ArgListTy Args;
17985 TargetLowering::ArgListEntry Entry;
17986 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
17987 EVT ArgVT = Op->getOperand(i).getValueType();
17988 assert(ArgVT.isInteger() && ArgVT.getSizeInBits() == 128 &&
17989 "Unexpected argument type for lowering");
17990 SDValue StackPtr = DAG.CreateStackTemporary(ArgVT, 16);
17991 Entry.Node = StackPtr;
17992 InChain = DAG.getStore(InChain, dl, Op->getOperand(i), StackPtr, MachinePointerInfo(),
17994 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
17995 Entry.Ty = PointerType::get(ArgTy,0);
17996 Entry.isSExt = false;
17997 Entry.isZExt = false;
17998 Args.push_back(Entry);
18001 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
18002 getPointerTy(DAG.getDataLayout()));
18004 TargetLowering::CallLoweringInfo CLI(DAG);
18005 CLI.setDebugLoc(dl).setChain(InChain)
18006 .setCallee(getLibcallCallingConv(LC),
18007 static_cast<EVT>(MVT::v2i64).getTypeForEVT(*DAG.getContext()),
18008 Callee, std::move(Args), 0)
18009 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
18011 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
18012 return DAG.getBitcast(VT, CallInfo.first);
18015 static SDValue LowerMUL_LOHI(SDValue Op, const X86Subtarget *Subtarget,
18016 SelectionDAG &DAG) {
18017 SDValue Op0 = Op.getOperand(0), Op1 = Op.getOperand(1);
18018 MVT VT = Op0.getSimpleValueType();
18021 assert((VT == MVT::v4i32 && Subtarget->hasSSE2()) ||
18022 (VT == MVT::v8i32 && Subtarget->hasInt256()));
18024 // PMULxD operations multiply each even value (starting at 0) of LHS with
18025 // the related value of RHS and produce a widen result.
18026 // E.g., PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
18027 // => <2 x i64> <ae|cg>
18029 // In other word, to have all the results, we need to perform two PMULxD:
18030 // 1. one with the even values.
18031 // 2. one with the odd values.
18032 // To achieve #2, with need to place the odd values at an even position.
18034 // Place the odd value at an even position (basically, shift all values 1
18035 // step to the left):
18036 const int Mask[] = {1, -1, 3, -1, 5, -1, 7, -1};
18037 // <a|b|c|d> => <b|undef|d|undef>
18038 SDValue Odd0 = DAG.getVectorShuffle(VT, dl, Op0, Op0, Mask);
18039 // <e|f|g|h> => <f|undef|h|undef>
18040 SDValue Odd1 = DAG.getVectorShuffle(VT, dl, Op1, Op1, Mask);
18042 // Emit two multiplies, one for the lower 2 ints and one for the higher 2
18044 MVT MulVT = VT == MVT::v4i32 ? MVT::v2i64 : MVT::v4i64;
18045 bool IsSigned = Op->getOpcode() == ISD::SMUL_LOHI;
18047 (!IsSigned || !Subtarget->hasSSE41()) ? X86ISD::PMULUDQ : X86ISD::PMULDQ;
18048 // PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
18049 // => <2 x i64> <ae|cg>
18050 SDValue Mul1 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Op0, Op1));
18051 // PMULUDQ <4 x i32> <b|undef|d|undef>, <4 x i32> <f|undef|h|undef>
18052 // => <2 x i64> <bf|dh>
18053 SDValue Mul2 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Odd0, Odd1));
18055 // Shuffle it back into the right order.
18056 SDValue Highs, Lows;
18057 if (VT == MVT::v8i32) {
18058 const int HighMask[] = {1, 9, 3, 11, 5, 13, 7, 15};
18059 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
18060 const int LowMask[] = {0, 8, 2, 10, 4, 12, 6, 14};
18061 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
18063 const int HighMask[] = {1, 5, 3, 7};
18064 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
18065 const int LowMask[] = {0, 4, 2, 6};
18066 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
18069 // If we have a signed multiply but no PMULDQ fix up the high parts of a
18070 // unsigned multiply.
18071 if (IsSigned && !Subtarget->hasSSE41()) {
18072 SDValue ShAmt = DAG.getConstant(
18074 DAG.getTargetLoweringInfo().getShiftAmountTy(VT, DAG.getDataLayout()));
18075 SDValue T1 = DAG.getNode(ISD::AND, dl, VT,
18076 DAG.getNode(ISD::SRA, dl, VT, Op0, ShAmt), Op1);
18077 SDValue T2 = DAG.getNode(ISD::AND, dl, VT,
18078 DAG.getNode(ISD::SRA, dl, VT, Op1, ShAmt), Op0);
18080 SDValue Fixup = DAG.getNode(ISD::ADD, dl, VT, T1, T2);
18081 Highs = DAG.getNode(ISD::SUB, dl, VT, Highs, Fixup);
18084 // The first result of MUL_LOHI is actually the low value, followed by the
18086 SDValue Ops[] = {Lows, Highs};
18087 return DAG.getMergeValues(Ops, dl);
18090 // Return true if the required (according to Opcode) shift-imm form is natively
18091 // supported by the Subtarget
18092 static bool SupportedVectorShiftWithImm(MVT VT, const X86Subtarget *Subtarget,
18094 if (VT.getScalarSizeInBits() < 16)
18097 if (VT.is512BitVector() &&
18098 (VT.getScalarSizeInBits() > 16 || Subtarget->hasBWI()))
18101 bool LShift = VT.is128BitVector() ||
18102 (VT.is256BitVector() && Subtarget->hasInt256());
18104 bool AShift = LShift && (Subtarget->hasVLX() ||
18105 (VT != MVT::v2i64 && VT != MVT::v4i64));
18106 return (Opcode == ISD::SRA) ? AShift : LShift;
18109 // The shift amount is a variable, but it is the same for all vector lanes.
18110 // These instructions are defined together with shift-immediate.
18112 bool SupportedVectorShiftWithBaseAmnt(MVT VT, const X86Subtarget *Subtarget,
18114 return SupportedVectorShiftWithImm(VT, Subtarget, Opcode);
18117 // Return true if the required (according to Opcode) variable-shift form is
18118 // natively supported by the Subtarget
18119 static bool SupportedVectorVarShift(MVT VT, const X86Subtarget *Subtarget,
18122 if (!Subtarget->hasInt256() || VT.getScalarSizeInBits() < 16)
18125 // vXi16 supported only on AVX-512, BWI
18126 if (VT.getScalarSizeInBits() == 16 && !Subtarget->hasBWI())
18129 if (VT.is512BitVector() || Subtarget->hasVLX())
18132 bool LShift = VT.is128BitVector() || VT.is256BitVector();
18133 bool AShift = LShift && VT != MVT::v2i64 && VT != MVT::v4i64;
18134 return (Opcode == ISD::SRA) ? AShift : LShift;
18137 static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
18138 const X86Subtarget *Subtarget) {
18139 MVT VT = Op.getSimpleValueType();
18141 SDValue R = Op.getOperand(0);
18142 SDValue Amt = Op.getOperand(1);
18144 unsigned X86Opc = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
18145 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
18147 auto ArithmeticShiftRight64 = [&](uint64_t ShiftAmt) {
18148 assert((VT == MVT::v2i64 || VT == MVT::v4i64) && "Unexpected SRA type");
18149 MVT ExVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() * 2);
18150 SDValue Ex = DAG.getBitcast(ExVT, R);
18152 if (ShiftAmt >= 32) {
18153 // Splat sign to upper i32 dst, and SRA upper i32 src to lower i32.
18155 getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex, 31, DAG);
18156 SDValue Lower = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex,
18157 ShiftAmt - 32, DAG);
18158 if (VT == MVT::v2i64)
18159 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {5, 1, 7, 3});
18160 if (VT == MVT::v4i64)
18161 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower,
18162 {9, 1, 11, 3, 13, 5, 15, 7});
18164 // SRA upper i32, SHL whole i64 and select lower i32.
18165 SDValue Upper = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex,
18168 getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt, DAG);
18169 Lower = DAG.getBitcast(ExVT, Lower);
18170 if (VT == MVT::v2i64)
18171 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {4, 1, 6, 3});
18172 if (VT == MVT::v4i64)
18173 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower,
18174 {8, 1, 10, 3, 12, 5, 14, 7});
18176 return DAG.getBitcast(VT, Ex);
18179 // Optimize shl/srl/sra with constant shift amount.
18180 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
18181 if (auto *ShiftConst = BVAmt->getConstantSplatNode()) {
18182 uint64_t ShiftAmt = ShiftConst->getZExtValue();
18184 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
18185 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
18187 // i64 SRA needs to be performed as partial shifts.
18188 if ((VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
18189 Op.getOpcode() == ISD::SRA && !Subtarget->hasXOP())
18190 return ArithmeticShiftRight64(ShiftAmt);
18192 if (VT == MVT::v16i8 || (Subtarget->hasInt256() && VT == MVT::v32i8)) {
18193 unsigned NumElts = VT.getVectorNumElements();
18194 MVT ShiftVT = MVT::getVectorVT(MVT::i16, NumElts / 2);
18196 // Simple i8 add case
18197 if (Op.getOpcode() == ISD::SHL && ShiftAmt == 1)
18198 return DAG.getNode(ISD::ADD, dl, VT, R, R);
18200 // ashr(R, 7) === cmp_slt(R, 0)
18201 if (Op.getOpcode() == ISD::SRA && ShiftAmt == 7) {
18202 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
18203 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
18206 // XOP can shift v16i8 directly instead of as shift v8i16 + mask.
18207 if (VT == MVT::v16i8 && Subtarget->hasXOP())
18210 if (Op.getOpcode() == ISD::SHL) {
18211 // Make a large shift.
18212 SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, ShiftVT,
18214 SHL = DAG.getBitcast(VT, SHL);
18215 // Zero out the rightmost bits.
18216 SmallVector<SDValue, 32> V(
18217 NumElts, DAG.getConstant(uint8_t(-1U << ShiftAmt), dl, MVT::i8));
18218 return DAG.getNode(ISD::AND, dl, VT, SHL,
18219 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
18221 if (Op.getOpcode() == ISD::SRL) {
18222 // Make a large shift.
18223 SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, ShiftVT,
18225 SRL = DAG.getBitcast(VT, SRL);
18226 // Zero out the leftmost bits.
18227 SmallVector<SDValue, 32> V(
18228 NumElts, DAG.getConstant(uint8_t(-1U) >> ShiftAmt, dl, MVT::i8));
18229 return DAG.getNode(ISD::AND, dl, VT, SRL,
18230 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
18232 if (Op.getOpcode() == ISD::SRA) {
18233 // ashr(R, Amt) === sub(xor(lshr(R, Amt), Mask), Mask)
18234 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
18235 SmallVector<SDValue, 32> V(NumElts,
18236 DAG.getConstant(128 >> ShiftAmt, dl,
18238 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V);
18239 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
18240 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
18243 llvm_unreachable("Unknown shift opcode.");
18248 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
18249 if (!Subtarget->is64Bit() && !Subtarget->hasXOP() &&
18250 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64))) {
18252 // Peek through any splat that was introduced for i64 shift vectorization.
18253 int SplatIndex = -1;
18254 if (ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt.getNode()))
18255 if (SVN->isSplat()) {
18256 SplatIndex = SVN->getSplatIndex();
18257 Amt = Amt.getOperand(0);
18258 assert(SplatIndex < (int)VT.getVectorNumElements() &&
18259 "Splat shuffle referencing second operand");
18262 if (Amt.getOpcode() != ISD::BITCAST ||
18263 Amt.getOperand(0).getOpcode() != ISD::BUILD_VECTOR)
18266 Amt = Amt.getOperand(0);
18267 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
18268 VT.getVectorNumElements();
18269 unsigned RatioInLog2 = Log2_32_Ceil(Ratio);
18270 uint64_t ShiftAmt = 0;
18271 unsigned BaseOp = (SplatIndex < 0 ? 0 : SplatIndex * Ratio);
18272 for (unsigned i = 0; i != Ratio; ++i) {
18273 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i + BaseOp));
18277 ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2)));
18280 // Check remaining shift amounts (if not a splat).
18281 if (SplatIndex < 0) {
18282 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
18283 uint64_t ShAmt = 0;
18284 for (unsigned j = 0; j != Ratio; ++j) {
18285 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i + j));
18289 ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2)));
18291 if (ShAmt != ShiftAmt)
18296 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
18297 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
18299 if (Op.getOpcode() == ISD::SRA)
18300 return ArithmeticShiftRight64(ShiftAmt);
18306 static SDValue LowerScalarVariableShift(SDValue Op, SelectionDAG &DAG,
18307 const X86Subtarget* Subtarget) {
18308 MVT VT = Op.getSimpleValueType();
18310 SDValue R = Op.getOperand(0);
18311 SDValue Amt = Op.getOperand(1);
18313 unsigned X86OpcI = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
18314 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
18316 unsigned X86OpcV = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHL :
18317 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRL : X86ISD::VSRA;
18319 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode())) {
18321 MVT EltVT = VT.getVectorElementType();
18323 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Amt)) {
18324 // Check if this build_vector node is doing a splat.
18325 // If so, then set BaseShAmt equal to the splat value.
18326 BaseShAmt = BV->getSplatValue();
18327 if (BaseShAmt && BaseShAmt.getOpcode() == ISD::UNDEF)
18328 BaseShAmt = SDValue();
18330 if (Amt.getOpcode() == ISD::EXTRACT_SUBVECTOR)
18331 Amt = Amt.getOperand(0);
18333 ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt);
18334 if (SVN && SVN->isSplat()) {
18335 unsigned SplatIdx = (unsigned)SVN->getSplatIndex();
18336 SDValue InVec = Amt.getOperand(0);
18337 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
18338 assert((SplatIdx < InVec.getSimpleValueType().getVectorNumElements()) &&
18339 "Unexpected shuffle index found!");
18340 BaseShAmt = InVec.getOperand(SplatIdx);
18341 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
18342 if (ConstantSDNode *C =
18343 dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
18344 if (C->getZExtValue() == SplatIdx)
18345 BaseShAmt = InVec.getOperand(1);
18350 // Avoid introducing an extract element from a shuffle.
18351 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InVec,
18352 DAG.getIntPtrConstant(SplatIdx, dl));
18356 if (BaseShAmt.getNode()) {
18357 assert(EltVT.bitsLE(MVT::i64) && "Unexpected element type!");
18358 if (EltVT != MVT::i64 && EltVT.bitsGT(MVT::i32))
18359 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, BaseShAmt);
18360 else if (EltVT.bitsLT(MVT::i32))
18361 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt);
18363 return getTargetVShiftNode(X86OpcI, dl, VT, R, BaseShAmt, DAG);
18367 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
18368 if (!Subtarget->is64Bit() && VT == MVT::v2i64 &&
18369 Amt.getOpcode() == ISD::BITCAST &&
18370 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
18371 Amt = Amt.getOperand(0);
18372 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
18373 VT.getVectorNumElements();
18374 std::vector<SDValue> Vals(Ratio);
18375 for (unsigned i = 0; i != Ratio; ++i)
18376 Vals[i] = Amt.getOperand(i);
18377 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
18378 for (unsigned j = 0; j != Ratio; ++j)
18379 if (Vals[j] != Amt.getOperand(i + j))
18383 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode()))
18384 return DAG.getNode(X86OpcV, dl, VT, R, Op.getOperand(1));
18389 static SDValue LowerShift(SDValue Op, const X86Subtarget* Subtarget,
18390 SelectionDAG &DAG) {
18391 MVT VT = Op.getSimpleValueType();
18393 SDValue R = Op.getOperand(0);
18394 SDValue Amt = Op.getOperand(1);
18396 assert(VT.isVector() && "Custom lowering only for vector shifts!");
18397 assert(Subtarget->hasSSE2() && "Only custom lower when we have SSE2!");
18399 if (SDValue V = LowerScalarImmediateShift(Op, DAG, Subtarget))
18402 if (SDValue V = LowerScalarVariableShift(Op, DAG, Subtarget))
18405 if (SupportedVectorVarShift(VT, Subtarget, Op.getOpcode()))
18408 // XOP has 128-bit variable logical/arithmetic shifts.
18409 // +ve/-ve Amt = shift left/right.
18410 if (Subtarget->hasXOP() &&
18411 (VT == MVT::v2i64 || VT == MVT::v4i32 ||
18412 VT == MVT::v8i16 || VT == MVT::v16i8)) {
18413 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SRA) {
18414 SDValue Zero = getZeroVector(VT, Subtarget, DAG, dl);
18415 Amt = DAG.getNode(ISD::SUB, dl, VT, Zero, Amt);
18417 if (Op.getOpcode() == ISD::SHL || Op.getOpcode() == ISD::SRL)
18418 return DAG.getNode(X86ISD::VPSHL, dl, VT, R, Amt);
18419 if (Op.getOpcode() == ISD::SRA)
18420 return DAG.getNode(X86ISD::VPSHA, dl, VT, R, Amt);
18423 // 2i64 vector logical shifts can efficiently avoid scalarization - do the
18424 // shifts per-lane and then shuffle the partial results back together.
18425 if (VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) {
18426 // Splat the shift amounts so the scalar shifts above will catch it.
18427 SDValue Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {0, 0});
18428 SDValue Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {1, 1});
18429 SDValue R0 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt0);
18430 SDValue R1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt1);
18431 return DAG.getVectorShuffle(VT, dl, R0, R1, {0, 3});
18434 // i64 vector arithmetic shift can be emulated with the transform:
18435 // M = lshr(SIGN_BIT, Amt)
18436 // ashr(R, Amt) === sub(xor(lshr(R, Amt), M), M)
18437 if ((VT == MVT::v2i64 || (VT == MVT::v4i64 && Subtarget->hasInt256())) &&
18438 Op.getOpcode() == ISD::SRA) {
18439 SDValue S = DAG.getConstant(APInt::getSignBit(64), dl, VT);
18440 SDValue M = DAG.getNode(ISD::SRL, dl, VT, S, Amt);
18441 R = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
18442 R = DAG.getNode(ISD::XOR, dl, VT, R, M);
18443 R = DAG.getNode(ISD::SUB, dl, VT, R, M);
18447 // If possible, lower this packed shift into a vector multiply instead of
18448 // expanding it into a sequence of scalar shifts.
18449 // Do this only if the vector shift count is a constant build_vector.
18450 if (Op.getOpcode() == ISD::SHL &&
18451 (VT == MVT::v8i16 || VT == MVT::v4i32 ||
18452 (Subtarget->hasInt256() && VT == MVT::v16i16)) &&
18453 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18454 SmallVector<SDValue, 8> Elts;
18455 MVT SVT = VT.getVectorElementType();
18456 unsigned SVTBits = SVT.getSizeInBits();
18457 APInt One(SVTBits, 1);
18458 unsigned NumElems = VT.getVectorNumElements();
18460 for (unsigned i=0; i !=NumElems; ++i) {
18461 SDValue Op = Amt->getOperand(i);
18462 if (Op->getOpcode() == ISD::UNDEF) {
18463 Elts.push_back(Op);
18467 ConstantSDNode *ND = cast<ConstantSDNode>(Op);
18468 APInt C(SVTBits, ND->getAPIntValue().getZExtValue());
18469 uint64_t ShAmt = C.getZExtValue();
18470 if (ShAmt >= SVTBits) {
18471 Elts.push_back(DAG.getUNDEF(SVT));
18474 Elts.push_back(DAG.getConstant(One.shl(ShAmt), dl, SVT));
18476 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
18477 return DAG.getNode(ISD::MUL, dl, VT, R, BV);
18480 // Lower SHL with variable shift amount.
18481 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
18482 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, dl, VT));
18484 Op = DAG.getNode(ISD::ADD, dl, VT, Op,
18485 DAG.getConstant(0x3f800000U, dl, VT));
18486 Op = DAG.getBitcast(MVT::v4f32, Op);
18487 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
18488 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
18491 // If possible, lower this shift as a sequence of two shifts by
18492 // constant plus a MOVSS/MOVSD instead of scalarizing it.
18494 // (v4i32 (srl A, (build_vector < X, Y, Y, Y>)))
18496 // Could be rewritten as:
18497 // (v4i32 (MOVSS (srl A, <Y,Y,Y,Y>), (srl A, <X,X,X,X>)))
18499 // The advantage is that the two shifts from the example would be
18500 // lowered as X86ISD::VSRLI nodes. This would be cheaper than scalarizing
18501 // the vector shift into four scalar shifts plus four pairs of vector
18503 if ((VT == MVT::v8i16 || VT == MVT::v4i32) &&
18504 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18505 unsigned TargetOpcode = X86ISD::MOVSS;
18506 bool CanBeSimplified;
18507 // The splat value for the first packed shift (the 'X' from the example).
18508 SDValue Amt1 = Amt->getOperand(0);
18509 // The splat value for the second packed shift (the 'Y' from the example).
18510 SDValue Amt2 = (VT == MVT::v4i32) ? Amt->getOperand(1) :
18511 Amt->getOperand(2);
18513 // See if it is possible to replace this node with a sequence of
18514 // two shifts followed by a MOVSS/MOVSD
18515 if (VT == MVT::v4i32) {
18516 // Check if it is legal to use a MOVSS.
18517 CanBeSimplified = Amt2 == Amt->getOperand(2) &&
18518 Amt2 == Amt->getOperand(3);
18519 if (!CanBeSimplified) {
18520 // Otherwise, check if we can still simplify this node using a MOVSD.
18521 CanBeSimplified = Amt1 == Amt->getOperand(1) &&
18522 Amt->getOperand(2) == Amt->getOperand(3);
18523 TargetOpcode = X86ISD::MOVSD;
18524 Amt2 = Amt->getOperand(2);
18527 // Do similar checks for the case where the machine value type
18529 CanBeSimplified = Amt1 == Amt->getOperand(1);
18530 for (unsigned i=3; i != 8 && CanBeSimplified; ++i)
18531 CanBeSimplified = Amt2 == Amt->getOperand(i);
18533 if (!CanBeSimplified) {
18534 TargetOpcode = X86ISD::MOVSD;
18535 CanBeSimplified = true;
18536 Amt2 = Amt->getOperand(4);
18537 for (unsigned i=0; i != 4 && CanBeSimplified; ++i)
18538 CanBeSimplified = Amt1 == Amt->getOperand(i);
18539 for (unsigned j=4; j != 8 && CanBeSimplified; ++j)
18540 CanBeSimplified = Amt2 == Amt->getOperand(j);
18544 if (CanBeSimplified && isa<ConstantSDNode>(Amt1) &&
18545 isa<ConstantSDNode>(Amt2)) {
18546 // Replace this node with two shifts followed by a MOVSS/MOVSD.
18547 MVT CastVT = MVT::v4i32;
18549 DAG.getConstant(cast<ConstantSDNode>(Amt1)->getAPIntValue(), dl, VT);
18550 SDValue Shift1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat1);
18552 DAG.getConstant(cast<ConstantSDNode>(Amt2)->getAPIntValue(), dl, VT);
18553 SDValue Shift2 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat2);
18554 if (TargetOpcode == X86ISD::MOVSD)
18555 CastVT = MVT::v2i64;
18556 SDValue BitCast1 = DAG.getBitcast(CastVT, Shift1);
18557 SDValue BitCast2 = DAG.getBitcast(CastVT, Shift2);
18558 SDValue Result = getTargetShuffleNode(TargetOpcode, dl, CastVT, BitCast2,
18560 return DAG.getBitcast(VT, Result);
18564 // v4i32 Non Uniform Shifts.
18565 // If the shift amount is constant we can shift each lane using the SSE2
18566 // immediate shifts, else we need to zero-extend each lane to the lower i64
18567 // and shift using the SSE2 variable shifts.
18568 // The separate results can then be blended together.
18569 if (VT == MVT::v4i32) {
18570 unsigned Opc = Op.getOpcode();
18571 SDValue Amt0, Amt1, Amt2, Amt3;
18572 if (ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18573 Amt0 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {0, 0, 0, 0});
18574 Amt1 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {1, 1, 1, 1});
18575 Amt2 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {2, 2, 2, 2});
18576 Amt3 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {3, 3, 3, 3});
18578 // ISD::SHL is handled above but we include it here for completeness.
18581 llvm_unreachable("Unknown target vector shift node");
18583 Opc = X86ISD::VSHL;
18586 Opc = X86ISD::VSRL;
18589 Opc = X86ISD::VSRA;
18592 // The SSE2 shifts use the lower i64 as the same shift amount for
18593 // all lanes and the upper i64 is ignored. These shuffle masks
18594 // optimally zero-extend each lanes on SSE2/SSE41/AVX targets.
18595 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
18596 Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Z, {0, 4, -1, -1});
18597 Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Z, {1, 5, -1, -1});
18598 Amt2 = DAG.getVectorShuffle(VT, dl, Amt, Z, {2, 6, -1, -1});
18599 Amt3 = DAG.getVectorShuffle(VT, dl, Amt, Z, {3, 7, -1, -1});
18602 SDValue R0 = DAG.getNode(Opc, dl, VT, R, Amt0);
18603 SDValue R1 = DAG.getNode(Opc, dl, VT, R, Amt1);
18604 SDValue R2 = DAG.getNode(Opc, dl, VT, R, Amt2);
18605 SDValue R3 = DAG.getNode(Opc, dl, VT, R, Amt3);
18606 SDValue R02 = DAG.getVectorShuffle(VT, dl, R0, R2, {0, -1, 6, -1});
18607 SDValue R13 = DAG.getVectorShuffle(VT, dl, R1, R3, {-1, 1, -1, 7});
18608 return DAG.getVectorShuffle(VT, dl, R02, R13, {0, 5, 2, 7});
18611 if (VT == MVT::v16i8 ||
18612 (VT == MVT::v32i8 && Subtarget->hasInt256() && !Subtarget->hasXOP())) {
18613 MVT ExtVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements() / 2);
18614 unsigned ShiftOpcode = Op->getOpcode();
18616 auto SignBitSelect = [&](MVT SelVT, SDValue Sel, SDValue V0, SDValue V1) {
18617 // On SSE41 targets we make use of the fact that VSELECT lowers
18618 // to PBLENDVB which selects bytes based just on the sign bit.
18619 if (Subtarget->hasSSE41()) {
18620 V0 = DAG.getBitcast(VT, V0);
18621 V1 = DAG.getBitcast(VT, V1);
18622 Sel = DAG.getBitcast(VT, Sel);
18623 return DAG.getBitcast(SelVT,
18624 DAG.getNode(ISD::VSELECT, dl, VT, Sel, V0, V1));
18626 // On pre-SSE41 targets we test for the sign bit by comparing to
18627 // zero - a negative value will set all bits of the lanes to true
18628 // and VSELECT uses that in its OR(AND(V0,C),AND(V1,~C)) lowering.
18629 SDValue Z = getZeroVector(SelVT, Subtarget, DAG, dl);
18630 SDValue C = DAG.getNode(X86ISD::PCMPGT, dl, SelVT, Z, Sel);
18631 return DAG.getNode(ISD::VSELECT, dl, SelVT, C, V0, V1);
18634 // Turn 'a' into a mask suitable for VSELECT: a = a << 5;
18635 // We can safely do this using i16 shifts as we're only interested in
18636 // the 3 lower bits of each byte.
18637 Amt = DAG.getBitcast(ExtVT, Amt);
18638 Amt = DAG.getNode(ISD::SHL, dl, ExtVT, Amt, DAG.getConstant(5, dl, ExtVT));
18639 Amt = DAG.getBitcast(VT, Amt);
18641 if (Op->getOpcode() == ISD::SHL || Op->getOpcode() == ISD::SRL) {
18642 // r = VSELECT(r, shift(r, 4), a);
18644 DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT));
18645 R = SignBitSelect(VT, Amt, M, R);
18648 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18650 // r = VSELECT(r, shift(r, 2), a);
18651 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT));
18652 R = SignBitSelect(VT, Amt, M, R);
18655 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18657 // return VSELECT(r, shift(r, 1), a);
18658 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT));
18659 R = SignBitSelect(VT, Amt, M, R);
18663 if (Op->getOpcode() == ISD::SRA) {
18664 // For SRA we need to unpack each byte to the higher byte of a i16 vector
18665 // so we can correctly sign extend. We don't care what happens to the
18667 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), Amt);
18668 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), Amt);
18669 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), R);
18670 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), R);
18671 ALo = DAG.getBitcast(ExtVT, ALo);
18672 AHi = DAG.getBitcast(ExtVT, AHi);
18673 RLo = DAG.getBitcast(ExtVT, RLo);
18674 RHi = DAG.getBitcast(ExtVT, RHi);
18676 // r = VSELECT(r, shift(r, 4), a);
18677 SDValue MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
18678 DAG.getConstant(4, dl, ExtVT));
18679 SDValue MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
18680 DAG.getConstant(4, dl, ExtVT));
18681 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
18682 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
18685 ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo);
18686 AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi);
18688 // r = VSELECT(r, shift(r, 2), a);
18689 MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
18690 DAG.getConstant(2, dl, ExtVT));
18691 MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
18692 DAG.getConstant(2, dl, ExtVT));
18693 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
18694 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
18697 ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo);
18698 AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi);
18700 // r = VSELECT(r, shift(r, 1), a);
18701 MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
18702 DAG.getConstant(1, dl, ExtVT));
18703 MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
18704 DAG.getConstant(1, dl, ExtVT));
18705 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
18706 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
18708 // Logical shift the result back to the lower byte, leaving a zero upper
18710 // meaning that we can safely pack with PACKUSWB.
18712 DAG.getNode(ISD::SRL, dl, ExtVT, RLo, DAG.getConstant(8, dl, ExtVT));
18714 DAG.getNode(ISD::SRL, dl, ExtVT, RHi, DAG.getConstant(8, dl, ExtVT));
18715 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
18719 // It's worth extending once and using the v8i32 shifts for 16-bit types, but
18720 // the extra overheads to get from v16i8 to v8i32 make the existing SSE
18721 // solution better.
18722 if (Subtarget->hasInt256() && VT == MVT::v8i16) {
18723 MVT ExtVT = MVT::v8i32;
18725 Op.getOpcode() == ISD::SRA ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
18726 R = DAG.getNode(ExtOpc, dl, ExtVT, R);
18727 Amt = DAG.getNode(ISD::ANY_EXTEND, dl, ExtVT, Amt);
18728 return DAG.getNode(ISD::TRUNCATE, dl, VT,
18729 DAG.getNode(Op.getOpcode(), dl, ExtVT, R, Amt));
18732 if (Subtarget->hasInt256() && !Subtarget->hasXOP() && VT == MVT::v16i16) {
18733 MVT ExtVT = MVT::v8i32;
18734 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
18735 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, Amt, Z);
18736 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, Amt, Z);
18737 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, R, R);
18738 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, R, R);
18739 ALo = DAG.getBitcast(ExtVT, ALo);
18740 AHi = DAG.getBitcast(ExtVT, AHi);
18741 RLo = DAG.getBitcast(ExtVT, RLo);
18742 RHi = DAG.getBitcast(ExtVT, RHi);
18743 SDValue Lo = DAG.getNode(Op.getOpcode(), dl, ExtVT, RLo, ALo);
18744 SDValue Hi = DAG.getNode(Op.getOpcode(), dl, ExtVT, RHi, AHi);
18745 Lo = DAG.getNode(ISD::SRL, dl, ExtVT, Lo, DAG.getConstant(16, dl, ExtVT));
18746 Hi = DAG.getNode(ISD::SRL, dl, ExtVT, Hi, DAG.getConstant(16, dl, ExtVT));
18747 return DAG.getNode(X86ISD::PACKUS, dl, VT, Lo, Hi);
18750 if (VT == MVT::v8i16) {
18751 unsigned ShiftOpcode = Op->getOpcode();
18753 auto SignBitSelect = [&](SDValue Sel, SDValue V0, SDValue V1) {
18754 // On SSE41 targets we make use of the fact that VSELECT lowers
18755 // to PBLENDVB which selects bytes based just on the sign bit.
18756 if (Subtarget->hasSSE41()) {
18757 MVT ExtVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() * 2);
18758 V0 = DAG.getBitcast(ExtVT, V0);
18759 V1 = DAG.getBitcast(ExtVT, V1);
18760 Sel = DAG.getBitcast(ExtVT, Sel);
18761 return DAG.getBitcast(
18762 VT, DAG.getNode(ISD::VSELECT, dl, ExtVT, Sel, V0, V1));
18764 // On pre-SSE41 targets we splat the sign bit - a negative value will
18765 // set all bits of the lanes to true and VSELECT uses that in
18766 // its OR(AND(V0,C),AND(V1,~C)) lowering.
18768 DAG.getNode(ISD::SRA, dl, VT, Sel, DAG.getConstant(15, dl, VT));
18769 return DAG.getNode(ISD::VSELECT, dl, VT, C, V0, V1);
18772 // Turn 'a' into a mask suitable for VSELECT: a = a << 12;
18773 if (Subtarget->hasSSE41()) {
18774 // On SSE41 targets we need to replicate the shift mask in both
18775 // bytes for PBLENDVB.
18778 DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(4, dl, VT)),
18779 DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT)));
18781 Amt = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT));
18784 // r = VSELECT(r, shift(r, 8), a);
18785 SDValue M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(8, dl, VT));
18786 R = SignBitSelect(Amt, M, R);
18789 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18791 // r = VSELECT(r, shift(r, 4), a);
18792 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT));
18793 R = SignBitSelect(Amt, M, R);
18796 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18798 // r = VSELECT(r, shift(r, 2), a);
18799 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT));
18800 R = SignBitSelect(Amt, M, R);
18803 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18805 // return VSELECT(r, shift(r, 1), a);
18806 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT));
18807 R = SignBitSelect(Amt, M, R);
18811 // Decompose 256-bit shifts into smaller 128-bit shifts.
18812 if (VT.is256BitVector()) {
18813 unsigned NumElems = VT.getVectorNumElements();
18814 MVT EltVT = VT.getVectorElementType();
18815 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
18817 // Extract the two vectors
18818 SDValue V1 = Extract128BitVector(R, 0, DAG, dl);
18819 SDValue V2 = Extract128BitVector(R, NumElems/2, DAG, dl);
18821 // Recreate the shift amount vectors
18822 SDValue Amt1, Amt2;
18823 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
18824 // Constant shift amount
18825 SmallVector<SDValue, 8> Ops(Amt->op_begin(), Amt->op_begin() + NumElems);
18826 ArrayRef<SDValue> Amt1Csts = makeArrayRef(Ops).slice(0, NumElems / 2);
18827 ArrayRef<SDValue> Amt2Csts = makeArrayRef(Ops).slice(NumElems / 2);
18829 Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt1Csts);
18830 Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt2Csts);
18832 // Variable shift amount
18833 Amt1 = Extract128BitVector(Amt, 0, DAG, dl);
18834 Amt2 = Extract128BitVector(Amt, NumElems/2, DAG, dl);
18837 // Issue new vector shifts for the smaller types
18838 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
18839 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
18841 // Concatenate the result back
18842 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
18848 static SDValue LowerRotate(SDValue Op, const X86Subtarget *Subtarget,
18849 SelectionDAG &DAG) {
18850 MVT VT = Op.getSimpleValueType();
18852 SDValue R = Op.getOperand(0);
18853 SDValue Amt = Op.getOperand(1);
18855 assert(VT.isVector() && "Custom lowering only for vector rotates!");
18856 assert(Subtarget->hasXOP() && "XOP support required for vector rotates!");
18857 assert((Op.getOpcode() == ISD::ROTL) && "Only ROTL supported");
18859 // XOP has 128-bit vector variable + immediate rotates.
18860 // +ve/-ve Amt = rotate left/right.
18862 // Split 256-bit integers.
18863 if (VT.is256BitVector())
18864 return Lower256IntArith(Op, DAG);
18866 assert(VT.is128BitVector() && "Only rotate 128-bit vectors!");
18868 // Attempt to rotate by immediate.
18869 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
18870 if (auto *RotateConst = BVAmt->getConstantSplatNode()) {
18871 uint64_t RotateAmt = RotateConst->getAPIntValue().getZExtValue();
18872 assert(RotateAmt < VT.getScalarSizeInBits() && "Rotation out of range");
18873 return DAG.getNode(X86ISD::VPROTI, DL, VT, R,
18874 DAG.getConstant(RotateAmt, DL, MVT::i8));
18878 // Use general rotate by variable (per-element).
18879 return DAG.getNode(X86ISD::VPROT, DL, VT, R, Amt);
18882 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
18883 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
18884 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
18885 // looks for this combo and may remove the "setcc" instruction if the "setcc"
18886 // has only one use.
18887 SDNode *N = Op.getNode();
18888 SDValue LHS = N->getOperand(0);
18889 SDValue RHS = N->getOperand(1);
18890 unsigned BaseOp = 0;
18893 switch (Op.getOpcode()) {
18894 default: llvm_unreachable("Unknown ovf instruction!");
18896 // A subtract of one will be selected as a INC. Note that INC doesn't
18897 // set CF, so we can't do this for UADDO.
18898 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
18900 BaseOp = X86ISD::INC;
18901 Cond = X86::COND_O;
18904 BaseOp = X86ISD::ADD;
18905 Cond = X86::COND_O;
18908 BaseOp = X86ISD::ADD;
18909 Cond = X86::COND_B;
18912 // A subtract of one will be selected as a DEC. Note that DEC doesn't
18913 // set CF, so we can't do this for USUBO.
18914 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
18916 BaseOp = X86ISD::DEC;
18917 Cond = X86::COND_O;
18920 BaseOp = X86ISD::SUB;
18921 Cond = X86::COND_O;
18924 BaseOp = X86ISD::SUB;
18925 Cond = X86::COND_B;
18928 BaseOp = N->getValueType(0) == MVT::i8 ? X86ISD::SMUL8 : X86ISD::SMUL;
18929 Cond = X86::COND_O;
18931 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
18932 if (N->getValueType(0) == MVT::i8) {
18933 BaseOp = X86ISD::UMUL8;
18934 Cond = X86::COND_O;
18937 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
18939 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
18942 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
18943 DAG.getConstant(X86::COND_O, DL, MVT::i32),
18944 SDValue(Sum.getNode(), 2));
18946 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
18950 // Also sets EFLAGS.
18951 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
18952 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
18955 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
18956 DAG.getConstant(Cond, DL, MVT::i32),
18957 SDValue(Sum.getNode(), 1));
18959 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
18962 /// Returns true if the operand type is exactly twice the native width, and
18963 /// the corresponding cmpxchg8b or cmpxchg16b instruction is available.
18964 /// Used to know whether to use cmpxchg8/16b when expanding atomic operations
18965 /// (otherwise we leave them alone to become __sync_fetch_and_... calls).
18966 bool X86TargetLowering::needsCmpXchgNb(Type *MemType) const {
18967 unsigned OpWidth = MemType->getPrimitiveSizeInBits();
18970 return !Subtarget->is64Bit(); // FIXME this should be Subtarget.hasCmpxchg8b
18971 else if (OpWidth == 128)
18972 return Subtarget->hasCmpxchg16b();
18977 bool X86TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
18978 return needsCmpXchgNb(SI->getValueOperand()->getType());
18981 // Note: this turns large loads into lock cmpxchg8b/16b.
18982 // FIXME: On 32 bits x86, fild/movq might be faster than lock cmpxchg8b.
18983 TargetLowering::AtomicExpansionKind
18984 X86TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
18985 auto PTy = cast<PointerType>(LI->getPointerOperand()->getType());
18986 return needsCmpXchgNb(PTy->getElementType()) ? AtomicExpansionKind::CmpXChg
18987 : AtomicExpansionKind::None;
18990 TargetLowering::AtomicExpansionKind
18991 X86TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
18992 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
18993 Type *MemType = AI->getType();
18995 // If the operand is too big, we must see if cmpxchg8/16b is available
18996 // and default to library calls otherwise.
18997 if (MemType->getPrimitiveSizeInBits() > NativeWidth) {
18998 return needsCmpXchgNb(MemType) ? AtomicExpansionKind::CmpXChg
18999 : AtomicExpansionKind::None;
19002 AtomicRMWInst::BinOp Op = AI->getOperation();
19005 llvm_unreachable("Unknown atomic operation");
19006 case AtomicRMWInst::Xchg:
19007 case AtomicRMWInst::Add:
19008 case AtomicRMWInst::Sub:
19009 // It's better to use xadd, xsub or xchg for these in all cases.
19010 return AtomicExpansionKind::None;
19011 case AtomicRMWInst::Or:
19012 case AtomicRMWInst::And:
19013 case AtomicRMWInst::Xor:
19014 // If the atomicrmw's result isn't actually used, we can just add a "lock"
19015 // prefix to a normal instruction for these operations.
19016 return !AI->use_empty() ? AtomicExpansionKind::CmpXChg
19017 : AtomicExpansionKind::None;
19018 case AtomicRMWInst::Nand:
19019 case AtomicRMWInst::Max:
19020 case AtomicRMWInst::Min:
19021 case AtomicRMWInst::UMax:
19022 case AtomicRMWInst::UMin:
19023 // These always require a non-trivial set of data operations on x86. We must
19024 // use a cmpxchg loop.
19025 return AtomicExpansionKind::CmpXChg;
19029 static bool hasMFENCE(const X86Subtarget& Subtarget) {
19030 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
19031 // no-sse2). There isn't any reason to disable it if the target processor
19033 return Subtarget.hasSSE2() || Subtarget.is64Bit();
19037 X86TargetLowering::lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const {
19038 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
19039 Type *MemType = AI->getType();
19040 // Accesses larger than the native width are turned into cmpxchg/libcalls, so
19041 // there is no benefit in turning such RMWs into loads, and it is actually
19042 // harmful as it introduces a mfence.
19043 if (MemType->getPrimitiveSizeInBits() > NativeWidth)
19046 auto Builder = IRBuilder<>(AI);
19047 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
19048 auto SynchScope = AI->getSynchScope();
19049 // We must restrict the ordering to avoid generating loads with Release or
19050 // ReleaseAcquire orderings.
19051 auto Order = AtomicCmpXchgInst::getStrongestFailureOrdering(AI->getOrdering());
19052 auto Ptr = AI->getPointerOperand();
19054 // Before the load we need a fence. Here is an example lifted from
19055 // http://www.hpl.hp.com/techreports/2012/HPL-2012-68.pdf showing why a fence
19058 // x.store(1, relaxed);
19059 // r1 = y.fetch_add(0, release);
19061 // y.fetch_add(42, acquire);
19062 // r2 = x.load(relaxed);
19063 // r1 = r2 = 0 is impossible, but becomes possible if the idempotent rmw is
19064 // lowered to just a load without a fence. A mfence flushes the store buffer,
19065 // making the optimization clearly correct.
19066 // FIXME: it is required if isAtLeastRelease(Order) but it is not clear
19067 // otherwise, we might be able to be more aggressive on relaxed idempotent
19068 // rmw. In practice, they do not look useful, so we don't try to be
19069 // especially clever.
19070 if (SynchScope == SingleThread)
19071 // FIXME: we could just insert an X86ISD::MEMBARRIER here, except we are at
19072 // the IR level, so we must wrap it in an intrinsic.
19075 if (!hasMFENCE(*Subtarget))
19076 // FIXME: it might make sense to use a locked operation here but on a
19077 // different cache-line to prevent cache-line bouncing. In practice it
19078 // is probably a small win, and x86 processors without mfence are rare
19079 // enough that we do not bother.
19083 llvm::Intrinsic::getDeclaration(M, Intrinsic::x86_sse2_mfence);
19084 Builder.CreateCall(MFence, {});
19086 // Finally we can emit the atomic load.
19087 LoadInst *Loaded = Builder.CreateAlignedLoad(Ptr,
19088 AI->getType()->getPrimitiveSizeInBits());
19089 Loaded->setAtomic(Order, SynchScope);
19090 AI->replaceAllUsesWith(Loaded);
19091 AI->eraseFromParent();
19095 static SDValue LowerATOMIC_FENCE(SDValue Op, const X86Subtarget *Subtarget,
19096 SelectionDAG &DAG) {
19098 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
19099 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
19100 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
19101 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
19103 // The only fence that needs an instruction is a sequentially-consistent
19104 // cross-thread fence.
19105 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
19106 if (hasMFENCE(*Subtarget))
19107 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
19109 SDValue Chain = Op.getOperand(0);
19110 SDValue Zero = DAG.getConstant(0, dl, MVT::i32);
19112 DAG.getRegister(X86::ESP, MVT::i32), // Base
19113 DAG.getTargetConstant(1, dl, MVT::i8), // Scale
19114 DAG.getRegister(0, MVT::i32), // Index
19115 DAG.getTargetConstant(0, dl, MVT::i32), // Disp
19116 DAG.getRegister(0, MVT::i32), // Segment.
19120 SDNode *Res = DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops);
19121 return SDValue(Res, 0);
19124 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
19125 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
19128 static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget *Subtarget,
19129 SelectionDAG &DAG) {
19130 MVT T = Op.getSimpleValueType();
19134 switch(T.SimpleTy) {
19135 default: llvm_unreachable("Invalid value type!");
19136 case MVT::i8: Reg = X86::AL; size = 1; break;
19137 case MVT::i16: Reg = X86::AX; size = 2; break;
19138 case MVT::i32: Reg = X86::EAX; size = 4; break;
19140 assert(Subtarget->is64Bit() && "Node not type legal!");
19141 Reg = X86::RAX; size = 8;
19144 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
19145 Op.getOperand(2), SDValue());
19146 SDValue Ops[] = { cpIn.getValue(0),
19149 DAG.getTargetConstant(size, DL, MVT::i8),
19150 cpIn.getValue(1) };
19151 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
19152 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
19153 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
19157 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
19158 SDValue EFLAGS = DAG.getCopyFromReg(cpOut.getValue(1), DL, X86::EFLAGS,
19159 MVT::i32, cpOut.getValue(2));
19160 SDValue Success = DAG.getNode(X86ISD::SETCC, DL, Op->getValueType(1),
19161 DAG.getConstant(X86::COND_E, DL, MVT::i8),
19164 DAG.ReplaceAllUsesOfValueWith(Op.getValue(0), cpOut);
19165 DAG.ReplaceAllUsesOfValueWith(Op.getValue(1), Success);
19166 DAG.ReplaceAllUsesOfValueWith(Op.getValue(2), EFLAGS.getValue(1));
19170 static SDValue LowerBITCAST(SDValue Op, const X86Subtarget *Subtarget,
19171 SelectionDAG &DAG) {
19172 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
19173 MVT DstVT = Op.getSimpleValueType();
19175 if (SrcVT == MVT::v2i32 || SrcVT == MVT::v4i16 || SrcVT == MVT::v8i8) {
19176 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19177 if (DstVT != MVT::f64)
19178 // This conversion needs to be expanded.
19181 SDValue InVec = Op->getOperand(0);
19183 unsigned NumElts = SrcVT.getVectorNumElements();
19184 MVT SVT = SrcVT.getVectorElementType();
19186 // Widen the vector in input in the case of MVT::v2i32.
19187 // Example: from MVT::v2i32 to MVT::v4i32.
19188 SmallVector<SDValue, 16> Elts;
19189 for (unsigned i = 0, e = NumElts; i != e; ++i)
19190 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT, InVec,
19191 DAG.getIntPtrConstant(i, dl)));
19193 // Explicitly mark the extra elements as Undef.
19194 Elts.append(NumElts, DAG.getUNDEF(SVT));
19196 EVT NewVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
19197 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Elts);
19198 SDValue ToV2F64 = DAG.getBitcast(MVT::v2f64, BV);
19199 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, ToV2F64,
19200 DAG.getIntPtrConstant(0, dl));
19203 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
19204 Subtarget->hasMMX() && "Unexpected custom BITCAST");
19205 assert((DstVT == MVT::i64 ||
19206 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
19207 "Unexpected custom BITCAST");
19208 // i64 <=> MMX conversions are Legal.
19209 if (SrcVT==MVT::i64 && DstVT.isVector())
19211 if (DstVT==MVT::i64 && SrcVT.isVector())
19213 // MMX <=> MMX conversions are Legal.
19214 if (SrcVT.isVector() && DstVT.isVector())
19216 // All other conversions need to be expanded.
19220 /// Compute the horizontal sum of bytes in V for the elements of VT.
19222 /// Requires V to be a byte vector and VT to be an integer vector type with
19223 /// wider elements than V's type. The width of the elements of VT determines
19224 /// how many bytes of V are summed horizontally to produce each element of the
19226 static SDValue LowerHorizontalByteSum(SDValue V, MVT VT,
19227 const X86Subtarget *Subtarget,
19228 SelectionDAG &DAG) {
19230 MVT ByteVecVT = V.getSimpleValueType();
19231 MVT EltVT = VT.getVectorElementType();
19232 int NumElts = VT.getVectorNumElements();
19233 assert(ByteVecVT.getVectorElementType() == MVT::i8 &&
19234 "Expected value to have byte element type.");
19235 assert(EltVT != MVT::i8 &&
19236 "Horizontal byte sum only makes sense for wider elements!");
19237 unsigned VecSize = VT.getSizeInBits();
19238 assert(ByteVecVT.getSizeInBits() == VecSize && "Cannot change vector size!");
19240 // PSADBW instruction horizontally add all bytes and leave the result in i64
19241 // chunks, thus directly computes the pop count for v2i64 and v4i64.
19242 if (EltVT == MVT::i64) {
19243 SDValue Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
19244 V = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT, V, Zeros);
19245 return DAG.getBitcast(VT, V);
19248 if (EltVT == MVT::i32) {
19249 // We unpack the low half and high half into i32s interleaved with zeros so
19250 // that we can use PSADBW to horizontally sum them. The most useful part of
19251 // this is that it lines up the results of two PSADBW instructions to be
19252 // two v2i64 vectors which concatenated are the 4 population counts. We can
19253 // then use PACKUSWB to shrink and concatenate them into a v4i32 again.
19254 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, DL);
19255 SDValue Low = DAG.getNode(X86ISD::UNPCKL, DL, VT, V, Zeros);
19256 SDValue High = DAG.getNode(X86ISD::UNPCKH, DL, VT, V, Zeros);
19258 // Do the horizontal sums into two v2i64s.
19259 Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
19260 Low = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT,
19261 DAG.getBitcast(ByteVecVT, Low), Zeros);
19262 High = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT,
19263 DAG.getBitcast(ByteVecVT, High), Zeros);
19265 // Merge them together.
19266 MVT ShortVecVT = MVT::getVectorVT(MVT::i16, VecSize / 16);
19267 V = DAG.getNode(X86ISD::PACKUS, DL, ByteVecVT,
19268 DAG.getBitcast(ShortVecVT, Low),
19269 DAG.getBitcast(ShortVecVT, High));
19271 return DAG.getBitcast(VT, V);
19274 // The only element type left is i16.
19275 assert(EltVT == MVT::i16 && "Unknown how to handle type");
19277 // To obtain pop count for each i16 element starting from the pop count for
19278 // i8 elements, shift the i16s left by 8, sum as i8s, and then shift as i16s
19279 // right by 8. It is important to shift as i16s as i8 vector shift isn't
19280 // directly supported.
19281 SmallVector<SDValue, 16> Shifters(NumElts, DAG.getConstant(8, DL, EltVT));
19282 SDValue Shifter = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters);
19283 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, DAG.getBitcast(VT, V), Shifter);
19284 V = DAG.getNode(ISD::ADD, DL, ByteVecVT, DAG.getBitcast(ByteVecVT, Shl),
19285 DAG.getBitcast(ByteVecVT, V));
19286 return DAG.getNode(ISD::SRL, DL, VT, DAG.getBitcast(VT, V), Shifter);
19289 static SDValue LowerVectorCTPOPInRegLUT(SDValue Op, SDLoc DL,
19290 const X86Subtarget *Subtarget,
19291 SelectionDAG &DAG) {
19292 MVT VT = Op.getSimpleValueType();
19293 MVT EltVT = VT.getVectorElementType();
19294 unsigned VecSize = VT.getSizeInBits();
19296 // Implement a lookup table in register by using an algorithm based on:
19297 // http://wm.ite.pl/articles/sse-popcount.html
19299 // The general idea is that every lower byte nibble in the input vector is an
19300 // index into a in-register pre-computed pop count table. We then split up the
19301 // input vector in two new ones: (1) a vector with only the shifted-right
19302 // higher nibbles for each byte and (2) a vector with the lower nibbles (and
19303 // masked out higher ones) for each byte. PSHUB is used separately with both
19304 // to index the in-register table. Next, both are added and the result is a
19305 // i8 vector where each element contains the pop count for input byte.
19307 // To obtain the pop count for elements != i8, we follow up with the same
19308 // approach and use additional tricks as described below.
19310 const int LUT[16] = {/* 0 */ 0, /* 1 */ 1, /* 2 */ 1, /* 3 */ 2,
19311 /* 4 */ 1, /* 5 */ 2, /* 6 */ 2, /* 7 */ 3,
19312 /* 8 */ 1, /* 9 */ 2, /* a */ 2, /* b */ 3,
19313 /* c */ 2, /* d */ 3, /* e */ 3, /* f */ 4};
19315 int NumByteElts = VecSize / 8;
19316 MVT ByteVecVT = MVT::getVectorVT(MVT::i8, NumByteElts);
19317 SDValue In = DAG.getBitcast(ByteVecVT, Op);
19318 SmallVector<SDValue, 16> LUTVec;
19319 for (int i = 0; i < NumByteElts; ++i)
19320 LUTVec.push_back(DAG.getConstant(LUT[i % 16], DL, MVT::i8));
19321 SDValue InRegLUT = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, LUTVec);
19322 SmallVector<SDValue, 16> Mask0F(NumByteElts,
19323 DAG.getConstant(0x0F, DL, MVT::i8));
19324 SDValue M0F = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Mask0F);
19327 SmallVector<SDValue, 16> Four(NumByteElts, DAG.getConstant(4, DL, MVT::i8));
19328 SDValue FourV = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Four);
19329 SDValue HighNibbles = DAG.getNode(ISD::SRL, DL, ByteVecVT, In, FourV);
19332 SDValue LowNibbles = DAG.getNode(ISD::AND, DL, ByteVecVT, In, M0F);
19334 // The input vector is used as the shuffle mask that index elements into the
19335 // LUT. After counting low and high nibbles, add the vector to obtain the
19336 // final pop count per i8 element.
19337 SDValue HighPopCnt =
19338 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, HighNibbles);
19339 SDValue LowPopCnt =
19340 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, LowNibbles);
19341 SDValue PopCnt = DAG.getNode(ISD::ADD, DL, ByteVecVT, HighPopCnt, LowPopCnt);
19343 if (EltVT == MVT::i8)
19346 return LowerHorizontalByteSum(PopCnt, VT, Subtarget, DAG);
19349 static SDValue LowerVectorCTPOPBitmath(SDValue Op, SDLoc DL,
19350 const X86Subtarget *Subtarget,
19351 SelectionDAG &DAG) {
19352 MVT VT = Op.getSimpleValueType();
19353 assert(VT.is128BitVector() &&
19354 "Only 128-bit vector bitmath lowering supported.");
19356 int VecSize = VT.getSizeInBits();
19357 MVT EltVT = VT.getVectorElementType();
19358 int Len = EltVT.getSizeInBits();
19360 // This is the vectorized version of the "best" algorithm from
19361 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
19362 // with a minor tweak to use a series of adds + shifts instead of vector
19363 // multiplications. Implemented for all integer vector types. We only use
19364 // this when we don't have SSSE3 which allows a LUT-based lowering that is
19365 // much faster, even faster than using native popcnt instructions.
19367 auto GetShift = [&](unsigned OpCode, SDValue V, int Shifter) {
19368 MVT VT = V.getSimpleValueType();
19369 SmallVector<SDValue, 32> Shifters(
19370 VT.getVectorNumElements(),
19371 DAG.getConstant(Shifter, DL, VT.getVectorElementType()));
19372 return DAG.getNode(OpCode, DL, VT, V,
19373 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters));
19375 auto GetMask = [&](SDValue V, APInt Mask) {
19376 MVT VT = V.getSimpleValueType();
19377 SmallVector<SDValue, 32> Masks(
19378 VT.getVectorNumElements(),
19379 DAG.getConstant(Mask, DL, VT.getVectorElementType()));
19380 return DAG.getNode(ISD::AND, DL, VT, V,
19381 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Masks));
19384 // We don't want to incur the implicit masks required to SRL vNi8 vectors on
19385 // x86, so set the SRL type to have elements at least i16 wide. This is
19386 // correct because all of our SRLs are followed immediately by a mask anyways
19387 // that handles any bits that sneak into the high bits of the byte elements.
19388 MVT SrlVT = Len > 8 ? VT : MVT::getVectorVT(MVT::i16, VecSize / 16);
19392 // v = v - ((v >> 1) & 0x55555555...)
19394 DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 1));
19395 SDValue And = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x55)));
19396 V = DAG.getNode(ISD::SUB, DL, VT, V, And);
19398 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
19399 SDValue AndLHS = GetMask(V, APInt::getSplat(Len, APInt(8, 0x33)));
19400 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 2));
19401 SDValue AndRHS = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x33)));
19402 V = DAG.getNode(ISD::ADD, DL, VT, AndLHS, AndRHS);
19404 // v = (v + (v >> 4)) & 0x0F0F0F0F...
19405 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 4));
19406 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, V, Srl);
19407 V = GetMask(Add, APInt::getSplat(Len, APInt(8, 0x0F)));
19409 // At this point, V contains the byte-wise population count, and we are
19410 // merely doing a horizontal sum if necessary to get the wider element
19412 if (EltVT == MVT::i8)
19415 return LowerHorizontalByteSum(
19416 DAG.getBitcast(MVT::getVectorVT(MVT::i8, VecSize / 8), V), VT, Subtarget,
19420 static SDValue LowerVectorCTPOP(SDValue Op, const X86Subtarget *Subtarget,
19421 SelectionDAG &DAG) {
19422 MVT VT = Op.getSimpleValueType();
19423 // FIXME: Need to add AVX-512 support here!
19424 assert((VT.is256BitVector() || VT.is128BitVector()) &&
19425 "Unknown CTPOP type to handle");
19426 SDLoc DL(Op.getNode());
19427 SDValue Op0 = Op.getOperand(0);
19429 if (!Subtarget->hasSSSE3()) {
19430 // We can't use the fast LUT approach, so fall back on vectorized bitmath.
19431 assert(VT.is128BitVector() && "Only 128-bit vectors supported in SSE!");
19432 return LowerVectorCTPOPBitmath(Op0, DL, Subtarget, DAG);
19435 if (VT.is256BitVector() && !Subtarget->hasInt256()) {
19436 unsigned NumElems = VT.getVectorNumElements();
19438 // Extract each 128-bit vector, compute pop count and concat the result.
19439 SDValue LHS = Extract128BitVector(Op0, 0, DAG, DL);
19440 SDValue RHS = Extract128BitVector(Op0, NumElems/2, DAG, DL);
19442 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT,
19443 LowerVectorCTPOPInRegLUT(LHS, DL, Subtarget, DAG),
19444 LowerVectorCTPOPInRegLUT(RHS, DL, Subtarget, DAG));
19447 return LowerVectorCTPOPInRegLUT(Op0, DL, Subtarget, DAG);
19450 static SDValue LowerCTPOP(SDValue Op, const X86Subtarget *Subtarget,
19451 SelectionDAG &DAG) {
19452 assert(Op.getSimpleValueType().isVector() &&
19453 "We only do custom lowering for vector population count.");
19454 return LowerVectorCTPOP(Op, Subtarget, DAG);
19457 static SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
19458 SDNode *Node = Op.getNode();
19460 EVT T = Node->getValueType(0);
19461 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
19462 DAG.getConstant(0, dl, T), Node->getOperand(2));
19463 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
19464 cast<AtomicSDNode>(Node)->getMemoryVT(),
19465 Node->getOperand(0),
19466 Node->getOperand(1), negOp,
19467 cast<AtomicSDNode>(Node)->getMemOperand(),
19468 cast<AtomicSDNode>(Node)->getOrdering(),
19469 cast<AtomicSDNode>(Node)->getSynchScope());
19472 static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) {
19473 SDNode *Node = Op.getNode();
19475 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
19477 // Convert seq_cst store -> xchg
19478 // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b)
19479 // FIXME: On 32-bit, store -> fist or movq would be more efficient
19480 // (The only way to get a 16-byte store is cmpxchg16b)
19481 // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment.
19482 if (cast<AtomicSDNode>(Node)->getOrdering() == SequentiallyConsistent ||
19483 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
19484 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
19485 cast<AtomicSDNode>(Node)->getMemoryVT(),
19486 Node->getOperand(0),
19487 Node->getOperand(1), Node->getOperand(2),
19488 cast<AtomicSDNode>(Node)->getMemOperand(),
19489 cast<AtomicSDNode>(Node)->getOrdering(),
19490 cast<AtomicSDNode>(Node)->getSynchScope());
19491 return Swap.getValue(1);
19493 // Other atomic stores have a simple pattern.
19497 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
19498 MVT VT = Op.getNode()->getSimpleValueType(0);
19500 // Let legalize expand this if it isn't a legal type yet.
19501 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
19504 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
19507 bool ExtraOp = false;
19508 switch (Op.getOpcode()) {
19509 default: llvm_unreachable("Invalid code");
19510 case ISD::ADDC: Opc = X86ISD::ADD; break;
19511 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
19512 case ISD::SUBC: Opc = X86ISD::SUB; break;
19513 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
19517 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
19519 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
19520 Op.getOperand(1), Op.getOperand(2));
19523 static SDValue LowerFSINCOS(SDValue Op, const X86Subtarget *Subtarget,
19524 SelectionDAG &DAG) {
19525 assert(Subtarget->isTargetDarwin() && Subtarget->is64Bit());
19527 // For MacOSX, we want to call an alternative entry point: __sincos_stret,
19528 // which returns the values as { float, float } (in XMM0) or
19529 // { double, double } (which is returned in XMM0, XMM1).
19531 SDValue Arg = Op.getOperand(0);
19532 EVT ArgVT = Arg.getValueType();
19533 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
19535 TargetLowering::ArgListTy Args;
19536 TargetLowering::ArgListEntry Entry;
19540 Entry.isSExt = false;
19541 Entry.isZExt = false;
19542 Args.push_back(Entry);
19544 bool isF64 = ArgVT == MVT::f64;
19545 // Only optimize x86_64 for now. i386 is a bit messy. For f32,
19546 // the small struct {f32, f32} is returned in (eax, edx). For f64,
19547 // the results are returned via SRet in memory.
19548 const char *LibcallName = isF64 ? "__sincos_stret" : "__sincosf_stret";
19549 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
19551 DAG.getExternalSymbol(LibcallName, TLI.getPointerTy(DAG.getDataLayout()));
19553 Type *RetTy = isF64
19554 ? (Type*)StructType::get(ArgTy, ArgTy, nullptr)
19555 : (Type*)VectorType::get(ArgTy, 4);
19557 TargetLowering::CallLoweringInfo CLI(DAG);
19558 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
19559 .setCallee(CallingConv::C, RetTy, Callee, std::move(Args), 0);
19561 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
19564 // Returned in xmm0 and xmm1.
19565 return CallResult.first;
19567 // Returned in bits 0:31 and 32:64 xmm0.
19568 SDValue SinVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
19569 CallResult.first, DAG.getIntPtrConstant(0, dl));
19570 SDValue CosVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
19571 CallResult.first, DAG.getIntPtrConstant(1, dl));
19572 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
19573 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys, SinVal, CosVal);
19576 static SDValue LowerMSCATTER(SDValue Op, const X86Subtarget *Subtarget,
19577 SelectionDAG &DAG) {
19578 assert(Subtarget->hasAVX512() &&
19579 "MGATHER/MSCATTER are supported on AVX-512 arch only");
19581 MaskedScatterSDNode *N = cast<MaskedScatterSDNode>(Op.getNode());
19582 MVT VT = N->getValue().getSimpleValueType();
19583 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported scatter op");
19586 // X86 scatter kills mask register, so its type should be added to
19587 // the list of return values
19588 if (N->getNumValues() == 1) {
19589 SDValue Index = N->getIndex();
19590 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
19591 !Index.getSimpleValueType().is512BitVector())
19592 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
19594 SDVTList VTs = DAG.getVTList(N->getMask().getValueType(), MVT::Other);
19595 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
19596 N->getOperand(3), Index };
19598 SDValue NewScatter = DAG.getMaskedScatter(VTs, VT, dl, Ops, N->getMemOperand());
19599 DAG.ReplaceAllUsesWith(Op, SDValue(NewScatter.getNode(), 1));
19600 return SDValue(NewScatter.getNode(), 0);
19605 static SDValue LowerMGATHER(SDValue Op, const X86Subtarget *Subtarget,
19606 SelectionDAG &DAG) {
19607 assert(Subtarget->hasAVX512() &&
19608 "MGATHER/MSCATTER are supported on AVX-512 arch only");
19610 MaskedGatherSDNode *N = cast<MaskedGatherSDNode>(Op.getNode());
19611 MVT VT = Op.getSimpleValueType();
19612 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported gather op");
19615 SDValue Index = N->getIndex();
19616 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
19617 !Index.getSimpleValueType().is512BitVector()) {
19618 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
19619 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
19620 N->getOperand(3), Index };
19621 DAG.UpdateNodeOperands(N, Ops);
19626 SDValue X86TargetLowering::LowerGC_TRANSITION_START(SDValue Op,
19627 SelectionDAG &DAG) const {
19628 // TODO: Eventually, the lowering of these nodes should be informed by or
19629 // deferred to the GC strategy for the function in which they appear. For
19630 // now, however, they must be lowered to something. Since they are logically
19631 // no-ops in the case of a null GC strategy (or a GC strategy which does not
19632 // require special handling for these nodes), lower them as literal NOOPs for
19634 SmallVector<SDValue, 2> Ops;
19636 Ops.push_back(Op.getOperand(0));
19637 if (Op->getGluedNode())
19638 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
19641 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
19642 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
19647 SDValue X86TargetLowering::LowerGC_TRANSITION_END(SDValue Op,
19648 SelectionDAG &DAG) const {
19649 // TODO: Eventually, the lowering of these nodes should be informed by or
19650 // deferred to the GC strategy for the function in which they appear. For
19651 // now, however, they must be lowered to something. Since they are logically
19652 // no-ops in the case of a null GC strategy (or a GC strategy which does not
19653 // require special handling for these nodes), lower them as literal NOOPs for
19655 SmallVector<SDValue, 2> Ops;
19657 Ops.push_back(Op.getOperand(0));
19658 if (Op->getGluedNode())
19659 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
19662 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
19663 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
19668 /// LowerOperation - Provide custom lowering hooks for some operations.
19670 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
19671 switch (Op.getOpcode()) {
19672 default: llvm_unreachable("Should not custom lower this!");
19673 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, Subtarget, DAG);
19674 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
19675 return LowerCMP_SWAP(Op, Subtarget, DAG);
19676 case ISD::CTPOP: return LowerCTPOP(Op, Subtarget, DAG);
19677 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
19678 case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG);
19679 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
19680 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, Subtarget, DAG);
19681 case ISD::VECTOR_SHUFFLE: return lowerVectorShuffle(Op, Subtarget, DAG);
19682 case ISD::VSELECT: return LowerVSELECT(Op, DAG);
19683 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
19684 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
19685 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op,Subtarget,DAG);
19686 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, Subtarget,DAG);
19687 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
19688 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
19689 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
19690 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
19691 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
19692 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
19693 case ISD::SHL_PARTS:
19694 case ISD::SRA_PARTS:
19695 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
19696 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
19697 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
19698 case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG);
19699 case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, Subtarget, DAG);
19700 case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, Subtarget, DAG);
19701 case ISD::ANY_EXTEND: return LowerANY_EXTEND(Op, Subtarget, DAG);
19702 case ISD::SIGN_EXTEND_VECTOR_INREG:
19703 return LowerSIGN_EXTEND_VECTOR_INREG(Op, Subtarget, DAG);
19704 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
19705 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
19706 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
19707 case ISD::LOAD: return LowerExtendedLoad(Op, Subtarget, DAG);
19709 case ISD::FNEG: return LowerFABSorFNEG(Op, DAG);
19710 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
19711 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
19712 case ISD::SETCC: return LowerSETCC(Op, DAG);
19713 case ISD::SETCCE: return LowerSETCCE(Op, DAG);
19714 case ISD::SELECT: return LowerSELECT(Op, DAG);
19715 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
19716 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
19717 case ISD::VASTART: return LowerVASTART(Op, DAG);
19718 case ISD::VAARG: return LowerVAARG(Op, DAG);
19719 case ISD::VACOPY: return LowerVACOPY(Op, Subtarget, DAG);
19720 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, Subtarget, DAG);
19721 case ISD::INTRINSIC_VOID:
19722 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, Subtarget, DAG);
19723 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
19724 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
19725 case ISD::FRAME_TO_ARGS_OFFSET:
19726 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
19727 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
19728 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
19729 case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
19730 case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
19731 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
19732 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
19733 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
19734 case ISD::CTLZ: return LowerCTLZ(Op, Subtarget, DAG);
19735 case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, Subtarget, DAG);
19737 case ISD::CTTZ_ZERO_UNDEF: return LowerCTTZ(Op, DAG);
19738 case ISD::MUL: return LowerMUL(Op, Subtarget, DAG);
19739 case ISD::UMUL_LOHI:
19740 case ISD::SMUL_LOHI: return LowerMUL_LOHI(Op, Subtarget, DAG);
19741 case ISD::ROTL: return LowerRotate(Op, Subtarget, DAG);
19744 case ISD::SHL: return LowerShift(Op, Subtarget, DAG);
19750 case ISD::UMULO: return LowerXALUO(Op, DAG);
19751 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, Subtarget,DAG);
19752 case ISD::BITCAST: return LowerBITCAST(Op, Subtarget, DAG);
19756 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
19757 case ISD::ADD: return LowerADD(Op, DAG);
19758 case ISD::SUB: return LowerSUB(Op, DAG);
19762 case ISD::UMIN: return LowerMINMAX(Op, DAG);
19763 case ISD::FSINCOS: return LowerFSINCOS(Op, Subtarget, DAG);
19764 case ISD::MGATHER: return LowerMGATHER(Op, Subtarget, DAG);
19765 case ISD::MSCATTER: return LowerMSCATTER(Op, Subtarget, DAG);
19766 case ISD::GC_TRANSITION_START:
19767 return LowerGC_TRANSITION_START(Op, DAG);
19768 case ISD::GC_TRANSITION_END: return LowerGC_TRANSITION_END(Op, DAG);
19772 /// ReplaceNodeResults - Replace a node with an illegal result type
19773 /// with a new node built out of custom code.
19774 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
19775 SmallVectorImpl<SDValue>&Results,
19776 SelectionDAG &DAG) const {
19778 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
19779 switch (N->getOpcode()) {
19781 llvm_unreachable("Do not know how to custom type legalize this operation!");
19782 // We might have generated v2f32 FMIN/FMAX operations. Widen them to v4f32.
19783 case X86ISD::FMINC:
19785 case X86ISD::FMAXC:
19786 case X86ISD::FMAX: {
19787 EVT VT = N->getValueType(0);
19788 assert(VT == MVT::v2f32 && "Unexpected type (!= v2f32) on FMIN/FMAX.");
19789 SDValue UNDEF = DAG.getUNDEF(VT);
19790 SDValue LHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
19791 N->getOperand(0), UNDEF);
19792 SDValue RHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
19793 N->getOperand(1), UNDEF);
19794 Results.push_back(DAG.getNode(N->getOpcode(), dl, MVT::v4f32, LHS, RHS));
19797 case ISD::SIGN_EXTEND_INREG:
19802 // We don't want to expand or promote these.
19809 case ISD::UDIVREM: {
19810 SDValue V = LowerWin64_i128OP(SDValue(N,0), DAG);
19811 Results.push_back(V);
19814 case ISD::FP_TO_SINT:
19815 case ISD::FP_TO_UINT: {
19816 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
19818 std::pair<SDValue,SDValue> Vals =
19819 FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true);
19820 SDValue FIST = Vals.first, StackSlot = Vals.second;
19821 if (FIST.getNode()) {
19822 EVT VT = N->getValueType(0);
19823 // Return a load from the stack slot.
19824 if (StackSlot.getNode())
19825 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
19826 MachinePointerInfo(),
19827 false, false, false, 0));
19829 Results.push_back(FIST);
19833 case ISD::UINT_TO_FP: {
19834 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19835 if (N->getOperand(0).getValueType() != MVT::v2i32 ||
19836 N->getValueType(0) != MVT::v2f32)
19838 SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64,
19840 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
19842 SDValue VBias = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2f64, Bias, Bias);
19843 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn,
19844 DAG.getBitcast(MVT::v2i64, VBias));
19845 Or = DAG.getBitcast(MVT::v2f64, Or);
19846 // TODO: Are there any fast-math-flags to propagate here?
19847 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias);
19848 Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub));
19851 case ISD::FP_ROUND: {
19852 if (!TLI.isTypeLegal(N->getOperand(0).getValueType()))
19854 SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0));
19855 Results.push_back(V);
19858 case ISD::FP_EXTEND: {
19859 // Right now, only MVT::v2f32 has OperationAction for FP_EXTEND.
19860 // No other ValueType for FP_EXTEND should reach this point.
19861 assert(N->getValueType(0) == MVT::v2f32 &&
19862 "Do not know how to legalize this Node");
19865 case ISD::INTRINSIC_W_CHAIN: {
19866 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
19868 default : llvm_unreachable("Do not know how to custom type "
19869 "legalize this intrinsic operation!");
19870 case Intrinsic::x86_rdtsc:
19871 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
19873 case Intrinsic::x86_rdtscp:
19874 return getReadTimeStampCounter(N, dl, X86ISD::RDTSCP_DAG, DAG, Subtarget,
19876 case Intrinsic::x86_rdpmc:
19877 return getReadPerformanceCounter(N, dl, DAG, Subtarget, Results);
19880 case ISD::READCYCLECOUNTER: {
19881 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
19884 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
19885 EVT T = N->getValueType(0);
19886 assert((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair");
19887 bool Regs64bit = T == MVT::i128;
19888 MVT HalfT = Regs64bit ? MVT::i64 : MVT::i32;
19889 SDValue cpInL, cpInH;
19890 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
19891 DAG.getConstant(0, dl, HalfT));
19892 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
19893 DAG.getConstant(1, dl, HalfT));
19894 cpInL = DAG.getCopyToReg(N->getOperand(0), dl,
19895 Regs64bit ? X86::RAX : X86::EAX,
19897 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl,
19898 Regs64bit ? X86::RDX : X86::EDX,
19899 cpInH, cpInL.getValue(1));
19900 SDValue swapInL, swapInH;
19901 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
19902 DAG.getConstant(0, dl, HalfT));
19903 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
19904 DAG.getConstant(1, dl, HalfT));
19905 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl,
19906 Regs64bit ? X86::RBX : X86::EBX,
19907 swapInL, cpInH.getValue(1));
19908 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl,
19909 Regs64bit ? X86::RCX : X86::ECX,
19910 swapInH, swapInL.getValue(1));
19911 SDValue Ops[] = { swapInH.getValue(0),
19913 swapInH.getValue(1) };
19914 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
19915 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
19916 unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_DAG :
19917 X86ISD::LCMPXCHG8_DAG;
19918 SDValue Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys, Ops, T, MMO);
19919 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl,
19920 Regs64bit ? X86::RAX : X86::EAX,
19921 HalfT, Result.getValue(1));
19922 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl,
19923 Regs64bit ? X86::RDX : X86::EDX,
19924 HalfT, cpOutL.getValue(2));
19925 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
19927 SDValue EFLAGS = DAG.getCopyFromReg(cpOutH.getValue(1), dl, X86::EFLAGS,
19928 MVT::i32, cpOutH.getValue(2));
19930 DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
19931 DAG.getConstant(X86::COND_E, dl, MVT::i8), EFLAGS);
19932 Success = DAG.getZExtOrTrunc(Success, dl, N->getValueType(1));
19934 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF));
19935 Results.push_back(Success);
19936 Results.push_back(EFLAGS.getValue(1));
19939 case ISD::ATOMIC_SWAP:
19940 case ISD::ATOMIC_LOAD_ADD:
19941 case ISD::ATOMIC_LOAD_SUB:
19942 case ISD::ATOMIC_LOAD_AND:
19943 case ISD::ATOMIC_LOAD_OR:
19944 case ISD::ATOMIC_LOAD_XOR:
19945 case ISD::ATOMIC_LOAD_NAND:
19946 case ISD::ATOMIC_LOAD_MIN:
19947 case ISD::ATOMIC_LOAD_MAX:
19948 case ISD::ATOMIC_LOAD_UMIN:
19949 case ISD::ATOMIC_LOAD_UMAX:
19950 case ISD::ATOMIC_LOAD: {
19951 // Delegate to generic TypeLegalization. Situations we can really handle
19952 // should have already been dealt with by AtomicExpandPass.cpp.
19955 case ISD::BITCAST: {
19956 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19957 EVT DstVT = N->getValueType(0);
19958 EVT SrcVT = N->getOperand(0)->getValueType(0);
19960 if (SrcVT != MVT::f64 ||
19961 (DstVT != MVT::v2i32 && DstVT != MVT::v4i16 && DstVT != MVT::v8i8))
19964 unsigned NumElts = DstVT.getVectorNumElements();
19965 EVT SVT = DstVT.getVectorElementType();
19966 EVT WiderVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
19967 SDValue Expanded = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
19968 MVT::v2f64, N->getOperand(0));
19969 SDValue ToVecInt = DAG.getBitcast(WiderVT, Expanded);
19971 if (ExperimentalVectorWideningLegalization) {
19972 // If we are legalizing vectors by widening, we already have the desired
19973 // legal vector type, just return it.
19974 Results.push_back(ToVecInt);
19978 SmallVector<SDValue, 8> Elts;
19979 for (unsigned i = 0, e = NumElts; i != e; ++i)
19980 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT,
19981 ToVecInt, DAG.getIntPtrConstant(i, dl)));
19983 Results.push_back(DAG.getNode(ISD::BUILD_VECTOR, dl, DstVT, Elts));
19988 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
19989 switch ((X86ISD::NodeType)Opcode) {
19990 case X86ISD::FIRST_NUMBER: break;
19991 case X86ISD::BSF: return "X86ISD::BSF";
19992 case X86ISD::BSR: return "X86ISD::BSR";
19993 case X86ISD::SHLD: return "X86ISD::SHLD";
19994 case X86ISD::SHRD: return "X86ISD::SHRD";
19995 case X86ISD::FAND: return "X86ISD::FAND";
19996 case X86ISD::FANDN: return "X86ISD::FANDN";
19997 case X86ISD::FOR: return "X86ISD::FOR";
19998 case X86ISD::FXOR: return "X86ISD::FXOR";
19999 case X86ISD::FILD: return "X86ISD::FILD";
20000 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
20001 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
20002 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
20003 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
20004 case X86ISD::FLD: return "X86ISD::FLD";
20005 case X86ISD::FST: return "X86ISD::FST";
20006 case X86ISD::CALL: return "X86ISD::CALL";
20007 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
20008 case X86ISD::RDTSCP_DAG: return "X86ISD::RDTSCP_DAG";
20009 case X86ISD::RDPMC_DAG: return "X86ISD::RDPMC_DAG";
20010 case X86ISD::BT: return "X86ISD::BT";
20011 case X86ISD::CMP: return "X86ISD::CMP";
20012 case X86ISD::COMI: return "X86ISD::COMI";
20013 case X86ISD::UCOMI: return "X86ISD::UCOMI";
20014 case X86ISD::CMPM: return "X86ISD::CMPM";
20015 case X86ISD::CMPMU: return "X86ISD::CMPMU";
20016 case X86ISD::CMPM_RND: return "X86ISD::CMPM_RND";
20017 case X86ISD::SETCC: return "X86ISD::SETCC";
20018 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
20019 case X86ISD::FSETCC: return "X86ISD::FSETCC";
20020 case X86ISD::FGETSIGNx86: return "X86ISD::FGETSIGNx86";
20021 case X86ISD::CMOV: return "X86ISD::CMOV";
20022 case X86ISD::BRCOND: return "X86ISD::BRCOND";
20023 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
20024 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
20025 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
20026 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
20027 case X86ISD::Wrapper: return "X86ISD::Wrapper";
20028 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
20029 case X86ISD::MOVDQ2Q: return "X86ISD::MOVDQ2Q";
20030 case X86ISD::MMX_MOVD2W: return "X86ISD::MMX_MOVD2W";
20031 case X86ISD::MMX_MOVW2D: return "X86ISD::MMX_MOVW2D";
20032 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
20033 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
20034 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
20035 case X86ISD::PINSRB: return "X86ISD::PINSRB";
20036 case X86ISD::PINSRW: return "X86ISD::PINSRW";
20037 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
20038 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
20039 case X86ISD::ANDNP: return "X86ISD::ANDNP";
20040 case X86ISD::PSIGN: return "X86ISD::PSIGN";
20041 case X86ISD::BLENDI: return "X86ISD::BLENDI";
20042 case X86ISD::SHRUNKBLEND: return "X86ISD::SHRUNKBLEND";
20043 case X86ISD::ADDUS: return "X86ISD::ADDUS";
20044 case X86ISD::SUBUS: return "X86ISD::SUBUS";
20045 case X86ISD::HADD: return "X86ISD::HADD";
20046 case X86ISD::HSUB: return "X86ISD::HSUB";
20047 case X86ISD::FHADD: return "X86ISD::FHADD";
20048 case X86ISD::FHSUB: return "X86ISD::FHSUB";
20049 case X86ISD::ABS: return "X86ISD::ABS";
20050 case X86ISD::CONFLICT: return "X86ISD::CONFLICT";
20051 case X86ISD::FMAX: return "X86ISD::FMAX";
20052 case X86ISD::FMAX_RND: return "X86ISD::FMAX_RND";
20053 case X86ISD::FMIN: return "X86ISD::FMIN";
20054 case X86ISD::FMIN_RND: return "X86ISD::FMIN_RND";
20055 case X86ISD::FMAXC: return "X86ISD::FMAXC";
20056 case X86ISD::FMINC: return "X86ISD::FMINC";
20057 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
20058 case X86ISD::FRCP: return "X86ISD::FRCP";
20059 case X86ISD::EXTRQI: return "X86ISD::EXTRQI";
20060 case X86ISD::INSERTQI: return "X86ISD::INSERTQI";
20061 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
20062 case X86ISD::TLSBASEADDR: return "X86ISD::TLSBASEADDR";
20063 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
20064 case X86ISD::EH_SJLJ_SETJMP: return "X86ISD::EH_SJLJ_SETJMP";
20065 case X86ISD::EH_SJLJ_LONGJMP: return "X86ISD::EH_SJLJ_LONGJMP";
20066 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
20067 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
20068 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
20069 case X86ISD::FNSTSW16r: return "X86ISD::FNSTSW16r";
20070 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
20071 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
20072 case X86ISD::LCMPXCHG16_DAG: return "X86ISD::LCMPXCHG16_DAG";
20073 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
20074 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
20075 case X86ISD::VZEXT: return "X86ISD::VZEXT";
20076 case X86ISD::VSEXT: return "X86ISD::VSEXT";
20077 case X86ISD::VTRUNC: return "X86ISD::VTRUNC";
20078 case X86ISD::VTRUNCS: return "X86ISD::VTRUNCS";
20079 case X86ISD::VTRUNCUS: return "X86ISD::VTRUNCUS";
20080 case X86ISD::VINSERT: return "X86ISD::VINSERT";
20081 case X86ISD::VFPEXT: return "X86ISD::VFPEXT";
20082 case X86ISD::VFPROUND: return "X86ISD::VFPROUND";
20083 case X86ISD::CVTDQ2PD: return "X86ISD::CVTDQ2PD";
20084 case X86ISD::CVTUDQ2PD: return "X86ISD::CVTUDQ2PD";
20085 case X86ISD::VSHLDQ: return "X86ISD::VSHLDQ";
20086 case X86ISD::VSRLDQ: return "X86ISD::VSRLDQ";
20087 case X86ISD::VSHL: return "X86ISD::VSHL";
20088 case X86ISD::VSRL: return "X86ISD::VSRL";
20089 case X86ISD::VSRA: return "X86ISD::VSRA";
20090 case X86ISD::VSHLI: return "X86ISD::VSHLI";
20091 case X86ISD::VSRLI: return "X86ISD::VSRLI";
20092 case X86ISD::VSRAI: return "X86ISD::VSRAI";
20093 case X86ISD::CMPP: return "X86ISD::CMPP";
20094 case X86ISD::PCMPEQ: return "X86ISD::PCMPEQ";
20095 case X86ISD::PCMPGT: return "X86ISD::PCMPGT";
20096 case X86ISD::PCMPEQM: return "X86ISD::PCMPEQM";
20097 case X86ISD::PCMPGTM: return "X86ISD::PCMPGTM";
20098 case X86ISD::ADD: return "X86ISD::ADD";
20099 case X86ISD::SUB: return "X86ISD::SUB";
20100 case X86ISD::ADC: return "X86ISD::ADC";
20101 case X86ISD::SBB: return "X86ISD::SBB";
20102 case X86ISD::SMUL: return "X86ISD::SMUL";
20103 case X86ISD::UMUL: return "X86ISD::UMUL";
20104 case X86ISD::SMUL8: return "X86ISD::SMUL8";
20105 case X86ISD::UMUL8: return "X86ISD::UMUL8";
20106 case X86ISD::SDIVREM8_SEXT_HREG: return "X86ISD::SDIVREM8_SEXT_HREG";
20107 case X86ISD::UDIVREM8_ZEXT_HREG: return "X86ISD::UDIVREM8_ZEXT_HREG";
20108 case X86ISD::INC: return "X86ISD::INC";
20109 case X86ISD::DEC: return "X86ISD::DEC";
20110 case X86ISD::OR: return "X86ISD::OR";
20111 case X86ISD::XOR: return "X86ISD::XOR";
20112 case X86ISD::AND: return "X86ISD::AND";
20113 case X86ISD::BEXTR: return "X86ISD::BEXTR";
20114 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
20115 case X86ISD::PTEST: return "X86ISD::PTEST";
20116 case X86ISD::TESTP: return "X86ISD::TESTP";
20117 case X86ISD::TESTM: return "X86ISD::TESTM";
20118 case X86ISD::TESTNM: return "X86ISD::TESTNM";
20119 case X86ISD::KORTEST: return "X86ISD::KORTEST";
20120 case X86ISD::KTEST: return "X86ISD::KTEST";
20121 case X86ISD::PACKSS: return "X86ISD::PACKSS";
20122 case X86ISD::PACKUS: return "X86ISD::PACKUS";
20123 case X86ISD::PALIGNR: return "X86ISD::PALIGNR";
20124 case X86ISD::VALIGN: return "X86ISD::VALIGN";
20125 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
20126 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
20127 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
20128 case X86ISD::SHUFP: return "X86ISD::SHUFP";
20129 case X86ISD::SHUF128: return "X86ISD::SHUF128";
20130 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
20131 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
20132 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
20133 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
20134 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
20135 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
20136 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
20137 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
20138 case X86ISD::MOVSD: return "X86ISD::MOVSD";
20139 case X86ISD::MOVSS: return "X86ISD::MOVSS";
20140 case X86ISD::UNPCKL: return "X86ISD::UNPCKL";
20141 case X86ISD::UNPCKH: return "X86ISD::UNPCKH";
20142 case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
20143 case X86ISD::VBROADCASTM: return "X86ISD::VBROADCASTM";
20144 case X86ISD::SUBV_BROADCAST: return "X86ISD::SUBV_BROADCAST";
20145 case X86ISD::VEXTRACT: return "X86ISD::VEXTRACT";
20146 case X86ISD::VPERMILPV: return "X86ISD::VPERMILPV";
20147 case X86ISD::VPERMILPI: return "X86ISD::VPERMILPI";
20148 case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128";
20149 case X86ISD::VPERMV: return "X86ISD::VPERMV";
20150 case X86ISD::VPERMV3: return "X86ISD::VPERMV3";
20151 case X86ISD::VPERMIV3: return "X86ISD::VPERMIV3";
20152 case X86ISD::VPERMI: return "X86ISD::VPERMI";
20153 case X86ISD::VPTERNLOG: return "X86ISD::VPTERNLOG";
20154 case X86ISD::VFIXUPIMM: return "X86ISD::VFIXUPIMM";
20155 case X86ISD::VRANGE: return "X86ISD::VRANGE";
20156 case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
20157 case X86ISD::PMULDQ: return "X86ISD::PMULDQ";
20158 case X86ISD::PSADBW: return "X86ISD::PSADBW";
20159 case X86ISD::DBPSADBW: return "X86ISD::DBPSADBW";
20160 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
20161 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
20162 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
20163 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
20164 case X86ISD::MFENCE: return "X86ISD::MFENCE";
20165 case X86ISD::SFENCE: return "X86ISD::SFENCE";
20166 case X86ISD::LFENCE: return "X86ISD::LFENCE";
20167 case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA";
20168 case X86ISD::SAHF: return "X86ISD::SAHF";
20169 case X86ISD::RDRAND: return "X86ISD::RDRAND";
20170 case X86ISD::RDSEED: return "X86ISD::RDSEED";
20171 case X86ISD::VPMADDUBSW: return "X86ISD::VPMADDUBSW";
20172 case X86ISD::VPMADDWD: return "X86ISD::VPMADDWD";
20173 case X86ISD::VPROT: return "X86ISD::VPROT";
20174 case X86ISD::VPROTI: return "X86ISD::VPROTI";
20175 case X86ISD::VPSHA: return "X86ISD::VPSHA";
20176 case X86ISD::VPSHL: return "X86ISD::VPSHL";
20177 case X86ISD::VPCOM: return "X86ISD::VPCOM";
20178 case X86ISD::VPCOMU: return "X86ISD::VPCOMU";
20179 case X86ISD::FMADD: return "X86ISD::FMADD";
20180 case X86ISD::FMSUB: return "X86ISD::FMSUB";
20181 case X86ISD::FNMADD: return "X86ISD::FNMADD";
20182 case X86ISD::FNMSUB: return "X86ISD::FNMSUB";
20183 case X86ISD::FMADDSUB: return "X86ISD::FMADDSUB";
20184 case X86ISD::FMSUBADD: return "X86ISD::FMSUBADD";
20185 case X86ISD::FMADD_RND: return "X86ISD::FMADD_RND";
20186 case X86ISD::FNMADD_RND: return "X86ISD::FNMADD_RND";
20187 case X86ISD::FMSUB_RND: return "X86ISD::FMSUB_RND";
20188 case X86ISD::FNMSUB_RND: return "X86ISD::FNMSUB_RND";
20189 case X86ISD::FMADDSUB_RND: return "X86ISD::FMADDSUB_RND";
20190 case X86ISD::FMSUBADD_RND: return "X86ISD::FMSUBADD_RND";
20191 case X86ISD::VRNDSCALE: return "X86ISD::VRNDSCALE";
20192 case X86ISD::VREDUCE: return "X86ISD::VREDUCE";
20193 case X86ISD::VGETMANT: return "X86ISD::VGETMANT";
20194 case X86ISD::PCMPESTRI: return "X86ISD::PCMPESTRI";
20195 case X86ISD::PCMPISTRI: return "X86ISD::PCMPISTRI";
20196 case X86ISD::XTEST: return "X86ISD::XTEST";
20197 case X86ISD::COMPRESS: return "X86ISD::COMPRESS";
20198 case X86ISD::EXPAND: return "X86ISD::EXPAND";
20199 case X86ISD::SELECT: return "X86ISD::SELECT";
20200 case X86ISD::ADDSUB: return "X86ISD::ADDSUB";
20201 case X86ISD::RCP28: return "X86ISD::RCP28";
20202 case X86ISD::EXP2: return "X86ISD::EXP2";
20203 case X86ISD::RSQRT28: return "X86ISD::RSQRT28";
20204 case X86ISD::FADD_RND: return "X86ISD::FADD_RND";
20205 case X86ISD::FSUB_RND: return "X86ISD::FSUB_RND";
20206 case X86ISD::FMUL_RND: return "X86ISD::FMUL_RND";
20207 case X86ISD::FDIV_RND: return "X86ISD::FDIV_RND";
20208 case X86ISD::FSQRT_RND: return "X86ISD::FSQRT_RND";
20209 case X86ISD::FGETEXP_RND: return "X86ISD::FGETEXP_RND";
20210 case X86ISD::SCALEF: return "X86ISD::SCALEF";
20211 case X86ISD::ADDS: return "X86ISD::ADDS";
20212 case X86ISD::SUBS: return "X86ISD::SUBS";
20213 case X86ISD::AVG: return "X86ISD::AVG";
20214 case X86ISD::MULHRS: return "X86ISD::MULHRS";
20215 case X86ISD::SINT_TO_FP_RND: return "X86ISD::SINT_TO_FP_RND";
20216 case X86ISD::UINT_TO_FP_RND: return "X86ISD::UINT_TO_FP_RND";
20217 case X86ISD::FP_TO_SINT_RND: return "X86ISD::FP_TO_SINT_RND";
20218 case X86ISD::FP_TO_UINT_RND: return "X86ISD::FP_TO_UINT_RND";
20219 case X86ISD::VFPCLASS: return "X86ISD::VFPCLASS";
20224 // isLegalAddressingMode - Return true if the addressing mode represented
20225 // by AM is legal for this target, for a load/store of the specified type.
20226 bool X86TargetLowering::isLegalAddressingMode(const DataLayout &DL,
20227 const AddrMode &AM, Type *Ty,
20228 unsigned AS) const {
20229 // X86 supports extremely general addressing modes.
20230 CodeModel::Model M = getTargetMachine().getCodeModel();
20231 Reloc::Model R = getTargetMachine().getRelocationModel();
20233 // X86 allows a sign-extended 32-bit immediate field as a displacement.
20234 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != nullptr))
20239 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
20241 // If a reference to this global requires an extra load, we can't fold it.
20242 if (isGlobalStubReference(GVFlags))
20245 // If BaseGV requires a register for the PIC base, we cannot also have a
20246 // BaseReg specified.
20247 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
20250 // If lower 4G is not available, then we must use rip-relative addressing.
20251 if ((M != CodeModel::Small || R != Reloc::Static) &&
20252 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
20256 switch (AM.Scale) {
20262 // These scales always work.
20267 // These scales are formed with basereg+scalereg. Only accept if there is
20272 default: // Other stuff never works.
20279 bool X86TargetLowering::isVectorShiftByScalarCheap(Type *Ty) const {
20280 unsigned Bits = Ty->getScalarSizeInBits();
20282 // 8-bit shifts are always expensive, but versions with a scalar amount aren't
20283 // particularly cheaper than those without.
20287 // On AVX2 there are new vpsllv[dq] instructions (and other shifts), that make
20288 // variable shifts just as cheap as scalar ones.
20289 if (Subtarget->hasInt256() && (Bits == 32 || Bits == 64))
20292 // Otherwise, it's significantly cheaper to shift by a scalar amount than by a
20293 // fully general vector.
20297 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
20298 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
20300 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
20301 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
20302 return NumBits1 > NumBits2;
20305 bool X86TargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
20306 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
20309 if (!isTypeLegal(EVT::getEVT(Ty1)))
20312 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
20314 // Assuming the caller doesn't have a zeroext or signext return parameter,
20315 // truncation all the way down to i1 is valid.
20319 bool X86TargetLowering::isLegalICmpImmediate(int64_t Imm) const {
20320 return isInt<32>(Imm);
20323 bool X86TargetLowering::isLegalAddImmediate(int64_t Imm) const {
20324 // Can also use sub to handle negated immediates.
20325 return isInt<32>(Imm);
20328 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
20329 if (!VT1.isInteger() || !VT2.isInteger())
20331 unsigned NumBits1 = VT1.getSizeInBits();
20332 unsigned NumBits2 = VT2.getSizeInBits();
20333 return NumBits1 > NumBits2;
20336 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
20337 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
20338 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
20341 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
20342 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
20343 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
20346 bool X86TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
20347 EVT VT1 = Val.getValueType();
20348 if (isZExtFree(VT1, VT2))
20351 if (Val.getOpcode() != ISD::LOAD)
20354 if (!VT1.isSimple() || !VT1.isInteger() ||
20355 !VT2.isSimple() || !VT2.isInteger())
20358 switch (VT1.getSimpleVT().SimpleTy) {
20363 // X86 has 8, 16, and 32-bit zero-extending loads.
20370 bool X86TargetLowering::isVectorLoadExtDesirable(SDValue) const { return true; }
20373 X86TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
20374 if (!(Subtarget->hasFMA() || Subtarget->hasFMA4() || Subtarget->hasAVX512()))
20377 VT = VT.getScalarType();
20379 if (!VT.isSimple())
20382 switch (VT.getSimpleVT().SimpleTy) {
20393 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
20394 // i16 instructions are longer (0x66 prefix) and potentially slower.
20395 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
20398 /// isShuffleMaskLegal - Targets can use this to indicate that they only
20399 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
20400 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
20401 /// are assumed to be legal.
20403 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
20405 if (!VT.isSimple())
20408 // Not for i1 vectors
20409 if (VT.getSimpleVT().getScalarType() == MVT::i1)
20412 // Very little shuffling can be done for 64-bit vectors right now.
20413 if (VT.getSimpleVT().getSizeInBits() == 64)
20416 // We only care that the types being shuffled are legal. The lowering can
20417 // handle any possible shuffle mask that results.
20418 return isTypeLegal(VT.getSimpleVT());
20422 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
20424 // Just delegate to the generic legality, clear masks aren't special.
20425 return isShuffleMaskLegal(Mask, VT);
20428 //===----------------------------------------------------------------------===//
20429 // X86 Scheduler Hooks
20430 //===----------------------------------------------------------------------===//
20432 /// Utility function to emit xbegin specifying the start of an RTM region.
20433 static MachineBasicBlock *EmitXBegin(MachineInstr *MI, MachineBasicBlock *MBB,
20434 const TargetInstrInfo *TII) {
20435 DebugLoc DL = MI->getDebugLoc();
20437 const BasicBlock *BB = MBB->getBasicBlock();
20438 MachineFunction::iterator I = ++MBB->getIterator();
20440 // For the v = xbegin(), we generate
20451 MachineBasicBlock *thisMBB = MBB;
20452 MachineFunction *MF = MBB->getParent();
20453 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
20454 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
20455 MF->insert(I, mainMBB);
20456 MF->insert(I, sinkMBB);
20458 // Transfer the remainder of BB and its successor edges to sinkMBB.
20459 sinkMBB->splice(sinkMBB->begin(), MBB,
20460 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
20461 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
20465 // # fallthrough to mainMBB
20466 // # abortion to sinkMBB
20467 BuildMI(thisMBB, DL, TII->get(X86::XBEGIN_4)).addMBB(sinkMBB);
20468 thisMBB->addSuccessor(mainMBB);
20469 thisMBB->addSuccessor(sinkMBB);
20473 BuildMI(mainMBB, DL, TII->get(X86::MOV32ri), X86::EAX).addImm(-1);
20474 mainMBB->addSuccessor(sinkMBB);
20477 // EAX is live into the sinkMBB
20478 sinkMBB->addLiveIn(X86::EAX);
20479 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
20480 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20483 MI->eraseFromParent();
20487 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
20488 // or XMM0_V32I8 in AVX all of this code can be replaced with that
20489 // in the .td file.
20490 static MachineBasicBlock *EmitPCMPSTRM(MachineInstr *MI, MachineBasicBlock *BB,
20491 const TargetInstrInfo *TII) {
20493 switch (MI->getOpcode()) {
20494 default: llvm_unreachable("illegal opcode!");
20495 case X86::PCMPISTRM128REG: Opc = X86::PCMPISTRM128rr; break;
20496 case X86::VPCMPISTRM128REG: Opc = X86::VPCMPISTRM128rr; break;
20497 case X86::PCMPISTRM128MEM: Opc = X86::PCMPISTRM128rm; break;
20498 case X86::VPCMPISTRM128MEM: Opc = X86::VPCMPISTRM128rm; break;
20499 case X86::PCMPESTRM128REG: Opc = X86::PCMPESTRM128rr; break;
20500 case X86::VPCMPESTRM128REG: Opc = X86::VPCMPESTRM128rr; break;
20501 case X86::PCMPESTRM128MEM: Opc = X86::PCMPESTRM128rm; break;
20502 case X86::VPCMPESTRM128MEM: Opc = X86::VPCMPESTRM128rm; break;
20505 DebugLoc dl = MI->getDebugLoc();
20506 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
20508 unsigned NumArgs = MI->getNumOperands();
20509 for (unsigned i = 1; i < NumArgs; ++i) {
20510 MachineOperand &Op = MI->getOperand(i);
20511 if (!(Op.isReg() && Op.isImplicit()))
20512 MIB.addOperand(Op);
20514 if (MI->hasOneMemOperand())
20515 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
20517 BuildMI(*BB, MI, dl,
20518 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20519 .addReg(X86::XMM0);
20521 MI->eraseFromParent();
20525 // FIXME: Custom handling because TableGen doesn't support multiple implicit
20526 // defs in an instruction pattern
20527 static MachineBasicBlock *EmitPCMPSTRI(MachineInstr *MI, MachineBasicBlock *BB,
20528 const TargetInstrInfo *TII) {
20530 switch (MI->getOpcode()) {
20531 default: llvm_unreachable("illegal opcode!");
20532 case X86::PCMPISTRIREG: Opc = X86::PCMPISTRIrr; break;
20533 case X86::VPCMPISTRIREG: Opc = X86::VPCMPISTRIrr; break;
20534 case X86::PCMPISTRIMEM: Opc = X86::PCMPISTRIrm; break;
20535 case X86::VPCMPISTRIMEM: Opc = X86::VPCMPISTRIrm; break;
20536 case X86::PCMPESTRIREG: Opc = X86::PCMPESTRIrr; break;
20537 case X86::VPCMPESTRIREG: Opc = X86::VPCMPESTRIrr; break;
20538 case X86::PCMPESTRIMEM: Opc = X86::PCMPESTRIrm; break;
20539 case X86::VPCMPESTRIMEM: Opc = X86::VPCMPESTRIrm; break;
20542 DebugLoc dl = MI->getDebugLoc();
20543 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
20545 unsigned NumArgs = MI->getNumOperands(); // remove the results
20546 for (unsigned i = 1; i < NumArgs; ++i) {
20547 MachineOperand &Op = MI->getOperand(i);
20548 if (!(Op.isReg() && Op.isImplicit()))
20549 MIB.addOperand(Op);
20551 if (MI->hasOneMemOperand())
20552 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
20554 BuildMI(*BB, MI, dl,
20555 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20558 MI->eraseFromParent();
20562 static MachineBasicBlock *EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB,
20563 const X86Subtarget *Subtarget) {
20564 DebugLoc dl = MI->getDebugLoc();
20565 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20566 // Address into RAX/EAX, other two args into ECX, EDX.
20567 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
20568 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
20569 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
20570 for (int i = 0; i < X86::AddrNumOperands; ++i)
20571 MIB.addOperand(MI->getOperand(i));
20573 unsigned ValOps = X86::AddrNumOperands;
20574 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
20575 .addReg(MI->getOperand(ValOps).getReg());
20576 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
20577 .addReg(MI->getOperand(ValOps+1).getReg());
20579 // The instruction doesn't actually take any operands though.
20580 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
20582 MI->eraseFromParent(); // The pseudo is gone now.
20586 MachineBasicBlock *
20587 X86TargetLowering::EmitVAARG64WithCustomInserter(MachineInstr *MI,
20588 MachineBasicBlock *MBB) const {
20589 // Emit va_arg instruction on X86-64.
20591 // Operands to this pseudo-instruction:
20592 // 0 ) Output : destination address (reg)
20593 // 1-5) Input : va_list address (addr, i64mem)
20594 // 6 ) ArgSize : Size (in bytes) of vararg type
20595 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
20596 // 8 ) Align : Alignment of type
20597 // 9 ) EFLAGS (implicit-def)
20599 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
20600 static_assert(X86::AddrNumOperands == 5,
20601 "VAARG_64 assumes 5 address operands");
20603 unsigned DestReg = MI->getOperand(0).getReg();
20604 MachineOperand &Base = MI->getOperand(1);
20605 MachineOperand &Scale = MI->getOperand(2);
20606 MachineOperand &Index = MI->getOperand(3);
20607 MachineOperand &Disp = MI->getOperand(4);
20608 MachineOperand &Segment = MI->getOperand(5);
20609 unsigned ArgSize = MI->getOperand(6).getImm();
20610 unsigned ArgMode = MI->getOperand(7).getImm();
20611 unsigned Align = MI->getOperand(8).getImm();
20613 // Memory Reference
20614 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
20615 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
20616 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
20618 // Machine Information
20619 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20620 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
20621 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
20622 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
20623 DebugLoc DL = MI->getDebugLoc();
20625 // struct va_list {
20628 // i64 overflow_area (address)
20629 // i64 reg_save_area (address)
20631 // sizeof(va_list) = 24
20632 // alignment(va_list) = 8
20634 unsigned TotalNumIntRegs = 6;
20635 unsigned TotalNumXMMRegs = 8;
20636 bool UseGPOffset = (ArgMode == 1);
20637 bool UseFPOffset = (ArgMode == 2);
20638 unsigned MaxOffset = TotalNumIntRegs * 8 +
20639 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
20641 /* Align ArgSize to a multiple of 8 */
20642 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
20643 bool NeedsAlign = (Align > 8);
20645 MachineBasicBlock *thisMBB = MBB;
20646 MachineBasicBlock *overflowMBB;
20647 MachineBasicBlock *offsetMBB;
20648 MachineBasicBlock *endMBB;
20650 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
20651 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
20652 unsigned OffsetReg = 0;
20654 if (!UseGPOffset && !UseFPOffset) {
20655 // If we only pull from the overflow region, we don't create a branch.
20656 // We don't need to alter control flow.
20657 OffsetDestReg = 0; // unused
20658 OverflowDestReg = DestReg;
20660 offsetMBB = nullptr;
20661 overflowMBB = thisMBB;
20664 // First emit code to check if gp_offset (or fp_offset) is below the bound.
20665 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
20666 // If not, pull from overflow_area. (branch to overflowMBB)
20671 // offsetMBB overflowMBB
20676 // Registers for the PHI in endMBB
20677 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
20678 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
20680 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
20681 MachineFunction *MF = MBB->getParent();
20682 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20683 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20684 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20686 MachineFunction::iterator MBBIter = ++MBB->getIterator();
20688 // Insert the new basic blocks
20689 MF->insert(MBBIter, offsetMBB);
20690 MF->insert(MBBIter, overflowMBB);
20691 MF->insert(MBBIter, endMBB);
20693 // Transfer the remainder of MBB and its successor edges to endMBB.
20694 endMBB->splice(endMBB->begin(), thisMBB,
20695 std::next(MachineBasicBlock::iterator(MI)), thisMBB->end());
20696 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
20698 // Make offsetMBB and overflowMBB successors of thisMBB
20699 thisMBB->addSuccessor(offsetMBB);
20700 thisMBB->addSuccessor(overflowMBB);
20702 // endMBB is a successor of both offsetMBB and overflowMBB
20703 offsetMBB->addSuccessor(endMBB);
20704 overflowMBB->addSuccessor(endMBB);
20706 // Load the offset value into a register
20707 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
20708 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
20712 .addDisp(Disp, UseFPOffset ? 4 : 0)
20713 .addOperand(Segment)
20714 .setMemRefs(MMOBegin, MMOEnd);
20716 // Check if there is enough room left to pull this argument.
20717 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
20719 .addImm(MaxOffset + 8 - ArgSizeA8);
20721 // Branch to "overflowMBB" if offset >= max
20722 // Fall through to "offsetMBB" otherwise
20723 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
20724 .addMBB(overflowMBB);
20727 // In offsetMBB, emit code to use the reg_save_area.
20729 assert(OffsetReg != 0);
20731 // Read the reg_save_area address.
20732 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
20733 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
20738 .addOperand(Segment)
20739 .setMemRefs(MMOBegin, MMOEnd);
20741 // Zero-extend the offset
20742 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
20743 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
20746 .addImm(X86::sub_32bit);
20748 // Add the offset to the reg_save_area to get the final address.
20749 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
20750 .addReg(OffsetReg64)
20751 .addReg(RegSaveReg);
20753 // Compute the offset for the next argument
20754 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
20755 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
20757 .addImm(UseFPOffset ? 16 : 8);
20759 // Store it back into the va_list.
20760 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
20764 .addDisp(Disp, UseFPOffset ? 4 : 0)
20765 .addOperand(Segment)
20766 .addReg(NextOffsetReg)
20767 .setMemRefs(MMOBegin, MMOEnd);
20770 BuildMI(offsetMBB, DL, TII->get(X86::JMP_1))
20775 // Emit code to use overflow area
20778 // Load the overflow_area address into a register.
20779 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
20780 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
20785 .addOperand(Segment)
20786 .setMemRefs(MMOBegin, MMOEnd);
20788 // If we need to align it, do so. Otherwise, just copy the address
20789 // to OverflowDestReg.
20791 // Align the overflow address
20792 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
20793 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
20795 // aligned_addr = (addr + (align-1)) & ~(align-1)
20796 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
20797 .addReg(OverflowAddrReg)
20800 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
20802 .addImm(~(uint64_t)(Align-1));
20804 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
20805 .addReg(OverflowAddrReg);
20808 // Compute the next overflow address after this argument.
20809 // (the overflow address should be kept 8-byte aligned)
20810 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
20811 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
20812 .addReg(OverflowDestReg)
20813 .addImm(ArgSizeA8);
20815 // Store the new overflow address.
20816 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
20821 .addOperand(Segment)
20822 .addReg(NextAddrReg)
20823 .setMemRefs(MMOBegin, MMOEnd);
20825 // If we branched, emit the PHI to the front of endMBB.
20827 BuildMI(*endMBB, endMBB->begin(), DL,
20828 TII->get(X86::PHI), DestReg)
20829 .addReg(OffsetDestReg).addMBB(offsetMBB)
20830 .addReg(OverflowDestReg).addMBB(overflowMBB);
20833 // Erase the pseudo instruction
20834 MI->eraseFromParent();
20839 MachineBasicBlock *
20840 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
20842 MachineBasicBlock *MBB) const {
20843 // Emit code to save XMM registers to the stack. The ABI says that the
20844 // number of registers to save is given in %al, so it's theoretically
20845 // possible to do an indirect jump trick to avoid saving all of them,
20846 // however this code takes a simpler approach and just executes all
20847 // of the stores if %al is non-zero. It's less code, and it's probably
20848 // easier on the hardware branch predictor, and stores aren't all that
20849 // expensive anyway.
20851 // Create the new basic blocks. One block contains all the XMM stores,
20852 // and one block is the final destination regardless of whether any
20853 // stores were performed.
20854 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
20855 MachineFunction *F = MBB->getParent();
20856 MachineFunction::iterator MBBIter = ++MBB->getIterator();
20857 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
20858 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
20859 F->insert(MBBIter, XMMSaveMBB);
20860 F->insert(MBBIter, EndMBB);
20862 // Transfer the remainder of MBB and its successor edges to EndMBB.
20863 EndMBB->splice(EndMBB->begin(), MBB,
20864 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
20865 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
20867 // The original block will now fall through to the XMM save block.
20868 MBB->addSuccessor(XMMSaveMBB);
20869 // The XMMSaveMBB will fall through to the end block.
20870 XMMSaveMBB->addSuccessor(EndMBB);
20872 // Now add the instructions.
20873 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20874 DebugLoc DL = MI->getDebugLoc();
20876 unsigned CountReg = MI->getOperand(0).getReg();
20877 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
20878 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
20880 if (!Subtarget->isCallingConvWin64(F->getFunction()->getCallingConv())) {
20881 // If %al is 0, branch around the XMM save block.
20882 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
20883 BuildMI(MBB, DL, TII->get(X86::JE_1)).addMBB(EndMBB);
20884 MBB->addSuccessor(EndMBB);
20887 // Make sure the last operand is EFLAGS, which gets clobbered by the branch
20888 // that was just emitted, but clearly shouldn't be "saved".
20889 assert((MI->getNumOperands() <= 3 ||
20890 !MI->getOperand(MI->getNumOperands() - 1).isReg() ||
20891 MI->getOperand(MI->getNumOperands() - 1).getReg() == X86::EFLAGS)
20892 && "Expected last argument to be EFLAGS");
20893 unsigned MOVOpc = Subtarget->hasFp256() ? X86::VMOVAPSmr : X86::MOVAPSmr;
20894 // In the XMM save block, save all the XMM argument registers.
20895 for (int i = 3, e = MI->getNumOperands() - 1; i != e; ++i) {
20896 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
20897 MachineMemOperand *MMO = F->getMachineMemOperand(
20898 MachinePointerInfo::getFixedStack(*F, RegSaveFrameIndex, Offset),
20899 MachineMemOperand::MOStore,
20900 /*Size=*/16, /*Align=*/16);
20901 BuildMI(XMMSaveMBB, DL, TII->get(MOVOpc))
20902 .addFrameIndex(RegSaveFrameIndex)
20903 .addImm(/*Scale=*/1)
20904 .addReg(/*IndexReg=*/0)
20905 .addImm(/*Disp=*/Offset)
20906 .addReg(/*Segment=*/0)
20907 .addReg(MI->getOperand(i).getReg())
20908 .addMemOperand(MMO);
20911 MI->eraseFromParent(); // The pseudo instruction is gone now.
20916 // The EFLAGS operand of SelectItr might be missing a kill marker
20917 // because there were multiple uses of EFLAGS, and ISel didn't know
20918 // which to mark. Figure out whether SelectItr should have had a
20919 // kill marker, and set it if it should. Returns the correct kill
20921 static bool checkAndUpdateEFLAGSKill(MachineBasicBlock::iterator SelectItr,
20922 MachineBasicBlock* BB,
20923 const TargetRegisterInfo* TRI) {
20924 // Scan forward through BB for a use/def of EFLAGS.
20925 MachineBasicBlock::iterator miI(std::next(SelectItr));
20926 for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) {
20927 const MachineInstr& mi = *miI;
20928 if (mi.readsRegister(X86::EFLAGS))
20930 if (mi.definesRegister(X86::EFLAGS))
20931 break; // Should have kill-flag - update below.
20934 // If we hit the end of the block, check whether EFLAGS is live into a
20936 if (miI == BB->end()) {
20937 for (MachineBasicBlock::succ_iterator sItr = BB->succ_begin(),
20938 sEnd = BB->succ_end();
20939 sItr != sEnd; ++sItr) {
20940 MachineBasicBlock* succ = *sItr;
20941 if (succ->isLiveIn(X86::EFLAGS))
20946 // We found a def, or hit the end of the basic block and EFLAGS wasn't live
20947 // out. SelectMI should have a kill flag on EFLAGS.
20948 SelectItr->addRegisterKilled(X86::EFLAGS, TRI);
20952 // Return true if it is OK for this CMOV pseudo-opcode to be cascaded
20953 // together with other CMOV pseudo-opcodes into a single basic-block with
20954 // conditional jump around it.
20955 static bool isCMOVPseudo(MachineInstr *MI) {
20956 switch (MI->getOpcode()) {
20957 case X86::CMOV_FR32:
20958 case X86::CMOV_FR64:
20959 case X86::CMOV_GR8:
20960 case X86::CMOV_GR16:
20961 case X86::CMOV_GR32:
20962 case X86::CMOV_RFP32:
20963 case X86::CMOV_RFP64:
20964 case X86::CMOV_RFP80:
20965 case X86::CMOV_V2F64:
20966 case X86::CMOV_V2I64:
20967 case X86::CMOV_V4F32:
20968 case X86::CMOV_V4F64:
20969 case X86::CMOV_V4I64:
20970 case X86::CMOV_V16F32:
20971 case X86::CMOV_V8F32:
20972 case X86::CMOV_V8F64:
20973 case X86::CMOV_V8I64:
20974 case X86::CMOV_V8I1:
20975 case X86::CMOV_V16I1:
20976 case X86::CMOV_V32I1:
20977 case X86::CMOV_V64I1:
20985 MachineBasicBlock *
20986 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
20987 MachineBasicBlock *BB) const {
20988 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20989 DebugLoc DL = MI->getDebugLoc();
20991 // To "insert" a SELECT_CC instruction, we actually have to insert the
20992 // diamond control-flow pattern. The incoming instruction knows the
20993 // destination vreg to set, the condition code register to branch on, the
20994 // true/false values to select between, and a branch opcode to use.
20995 const BasicBlock *LLVM_BB = BB->getBasicBlock();
20996 MachineFunction::iterator It = ++BB->getIterator();
21001 // cmpTY ccX, r1, r2
21003 // fallthrough --> copy0MBB
21004 MachineBasicBlock *thisMBB = BB;
21005 MachineFunction *F = BB->getParent();
21007 // This code lowers all pseudo-CMOV instructions. Generally it lowers these
21008 // as described above, by inserting a BB, and then making a PHI at the join
21009 // point to select the true and false operands of the CMOV in the PHI.
21011 // The code also handles two different cases of multiple CMOV opcodes
21015 // In this case, there are multiple CMOVs in a row, all which are based on
21016 // the same condition setting (or the exact opposite condition setting).
21017 // In this case we can lower all the CMOVs using a single inserted BB, and
21018 // then make a number of PHIs at the join point to model the CMOVs. The only
21019 // trickiness here, is that in a case like:
21021 // t2 = CMOV cond1 t1, f1
21022 // t3 = CMOV cond1 t2, f2
21024 // when rewriting this into PHIs, we have to perform some renaming on the
21025 // temps since you cannot have a PHI operand refer to a PHI result earlier
21026 // in the same block. The "simple" but wrong lowering would be:
21028 // t2 = PHI t1(BB1), f1(BB2)
21029 // t3 = PHI t2(BB1), f2(BB2)
21031 // but clearly t2 is not defined in BB1, so that is incorrect. The proper
21032 // renaming is to note that on the path through BB1, t2 is really just a
21033 // copy of t1, and do that renaming, properly generating:
21035 // t2 = PHI t1(BB1), f1(BB2)
21036 // t3 = PHI t1(BB1), f2(BB2)
21038 // Case 2, we lower cascaded CMOVs such as
21040 // (CMOV (CMOV F, T, cc1), T, cc2)
21042 // to two successives branches. For that, we look for another CMOV as the
21043 // following instruction.
21045 // Without this, we would add a PHI between the two jumps, which ends up
21046 // creating a few copies all around. For instance, for
21048 // (sitofp (zext (fcmp une)))
21050 // we would generate:
21052 // ucomiss %xmm1, %xmm0
21053 // movss <1.0f>, %xmm0
21054 // movaps %xmm0, %xmm1
21056 // xorps %xmm1, %xmm1
21059 // movaps %xmm1, %xmm0
21063 // because this custom-inserter would have generated:
21075 // A: X = ...; Y = ...
21077 // C: Z = PHI [X, A], [Y, B]
21079 // E: PHI [X, C], [Z, D]
21081 // If we lower both CMOVs in a single step, we can instead generate:
21093 // A: X = ...; Y = ...
21095 // E: PHI [X, A], [X, C], [Y, D]
21097 // Which, in our sitofp/fcmp example, gives us something like:
21099 // ucomiss %xmm1, %xmm0
21100 // movss <1.0f>, %xmm0
21103 // xorps %xmm0, %xmm0
21107 MachineInstr *CascadedCMOV = nullptr;
21108 MachineInstr *LastCMOV = MI;
21109 X86::CondCode CC = X86::CondCode(MI->getOperand(3).getImm());
21110 X86::CondCode OppCC = X86::GetOppositeBranchCondition(CC);
21111 MachineBasicBlock::iterator NextMIIt =
21112 std::next(MachineBasicBlock::iterator(MI));
21114 // Check for case 1, where there are multiple CMOVs with the same condition
21115 // first. Of the two cases of multiple CMOV lowerings, case 1 reduces the
21116 // number of jumps the most.
21118 if (isCMOVPseudo(MI)) {
21119 // See if we have a string of CMOVS with the same condition.
21120 while (NextMIIt != BB->end() &&
21121 isCMOVPseudo(NextMIIt) &&
21122 (NextMIIt->getOperand(3).getImm() == CC ||
21123 NextMIIt->getOperand(3).getImm() == OppCC)) {
21124 LastCMOV = &*NextMIIt;
21129 // This checks for case 2, but only do this if we didn't already find
21130 // case 1, as indicated by LastCMOV == MI.
21131 if (LastCMOV == MI &&
21132 NextMIIt != BB->end() && NextMIIt->getOpcode() == MI->getOpcode() &&
21133 NextMIIt->getOperand(2).getReg() == MI->getOperand(2).getReg() &&
21134 NextMIIt->getOperand(1).getReg() == MI->getOperand(0).getReg()) {
21135 CascadedCMOV = &*NextMIIt;
21138 MachineBasicBlock *jcc1MBB = nullptr;
21140 // If we have a cascaded CMOV, we lower it to two successive branches to
21141 // the same block. EFLAGS is used by both, so mark it as live in the second.
21142 if (CascadedCMOV) {
21143 jcc1MBB = F->CreateMachineBasicBlock(LLVM_BB);
21144 F->insert(It, jcc1MBB);
21145 jcc1MBB->addLiveIn(X86::EFLAGS);
21148 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
21149 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
21150 F->insert(It, copy0MBB);
21151 F->insert(It, sinkMBB);
21153 // If the EFLAGS register isn't dead in the terminator, then claim that it's
21154 // live into the sink and copy blocks.
21155 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
21157 MachineInstr *LastEFLAGSUser = CascadedCMOV ? CascadedCMOV : LastCMOV;
21158 if (!LastEFLAGSUser->killsRegister(X86::EFLAGS) &&
21159 !checkAndUpdateEFLAGSKill(LastEFLAGSUser, BB, TRI)) {
21160 copy0MBB->addLiveIn(X86::EFLAGS);
21161 sinkMBB->addLiveIn(X86::EFLAGS);
21164 // Transfer the remainder of BB and its successor edges to sinkMBB.
21165 sinkMBB->splice(sinkMBB->begin(), BB,
21166 std::next(MachineBasicBlock::iterator(LastCMOV)), BB->end());
21167 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
21169 // Add the true and fallthrough blocks as its successors.
21170 if (CascadedCMOV) {
21171 // The fallthrough block may be jcc1MBB, if we have a cascaded CMOV.
21172 BB->addSuccessor(jcc1MBB);
21174 // In that case, jcc1MBB will itself fallthrough the copy0MBB, and
21175 // jump to the sinkMBB.
21176 jcc1MBB->addSuccessor(copy0MBB);
21177 jcc1MBB->addSuccessor(sinkMBB);
21179 BB->addSuccessor(copy0MBB);
21182 // The true block target of the first (or only) branch is always sinkMBB.
21183 BB->addSuccessor(sinkMBB);
21185 // Create the conditional branch instruction.
21186 unsigned Opc = X86::GetCondBranchFromCond(CC);
21187 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
21189 if (CascadedCMOV) {
21190 unsigned Opc2 = X86::GetCondBranchFromCond(
21191 (X86::CondCode)CascadedCMOV->getOperand(3).getImm());
21192 BuildMI(jcc1MBB, DL, TII->get(Opc2)).addMBB(sinkMBB);
21196 // %FalseValue = ...
21197 // # fallthrough to sinkMBB
21198 copy0MBB->addSuccessor(sinkMBB);
21201 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
21203 MachineBasicBlock::iterator MIItBegin = MachineBasicBlock::iterator(MI);
21204 MachineBasicBlock::iterator MIItEnd =
21205 std::next(MachineBasicBlock::iterator(LastCMOV));
21206 MachineBasicBlock::iterator SinkInsertionPoint = sinkMBB->begin();
21207 DenseMap<unsigned, std::pair<unsigned, unsigned>> RegRewriteTable;
21208 MachineInstrBuilder MIB;
21210 // As we are creating the PHIs, we have to be careful if there is more than
21211 // one. Later CMOVs may reference the results of earlier CMOVs, but later
21212 // PHIs have to reference the individual true/false inputs from earlier PHIs.
21213 // That also means that PHI construction must work forward from earlier to
21214 // later, and that the code must maintain a mapping from earlier PHI's
21215 // destination registers, and the registers that went into the PHI.
21217 for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd; ++MIIt) {
21218 unsigned DestReg = MIIt->getOperand(0).getReg();
21219 unsigned Op1Reg = MIIt->getOperand(1).getReg();
21220 unsigned Op2Reg = MIIt->getOperand(2).getReg();
21222 // If this CMOV we are generating is the opposite condition from
21223 // the jump we generated, then we have to swap the operands for the
21224 // PHI that is going to be generated.
21225 if (MIIt->getOperand(3).getImm() == OppCC)
21226 std::swap(Op1Reg, Op2Reg);
21228 if (RegRewriteTable.find(Op1Reg) != RegRewriteTable.end())
21229 Op1Reg = RegRewriteTable[Op1Reg].first;
21231 if (RegRewriteTable.find(Op2Reg) != RegRewriteTable.end())
21232 Op2Reg = RegRewriteTable[Op2Reg].second;
21234 MIB = BuildMI(*sinkMBB, SinkInsertionPoint, DL,
21235 TII->get(X86::PHI), DestReg)
21236 .addReg(Op1Reg).addMBB(copy0MBB)
21237 .addReg(Op2Reg).addMBB(thisMBB);
21239 // Add this PHI to the rewrite table.
21240 RegRewriteTable[DestReg] = std::make_pair(Op1Reg, Op2Reg);
21243 // If we have a cascaded CMOV, the second Jcc provides the same incoming
21244 // value as the first Jcc (the True operand of the SELECT_CC/CMOV nodes).
21245 if (CascadedCMOV) {
21246 MIB.addReg(MI->getOperand(2).getReg()).addMBB(jcc1MBB);
21247 // Copy the PHI result to the register defined by the second CMOV.
21248 BuildMI(*sinkMBB, std::next(MachineBasicBlock::iterator(MIB.getInstr())),
21249 DL, TII->get(TargetOpcode::COPY),
21250 CascadedCMOV->getOperand(0).getReg())
21251 .addReg(MI->getOperand(0).getReg());
21252 CascadedCMOV->eraseFromParent();
21255 // Now remove the CMOV(s).
21256 for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd; )
21257 (MIIt++)->eraseFromParent();
21262 MachineBasicBlock *
21263 X86TargetLowering::EmitLoweredAtomicFP(MachineInstr *MI,
21264 MachineBasicBlock *BB) const {
21265 // Combine the following atomic floating-point modification pattern:
21266 // a.store(reg OP a.load(acquire), release)
21267 // Transform them into:
21268 // OPss (%gpr), %xmm
21269 // movss %xmm, (%gpr)
21270 // Or sd equivalent for 64-bit operations.
21272 switch (MI->getOpcode()) {
21273 default: llvm_unreachable("unexpected instr type for EmitLoweredAtomicFP");
21274 case X86::RELEASE_FADD32mr: MOp = X86::MOVSSmr; FOp = X86::ADDSSrm; break;
21275 case X86::RELEASE_FADD64mr: MOp = X86::MOVSDmr; FOp = X86::ADDSDrm; break;
21277 const X86InstrInfo *TII = Subtarget->getInstrInfo();
21278 DebugLoc DL = MI->getDebugLoc();
21279 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
21280 MachineOperand MSrc = MI->getOperand(0);
21281 unsigned VSrc = MI->getOperand(5).getReg();
21282 const MachineOperand &Disp = MI->getOperand(3);
21283 MachineOperand ZeroDisp = MachineOperand::CreateImm(0);
21284 bool hasDisp = Disp.isGlobal() || Disp.isImm();
21285 if (hasDisp && MSrc.isReg())
21286 MSrc.setIsKill(false);
21287 MachineInstrBuilder MIM = BuildMI(*BB, MI, DL, TII->get(MOp))
21288 .addOperand(/*Base=*/MSrc)
21289 .addImm(/*Scale=*/1)
21290 .addReg(/*Index=*/0)
21291 .addDisp(hasDisp ? Disp : ZeroDisp, /*off=*/0)
21293 MachineInstr *MIO = BuildMI(*BB, (MachineInstr *)MIM, DL, TII->get(FOp),
21294 MRI.createVirtualRegister(MRI.getRegClass(VSrc)))
21296 .addOperand(/*Base=*/MSrc)
21297 .addImm(/*Scale=*/1)
21298 .addReg(/*Index=*/0)
21299 .addDisp(hasDisp ? Disp : ZeroDisp, /*off=*/0)
21300 .addReg(/*Segment=*/0);
21301 MIM.addReg(MIO->getOperand(0).getReg(), RegState::Kill);
21302 MI->eraseFromParent(); // The pseudo instruction is gone now.
21306 MachineBasicBlock *
21307 X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI,
21308 MachineBasicBlock *BB) const {
21309 MachineFunction *MF = BB->getParent();
21310 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21311 DebugLoc DL = MI->getDebugLoc();
21312 const BasicBlock *LLVM_BB = BB->getBasicBlock();
21314 assert(MF->shouldSplitStack());
21316 const bool Is64Bit = Subtarget->is64Bit();
21317 const bool IsLP64 = Subtarget->isTarget64BitLP64();
21319 const unsigned TlsReg = Is64Bit ? X86::FS : X86::GS;
21320 const unsigned TlsOffset = IsLP64 ? 0x70 : Is64Bit ? 0x40 : 0x30;
21323 // ... [Till the alloca]
21324 // If stacklet is not large enough, jump to mallocMBB
21327 // Allocate by subtracting from RSP
21328 // Jump to continueMBB
21331 // Allocate by call to runtime
21335 // [rest of original BB]
21338 MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21339 MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21340 MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21342 MachineRegisterInfo &MRI = MF->getRegInfo();
21343 const TargetRegisterClass *AddrRegClass =
21344 getRegClassFor(getPointerTy(MF->getDataLayout()));
21346 unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass),
21347 bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass),
21348 tmpSPVReg = MRI.createVirtualRegister(AddrRegClass),
21349 SPLimitVReg = MRI.createVirtualRegister(AddrRegClass),
21350 sizeVReg = MI->getOperand(1).getReg(),
21351 physSPReg = IsLP64 || Subtarget->isTargetNaCl64() ? X86::RSP : X86::ESP;
21353 MachineFunction::iterator MBBIter = ++BB->getIterator();
21355 MF->insert(MBBIter, bumpMBB);
21356 MF->insert(MBBIter, mallocMBB);
21357 MF->insert(MBBIter, continueMBB);
21359 continueMBB->splice(continueMBB->begin(), BB,
21360 std::next(MachineBasicBlock::iterator(MI)), BB->end());
21361 continueMBB->transferSuccessorsAndUpdatePHIs(BB);
21363 // Add code to the main basic block to check if the stack limit has been hit,
21364 // and if so, jump to mallocMBB otherwise to bumpMBB.
21365 BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg);
21366 BuildMI(BB, DL, TII->get(IsLP64 ? X86::SUB64rr:X86::SUB32rr), SPLimitVReg)
21367 .addReg(tmpSPVReg).addReg(sizeVReg);
21368 BuildMI(BB, DL, TII->get(IsLP64 ? X86::CMP64mr:X86::CMP32mr))
21369 .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg)
21370 .addReg(SPLimitVReg);
21371 BuildMI(BB, DL, TII->get(X86::JG_1)).addMBB(mallocMBB);
21373 // bumpMBB simply decreases the stack pointer, since we know the current
21374 // stacklet has enough space.
21375 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg)
21376 .addReg(SPLimitVReg);
21377 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg)
21378 .addReg(SPLimitVReg);
21379 BuildMI(bumpMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
21381 // Calls into a routine in libgcc to allocate more space from the heap.
21382 const uint32_t *RegMask =
21383 Subtarget->getRegisterInfo()->getCallPreservedMask(*MF, CallingConv::C);
21385 BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
21387 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
21388 .addExternalSymbol("__morestack_allocate_stack_space")
21389 .addRegMask(RegMask)
21390 .addReg(X86::RDI, RegState::Implicit)
21391 .addReg(X86::RAX, RegState::ImplicitDefine);
21392 } else if (Is64Bit) {
21393 BuildMI(mallocMBB, DL, TII->get(X86::MOV32rr), X86::EDI)
21395 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
21396 .addExternalSymbol("__morestack_allocate_stack_space")
21397 .addRegMask(RegMask)
21398 .addReg(X86::EDI, RegState::Implicit)
21399 .addReg(X86::EAX, RegState::ImplicitDefine);
21401 BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg)
21403 BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg);
21404 BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32))
21405 .addExternalSymbol("__morestack_allocate_stack_space")
21406 .addRegMask(RegMask)
21407 .addReg(X86::EAX, RegState::ImplicitDefine);
21411 BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg)
21414 BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg)
21415 .addReg(IsLP64 ? X86::RAX : X86::EAX);
21416 BuildMI(mallocMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
21418 // Set up the CFG correctly.
21419 BB->addSuccessor(bumpMBB);
21420 BB->addSuccessor(mallocMBB);
21421 mallocMBB->addSuccessor(continueMBB);
21422 bumpMBB->addSuccessor(continueMBB);
21424 // Take care of the PHI nodes.
21425 BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI),
21426 MI->getOperand(0).getReg())
21427 .addReg(mallocPtrVReg).addMBB(mallocMBB)
21428 .addReg(bumpSPPtrVReg).addMBB(bumpMBB);
21430 // Delete the original pseudo instruction.
21431 MI->eraseFromParent();
21434 return continueMBB;
21437 MachineBasicBlock *
21438 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
21439 MachineBasicBlock *BB) const {
21440 assert(!Subtarget->isTargetMachO());
21441 DebugLoc DL = MI->getDebugLoc();
21442 MachineInstr *ResumeMI = Subtarget->getFrameLowering()->emitStackProbe(
21443 *BB->getParent(), *BB, MI, DL, false);
21444 MachineBasicBlock *ResumeBB = ResumeMI->getParent();
21445 MI->eraseFromParent(); // The pseudo instruction is gone now.
21449 MachineBasicBlock *
21450 X86TargetLowering::EmitLoweredCatchRet(MachineInstr *MI,
21451 MachineBasicBlock *BB) const {
21452 MachineFunction *MF = BB->getParent();
21453 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
21454 MachineBasicBlock *TargetMBB = MI->getOperand(0).getMBB();
21455 DebugLoc DL = MI->getDebugLoc();
21457 assert(!isAsynchronousEHPersonality(
21458 classifyEHPersonality(MF->getFunction()->getPersonalityFn())) &&
21459 "SEH does not use catchret!");
21461 // Only 32-bit EH needs to worry about manually restoring stack pointers.
21462 if (!Subtarget->is32Bit())
21465 // C++ EH creates a new target block to hold the restore code, and wires up
21466 // the new block to the return destination with a normal JMP_4.
21467 MachineBasicBlock *RestoreMBB =
21468 MF->CreateMachineBasicBlock(BB->getBasicBlock());
21469 assert(BB->succ_size() == 1);
21470 MF->insert(std::next(BB->getIterator()), RestoreMBB);
21471 RestoreMBB->transferSuccessorsAndUpdatePHIs(BB);
21472 BB->addSuccessor(RestoreMBB);
21473 MI->getOperand(0).setMBB(RestoreMBB);
21475 auto RestoreMBBI = RestoreMBB->begin();
21476 BuildMI(*RestoreMBB, RestoreMBBI, DL, TII.get(X86::EH_RESTORE));
21477 BuildMI(*RestoreMBB, RestoreMBBI, DL, TII.get(X86::JMP_4)).addMBB(TargetMBB);
21481 MachineBasicBlock *
21482 X86TargetLowering::EmitLoweredCatchPad(MachineInstr *MI,
21483 MachineBasicBlock *BB) const {
21484 MachineFunction *MF = BB->getParent();
21485 const Constant *PerFn = MF->getFunction()->getPersonalityFn();
21486 bool IsSEH = isAsynchronousEHPersonality(classifyEHPersonality(PerFn));
21487 // Only 32-bit SEH requires special handling for catchpad.
21488 if (IsSEH && Subtarget->is32Bit()) {
21489 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
21490 DebugLoc DL = MI->getDebugLoc();
21491 BuildMI(*BB, MI, DL, TII.get(X86::EH_RESTORE));
21493 MI->eraseFromParent();
21497 MachineBasicBlock *
21498 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
21499 MachineBasicBlock *BB) const {
21500 // This is pretty easy. We're taking the value that we received from
21501 // our load from the relocation, sticking it in either RDI (x86-64)
21502 // or EAX and doing an indirect call. The return value will then
21503 // be in the normal return register.
21504 MachineFunction *F = BB->getParent();
21505 const X86InstrInfo *TII = Subtarget->getInstrInfo();
21506 DebugLoc DL = MI->getDebugLoc();
21508 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
21509 assert(MI->getOperand(3).isGlobal() && "This should be a global");
21511 // Get a register mask for the lowered call.
21512 // FIXME: The 32-bit calls have non-standard calling conventions. Use a
21513 // proper register mask.
21514 const uint32_t *RegMask =
21515 Subtarget->is64Bit() ?
21516 Subtarget->getRegisterInfo()->getDarwinTLSCallPreservedMask() :
21517 Subtarget->getRegisterInfo()->getCallPreservedMask(*F, CallingConv::C);
21518 if (Subtarget->is64Bit()) {
21519 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
21520 TII->get(X86::MOV64rm), X86::RDI)
21522 .addImm(0).addReg(0)
21523 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
21524 MI->getOperand(3).getTargetFlags())
21526 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
21527 addDirectMem(MIB, X86::RDI);
21528 MIB.addReg(X86::RAX, RegState::ImplicitDefine).addRegMask(RegMask);
21529 } else if (F->getTarget().getRelocationModel() != Reloc::PIC_) {
21530 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
21531 TII->get(X86::MOV32rm), X86::EAX)
21533 .addImm(0).addReg(0)
21534 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
21535 MI->getOperand(3).getTargetFlags())
21537 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
21538 addDirectMem(MIB, X86::EAX);
21539 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
21541 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
21542 TII->get(X86::MOV32rm), X86::EAX)
21543 .addReg(TII->getGlobalBaseReg(F))
21544 .addImm(0).addReg(0)
21545 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
21546 MI->getOperand(3).getTargetFlags())
21548 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
21549 addDirectMem(MIB, X86::EAX);
21550 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
21553 MI->eraseFromParent(); // The pseudo instruction is gone now.
21557 MachineBasicBlock *
21558 X86TargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
21559 MachineBasicBlock *MBB) const {
21560 DebugLoc DL = MI->getDebugLoc();
21561 MachineFunction *MF = MBB->getParent();
21562 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21563 MachineRegisterInfo &MRI = MF->getRegInfo();
21565 const BasicBlock *BB = MBB->getBasicBlock();
21566 MachineFunction::iterator I = ++MBB->getIterator();
21568 // Memory Reference
21569 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
21570 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
21573 unsigned MemOpndSlot = 0;
21575 unsigned CurOp = 0;
21577 DstReg = MI->getOperand(CurOp++).getReg();
21578 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
21579 assert(RC->hasType(MVT::i32) && "Invalid destination!");
21580 unsigned mainDstReg = MRI.createVirtualRegister(RC);
21581 unsigned restoreDstReg = MRI.createVirtualRegister(RC);
21583 MemOpndSlot = CurOp;
21585 MVT PVT = getPointerTy(MF->getDataLayout());
21586 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
21587 "Invalid Pointer Size!");
21589 // For v = setjmp(buf), we generate
21592 // buf[LabelOffset] = restoreMBB <-- takes address of restoreMBB
21593 // SjLjSetup restoreMBB
21599 // v = phi(main, restore)
21602 // if base pointer being used, load it from frame
21605 MachineBasicBlock *thisMBB = MBB;
21606 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
21607 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
21608 MachineBasicBlock *restoreMBB = MF->CreateMachineBasicBlock(BB);
21609 MF->insert(I, mainMBB);
21610 MF->insert(I, sinkMBB);
21611 MF->push_back(restoreMBB);
21612 restoreMBB->setHasAddressTaken();
21614 MachineInstrBuilder MIB;
21616 // Transfer the remainder of BB and its successor edges to sinkMBB.
21617 sinkMBB->splice(sinkMBB->begin(), MBB,
21618 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
21619 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
21622 unsigned PtrStoreOpc = 0;
21623 unsigned LabelReg = 0;
21624 const int64_t LabelOffset = 1 * PVT.getStoreSize();
21625 Reloc::Model RM = MF->getTarget().getRelocationModel();
21626 bool UseImmLabel = (MF->getTarget().getCodeModel() == CodeModel::Small) &&
21627 (RM == Reloc::Static || RM == Reloc::DynamicNoPIC);
21629 // Prepare IP either in reg or imm.
21630 if (!UseImmLabel) {
21631 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr;
21632 const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
21633 LabelReg = MRI.createVirtualRegister(PtrRC);
21634 if (Subtarget->is64Bit()) {
21635 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA64r), LabelReg)
21639 .addMBB(restoreMBB)
21642 const X86InstrInfo *XII = static_cast<const X86InstrInfo*>(TII);
21643 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA32r), LabelReg)
21644 .addReg(XII->getGlobalBaseReg(MF))
21647 .addMBB(restoreMBB, Subtarget->ClassifyBlockAddressReference())
21651 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi;
21653 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PtrStoreOpc));
21654 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
21655 if (i == X86::AddrDisp)
21656 MIB.addDisp(MI->getOperand(MemOpndSlot + i), LabelOffset);
21658 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
21661 MIB.addReg(LabelReg);
21663 MIB.addMBB(restoreMBB);
21664 MIB.setMemRefs(MMOBegin, MMOEnd);
21666 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
21667 .addMBB(restoreMBB);
21669 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
21670 MIB.addRegMask(RegInfo->getNoPreservedMask());
21671 thisMBB->addSuccessor(mainMBB);
21672 thisMBB->addSuccessor(restoreMBB);
21676 BuildMI(mainMBB, DL, TII->get(X86::MOV32r0), mainDstReg);
21677 mainMBB->addSuccessor(sinkMBB);
21680 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
21681 TII->get(X86::PHI), DstReg)
21682 .addReg(mainDstReg).addMBB(mainMBB)
21683 .addReg(restoreDstReg).addMBB(restoreMBB);
21686 if (RegInfo->hasBasePointer(*MF)) {
21687 const bool Uses64BitFramePtr =
21688 Subtarget->isTarget64BitLP64() || Subtarget->isTargetNaCl64();
21689 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
21690 X86FI->setRestoreBasePointer(MF);
21691 unsigned FramePtr = RegInfo->getFrameRegister(*MF);
21692 unsigned BasePtr = RegInfo->getBaseRegister();
21693 unsigned Opm = Uses64BitFramePtr ? X86::MOV64rm : X86::MOV32rm;
21694 addRegOffset(BuildMI(restoreMBB, DL, TII->get(Opm), BasePtr),
21695 FramePtr, true, X86FI->getRestoreBasePointerOffset())
21696 .setMIFlag(MachineInstr::FrameSetup);
21698 BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1);
21699 BuildMI(restoreMBB, DL, TII->get(X86::JMP_1)).addMBB(sinkMBB);
21700 restoreMBB->addSuccessor(sinkMBB);
21702 MI->eraseFromParent();
21706 MachineBasicBlock *
21707 X86TargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
21708 MachineBasicBlock *MBB) const {
21709 DebugLoc DL = MI->getDebugLoc();
21710 MachineFunction *MF = MBB->getParent();
21711 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21712 MachineRegisterInfo &MRI = MF->getRegInfo();
21714 // Memory Reference
21715 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
21716 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
21718 MVT PVT = getPointerTy(MF->getDataLayout());
21719 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
21720 "Invalid Pointer Size!");
21722 const TargetRegisterClass *RC =
21723 (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass;
21724 unsigned Tmp = MRI.createVirtualRegister(RC);
21725 // Since FP is only updated here but NOT referenced, it's treated as GPR.
21726 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
21727 unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
21728 unsigned SP = RegInfo->getStackRegister();
21730 MachineInstrBuilder MIB;
21732 const int64_t LabelOffset = 1 * PVT.getStoreSize();
21733 const int64_t SPOffset = 2 * PVT.getStoreSize();
21735 unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm;
21736 unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r;
21739 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP);
21740 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
21741 MIB.addOperand(MI->getOperand(i));
21742 MIB.setMemRefs(MMOBegin, MMOEnd);
21744 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp);
21745 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
21746 if (i == X86::AddrDisp)
21747 MIB.addDisp(MI->getOperand(i), LabelOffset);
21749 MIB.addOperand(MI->getOperand(i));
21751 MIB.setMemRefs(MMOBegin, MMOEnd);
21753 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP);
21754 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
21755 if (i == X86::AddrDisp)
21756 MIB.addDisp(MI->getOperand(i), SPOffset);
21758 MIB.addOperand(MI->getOperand(i));
21760 MIB.setMemRefs(MMOBegin, MMOEnd);
21762 BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp);
21764 MI->eraseFromParent();
21768 // Replace 213-type (isel default) FMA3 instructions with 231-type for
21769 // accumulator loops. Writing back to the accumulator allows the coalescer
21770 // to remove extra copies in the loop.
21771 // FIXME: Do this on AVX512. We don't support 231 variants yet (PR23937).
21772 MachineBasicBlock *
21773 X86TargetLowering::emitFMA3Instr(MachineInstr *MI,
21774 MachineBasicBlock *MBB) const {
21775 MachineOperand &AddendOp = MI->getOperand(3);
21777 // Bail out early if the addend isn't a register - we can't switch these.
21778 if (!AddendOp.isReg())
21781 MachineFunction &MF = *MBB->getParent();
21782 MachineRegisterInfo &MRI = MF.getRegInfo();
21784 // Check whether the addend is defined by a PHI:
21785 assert(MRI.hasOneDef(AddendOp.getReg()) && "Multiple defs in SSA?");
21786 MachineInstr &AddendDef = *MRI.def_instr_begin(AddendOp.getReg());
21787 if (!AddendDef.isPHI())
21790 // Look for the following pattern:
21792 // %addend = phi [%entry, 0], [%loop, %result]
21794 // %result<tied1> = FMA213 %m2<tied0>, %m1, %addend
21798 // %addend = phi [%entry, 0], [%loop, %result]
21800 // %result<tied1> = FMA231 %addend<tied0>, %m1, %m2
21802 for (unsigned i = 1, e = AddendDef.getNumOperands(); i < e; i += 2) {
21803 assert(AddendDef.getOperand(i).isReg());
21804 MachineOperand PHISrcOp = AddendDef.getOperand(i);
21805 MachineInstr &PHISrcInst = *MRI.def_instr_begin(PHISrcOp.getReg());
21806 if (&PHISrcInst == MI) {
21807 // Found a matching instruction.
21808 unsigned NewFMAOpc = 0;
21809 switch (MI->getOpcode()) {
21810 case X86::VFMADDPDr213r: NewFMAOpc = X86::VFMADDPDr231r; break;
21811 case X86::VFMADDPSr213r: NewFMAOpc = X86::VFMADDPSr231r; break;
21812 case X86::VFMADDSDr213r: NewFMAOpc = X86::VFMADDSDr231r; break;
21813 case X86::VFMADDSSr213r: NewFMAOpc = X86::VFMADDSSr231r; break;
21814 case X86::VFMSUBPDr213r: NewFMAOpc = X86::VFMSUBPDr231r; break;
21815 case X86::VFMSUBPSr213r: NewFMAOpc = X86::VFMSUBPSr231r; break;
21816 case X86::VFMSUBSDr213r: NewFMAOpc = X86::VFMSUBSDr231r; break;
21817 case X86::VFMSUBSSr213r: NewFMAOpc = X86::VFMSUBSSr231r; break;
21818 case X86::VFNMADDPDr213r: NewFMAOpc = X86::VFNMADDPDr231r; break;
21819 case X86::VFNMADDPSr213r: NewFMAOpc = X86::VFNMADDPSr231r; break;
21820 case X86::VFNMADDSDr213r: NewFMAOpc = X86::VFNMADDSDr231r; break;
21821 case X86::VFNMADDSSr213r: NewFMAOpc = X86::VFNMADDSSr231r; break;
21822 case X86::VFNMSUBPDr213r: NewFMAOpc = X86::VFNMSUBPDr231r; break;
21823 case X86::VFNMSUBPSr213r: NewFMAOpc = X86::VFNMSUBPSr231r; break;
21824 case X86::VFNMSUBSDr213r: NewFMAOpc = X86::VFNMSUBSDr231r; break;
21825 case X86::VFNMSUBSSr213r: NewFMAOpc = X86::VFNMSUBSSr231r; break;
21826 case X86::VFMADDSUBPDr213r: NewFMAOpc = X86::VFMADDSUBPDr231r; break;
21827 case X86::VFMADDSUBPSr213r: NewFMAOpc = X86::VFMADDSUBPSr231r; break;
21828 case X86::VFMSUBADDPDr213r: NewFMAOpc = X86::VFMSUBADDPDr231r; break;
21829 case X86::VFMSUBADDPSr213r: NewFMAOpc = X86::VFMSUBADDPSr231r; break;
21831 case X86::VFMADDPDr213rY: NewFMAOpc = X86::VFMADDPDr231rY; break;
21832 case X86::VFMADDPSr213rY: NewFMAOpc = X86::VFMADDPSr231rY; break;
21833 case X86::VFMSUBPDr213rY: NewFMAOpc = X86::VFMSUBPDr231rY; break;
21834 case X86::VFMSUBPSr213rY: NewFMAOpc = X86::VFMSUBPSr231rY; break;
21835 case X86::VFNMADDPDr213rY: NewFMAOpc = X86::VFNMADDPDr231rY; break;
21836 case X86::VFNMADDPSr213rY: NewFMAOpc = X86::VFNMADDPSr231rY; break;
21837 case X86::VFNMSUBPDr213rY: NewFMAOpc = X86::VFNMSUBPDr231rY; break;
21838 case X86::VFNMSUBPSr213rY: NewFMAOpc = X86::VFNMSUBPSr231rY; break;
21839 case X86::VFMADDSUBPDr213rY: NewFMAOpc = X86::VFMADDSUBPDr231rY; break;
21840 case X86::VFMADDSUBPSr213rY: NewFMAOpc = X86::VFMADDSUBPSr231rY; break;
21841 case X86::VFMSUBADDPDr213rY: NewFMAOpc = X86::VFMSUBADDPDr231rY; break;
21842 case X86::VFMSUBADDPSr213rY: NewFMAOpc = X86::VFMSUBADDPSr231rY; break;
21843 default: llvm_unreachable("Unrecognized FMA variant.");
21846 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
21847 MachineInstrBuilder MIB =
21848 BuildMI(MF, MI->getDebugLoc(), TII.get(NewFMAOpc))
21849 .addOperand(MI->getOperand(0))
21850 .addOperand(MI->getOperand(3))
21851 .addOperand(MI->getOperand(2))
21852 .addOperand(MI->getOperand(1));
21853 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
21854 MI->eraseFromParent();
21861 MachineBasicBlock *
21862 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
21863 MachineBasicBlock *BB) const {
21864 switch (MI->getOpcode()) {
21865 default: llvm_unreachable("Unexpected instr type to insert");
21866 case X86::TAILJMPd64:
21867 case X86::TAILJMPr64:
21868 case X86::TAILJMPm64:
21869 case X86::TAILJMPd64_REX:
21870 case X86::TAILJMPr64_REX:
21871 case X86::TAILJMPm64_REX:
21872 llvm_unreachable("TAILJMP64 would not be touched here.");
21873 case X86::TCRETURNdi64:
21874 case X86::TCRETURNri64:
21875 case X86::TCRETURNmi64:
21877 case X86::WIN_ALLOCA:
21878 return EmitLoweredWinAlloca(MI, BB);
21879 case X86::CATCHRET:
21880 return EmitLoweredCatchRet(MI, BB);
21881 case X86::CATCHPAD:
21882 return EmitLoweredCatchPad(MI, BB);
21883 case X86::SEG_ALLOCA_32:
21884 case X86::SEG_ALLOCA_64:
21885 return EmitLoweredSegAlloca(MI, BB);
21886 case X86::TLSCall_32:
21887 case X86::TLSCall_64:
21888 return EmitLoweredTLSCall(MI, BB);
21889 case X86::CMOV_FR32:
21890 case X86::CMOV_FR64:
21891 case X86::CMOV_GR8:
21892 case X86::CMOV_GR16:
21893 case X86::CMOV_GR32:
21894 case X86::CMOV_RFP32:
21895 case X86::CMOV_RFP64:
21896 case X86::CMOV_RFP80:
21897 case X86::CMOV_V2F64:
21898 case X86::CMOV_V2I64:
21899 case X86::CMOV_V4F32:
21900 case X86::CMOV_V4F64:
21901 case X86::CMOV_V4I64:
21902 case X86::CMOV_V16F32:
21903 case X86::CMOV_V8F32:
21904 case X86::CMOV_V8F64:
21905 case X86::CMOV_V8I64:
21906 case X86::CMOV_V8I1:
21907 case X86::CMOV_V16I1:
21908 case X86::CMOV_V32I1:
21909 case X86::CMOV_V64I1:
21910 return EmitLoweredSelect(MI, BB);
21912 case X86::RELEASE_FADD32mr:
21913 case X86::RELEASE_FADD64mr:
21914 return EmitLoweredAtomicFP(MI, BB);
21916 case X86::FP32_TO_INT16_IN_MEM:
21917 case X86::FP32_TO_INT32_IN_MEM:
21918 case X86::FP32_TO_INT64_IN_MEM:
21919 case X86::FP64_TO_INT16_IN_MEM:
21920 case X86::FP64_TO_INT32_IN_MEM:
21921 case X86::FP64_TO_INT64_IN_MEM:
21922 case X86::FP80_TO_INT16_IN_MEM:
21923 case X86::FP80_TO_INT32_IN_MEM:
21924 case X86::FP80_TO_INT64_IN_MEM: {
21925 MachineFunction *F = BB->getParent();
21926 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21927 DebugLoc DL = MI->getDebugLoc();
21929 // Change the floating point control register to use "round towards zero"
21930 // mode when truncating to an integer value.
21931 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
21932 addFrameReference(BuildMI(*BB, MI, DL,
21933 TII->get(X86::FNSTCW16m)), CWFrameIdx);
21935 // Load the old value of the high byte of the control word...
21937 F->getRegInfo().createVirtualRegister(&X86::GR16RegClass);
21938 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
21941 // Set the high part to be round to zero...
21942 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
21945 // Reload the modified control word now...
21946 addFrameReference(BuildMI(*BB, MI, DL,
21947 TII->get(X86::FLDCW16m)), CWFrameIdx);
21949 // Restore the memory image of control word to original value
21950 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
21953 // Get the X86 opcode to use.
21955 switch (MI->getOpcode()) {
21956 default: llvm_unreachable("illegal opcode!");
21957 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
21958 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
21959 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
21960 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
21961 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
21962 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
21963 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
21964 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
21965 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
21969 MachineOperand &Op = MI->getOperand(0);
21971 AM.BaseType = X86AddressMode::RegBase;
21972 AM.Base.Reg = Op.getReg();
21974 AM.BaseType = X86AddressMode::FrameIndexBase;
21975 AM.Base.FrameIndex = Op.getIndex();
21977 Op = MI->getOperand(1);
21979 AM.Scale = Op.getImm();
21980 Op = MI->getOperand(2);
21982 AM.IndexReg = Op.getImm();
21983 Op = MI->getOperand(3);
21984 if (Op.isGlobal()) {
21985 AM.GV = Op.getGlobal();
21987 AM.Disp = Op.getImm();
21989 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
21990 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
21992 // Reload the original control word now.
21993 addFrameReference(BuildMI(*BB, MI, DL,
21994 TII->get(X86::FLDCW16m)), CWFrameIdx);
21996 MI->eraseFromParent(); // The pseudo instruction is gone now.
21999 // String/text processing lowering.
22000 case X86::PCMPISTRM128REG:
22001 case X86::VPCMPISTRM128REG:
22002 case X86::PCMPISTRM128MEM:
22003 case X86::VPCMPISTRM128MEM:
22004 case X86::PCMPESTRM128REG:
22005 case X86::VPCMPESTRM128REG:
22006 case X86::PCMPESTRM128MEM:
22007 case X86::VPCMPESTRM128MEM:
22008 assert(Subtarget->hasSSE42() &&
22009 "Target must have SSE4.2 or AVX features enabled");
22010 return EmitPCMPSTRM(MI, BB, Subtarget->getInstrInfo());
22012 // String/text processing lowering.
22013 case X86::PCMPISTRIREG:
22014 case X86::VPCMPISTRIREG:
22015 case X86::PCMPISTRIMEM:
22016 case X86::VPCMPISTRIMEM:
22017 case X86::PCMPESTRIREG:
22018 case X86::VPCMPESTRIREG:
22019 case X86::PCMPESTRIMEM:
22020 case X86::VPCMPESTRIMEM:
22021 assert(Subtarget->hasSSE42() &&
22022 "Target must have SSE4.2 or AVX features enabled");
22023 return EmitPCMPSTRI(MI, BB, Subtarget->getInstrInfo());
22025 // Thread synchronization.
22027 return EmitMonitor(MI, BB, Subtarget);
22031 return EmitXBegin(MI, BB, Subtarget->getInstrInfo());
22033 case X86::VASTART_SAVE_XMM_REGS:
22034 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
22036 case X86::VAARG_64:
22037 return EmitVAARG64WithCustomInserter(MI, BB);
22039 case X86::EH_SjLj_SetJmp32:
22040 case X86::EH_SjLj_SetJmp64:
22041 return emitEHSjLjSetJmp(MI, BB);
22043 case X86::EH_SjLj_LongJmp32:
22044 case X86::EH_SjLj_LongJmp64:
22045 return emitEHSjLjLongJmp(MI, BB);
22047 case TargetOpcode::STATEPOINT:
22048 // As an implementation detail, STATEPOINT shares the STACKMAP format at
22049 // this point in the process. We diverge later.
22050 return emitPatchPoint(MI, BB);
22052 case TargetOpcode::STACKMAP:
22053 case TargetOpcode::PATCHPOINT:
22054 return emitPatchPoint(MI, BB);
22056 case X86::VFMADDPDr213r:
22057 case X86::VFMADDPSr213r:
22058 case X86::VFMADDSDr213r:
22059 case X86::VFMADDSSr213r:
22060 case X86::VFMSUBPDr213r:
22061 case X86::VFMSUBPSr213r:
22062 case X86::VFMSUBSDr213r:
22063 case X86::VFMSUBSSr213r:
22064 case X86::VFNMADDPDr213r:
22065 case X86::VFNMADDPSr213r:
22066 case X86::VFNMADDSDr213r:
22067 case X86::VFNMADDSSr213r:
22068 case X86::VFNMSUBPDr213r:
22069 case X86::VFNMSUBPSr213r:
22070 case X86::VFNMSUBSDr213r:
22071 case X86::VFNMSUBSSr213r:
22072 case X86::VFMADDSUBPDr213r:
22073 case X86::VFMADDSUBPSr213r:
22074 case X86::VFMSUBADDPDr213r:
22075 case X86::VFMSUBADDPSr213r:
22076 case X86::VFMADDPDr213rY:
22077 case X86::VFMADDPSr213rY:
22078 case X86::VFMSUBPDr213rY:
22079 case X86::VFMSUBPSr213rY:
22080 case X86::VFNMADDPDr213rY:
22081 case X86::VFNMADDPSr213rY:
22082 case X86::VFNMSUBPDr213rY:
22083 case X86::VFNMSUBPSr213rY:
22084 case X86::VFMADDSUBPDr213rY:
22085 case X86::VFMADDSUBPSr213rY:
22086 case X86::VFMSUBADDPDr213rY:
22087 case X86::VFMSUBADDPSr213rY:
22088 return emitFMA3Instr(MI, BB);
22092 //===----------------------------------------------------------------------===//
22093 // X86 Optimization Hooks
22094 //===----------------------------------------------------------------------===//
22096 void X86TargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
22099 const SelectionDAG &DAG,
22100 unsigned Depth) const {
22101 unsigned BitWidth = KnownZero.getBitWidth();
22102 unsigned Opc = Op.getOpcode();
22103 assert((Opc >= ISD::BUILTIN_OP_END ||
22104 Opc == ISD::INTRINSIC_WO_CHAIN ||
22105 Opc == ISD::INTRINSIC_W_CHAIN ||
22106 Opc == ISD::INTRINSIC_VOID) &&
22107 "Should use MaskedValueIsZero if you don't know whether Op"
22108 " is a target node!");
22110 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
22124 // These nodes' second result is a boolean.
22125 if (Op.getResNo() == 0)
22128 case X86ISD::SETCC:
22129 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
22131 case ISD::INTRINSIC_WO_CHAIN: {
22132 unsigned IntId = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
22133 unsigned NumLoBits = 0;
22136 case Intrinsic::x86_sse_movmsk_ps:
22137 case Intrinsic::x86_avx_movmsk_ps_256:
22138 case Intrinsic::x86_sse2_movmsk_pd:
22139 case Intrinsic::x86_avx_movmsk_pd_256:
22140 case Intrinsic::x86_mmx_pmovmskb:
22141 case Intrinsic::x86_sse2_pmovmskb_128:
22142 case Intrinsic::x86_avx2_pmovmskb: {
22143 // High bits of movmskp{s|d}, pmovmskb are known zero.
22145 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
22146 case Intrinsic::x86_sse_movmsk_ps: NumLoBits = 4; break;
22147 case Intrinsic::x86_avx_movmsk_ps_256: NumLoBits = 8; break;
22148 case Intrinsic::x86_sse2_movmsk_pd: NumLoBits = 2; break;
22149 case Intrinsic::x86_avx_movmsk_pd_256: NumLoBits = 4; break;
22150 case Intrinsic::x86_mmx_pmovmskb: NumLoBits = 8; break;
22151 case Intrinsic::x86_sse2_pmovmskb_128: NumLoBits = 16; break;
22152 case Intrinsic::x86_avx2_pmovmskb: NumLoBits = 32; break;
22154 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - NumLoBits);
22163 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(
22165 const SelectionDAG &,
22166 unsigned Depth) const {
22167 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
22168 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
22169 return Op.getValueType().getScalarSizeInBits();
22175 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
22176 /// node is a GlobalAddress + offset.
22177 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
22178 const GlobalValue* &GA,
22179 int64_t &Offset) const {
22180 if (N->getOpcode() == X86ISD::Wrapper) {
22181 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
22182 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
22183 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
22187 return TargetLowering::isGAPlusOffset(N, GA, Offset);
22190 /// isShuffleHigh128VectorInsertLow - Checks whether the shuffle node is the
22191 /// same as extracting the high 128-bit part of 256-bit vector and then
22192 /// inserting the result into the low part of a new 256-bit vector
22193 static bool isShuffleHigh128VectorInsertLow(ShuffleVectorSDNode *SVOp) {
22194 EVT VT = SVOp->getValueType(0);
22195 unsigned NumElems = VT.getVectorNumElements();
22197 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
22198 for (unsigned i = 0, j = NumElems/2; i != NumElems/2; ++i, ++j)
22199 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
22200 SVOp->getMaskElt(j) >= 0)
22206 /// isShuffleLow128VectorInsertHigh - Checks whether the shuffle node is the
22207 /// same as extracting the low 128-bit part of 256-bit vector and then
22208 /// inserting the result into the high part of a new 256-bit vector
22209 static bool isShuffleLow128VectorInsertHigh(ShuffleVectorSDNode *SVOp) {
22210 EVT VT = SVOp->getValueType(0);
22211 unsigned NumElems = VT.getVectorNumElements();
22213 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
22214 for (unsigned i = NumElems/2, j = 0; i != NumElems; ++i, ++j)
22215 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
22216 SVOp->getMaskElt(j) >= 0)
22222 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
22223 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
22224 TargetLowering::DAGCombinerInfo &DCI,
22225 const X86Subtarget* Subtarget) {
22227 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
22228 SDValue V1 = SVOp->getOperand(0);
22229 SDValue V2 = SVOp->getOperand(1);
22230 EVT VT = SVOp->getValueType(0);
22231 unsigned NumElems = VT.getVectorNumElements();
22233 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
22234 V2.getOpcode() == ISD::CONCAT_VECTORS) {
22238 // V UNDEF BUILD_VECTOR UNDEF
22240 // CONCAT_VECTOR CONCAT_VECTOR
22243 // RESULT: V + zero extended
22245 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
22246 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
22247 V1.getOperand(1).getOpcode() != ISD::UNDEF)
22250 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
22253 // To match the shuffle mask, the first half of the mask should
22254 // be exactly the first vector, and all the rest a splat with the
22255 // first element of the second one.
22256 for (unsigned i = 0; i != NumElems/2; ++i)
22257 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
22258 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
22261 // If V1 is coming from a vector load then just fold to a VZEXT_LOAD.
22262 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(V1.getOperand(0))) {
22263 if (Ld->hasNUsesOfValue(1, 0)) {
22264 SDVTList Tys = DAG.getVTList(MVT::v4i64, MVT::Other);
22265 SDValue Ops[] = { Ld->getChain(), Ld->getBasePtr() };
22267 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops,
22269 Ld->getPointerInfo(),
22270 Ld->getAlignment(),
22271 false/*isVolatile*/, true/*ReadMem*/,
22272 false/*WriteMem*/);
22274 // Make sure the newly-created LOAD is in the same position as Ld in
22275 // terms of dependency. We create a TokenFactor for Ld and ResNode,
22276 // and update uses of Ld's output chain to use the TokenFactor.
22277 if (Ld->hasAnyUseOfValue(1)) {
22278 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
22279 SDValue(Ld, 1), SDValue(ResNode.getNode(), 1));
22280 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), NewChain);
22281 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(Ld, 1),
22282 SDValue(ResNode.getNode(), 1));
22285 return DAG.getBitcast(VT, ResNode);
22289 // Emit a zeroed vector and insert the desired subvector on its
22291 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
22292 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0), 0, DAG, dl);
22293 return DCI.CombineTo(N, InsV);
22296 //===--------------------------------------------------------------------===//
22297 // Combine some shuffles into subvector extracts and inserts:
22300 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
22301 if (isShuffleHigh128VectorInsertLow(SVOp)) {
22302 SDValue V = Extract128BitVector(V1, NumElems/2, DAG, dl);
22303 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, 0, DAG, dl);
22304 return DCI.CombineTo(N, InsV);
22307 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
22308 if (isShuffleLow128VectorInsertHigh(SVOp)) {
22309 SDValue V = Extract128BitVector(V1, 0, DAG, dl);
22310 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, NumElems/2, DAG, dl);
22311 return DCI.CombineTo(N, InsV);
22317 /// \brief Combine an arbitrary chain of shuffles into a single instruction if
22320 /// This is the leaf of the recursive combinine below. When we have found some
22321 /// chain of single-use x86 shuffle instructions and accumulated the combined
22322 /// shuffle mask represented by them, this will try to pattern match that mask
22323 /// into either a single instruction if there is a special purpose instruction
22324 /// for this operation, or into a PSHUFB instruction which is a fully general
22325 /// instruction but should only be used to replace chains over a certain depth.
22326 static bool combineX86ShuffleChain(SDValue Op, SDValue Root, ArrayRef<int> Mask,
22327 int Depth, bool HasPSHUFB, SelectionDAG &DAG,
22328 TargetLowering::DAGCombinerInfo &DCI,
22329 const X86Subtarget *Subtarget) {
22330 assert(!Mask.empty() && "Cannot combine an empty shuffle mask!");
22332 // Find the operand that enters the chain. Note that multiple uses are OK
22333 // here, we're not going to remove the operand we find.
22334 SDValue Input = Op.getOperand(0);
22335 while (Input.getOpcode() == ISD::BITCAST)
22336 Input = Input.getOperand(0);
22338 MVT VT = Input.getSimpleValueType();
22339 MVT RootVT = Root.getSimpleValueType();
22342 if (Mask.size() == 1) {
22343 int Index = Mask[0];
22344 assert((Index >= 0 || Index == SM_SentinelUndef ||
22345 Index == SM_SentinelZero) &&
22346 "Invalid shuffle index found!");
22348 // We may end up with an accumulated mask of size 1 as a result of
22349 // widening of shuffle operands (see function canWidenShuffleElements).
22350 // If the only shuffle index is equal to SM_SentinelZero then propagate
22351 // a zero vector. Otherwise, the combine shuffle mask is a no-op shuffle
22352 // mask, and therefore the entire chain of shuffles can be folded away.
22353 if (Index == SM_SentinelZero)
22354 DCI.CombineTo(Root.getNode(), getZeroVector(RootVT, Subtarget, DAG, DL));
22356 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Input),
22361 // Use the float domain if the operand type is a floating point type.
22362 bool FloatDomain = VT.isFloatingPoint();
22364 // For floating point shuffles, we don't have free copies in the shuffle
22365 // instructions or the ability to load as part of the instruction, so
22366 // canonicalize their shuffles to UNPCK or MOV variants.
22368 // Note that even with AVX we prefer the PSHUFD form of shuffle for integer
22369 // vectors because it can have a load folded into it that UNPCK cannot. This
22370 // doesn't preclude something switching to the shorter encoding post-RA.
22372 // FIXME: Should teach these routines about AVX vector widths.
22373 if (FloatDomain && VT.is128BitVector()) {
22374 if (Mask.equals({0, 0}) || Mask.equals({1, 1})) {
22375 bool Lo = Mask.equals({0, 0});
22378 // Check if we have SSE3 which will let us use MOVDDUP. That instruction
22379 // is no slower than UNPCKLPD but has the option to fold the input operand
22380 // into even an unaligned memory load.
22381 if (Lo && Subtarget->hasSSE3()) {
22382 Shuffle = X86ISD::MOVDDUP;
22383 ShuffleVT = MVT::v2f64;
22385 // We have MOVLHPS and MOVHLPS throughout SSE and they encode smaller
22386 // than the UNPCK variants.
22387 Shuffle = Lo ? X86ISD::MOVLHPS : X86ISD::MOVHLPS;
22388 ShuffleVT = MVT::v4f32;
22390 if (Depth == 1 && Root->getOpcode() == Shuffle)
22391 return false; // Nothing to do!
22392 Op = DAG.getBitcast(ShuffleVT, Input);
22393 DCI.AddToWorklist(Op.getNode());
22394 if (Shuffle == X86ISD::MOVDDUP)
22395 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
22397 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
22398 DCI.AddToWorklist(Op.getNode());
22399 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22403 if (Subtarget->hasSSE3() &&
22404 (Mask.equals({0, 0, 2, 2}) || Mask.equals({1, 1, 3, 3}))) {
22405 bool Lo = Mask.equals({0, 0, 2, 2});
22406 unsigned Shuffle = Lo ? X86ISD::MOVSLDUP : X86ISD::MOVSHDUP;
22407 MVT ShuffleVT = MVT::v4f32;
22408 if (Depth == 1 && Root->getOpcode() == Shuffle)
22409 return false; // Nothing to do!
22410 Op = DAG.getBitcast(ShuffleVT, Input);
22411 DCI.AddToWorklist(Op.getNode());
22412 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
22413 DCI.AddToWorklist(Op.getNode());
22414 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22418 if (Mask.equals({0, 0, 1, 1}) || Mask.equals({2, 2, 3, 3})) {
22419 bool Lo = Mask.equals({0, 0, 1, 1});
22420 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
22421 MVT ShuffleVT = MVT::v4f32;
22422 if (Depth == 1 && Root->getOpcode() == Shuffle)
22423 return false; // Nothing to do!
22424 Op = DAG.getBitcast(ShuffleVT, Input);
22425 DCI.AddToWorklist(Op.getNode());
22426 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
22427 DCI.AddToWorklist(Op.getNode());
22428 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22434 // We always canonicalize the 8 x i16 and 16 x i8 shuffles into their UNPCK
22435 // variants as none of these have single-instruction variants that are
22436 // superior to the UNPCK formulation.
22437 if (!FloatDomain && VT.is128BitVector() &&
22438 (Mask.equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
22439 Mask.equals({4, 4, 5, 5, 6, 6, 7, 7}) ||
22440 Mask.equals({0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7}) ||
22442 {8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14, 14, 15, 15}))) {
22443 bool Lo = Mask[0] == 0;
22444 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
22445 if (Depth == 1 && Root->getOpcode() == Shuffle)
22446 return false; // Nothing to do!
22448 switch (Mask.size()) {
22450 ShuffleVT = MVT::v8i16;
22453 ShuffleVT = MVT::v16i8;
22456 llvm_unreachable("Impossible mask size!");
22458 Op = DAG.getBitcast(ShuffleVT, Input);
22459 DCI.AddToWorklist(Op.getNode());
22460 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
22461 DCI.AddToWorklist(Op.getNode());
22462 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22467 // Don't try to re-form single instruction chains under any circumstances now
22468 // that we've done encoding canonicalization for them.
22472 // If we have 3 or more shuffle instructions or a chain involving PSHUFB, we
22473 // can replace them with a single PSHUFB instruction profitably. Intel's
22474 // manuals suggest only using PSHUFB if doing so replacing 5 instructions, but
22475 // in practice PSHUFB tends to be *very* fast so we're more aggressive.
22476 if ((Depth >= 3 || HasPSHUFB) && Subtarget->hasSSSE3()) {
22477 SmallVector<SDValue, 16> PSHUFBMask;
22478 int NumBytes = VT.getSizeInBits() / 8;
22479 int Ratio = NumBytes / Mask.size();
22480 for (int i = 0; i < NumBytes; ++i) {
22481 if (Mask[i / Ratio] == SM_SentinelUndef) {
22482 PSHUFBMask.push_back(DAG.getUNDEF(MVT::i8));
22485 int M = Mask[i / Ratio] != SM_SentinelZero
22486 ? Ratio * Mask[i / Ratio] + i % Ratio
22488 PSHUFBMask.push_back(DAG.getConstant(M, DL, MVT::i8));
22490 MVT ByteVT = MVT::getVectorVT(MVT::i8, NumBytes);
22491 Op = DAG.getBitcast(ByteVT, Input);
22492 DCI.AddToWorklist(Op.getNode());
22493 SDValue PSHUFBMaskOp =
22494 DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVT, PSHUFBMask);
22495 DCI.AddToWorklist(PSHUFBMaskOp.getNode());
22496 Op = DAG.getNode(X86ISD::PSHUFB, DL, ByteVT, Op, PSHUFBMaskOp);
22497 DCI.AddToWorklist(Op.getNode());
22498 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22503 // Failed to find any combines.
22507 /// \brief Fully generic combining of x86 shuffle instructions.
22509 /// This should be the last combine run over the x86 shuffle instructions. Once
22510 /// they have been fully optimized, this will recursively consider all chains
22511 /// of single-use shuffle instructions, build a generic model of the cumulative
22512 /// shuffle operation, and check for simpler instructions which implement this
22513 /// operation. We use this primarily for two purposes:
22515 /// 1) Collapse generic shuffles to specialized single instructions when
22516 /// equivalent. In most cases, this is just an encoding size win, but
22517 /// sometimes we will collapse multiple generic shuffles into a single
22518 /// special-purpose shuffle.
22519 /// 2) Look for sequences of shuffle instructions with 3 or more total
22520 /// instructions, and replace them with the slightly more expensive SSSE3
22521 /// PSHUFB instruction if available. We do this as the last combining step
22522 /// to ensure we avoid using PSHUFB if we can implement the shuffle with
22523 /// a suitable short sequence of other instructions. The PHUFB will either
22524 /// use a register or have to read from memory and so is slightly (but only
22525 /// slightly) more expensive than the other shuffle instructions.
22527 /// Because this is inherently a quadratic operation (for each shuffle in
22528 /// a chain, we recurse up the chain), the depth is limited to 8 instructions.
22529 /// This should never be an issue in practice as the shuffle lowering doesn't
22530 /// produce sequences of more than 8 instructions.
22532 /// FIXME: We will currently miss some cases where the redundant shuffling
22533 /// would simplify under the threshold for PSHUFB formation because of
22534 /// combine-ordering. To fix this, we should do the redundant instruction
22535 /// combining in this recursive walk.
22536 static bool combineX86ShufflesRecursively(SDValue Op, SDValue Root,
22537 ArrayRef<int> RootMask,
22538 int Depth, bool HasPSHUFB,
22540 TargetLowering::DAGCombinerInfo &DCI,
22541 const X86Subtarget *Subtarget) {
22542 // Bound the depth of our recursive combine because this is ultimately
22543 // quadratic in nature.
22547 // Directly rip through bitcasts to find the underlying operand.
22548 while (Op.getOpcode() == ISD::BITCAST && Op.getOperand(0).hasOneUse())
22549 Op = Op.getOperand(0);
22551 MVT VT = Op.getSimpleValueType();
22552 if (!VT.isVector())
22553 return false; // Bail if we hit a non-vector.
22555 assert(Root.getSimpleValueType().isVector() &&
22556 "Shuffles operate on vector types!");
22557 assert(VT.getSizeInBits() == Root.getSimpleValueType().getSizeInBits() &&
22558 "Can only combine shuffles of the same vector register size.");
22560 if (!isTargetShuffle(Op.getOpcode()))
22562 SmallVector<int, 16> OpMask;
22564 bool HaveMask = getTargetShuffleMask(Op.getNode(), VT, OpMask, IsUnary);
22565 // We only can combine unary shuffles which we can decode the mask for.
22566 if (!HaveMask || !IsUnary)
22569 assert(VT.getVectorNumElements() == OpMask.size() &&
22570 "Different mask size from vector size!");
22571 assert(((RootMask.size() > OpMask.size() &&
22572 RootMask.size() % OpMask.size() == 0) ||
22573 (OpMask.size() > RootMask.size() &&
22574 OpMask.size() % RootMask.size() == 0) ||
22575 OpMask.size() == RootMask.size()) &&
22576 "The smaller number of elements must divide the larger.");
22577 int RootRatio = std::max<int>(1, OpMask.size() / RootMask.size());
22578 int OpRatio = std::max<int>(1, RootMask.size() / OpMask.size());
22579 assert(((RootRatio == 1 && OpRatio == 1) ||
22580 (RootRatio == 1) != (OpRatio == 1)) &&
22581 "Must not have a ratio for both incoming and op masks!");
22583 SmallVector<int, 16> Mask;
22584 Mask.reserve(std::max(OpMask.size(), RootMask.size()));
22586 // Merge this shuffle operation's mask into our accumulated mask. Note that
22587 // this shuffle's mask will be the first applied to the input, followed by the
22588 // root mask to get us all the way to the root value arrangement. The reason
22589 // for this order is that we are recursing up the operation chain.
22590 for (int i = 0, e = std::max(OpMask.size(), RootMask.size()); i < e; ++i) {
22591 int RootIdx = i / RootRatio;
22592 if (RootMask[RootIdx] < 0) {
22593 // This is a zero or undef lane, we're done.
22594 Mask.push_back(RootMask[RootIdx]);
22598 int RootMaskedIdx = RootMask[RootIdx] * RootRatio + i % RootRatio;
22599 int OpIdx = RootMaskedIdx / OpRatio;
22600 if (OpMask[OpIdx] < 0) {
22601 // The incoming lanes are zero or undef, it doesn't matter which ones we
22603 Mask.push_back(OpMask[OpIdx]);
22607 // Ok, we have non-zero lanes, map them through.
22608 Mask.push_back(OpMask[OpIdx] * OpRatio +
22609 RootMaskedIdx % OpRatio);
22612 // See if we can recurse into the operand to combine more things.
22613 switch (Op.getOpcode()) {
22614 case X86ISD::PSHUFB:
22616 case X86ISD::PSHUFD:
22617 case X86ISD::PSHUFHW:
22618 case X86ISD::PSHUFLW:
22619 if (Op.getOperand(0).hasOneUse() &&
22620 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
22621 HasPSHUFB, DAG, DCI, Subtarget))
22625 case X86ISD::UNPCKL:
22626 case X86ISD::UNPCKH:
22627 assert(Op.getOperand(0) == Op.getOperand(1) &&
22628 "We only combine unary shuffles!");
22629 // We can't check for single use, we have to check that this shuffle is the
22631 if (Op->isOnlyUserOf(Op.getOperand(0).getNode()) &&
22632 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
22633 HasPSHUFB, DAG, DCI, Subtarget))
22638 // Minor canonicalization of the accumulated shuffle mask to make it easier
22639 // to match below. All this does is detect masks with squential pairs of
22640 // elements, and shrink them to the half-width mask. It does this in a loop
22641 // so it will reduce the size of the mask to the minimal width mask which
22642 // performs an equivalent shuffle.
22643 SmallVector<int, 16> WidenedMask;
22644 while (Mask.size() > 1 && canWidenShuffleElements(Mask, WidenedMask)) {
22645 Mask = std::move(WidenedMask);
22646 WidenedMask.clear();
22649 return combineX86ShuffleChain(Op, Root, Mask, Depth, HasPSHUFB, DAG, DCI,
22653 /// \brief Get the PSHUF-style mask from PSHUF node.
22655 /// This is a very minor wrapper around getTargetShuffleMask to easy forming v4
22656 /// PSHUF-style masks that can be reused with such instructions.
22657 static SmallVector<int, 4> getPSHUFShuffleMask(SDValue N) {
22658 MVT VT = N.getSimpleValueType();
22659 SmallVector<int, 4> Mask;
22661 bool HaveMask = getTargetShuffleMask(N.getNode(), VT, Mask, IsUnary);
22665 // If we have more than 128-bits, only the low 128-bits of shuffle mask
22666 // matter. Check that the upper masks are repeats and remove them.
22667 if (VT.getSizeInBits() > 128) {
22668 int LaneElts = 128 / VT.getScalarSizeInBits();
22670 for (int i = 1, NumLanes = VT.getSizeInBits() / 128; i < NumLanes; ++i)
22671 for (int j = 0; j < LaneElts; ++j)
22672 assert(Mask[j] == Mask[i * LaneElts + j] - (LaneElts * i) &&
22673 "Mask doesn't repeat in high 128-bit lanes!");
22675 Mask.resize(LaneElts);
22678 switch (N.getOpcode()) {
22679 case X86ISD::PSHUFD:
22681 case X86ISD::PSHUFLW:
22684 case X86ISD::PSHUFHW:
22685 Mask.erase(Mask.begin(), Mask.begin() + 4);
22686 for (int &M : Mask)
22690 llvm_unreachable("No valid shuffle instruction found!");
22694 /// \brief Search for a combinable shuffle across a chain ending in pshufd.
22696 /// We walk up the chain and look for a combinable shuffle, skipping over
22697 /// shuffles that we could hoist this shuffle's transformation past without
22698 /// altering anything.
22700 combineRedundantDWordShuffle(SDValue N, MutableArrayRef<int> Mask,
22702 TargetLowering::DAGCombinerInfo &DCI) {
22703 assert(N.getOpcode() == X86ISD::PSHUFD &&
22704 "Called with something other than an x86 128-bit half shuffle!");
22707 // Walk up a single-use chain looking for a combinable shuffle. Keep a stack
22708 // of the shuffles in the chain so that we can form a fresh chain to replace
22710 SmallVector<SDValue, 8> Chain;
22711 SDValue V = N.getOperand(0);
22712 for (; V.hasOneUse(); V = V.getOperand(0)) {
22713 switch (V.getOpcode()) {
22715 return SDValue(); // Nothing combined!
22718 // Skip bitcasts as we always know the type for the target specific
22722 case X86ISD::PSHUFD:
22723 // Found another dword shuffle.
22726 case X86ISD::PSHUFLW:
22727 // Check that the low words (being shuffled) are the identity in the
22728 // dword shuffle, and the high words are self-contained.
22729 if (Mask[0] != 0 || Mask[1] != 1 ||
22730 !(Mask[2] >= 2 && Mask[2] < 4 && Mask[3] >= 2 && Mask[3] < 4))
22733 Chain.push_back(V);
22736 case X86ISD::PSHUFHW:
22737 // Check that the high words (being shuffled) are the identity in the
22738 // dword shuffle, and the low words are self-contained.
22739 if (Mask[2] != 2 || Mask[3] != 3 ||
22740 !(Mask[0] >= 0 && Mask[0] < 2 && Mask[1] >= 0 && Mask[1] < 2))
22743 Chain.push_back(V);
22746 case X86ISD::UNPCKL:
22747 case X86ISD::UNPCKH:
22748 // For either i8 -> i16 or i16 -> i32 unpacks, we can combine a dword
22749 // shuffle into a preceding word shuffle.
22750 if (V.getSimpleValueType().getVectorElementType() != MVT::i8 &&
22751 V.getSimpleValueType().getVectorElementType() != MVT::i16)
22754 // Search for a half-shuffle which we can combine with.
22755 unsigned CombineOp =
22756 V.getOpcode() == X86ISD::UNPCKL ? X86ISD::PSHUFLW : X86ISD::PSHUFHW;
22757 if (V.getOperand(0) != V.getOperand(1) ||
22758 !V->isOnlyUserOf(V.getOperand(0).getNode()))
22760 Chain.push_back(V);
22761 V = V.getOperand(0);
22763 switch (V.getOpcode()) {
22765 return SDValue(); // Nothing to combine.
22767 case X86ISD::PSHUFLW:
22768 case X86ISD::PSHUFHW:
22769 if (V.getOpcode() == CombineOp)
22772 Chain.push_back(V);
22776 V = V.getOperand(0);
22780 } while (V.hasOneUse());
22783 // Break out of the loop if we break out of the switch.
22787 if (!V.hasOneUse())
22788 // We fell out of the loop without finding a viable combining instruction.
22791 // Merge this node's mask and our incoming mask.
22792 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22793 for (int &M : Mask)
22795 V = DAG.getNode(V.getOpcode(), DL, V.getValueType(), V.getOperand(0),
22796 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
22798 // Rebuild the chain around this new shuffle.
22799 while (!Chain.empty()) {
22800 SDValue W = Chain.pop_back_val();
22802 if (V.getValueType() != W.getOperand(0).getValueType())
22803 V = DAG.getBitcast(W.getOperand(0).getValueType(), V);
22805 switch (W.getOpcode()) {
22807 llvm_unreachable("Only PSHUF and UNPCK instructions get here!");
22809 case X86ISD::UNPCKL:
22810 case X86ISD::UNPCKH:
22811 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, V);
22814 case X86ISD::PSHUFD:
22815 case X86ISD::PSHUFLW:
22816 case X86ISD::PSHUFHW:
22817 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, W.getOperand(1));
22821 if (V.getValueType() != N.getValueType())
22822 V = DAG.getBitcast(N.getValueType(), V);
22824 // Return the new chain to replace N.
22828 /// \brief Search for a combinable shuffle across a chain ending in pshuflw or
22831 /// We walk up the chain, skipping shuffles of the other half and looking
22832 /// through shuffles which switch halves trying to find a shuffle of the same
22833 /// pair of dwords.
22834 static bool combineRedundantHalfShuffle(SDValue N, MutableArrayRef<int> Mask,
22836 TargetLowering::DAGCombinerInfo &DCI) {
22838 (N.getOpcode() == X86ISD::PSHUFLW || N.getOpcode() == X86ISD::PSHUFHW) &&
22839 "Called with something other than an x86 128-bit half shuffle!");
22841 unsigned CombineOpcode = N.getOpcode();
22843 // Walk up a single-use chain looking for a combinable shuffle.
22844 SDValue V = N.getOperand(0);
22845 for (; V.hasOneUse(); V = V.getOperand(0)) {
22846 switch (V.getOpcode()) {
22848 return false; // Nothing combined!
22851 // Skip bitcasts as we always know the type for the target specific
22855 case X86ISD::PSHUFLW:
22856 case X86ISD::PSHUFHW:
22857 if (V.getOpcode() == CombineOpcode)
22860 // Other-half shuffles are no-ops.
22863 // Break out of the loop if we break out of the switch.
22867 if (!V.hasOneUse())
22868 // We fell out of the loop without finding a viable combining instruction.
22871 // Combine away the bottom node as its shuffle will be accumulated into
22872 // a preceding shuffle.
22873 DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
22875 // Record the old value.
22878 // Merge this node's mask and our incoming mask (adjusted to account for all
22879 // the pshufd instructions encountered).
22880 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22881 for (int &M : Mask)
22883 V = DAG.getNode(V.getOpcode(), DL, MVT::v8i16, V.getOperand(0),
22884 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
22886 // Check that the shuffles didn't cancel each other out. If not, we need to
22887 // combine to the new one.
22889 // Replace the combinable shuffle with the combined one, updating all users
22890 // so that we re-evaluate the chain here.
22891 DCI.CombineTo(Old.getNode(), V, /*AddTo*/ true);
22896 /// \brief Try to combine x86 target specific shuffles.
22897 static SDValue PerformTargetShuffleCombine(SDValue N, SelectionDAG &DAG,
22898 TargetLowering::DAGCombinerInfo &DCI,
22899 const X86Subtarget *Subtarget) {
22901 MVT VT = N.getSimpleValueType();
22902 SmallVector<int, 4> Mask;
22904 switch (N.getOpcode()) {
22905 case X86ISD::PSHUFD:
22906 case X86ISD::PSHUFLW:
22907 case X86ISD::PSHUFHW:
22908 Mask = getPSHUFShuffleMask(N);
22909 assert(Mask.size() == 4);
22911 case X86ISD::UNPCKL: {
22912 // Combine X86ISD::UNPCKL and ISD::VECTOR_SHUFFLE into X86ISD::UNPCKH, in
22913 // which X86ISD::UNPCKL has a ISD::UNDEF operand, and ISD::VECTOR_SHUFFLE
22914 // moves upper half elements into the lower half part. For example:
22916 // t2: v16i8 = vector_shuffle<8,9,10,11,12,13,14,15,u,u,u,u,u,u,u,u> t1,
22918 // t3: v16i8 = X86ISD::UNPCKL undef:v16i8, t2
22920 // will be combined to:
22922 // t3: v16i8 = X86ISD::UNPCKH undef:v16i8, t1
22924 // This is only for 128-bit vectors. From SSE4.1 onward this combine may not
22925 // happen due to advanced instructions.
22926 if (!VT.is128BitVector())
22929 auto Op0 = N.getOperand(0);
22930 auto Op1 = N.getOperand(1);
22931 if (Op0.getOpcode() == ISD::UNDEF &&
22932 Op1.getNode()->getOpcode() == ISD::VECTOR_SHUFFLE) {
22933 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Op1.getNode())->getMask();
22935 unsigned NumElts = VT.getVectorNumElements();
22936 SmallVector<int, 8> ExpectedMask(NumElts, -1);
22937 std::iota(ExpectedMask.begin(), ExpectedMask.begin() + NumElts / 2,
22940 auto ShufOp = Op1.getOperand(0);
22941 if (isShuffleEquivalent(Op1, ShufOp, Mask, ExpectedMask))
22942 return DAG.getNode(X86ISD::UNPCKH, DL, VT, N.getOperand(0), ShufOp);
22950 // Nuke no-op shuffles that show up after combining.
22951 if (isNoopShuffleMask(Mask))
22952 return DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
22954 // Look for simplifications involving one or two shuffle instructions.
22955 SDValue V = N.getOperand(0);
22956 switch (N.getOpcode()) {
22959 case X86ISD::PSHUFLW:
22960 case X86ISD::PSHUFHW:
22961 assert(VT.getVectorElementType() == MVT::i16 && "Bad word shuffle type!");
22963 if (combineRedundantHalfShuffle(N, Mask, DAG, DCI))
22964 return SDValue(); // We combined away this shuffle, so we're done.
22966 // See if this reduces to a PSHUFD which is no more expensive and can
22967 // combine with more operations. Note that it has to at least flip the
22968 // dwords as otherwise it would have been removed as a no-op.
22969 if (makeArrayRef(Mask).equals({2, 3, 0, 1})) {
22970 int DMask[] = {0, 1, 2, 3};
22971 int DOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 2;
22972 DMask[DOffset + 0] = DOffset + 1;
22973 DMask[DOffset + 1] = DOffset + 0;
22974 MVT DVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
22975 V = DAG.getBitcast(DVT, V);
22976 DCI.AddToWorklist(V.getNode());
22977 V = DAG.getNode(X86ISD::PSHUFD, DL, DVT, V,
22978 getV4X86ShuffleImm8ForMask(DMask, DL, DAG));
22979 DCI.AddToWorklist(V.getNode());
22980 return DAG.getBitcast(VT, V);
22983 // Look for shuffle patterns which can be implemented as a single unpack.
22984 // FIXME: This doesn't handle the location of the PSHUFD generically, and
22985 // only works when we have a PSHUFD followed by two half-shuffles.
22986 if (Mask[0] == Mask[1] && Mask[2] == Mask[3] &&
22987 (V.getOpcode() == X86ISD::PSHUFLW ||
22988 V.getOpcode() == X86ISD::PSHUFHW) &&
22989 V.getOpcode() != N.getOpcode() &&
22991 SDValue D = V.getOperand(0);
22992 while (D.getOpcode() == ISD::BITCAST && D.hasOneUse())
22993 D = D.getOperand(0);
22994 if (D.getOpcode() == X86ISD::PSHUFD && D.hasOneUse()) {
22995 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22996 SmallVector<int, 4> DMask = getPSHUFShuffleMask(D);
22997 int NOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
22998 int VOffset = V.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
23000 for (int i = 0; i < 4; ++i) {
23001 WordMask[i + NOffset] = Mask[i] + NOffset;
23002 WordMask[i + VOffset] = VMask[i] + VOffset;
23004 // Map the word mask through the DWord mask.
23006 for (int i = 0; i < 8; ++i)
23007 MappedMask[i] = 2 * DMask[WordMask[i] / 2] + WordMask[i] % 2;
23008 if (makeArrayRef(MappedMask).equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
23009 makeArrayRef(MappedMask).equals({4, 4, 5, 5, 6, 6, 7, 7})) {
23010 // We can replace all three shuffles with an unpack.
23011 V = DAG.getBitcast(VT, D.getOperand(0));
23012 DCI.AddToWorklist(V.getNode());
23013 return DAG.getNode(MappedMask[0] == 0 ? X86ISD::UNPCKL
23022 case X86ISD::PSHUFD:
23023 if (SDValue NewN = combineRedundantDWordShuffle(N, Mask, DAG, DCI))
23032 /// \brief Try to combine a shuffle into a target-specific add-sub node.
23034 /// We combine this directly on the abstract vector shuffle nodes so it is
23035 /// easier to generically match. We also insert dummy vector shuffle nodes for
23036 /// the operands which explicitly discard the lanes which are unused by this
23037 /// operation to try to flow through the rest of the combiner the fact that
23038 /// they're unused.
23039 static SDValue combineShuffleToAddSub(SDNode *N, SelectionDAG &DAG) {
23041 EVT VT = N->getValueType(0);
23043 // We only handle target-independent shuffles.
23044 // FIXME: It would be easy and harmless to use the target shuffle mask
23045 // extraction tool to support more.
23046 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
23049 auto *SVN = cast<ShuffleVectorSDNode>(N);
23050 ArrayRef<int> Mask = SVN->getMask();
23051 SDValue V1 = N->getOperand(0);
23052 SDValue V2 = N->getOperand(1);
23054 // We require the first shuffle operand to be the SUB node, and the second to
23055 // be the ADD node.
23056 // FIXME: We should support the commuted patterns.
23057 if (V1->getOpcode() != ISD::FSUB || V2->getOpcode() != ISD::FADD)
23060 // If there are other uses of these operations we can't fold them.
23061 if (!V1->hasOneUse() || !V2->hasOneUse())
23064 // Ensure that both operations have the same operands. Note that we can
23065 // commute the FADD operands.
23066 SDValue LHS = V1->getOperand(0), RHS = V1->getOperand(1);
23067 if ((V2->getOperand(0) != LHS || V2->getOperand(1) != RHS) &&
23068 (V2->getOperand(0) != RHS || V2->getOperand(1) != LHS))
23071 // We're looking for blends between FADD and FSUB nodes. We insist on these
23072 // nodes being lined up in a specific expected pattern.
23073 if (!(isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
23074 isShuffleEquivalent(V1, V2, Mask, {0, 5, 2, 7}) ||
23075 isShuffleEquivalent(V1, V2, Mask, {0, 9, 2, 11, 4, 13, 6, 15})))
23078 // Only specific types are legal at this point, assert so we notice if and
23079 // when these change.
23080 assert((VT == MVT::v4f32 || VT == MVT::v2f64 || VT == MVT::v8f32 ||
23081 VT == MVT::v4f64) &&
23082 "Unknown vector type encountered!");
23084 return DAG.getNode(X86ISD::ADDSUB, DL, VT, LHS, RHS);
23087 /// PerformShuffleCombine - Performs several different shuffle combines.
23088 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
23089 TargetLowering::DAGCombinerInfo &DCI,
23090 const X86Subtarget *Subtarget) {
23092 SDValue N0 = N->getOperand(0);
23093 SDValue N1 = N->getOperand(1);
23094 EVT VT = N->getValueType(0);
23096 // Don't create instructions with illegal types after legalize types has run.
23097 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23098 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
23101 // If we have legalized the vector types, look for blends of FADD and FSUB
23102 // nodes that we can fuse into an ADDSUB node.
23103 if (TLI.isTypeLegal(VT) && Subtarget->hasSSE3())
23104 if (SDValue AddSub = combineShuffleToAddSub(N, DAG))
23107 // Combine 256-bit vector shuffles. This is only profitable when in AVX mode
23108 if (Subtarget->hasFp256() && VT.is256BitVector() &&
23109 N->getOpcode() == ISD::VECTOR_SHUFFLE)
23110 return PerformShuffleCombine256(N, DAG, DCI, Subtarget);
23112 // During Type Legalization, when promoting illegal vector types,
23113 // the backend might introduce new shuffle dag nodes and bitcasts.
23115 // This code performs the following transformation:
23116 // fold: (shuffle (bitcast (BINOP A, B)), Undef, <Mask>) ->
23117 // (shuffle (BINOP (bitcast A), (bitcast B)), Undef, <Mask>)
23119 // We do this only if both the bitcast and the BINOP dag nodes have
23120 // one use. Also, perform this transformation only if the new binary
23121 // operation is legal. This is to avoid introducing dag nodes that
23122 // potentially need to be further expanded (or custom lowered) into a
23123 // less optimal sequence of dag nodes.
23124 if (!DCI.isBeforeLegalize() && DCI.isBeforeLegalizeOps() &&
23125 N1.getOpcode() == ISD::UNDEF && N0.hasOneUse() &&
23126 N0.getOpcode() == ISD::BITCAST) {
23127 SDValue BC0 = N0.getOperand(0);
23128 EVT SVT = BC0.getValueType();
23129 unsigned Opcode = BC0.getOpcode();
23130 unsigned NumElts = VT.getVectorNumElements();
23132 if (BC0.hasOneUse() && SVT.isVector() &&
23133 SVT.getVectorNumElements() * 2 == NumElts &&
23134 TLI.isOperationLegal(Opcode, VT)) {
23135 bool CanFold = false;
23147 unsigned SVTNumElts = SVT.getVectorNumElements();
23148 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
23149 for (unsigned i = 0, e = SVTNumElts; i != e && CanFold; ++i)
23150 CanFold = SVOp->getMaskElt(i) == (int)(i * 2);
23151 for (unsigned i = SVTNumElts, e = NumElts; i != e && CanFold; ++i)
23152 CanFold = SVOp->getMaskElt(i) < 0;
23155 SDValue BC00 = DAG.getBitcast(VT, BC0.getOperand(0));
23156 SDValue BC01 = DAG.getBitcast(VT, BC0.getOperand(1));
23157 SDValue NewBinOp = DAG.getNode(BC0.getOpcode(), dl, VT, BC00, BC01);
23158 return DAG.getVectorShuffle(VT, dl, NewBinOp, N1, &SVOp->getMask()[0]);
23163 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
23164 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
23165 // consecutive, non-overlapping, and in the right order.
23166 SmallVector<SDValue, 16> Elts;
23167 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
23168 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
23170 if (SDValue LD = EltsFromConsecutiveLoads(VT, Elts, dl, DAG, true))
23173 if (isTargetShuffle(N->getOpcode())) {
23175 PerformTargetShuffleCombine(SDValue(N, 0), DAG, DCI, Subtarget);
23176 if (Shuffle.getNode())
23179 // Try recursively combining arbitrary sequences of x86 shuffle
23180 // instructions into higher-order shuffles. We do this after combining
23181 // specific PSHUF instruction sequences into their minimal form so that we
23182 // can evaluate how many specialized shuffle instructions are involved in
23183 // a particular chain.
23184 SmallVector<int, 1> NonceMask; // Just a placeholder.
23185 NonceMask.push_back(0);
23186 if (combineX86ShufflesRecursively(SDValue(N, 0), SDValue(N, 0), NonceMask,
23187 /*Depth*/ 1, /*HasPSHUFB*/ false, DAG,
23189 return SDValue(); // This routine will use CombineTo to replace N.
23195 /// XFormVExtractWithShuffleIntoLoad - Check if a vector extract from a target
23196 /// specific shuffle of a load can be folded into a single element load.
23197 /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
23198 /// shuffles have been custom lowered so we need to handle those here.
23199 static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
23200 TargetLowering::DAGCombinerInfo &DCI) {
23201 if (DCI.isBeforeLegalizeOps())
23204 SDValue InVec = N->getOperand(0);
23205 SDValue EltNo = N->getOperand(1);
23207 if (!isa<ConstantSDNode>(EltNo))
23210 EVT OriginalVT = InVec.getValueType();
23212 if (InVec.getOpcode() == ISD::BITCAST) {
23213 // Don't duplicate a load with other uses.
23214 if (!InVec.hasOneUse())
23216 EVT BCVT = InVec.getOperand(0).getValueType();
23217 if (!BCVT.isVector() ||
23218 BCVT.getVectorNumElements() != OriginalVT.getVectorNumElements())
23220 InVec = InVec.getOperand(0);
23223 EVT CurrentVT = InVec.getValueType();
23225 if (!isTargetShuffle(InVec.getOpcode()))
23228 // Don't duplicate a load with other uses.
23229 if (!InVec.hasOneUse())
23232 SmallVector<int, 16> ShuffleMask;
23234 if (!getTargetShuffleMask(InVec.getNode(), CurrentVT.getSimpleVT(),
23235 ShuffleMask, UnaryShuffle))
23238 // Select the input vector, guarding against out of range extract vector.
23239 unsigned NumElems = CurrentVT.getVectorNumElements();
23240 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
23241 int Idx = (Elt > (int)NumElems) ? -1 : ShuffleMask[Elt];
23242 SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0)
23243 : InVec.getOperand(1);
23245 // If inputs to shuffle are the same for both ops, then allow 2 uses
23246 unsigned AllowedUses = InVec.getNumOperands() > 1 &&
23247 InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
23249 if (LdNode.getOpcode() == ISD::BITCAST) {
23250 // Don't duplicate a load with other uses.
23251 if (!LdNode.getNode()->hasNUsesOfValue(AllowedUses, 0))
23254 AllowedUses = 1; // only allow 1 load use if we have a bitcast
23255 LdNode = LdNode.getOperand(0);
23258 if (!ISD::isNormalLoad(LdNode.getNode()))
23261 LoadSDNode *LN0 = cast<LoadSDNode>(LdNode);
23263 if (!LN0 ||!LN0->hasNUsesOfValue(AllowedUses, 0) || LN0->isVolatile())
23266 EVT EltVT = N->getValueType(0);
23267 // If there's a bitcast before the shuffle, check if the load type and
23268 // alignment is valid.
23269 unsigned Align = LN0->getAlignment();
23270 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23271 unsigned NewAlign = DAG.getDataLayout().getABITypeAlignment(
23272 EltVT.getTypeForEVT(*DAG.getContext()));
23274 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, EltVT))
23277 // All checks match so transform back to vector_shuffle so that DAG combiner
23278 // can finish the job
23281 // Create shuffle node taking into account the case that its a unary shuffle
23282 SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(CurrentVT)
23283 : InVec.getOperand(1);
23284 Shuffle = DAG.getVectorShuffle(CurrentVT, dl,
23285 InVec.getOperand(0), Shuffle,
23287 Shuffle = DAG.getBitcast(OriginalVT, Shuffle);
23288 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle,
23292 static SDValue PerformBITCASTCombine(SDNode *N, SelectionDAG &DAG,
23293 const X86Subtarget *Subtarget) {
23294 SDValue N0 = N->getOperand(0);
23295 EVT VT = N->getValueType(0);
23297 // Detect bitcasts between i32 to x86mmx low word. Since MMX types are
23298 // special and don't usually play with other vector types, it's better to
23299 // handle them early to be sure we emit efficient code by avoiding
23300 // store-load conversions.
23301 if (VT == MVT::x86mmx && N0.getOpcode() == ISD::BUILD_VECTOR &&
23302 N0.getValueType() == MVT::v2i32 &&
23303 isa<ConstantSDNode>(N0.getOperand(1))) {
23304 SDValue N00 = N0->getOperand(0);
23305 if (N0.getConstantOperandVal(1) == 0 && N00.getValueType() == MVT::i32)
23306 return DAG.getNode(X86ISD::MMX_MOVW2D, SDLoc(N00), VT, N00);
23309 // Convert a bitcasted integer logic operation that has one bitcasted
23310 // floating-point operand and one constant operand into a floating-point
23311 // logic operation. This may create a load of the constant, but that is
23312 // cheaper than materializing the constant in an integer register and
23313 // transferring it to an SSE register or transferring the SSE operand to
23314 // integer register and back.
23316 switch (N0.getOpcode()) {
23317 case ISD::AND: FPOpcode = X86ISD::FAND; break;
23318 case ISD::OR: FPOpcode = X86ISD::FOR; break;
23319 case ISD::XOR: FPOpcode = X86ISD::FXOR; break;
23320 default: return SDValue();
23322 if (((Subtarget->hasSSE1() && VT == MVT::f32) ||
23323 (Subtarget->hasSSE2() && VT == MVT::f64)) &&
23324 isa<ConstantSDNode>(N0.getOperand(1)) &&
23325 N0.getOperand(0).getOpcode() == ISD::BITCAST &&
23326 N0.getOperand(0).getOperand(0).getValueType() == VT) {
23327 SDValue N000 = N0.getOperand(0).getOperand(0);
23328 SDValue FPConst = DAG.getBitcast(VT, N0.getOperand(1));
23329 return DAG.getNode(FPOpcode, SDLoc(N0), VT, N000, FPConst);
23335 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
23336 /// generation and convert it from being a bunch of shuffles and extracts
23337 /// into a somewhat faster sequence. For i686, the best sequence is apparently
23338 /// storing the value and loading scalars back, while for x64 we should
23339 /// use 64-bit extracts and shifts.
23340 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
23341 TargetLowering::DAGCombinerInfo &DCI) {
23342 if (SDValue NewOp = XFormVExtractWithShuffleIntoLoad(N, DAG, DCI))
23345 SDValue InputVector = N->getOperand(0);
23346 SDLoc dl(InputVector);
23347 // Detect mmx to i32 conversion through a v2i32 elt extract.
23348 if (InputVector.getOpcode() == ISD::BITCAST && InputVector.hasOneUse() &&
23349 N->getValueType(0) == MVT::i32 &&
23350 InputVector.getValueType() == MVT::v2i32) {
23352 // The bitcast source is a direct mmx result.
23353 SDValue MMXSrc = InputVector.getNode()->getOperand(0);
23354 if (MMXSrc.getValueType() == MVT::x86mmx)
23355 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
23356 N->getValueType(0),
23357 InputVector.getNode()->getOperand(0));
23359 // The mmx is indirect: (i64 extract_elt (v1i64 bitcast (x86mmx ...))).
23360 if (MMXSrc.getOpcode() == ISD::EXTRACT_VECTOR_ELT && MMXSrc.hasOneUse() &&
23361 MMXSrc.getValueType() == MVT::i64) {
23362 SDValue MMXSrcOp = MMXSrc.getOperand(0);
23363 if (MMXSrcOp.hasOneUse() && MMXSrcOp.getOpcode() == ISD::BITCAST &&
23364 MMXSrcOp.getValueType() == MVT::v1i64 &&
23365 MMXSrcOp.getOperand(0).getValueType() == MVT::x86mmx)
23366 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
23367 N->getValueType(0), MMXSrcOp.getOperand(0));
23371 EVT VT = N->getValueType(0);
23373 if (VT == MVT::i1 && isa<ConstantSDNode>(N->getOperand(1)) &&
23374 InputVector.getOpcode() == ISD::BITCAST &&
23375 isa<ConstantSDNode>(InputVector.getOperand(0))) {
23376 uint64_t ExtractedElt =
23377 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
23378 uint64_t InputValue =
23379 cast<ConstantSDNode>(InputVector.getOperand(0))->getZExtValue();
23380 uint64_t Res = (InputValue >> ExtractedElt) & 1;
23381 return DAG.getConstant(Res, dl, MVT::i1);
23383 // Only operate on vectors of 4 elements, where the alternative shuffling
23384 // gets to be more expensive.
23385 if (InputVector.getValueType() != MVT::v4i32)
23388 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
23389 // single use which is a sign-extend or zero-extend, and all elements are
23391 SmallVector<SDNode *, 4> Uses;
23392 unsigned ExtractedElements = 0;
23393 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
23394 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
23395 if (UI.getUse().getResNo() != InputVector.getResNo())
23398 SDNode *Extract = *UI;
23399 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
23402 if (Extract->getValueType(0) != MVT::i32)
23404 if (!Extract->hasOneUse())
23406 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
23407 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
23409 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
23412 // Record which element was extracted.
23413 ExtractedElements |=
23414 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
23416 Uses.push_back(Extract);
23419 // If not all the elements were used, this may not be worthwhile.
23420 if (ExtractedElements != 15)
23423 // Ok, we've now decided to do the transformation.
23424 // If 64-bit shifts are legal, use the extract-shift sequence,
23425 // otherwise bounce the vector off the cache.
23426 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23429 if (TLI.isOperationLegal(ISD::SRA, MVT::i64)) {
23430 SDValue Cst = DAG.getBitcast(MVT::v2i64, InputVector);
23431 auto &DL = DAG.getDataLayout();
23432 EVT VecIdxTy = DAG.getTargetLoweringInfo().getVectorIdxTy(DL);
23433 SDValue BottomHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
23434 DAG.getConstant(0, dl, VecIdxTy));
23435 SDValue TopHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
23436 DAG.getConstant(1, dl, VecIdxTy));
23438 SDValue ShAmt = DAG.getConstant(
23439 32, dl, DAG.getTargetLoweringInfo().getShiftAmountTy(MVT::i64, DL));
23440 Vals[0] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BottomHalf);
23441 Vals[1] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
23442 DAG.getNode(ISD::SRA, dl, MVT::i64, BottomHalf, ShAmt));
23443 Vals[2] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, TopHalf);
23444 Vals[3] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
23445 DAG.getNode(ISD::SRA, dl, MVT::i64, TopHalf, ShAmt));
23447 // Store the value to a temporary stack slot.
23448 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
23449 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
23450 MachinePointerInfo(), false, false, 0);
23452 EVT ElementType = InputVector.getValueType().getVectorElementType();
23453 unsigned EltSize = ElementType.getSizeInBits() / 8;
23455 // Replace each use (extract) with a load of the appropriate element.
23456 for (unsigned i = 0; i < 4; ++i) {
23457 uint64_t Offset = EltSize * i;
23458 auto PtrVT = TLI.getPointerTy(DAG.getDataLayout());
23459 SDValue OffsetVal = DAG.getConstant(Offset, dl, PtrVT);
23461 SDValue ScalarAddr =
23462 DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, OffsetVal);
23464 // Load the scalar.
23465 Vals[i] = DAG.getLoad(ElementType, dl, Ch,
23466 ScalarAddr, MachinePointerInfo(),
23467 false, false, false, 0);
23472 // Replace the extracts
23473 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
23474 UE = Uses.end(); UI != UE; ++UI) {
23475 SDNode *Extract = *UI;
23477 SDValue Idx = Extract->getOperand(1);
23478 uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
23479 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), Vals[IdxVal]);
23482 // The replacement was made in place; don't return anything.
23487 transformVSELECTtoBlendVECTOR_SHUFFLE(SDNode *N, SelectionDAG &DAG,
23488 const X86Subtarget *Subtarget) {
23490 SDValue Cond = N->getOperand(0);
23491 SDValue LHS = N->getOperand(1);
23492 SDValue RHS = N->getOperand(2);
23494 if (Cond.getOpcode() == ISD::SIGN_EXTEND) {
23495 SDValue CondSrc = Cond->getOperand(0);
23496 if (CondSrc->getOpcode() == ISD::SIGN_EXTEND_INREG)
23497 Cond = CondSrc->getOperand(0);
23500 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
23503 // A vselect where all conditions and data are constants can be optimized into
23504 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
23505 if (ISD::isBuildVectorOfConstantSDNodes(LHS.getNode()) &&
23506 ISD::isBuildVectorOfConstantSDNodes(RHS.getNode()))
23509 unsigned MaskValue = 0;
23510 if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue))
23513 MVT VT = N->getSimpleValueType(0);
23514 unsigned NumElems = VT.getVectorNumElements();
23515 SmallVector<int, 8> ShuffleMask(NumElems, -1);
23516 for (unsigned i = 0; i < NumElems; ++i) {
23517 // Be sure we emit undef where we can.
23518 if (Cond.getOperand(i)->getOpcode() == ISD::UNDEF)
23519 ShuffleMask[i] = -1;
23521 ShuffleMask[i] = i + NumElems * ((MaskValue >> i) & 1);
23524 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23525 if (!TLI.isShuffleMaskLegal(ShuffleMask, VT))
23527 return DAG.getVectorShuffle(VT, dl, LHS, RHS, &ShuffleMask[0]);
23530 /// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT
23532 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
23533 TargetLowering::DAGCombinerInfo &DCI,
23534 const X86Subtarget *Subtarget) {
23536 SDValue Cond = N->getOperand(0);
23537 // Get the LHS/RHS of the select.
23538 SDValue LHS = N->getOperand(1);
23539 SDValue RHS = N->getOperand(2);
23540 EVT VT = LHS.getValueType();
23541 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23543 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
23544 // instructions match the semantics of the common C idiom x<y?x:y but not
23545 // x<=y?x:y, because of how they handle negative zero (which can be
23546 // ignored in unsafe-math mode).
23547 // We also try to create v2f32 min/max nodes, which we later widen to v4f32.
23548 if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() &&
23549 VT != MVT::f80 && (TLI.isTypeLegal(VT) || VT == MVT::v2f32) &&
23550 (Subtarget->hasSSE2() ||
23551 (Subtarget->hasSSE1() && VT.getScalarType() == MVT::f32))) {
23552 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23554 unsigned Opcode = 0;
23555 // Check for x CC y ? x : y.
23556 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
23557 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
23561 // Converting this to a min would handle NaNs incorrectly, and swapping
23562 // the operands would cause it to handle comparisons between positive
23563 // and negative zero incorrectly.
23564 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
23565 if (!DAG.getTarget().Options.UnsafeFPMath &&
23566 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
23568 std::swap(LHS, RHS);
23570 Opcode = X86ISD::FMIN;
23573 // Converting this to a min would handle comparisons between positive
23574 // and negative zero incorrectly.
23575 if (!DAG.getTarget().Options.UnsafeFPMath &&
23576 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
23578 Opcode = X86ISD::FMIN;
23581 // Converting this to a min would handle both negative zeros and NaNs
23582 // incorrectly, but we can swap the operands to fix both.
23583 std::swap(LHS, RHS);
23587 Opcode = X86ISD::FMIN;
23591 // Converting this to a max would handle comparisons between positive
23592 // and negative zero incorrectly.
23593 if (!DAG.getTarget().Options.UnsafeFPMath &&
23594 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
23596 Opcode = X86ISD::FMAX;
23599 // Converting this to a max would handle NaNs incorrectly, and swapping
23600 // the operands would cause it to handle comparisons between positive
23601 // and negative zero incorrectly.
23602 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
23603 if (!DAG.getTarget().Options.UnsafeFPMath &&
23604 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
23606 std::swap(LHS, RHS);
23608 Opcode = X86ISD::FMAX;
23611 // Converting this to a max would handle both negative zeros and NaNs
23612 // incorrectly, but we can swap the operands to fix both.
23613 std::swap(LHS, RHS);
23617 Opcode = X86ISD::FMAX;
23620 // Check for x CC y ? y : x -- a min/max with reversed arms.
23621 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
23622 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
23626 // Converting this to a min would handle comparisons between positive
23627 // and negative zero incorrectly, and swapping the operands would
23628 // cause it to handle NaNs incorrectly.
23629 if (!DAG.getTarget().Options.UnsafeFPMath &&
23630 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
23631 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
23633 std::swap(LHS, RHS);
23635 Opcode = X86ISD::FMIN;
23638 // Converting this to a min would handle NaNs incorrectly.
23639 if (!DAG.getTarget().Options.UnsafeFPMath &&
23640 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
23642 Opcode = X86ISD::FMIN;
23645 // Converting this to a min would handle both negative zeros and NaNs
23646 // incorrectly, but we can swap the operands to fix both.
23647 std::swap(LHS, RHS);
23651 Opcode = X86ISD::FMIN;
23655 // Converting this to a max would handle NaNs incorrectly.
23656 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
23658 Opcode = X86ISD::FMAX;
23661 // Converting this to a max would handle comparisons between positive
23662 // and negative zero incorrectly, and swapping the operands would
23663 // cause it to handle NaNs incorrectly.
23664 if (!DAG.getTarget().Options.UnsafeFPMath &&
23665 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
23666 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
23668 std::swap(LHS, RHS);
23670 Opcode = X86ISD::FMAX;
23673 // Converting this to a max would handle both negative zeros and NaNs
23674 // incorrectly, but we can swap the operands to fix both.
23675 std::swap(LHS, RHS);
23679 Opcode = X86ISD::FMAX;
23685 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
23688 EVT CondVT = Cond.getValueType();
23689 if (Subtarget->hasAVX512() && VT.isVector() && CondVT.isVector() &&
23690 CondVT.getVectorElementType() == MVT::i1) {
23691 // v16i8 (select v16i1, v16i8, v16i8) does not have a proper
23692 // lowering on KNL. In this case we convert it to
23693 // v16i8 (select v16i8, v16i8, v16i8) and use AVX instruction.
23694 // The same situation for all 128 and 256-bit vectors of i8 and i16.
23695 // Since SKX these selects have a proper lowering.
23696 EVT OpVT = LHS.getValueType();
23697 if ((OpVT.is128BitVector() || OpVT.is256BitVector()) &&
23698 (OpVT.getVectorElementType() == MVT::i8 ||
23699 OpVT.getVectorElementType() == MVT::i16) &&
23700 !(Subtarget->hasBWI() && Subtarget->hasVLX())) {
23701 Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, OpVT, Cond);
23702 DCI.AddToWorklist(Cond.getNode());
23703 return DAG.getNode(N->getOpcode(), DL, OpVT, Cond, LHS, RHS);
23706 // If this is a select between two integer constants, try to do some
23708 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
23709 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
23710 // Don't do this for crazy integer types.
23711 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
23712 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
23713 // so that TrueC (the true value) is larger than FalseC.
23714 bool NeedsCondInvert = false;
23716 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
23717 // Efficiently invertible.
23718 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
23719 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
23720 isa<ConstantSDNode>(Cond.getOperand(1))))) {
23721 NeedsCondInvert = true;
23722 std::swap(TrueC, FalseC);
23725 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
23726 if (FalseC->getAPIntValue() == 0 &&
23727 TrueC->getAPIntValue().isPowerOf2()) {
23728 if (NeedsCondInvert) // Invert the condition if needed.
23729 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
23730 DAG.getConstant(1, DL, Cond.getValueType()));
23732 // Zero extend the condition if needed.
23733 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
23735 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
23736 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
23737 DAG.getConstant(ShAmt, DL, MVT::i8));
23740 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
23741 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
23742 if (NeedsCondInvert) // Invert the condition if needed.
23743 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
23744 DAG.getConstant(1, DL, Cond.getValueType()));
23746 // Zero extend the condition if needed.
23747 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
23748 FalseC->getValueType(0), Cond);
23749 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23750 SDValue(FalseC, 0));
23753 // Optimize cases that will turn into an LEA instruction. This requires
23754 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
23755 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
23756 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
23757 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
23759 bool isFastMultiplier = false;
23761 switch ((unsigned char)Diff) {
23763 case 1: // result = add base, cond
23764 case 2: // result = lea base( , cond*2)
23765 case 3: // result = lea base(cond, cond*2)
23766 case 4: // result = lea base( , cond*4)
23767 case 5: // result = lea base(cond, cond*4)
23768 case 8: // result = lea base( , cond*8)
23769 case 9: // result = lea base(cond, cond*8)
23770 isFastMultiplier = true;
23775 if (isFastMultiplier) {
23776 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
23777 if (NeedsCondInvert) // Invert the condition if needed.
23778 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
23779 DAG.getConstant(1, DL, Cond.getValueType()));
23781 // Zero extend the condition if needed.
23782 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
23784 // Scale the condition by the difference.
23786 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
23787 DAG.getConstant(Diff, DL,
23788 Cond.getValueType()));
23790 // Add the base if non-zero.
23791 if (FalseC->getAPIntValue() != 0)
23792 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23793 SDValue(FalseC, 0));
23800 // Canonicalize max and min:
23801 // (x > y) ? x : y -> (x >= y) ? x : y
23802 // (x < y) ? x : y -> (x <= y) ? x : y
23803 // This allows use of COND_S / COND_NS (see TranslateX86CC) which eliminates
23804 // the need for an extra compare
23805 // against zero. e.g.
23806 // (x - y) > 0 : (x - y) ? 0 -> (x - y) >= 0 : (x - y) ? 0
23808 // testl %edi, %edi
23810 // cmovgl %edi, %eax
23814 // cmovsl %eax, %edi
23815 if (N->getOpcode() == ISD::SELECT && Cond.getOpcode() == ISD::SETCC &&
23816 DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
23817 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
23818 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23823 ISD::CondCode NewCC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGE;
23824 Cond = DAG.getSetCC(SDLoc(Cond), Cond.getValueType(),
23825 Cond.getOperand(0), Cond.getOperand(1), NewCC);
23826 return DAG.getNode(ISD::SELECT, DL, VT, Cond, LHS, RHS);
23831 // Early exit check
23832 if (!TLI.isTypeLegal(VT))
23835 // Match VSELECTs into subs with unsigned saturation.
23836 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
23837 // psubus is available in SSE2 and AVX2 for i8 and i16 vectors.
23838 ((Subtarget->hasSSE2() && (VT == MVT::v16i8 || VT == MVT::v8i16)) ||
23839 (Subtarget->hasAVX2() && (VT == MVT::v32i8 || VT == MVT::v16i16)))) {
23840 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23842 // Check if one of the arms of the VSELECT is a zero vector. If it's on the
23843 // left side invert the predicate to simplify logic below.
23845 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
23847 CC = ISD::getSetCCInverse(CC, true);
23848 } else if (ISD::isBuildVectorAllZeros(RHS.getNode())) {
23852 if (Other.getNode() && Other->getNumOperands() == 2 &&
23853 DAG.isEqualTo(Other->getOperand(0), Cond.getOperand(0))) {
23854 SDValue OpLHS = Other->getOperand(0), OpRHS = Other->getOperand(1);
23855 SDValue CondRHS = Cond->getOperand(1);
23857 // Look for a general sub with unsigned saturation first.
23858 // x >= y ? x-y : 0 --> subus x, y
23859 // x > y ? x-y : 0 --> subus x, y
23860 if ((CC == ISD::SETUGE || CC == ISD::SETUGT) &&
23861 Other->getOpcode() == ISD::SUB && DAG.isEqualTo(OpRHS, CondRHS))
23862 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
23864 if (auto *OpRHSBV = dyn_cast<BuildVectorSDNode>(OpRHS))
23865 if (auto *OpRHSConst = OpRHSBV->getConstantSplatNode()) {
23866 if (auto *CondRHSBV = dyn_cast<BuildVectorSDNode>(CondRHS))
23867 if (auto *CondRHSConst = CondRHSBV->getConstantSplatNode())
23868 // If the RHS is a constant we have to reverse the const
23869 // canonicalization.
23870 // x > C-1 ? x+-C : 0 --> subus x, C
23871 if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD &&
23872 CondRHSConst->getAPIntValue() ==
23873 (-OpRHSConst->getAPIntValue() - 1))
23874 return DAG.getNode(
23875 X86ISD::SUBUS, DL, VT, OpLHS,
23876 DAG.getConstant(-OpRHSConst->getAPIntValue(), DL, VT));
23878 // Another special case: If C was a sign bit, the sub has been
23879 // canonicalized into a xor.
23880 // FIXME: Would it be better to use computeKnownBits to determine
23881 // whether it's safe to decanonicalize the xor?
23882 // x s< 0 ? x^C : 0 --> subus x, C
23883 if (CC == ISD::SETLT && Other->getOpcode() == ISD::XOR &&
23884 ISD::isBuildVectorAllZeros(CondRHS.getNode()) &&
23885 OpRHSConst->getAPIntValue().isSignBit())
23886 // Note that we have to rebuild the RHS constant here to ensure we
23887 // don't rely on particular values of undef lanes.
23888 return DAG.getNode(
23889 X86ISD::SUBUS, DL, VT, OpLHS,
23890 DAG.getConstant(OpRHSConst->getAPIntValue(), DL, VT));
23895 // Simplify vector selection if condition value type matches vselect
23897 if (N->getOpcode() == ISD::VSELECT && CondVT == VT) {
23898 assert(Cond.getValueType().isVector() &&
23899 "vector select expects a vector selector!");
23901 bool TValIsAllOnes = ISD::isBuildVectorAllOnes(LHS.getNode());
23902 bool FValIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
23904 // Try invert the condition if true value is not all 1s and false value
23906 if (!TValIsAllOnes && !FValIsAllZeros &&
23907 // Check if the selector will be produced by CMPP*/PCMP*
23908 Cond.getOpcode() == ISD::SETCC &&
23909 // Check if SETCC has already been promoted
23910 TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT) ==
23912 bool TValIsAllZeros = ISD::isBuildVectorAllZeros(LHS.getNode());
23913 bool FValIsAllOnes = ISD::isBuildVectorAllOnes(RHS.getNode());
23915 if (TValIsAllZeros || FValIsAllOnes) {
23916 SDValue CC = Cond.getOperand(2);
23917 ISD::CondCode NewCC =
23918 ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
23919 Cond.getOperand(0).getValueType().isInteger());
23920 Cond = DAG.getSetCC(DL, CondVT, Cond.getOperand(0), Cond.getOperand(1), NewCC);
23921 std::swap(LHS, RHS);
23922 TValIsAllOnes = FValIsAllOnes;
23923 FValIsAllZeros = TValIsAllZeros;
23927 if (TValIsAllOnes || FValIsAllZeros) {
23930 if (TValIsAllOnes && FValIsAllZeros)
23932 else if (TValIsAllOnes)
23934 DAG.getNode(ISD::OR, DL, CondVT, Cond, DAG.getBitcast(CondVT, RHS));
23935 else if (FValIsAllZeros)
23936 Ret = DAG.getNode(ISD::AND, DL, CondVT, Cond,
23937 DAG.getBitcast(CondVT, LHS));
23939 return DAG.getBitcast(VT, Ret);
23943 // We should generate an X86ISD::BLENDI from a vselect if its argument
23944 // is a sign_extend_inreg of an any_extend of a BUILD_VECTOR of
23945 // constants. This specific pattern gets generated when we split a
23946 // selector for a 512 bit vector in a machine without AVX512 (but with
23947 // 256-bit vectors), during legalization:
23949 // (vselect (sign_extend (any_extend (BUILD_VECTOR)) i1) LHS RHS)
23951 // Iff we find this pattern and the build_vectors are built from
23952 // constants, we translate the vselect into a shuffle_vector that we
23953 // know will be matched by LowerVECTOR_SHUFFLEtoBlend.
23954 if ((N->getOpcode() == ISD::VSELECT ||
23955 N->getOpcode() == X86ISD::SHRUNKBLEND) &&
23956 !DCI.isBeforeLegalize() && !VT.is512BitVector()) {
23957 SDValue Shuffle = transformVSELECTtoBlendVECTOR_SHUFFLE(N, DAG, Subtarget);
23958 if (Shuffle.getNode())
23962 // If this is a *dynamic* select (non-constant condition) and we can match
23963 // this node with one of the variable blend instructions, restructure the
23964 // condition so that the blends can use the high bit of each element and use
23965 // SimplifyDemandedBits to simplify the condition operand.
23966 if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() &&
23967 !DCI.isBeforeLegalize() &&
23968 !ISD::isBuildVectorOfConstantSDNodes(Cond.getNode())) {
23969 unsigned BitWidth = Cond.getValueType().getScalarSizeInBits();
23971 // Don't optimize vector selects that map to mask-registers.
23975 // We can only handle the cases where VSELECT is directly legal on the
23976 // subtarget. We custom lower VSELECT nodes with constant conditions and
23977 // this makes it hard to see whether a dynamic VSELECT will correctly
23978 // lower, so we both check the operation's status and explicitly handle the
23979 // cases where a *dynamic* blend will fail even though a constant-condition
23980 // blend could be custom lowered.
23981 // FIXME: We should find a better way to handle this class of problems.
23982 // Potentially, we should combine constant-condition vselect nodes
23983 // pre-legalization into shuffles and not mark as many types as custom
23985 if (!TLI.isOperationLegalOrCustom(ISD::VSELECT, VT))
23987 // FIXME: We don't support i16-element blends currently. We could and
23988 // should support them by making *all* the bits in the condition be set
23989 // rather than just the high bit and using an i8-element blend.
23990 if (VT.getVectorElementType() == MVT::i16)
23992 // Dynamic blending was only available from SSE4.1 onward.
23993 if (VT.is128BitVector() && !Subtarget->hasSSE41())
23995 // Byte blends are only available in AVX2
23996 if (VT == MVT::v32i8 && !Subtarget->hasAVX2())
23999 assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size");
24000 APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);
24002 APInt KnownZero, KnownOne;
24003 TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(),
24004 DCI.isBeforeLegalizeOps());
24005 if (TLO.ShrinkDemandedConstant(Cond, DemandedMask) ||
24006 TLI.SimplifyDemandedBits(Cond, DemandedMask, KnownZero, KnownOne,
24008 // If we changed the computation somewhere in the DAG, this change
24009 // will affect all users of Cond.
24010 // Make sure it is fine and update all the nodes so that we do not
24011 // use the generic VSELECT anymore. Otherwise, we may perform
24012 // wrong optimizations as we messed up with the actual expectation
24013 // for the vector boolean values.
24014 if (Cond != TLO.Old) {
24015 // Check all uses of that condition operand to check whether it will be
24016 // consumed by non-BLEND instructions, which may depend on all bits are
24018 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
24020 if (I->getOpcode() != ISD::VSELECT)
24021 // TODO: Add other opcodes eventually lowered into BLEND.
24024 // Update all the users of the condition, before committing the change,
24025 // so that the VSELECT optimizations that expect the correct vector
24026 // boolean value will not be triggered.
24027 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
24029 DAG.ReplaceAllUsesOfValueWith(
24031 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(*I), I->getValueType(0),
24032 Cond, I->getOperand(1), I->getOperand(2)));
24033 DCI.CommitTargetLoweringOpt(TLO);
24036 // At this point, only Cond is changed. Change the condition
24037 // just for N to keep the opportunity to optimize all other
24038 // users their own way.
24039 DAG.ReplaceAllUsesOfValueWith(
24041 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(N), N->getValueType(0),
24042 TLO.New, N->getOperand(1), N->getOperand(2)));
24050 // Check whether a boolean test is testing a boolean value generated by
24051 // X86ISD::SETCC. If so, return the operand of that SETCC and proper condition
24054 // Simplify the following patterns:
24055 // (Op (CMP (SETCC Cond EFLAGS) 1) EQ) or
24056 // (Op (CMP (SETCC Cond EFLAGS) 0) NEQ)
24057 // to (Op EFLAGS Cond)
24059 // (Op (CMP (SETCC Cond EFLAGS) 0) EQ) or
24060 // (Op (CMP (SETCC Cond EFLAGS) 1) NEQ)
24061 // to (Op EFLAGS !Cond)
24063 // where Op could be BRCOND or CMOV.
24065 static SDValue checkBoolTestSetCCCombine(SDValue Cmp, X86::CondCode &CC) {
24066 // Quit if not CMP and SUB with its value result used.
24067 if (Cmp.getOpcode() != X86ISD::CMP &&
24068 (Cmp.getOpcode() != X86ISD::SUB || Cmp.getNode()->hasAnyUseOfValue(0)))
24071 // Quit if not used as a boolean value.
24072 if (CC != X86::COND_E && CC != X86::COND_NE)
24075 // Check CMP operands. One of them should be 0 or 1 and the other should be
24076 // an SetCC or extended from it.
24077 SDValue Op1 = Cmp.getOperand(0);
24078 SDValue Op2 = Cmp.getOperand(1);
24081 const ConstantSDNode* C = nullptr;
24082 bool needOppositeCond = (CC == X86::COND_E);
24083 bool checkAgainstTrue = false; // Is it a comparison against 1?
24085 if ((C = dyn_cast<ConstantSDNode>(Op1)))
24087 else if ((C = dyn_cast<ConstantSDNode>(Op2)))
24089 else // Quit if all operands are not constants.
24092 if (C->getZExtValue() == 1) {
24093 needOppositeCond = !needOppositeCond;
24094 checkAgainstTrue = true;
24095 } else if (C->getZExtValue() != 0)
24096 // Quit if the constant is neither 0 or 1.
24099 bool truncatedToBoolWithAnd = false;
24100 // Skip (zext $x), (trunc $x), or (and $x, 1) node.
24101 while (SetCC.getOpcode() == ISD::ZERO_EXTEND ||
24102 SetCC.getOpcode() == ISD::TRUNCATE ||
24103 SetCC.getOpcode() == ISD::AND) {
24104 if (SetCC.getOpcode() == ISD::AND) {
24106 ConstantSDNode *CS;
24107 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(0))) &&
24108 CS->getZExtValue() == 1)
24110 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(1))) &&
24111 CS->getZExtValue() == 1)
24115 SetCC = SetCC.getOperand(OpIdx);
24116 truncatedToBoolWithAnd = true;
24118 SetCC = SetCC.getOperand(0);
24121 switch (SetCC.getOpcode()) {
24122 case X86ISD::SETCC_CARRY:
24123 // Since SETCC_CARRY gives output based on R = CF ? ~0 : 0, it's unsafe to
24124 // simplify it if the result of SETCC_CARRY is not canonicalized to 0 or 1,
24125 // i.e. it's a comparison against true but the result of SETCC_CARRY is not
24126 // truncated to i1 using 'and'.
24127 if (checkAgainstTrue && !truncatedToBoolWithAnd)
24129 assert(X86::CondCode(SetCC.getConstantOperandVal(0)) == X86::COND_B &&
24130 "Invalid use of SETCC_CARRY!");
24132 case X86ISD::SETCC:
24133 // Set the condition code or opposite one if necessary.
24134 CC = X86::CondCode(SetCC.getConstantOperandVal(0));
24135 if (needOppositeCond)
24136 CC = X86::GetOppositeBranchCondition(CC);
24137 return SetCC.getOperand(1);
24138 case X86ISD::CMOV: {
24139 // Check whether false/true value has canonical one, i.e. 0 or 1.
24140 ConstantSDNode *FVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(0));
24141 ConstantSDNode *TVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(1));
24142 // Quit if true value is not a constant.
24145 // Quit if false value is not a constant.
24147 SDValue Op = SetCC.getOperand(0);
24148 // Skip 'zext' or 'trunc' node.
24149 if (Op.getOpcode() == ISD::ZERO_EXTEND ||
24150 Op.getOpcode() == ISD::TRUNCATE)
24151 Op = Op.getOperand(0);
24152 // A special case for rdrand/rdseed, where 0 is set if false cond is
24154 if ((Op.getOpcode() != X86ISD::RDRAND &&
24155 Op.getOpcode() != X86ISD::RDSEED) || Op.getResNo() != 0)
24158 // Quit if false value is not the constant 0 or 1.
24159 bool FValIsFalse = true;
24160 if (FVal && FVal->getZExtValue() != 0) {
24161 if (FVal->getZExtValue() != 1)
24163 // If FVal is 1, opposite cond is needed.
24164 needOppositeCond = !needOppositeCond;
24165 FValIsFalse = false;
24167 // Quit if TVal is not the constant opposite of FVal.
24168 if (FValIsFalse && TVal->getZExtValue() != 1)
24170 if (!FValIsFalse && TVal->getZExtValue() != 0)
24172 CC = X86::CondCode(SetCC.getConstantOperandVal(2));
24173 if (needOppositeCond)
24174 CC = X86::GetOppositeBranchCondition(CC);
24175 return SetCC.getOperand(3);
24182 /// Check whether Cond is an AND/OR of SETCCs off of the same EFLAGS.
24184 /// (X86or (X86setcc) (X86setcc))
24185 /// (X86cmp (and (X86setcc) (X86setcc)), 0)
24186 static bool checkBoolTestAndOrSetCCCombine(SDValue Cond, X86::CondCode &CC0,
24187 X86::CondCode &CC1, SDValue &Flags,
24189 if (Cond->getOpcode() == X86ISD::CMP) {
24190 ConstantSDNode *CondOp1C = dyn_cast<ConstantSDNode>(Cond->getOperand(1));
24191 if (!CondOp1C || !CondOp1C->isNullValue())
24194 Cond = Cond->getOperand(0);
24199 SDValue SetCC0, SetCC1;
24200 switch (Cond->getOpcode()) {
24201 default: return false;
24208 SetCC0 = Cond->getOperand(0);
24209 SetCC1 = Cond->getOperand(1);
24213 // Make sure we have SETCC nodes, using the same flags value.
24214 if (SetCC0.getOpcode() != X86ISD::SETCC ||
24215 SetCC1.getOpcode() != X86ISD::SETCC ||
24216 SetCC0->getOperand(1) != SetCC1->getOperand(1))
24219 CC0 = (X86::CondCode)SetCC0->getConstantOperandVal(0);
24220 CC1 = (X86::CondCode)SetCC1->getConstantOperandVal(0);
24221 Flags = SetCC0->getOperand(1);
24225 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
24226 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
24227 TargetLowering::DAGCombinerInfo &DCI,
24228 const X86Subtarget *Subtarget) {
24231 // If the flag operand isn't dead, don't touch this CMOV.
24232 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
24235 SDValue FalseOp = N->getOperand(0);
24236 SDValue TrueOp = N->getOperand(1);
24237 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
24238 SDValue Cond = N->getOperand(3);
24240 if (CC == X86::COND_E || CC == X86::COND_NE) {
24241 switch (Cond.getOpcode()) {
24245 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
24246 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
24247 return (CC == X86::COND_E) ? FalseOp : TrueOp;
24253 Flags = checkBoolTestSetCCCombine(Cond, CC);
24254 if (Flags.getNode() &&
24255 // Extra check as FCMOV only supports a subset of X86 cond.
24256 (FalseOp.getValueType() != MVT::f80 || hasFPCMov(CC))) {
24257 SDValue Ops[] = { FalseOp, TrueOp,
24258 DAG.getConstant(CC, DL, MVT::i8), Flags };
24259 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
24262 // If this is a select between two integer constants, try to do some
24263 // optimizations. Note that the operands are ordered the opposite of SELECT
24265 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
24266 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
24267 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
24268 // larger than FalseC (the false value).
24269 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
24270 CC = X86::GetOppositeBranchCondition(CC);
24271 std::swap(TrueC, FalseC);
24272 std::swap(TrueOp, FalseOp);
24275 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
24276 // This is efficient for any integer data type (including i8/i16) and
24278 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
24279 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
24280 DAG.getConstant(CC, DL, MVT::i8), Cond);
24282 // Zero extend the condition if needed.
24283 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
24285 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
24286 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
24287 DAG.getConstant(ShAmt, DL, MVT::i8));
24288 if (N->getNumValues() == 2) // Dead flag value?
24289 return DCI.CombineTo(N, Cond, SDValue());
24293 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
24294 // for any integer data type, including i8/i16.
24295 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
24296 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
24297 DAG.getConstant(CC, DL, MVT::i8), Cond);
24299 // Zero extend the condition if needed.
24300 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
24301 FalseC->getValueType(0), Cond);
24302 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
24303 SDValue(FalseC, 0));
24305 if (N->getNumValues() == 2) // Dead flag value?
24306 return DCI.CombineTo(N, Cond, SDValue());
24310 // Optimize cases that will turn into an LEA instruction. This requires
24311 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
24312 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
24313 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
24314 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
24316 bool isFastMultiplier = false;
24318 switch ((unsigned char)Diff) {
24320 case 1: // result = add base, cond
24321 case 2: // result = lea base( , cond*2)
24322 case 3: // result = lea base(cond, cond*2)
24323 case 4: // result = lea base( , cond*4)
24324 case 5: // result = lea base(cond, cond*4)
24325 case 8: // result = lea base( , cond*8)
24326 case 9: // result = lea base(cond, cond*8)
24327 isFastMultiplier = true;
24332 if (isFastMultiplier) {
24333 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
24334 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
24335 DAG.getConstant(CC, DL, MVT::i8), Cond);
24336 // Zero extend the condition if needed.
24337 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
24339 // Scale the condition by the difference.
24341 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
24342 DAG.getConstant(Diff, DL, Cond.getValueType()));
24344 // Add the base if non-zero.
24345 if (FalseC->getAPIntValue() != 0)
24346 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
24347 SDValue(FalseC, 0));
24348 if (N->getNumValues() == 2) // Dead flag value?
24349 return DCI.CombineTo(N, Cond, SDValue());
24356 // Handle these cases:
24357 // (select (x != c), e, c) -> select (x != c), e, x),
24358 // (select (x == c), c, e) -> select (x == c), x, e)
24359 // where the c is an integer constant, and the "select" is the combination
24360 // of CMOV and CMP.
24362 // The rationale for this change is that the conditional-move from a constant
24363 // needs two instructions, however, conditional-move from a register needs
24364 // only one instruction.
24366 // CAVEAT: By replacing a constant with a symbolic value, it may obscure
24367 // some instruction-combining opportunities. This opt needs to be
24368 // postponed as late as possible.
24370 if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) {
24371 // the DCI.xxxx conditions are provided to postpone the optimization as
24372 // late as possible.
24374 ConstantSDNode *CmpAgainst = nullptr;
24375 if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) &&
24376 (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) &&
24377 !isa<ConstantSDNode>(Cond.getOperand(0))) {
24379 if (CC == X86::COND_NE &&
24380 CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) {
24381 CC = X86::GetOppositeBranchCondition(CC);
24382 std::swap(TrueOp, FalseOp);
24385 if (CC == X86::COND_E &&
24386 CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) {
24387 SDValue Ops[] = { FalseOp, Cond.getOperand(0),
24388 DAG.getConstant(CC, DL, MVT::i8), Cond };
24389 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops);
24394 // Fold and/or of setcc's to double CMOV:
24395 // (CMOV F, T, ((cc1 | cc2) != 0)) -> (CMOV (CMOV F, T, cc1), T, cc2)
24396 // (CMOV F, T, ((cc1 & cc2) != 0)) -> (CMOV (CMOV T, F, !cc1), F, !cc2)
24398 // This combine lets us generate:
24399 // cmovcc1 (jcc1 if we don't have CMOV)
24405 // cmovne (jne if we don't have CMOV)
24406 // When we can't use the CMOV instruction, it might increase branch
24408 // When we can use CMOV, or when there is no mispredict, this improves
24409 // throughput and reduces register pressure.
24411 if (CC == X86::COND_NE) {
24413 X86::CondCode CC0, CC1;
24415 if (checkBoolTestAndOrSetCCCombine(Cond, CC0, CC1, Flags, isAndSetCC)) {
24417 std::swap(FalseOp, TrueOp);
24418 CC0 = X86::GetOppositeBranchCondition(CC0);
24419 CC1 = X86::GetOppositeBranchCondition(CC1);
24422 SDValue LOps[] = {FalseOp, TrueOp, DAG.getConstant(CC0, DL, MVT::i8),
24424 SDValue LCMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), LOps);
24425 SDValue Ops[] = {LCMOV, TrueOp, DAG.getConstant(CC1, DL, MVT::i8), Flags};
24426 SDValue CMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
24427 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SDValue(CMOV.getNode(), 1));
24435 /// PerformMulCombine - Optimize a single multiply with constant into two
24436 /// in order to implement it with two cheaper instructions, e.g.
24437 /// LEA + SHL, LEA + LEA.
24438 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
24439 TargetLowering::DAGCombinerInfo &DCI) {
24440 // An imul is usually smaller than the alternative sequence.
24441 if (DAG.getMachineFunction().getFunction()->optForMinSize())
24444 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
24447 EVT VT = N->getValueType(0);
24448 if (VT != MVT::i64 && VT != MVT::i32)
24451 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
24454 uint64_t MulAmt = C->getZExtValue();
24455 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
24458 uint64_t MulAmt1 = 0;
24459 uint64_t MulAmt2 = 0;
24460 if ((MulAmt % 9) == 0) {
24462 MulAmt2 = MulAmt / 9;
24463 } else if ((MulAmt % 5) == 0) {
24465 MulAmt2 = MulAmt / 5;
24466 } else if ((MulAmt % 3) == 0) {
24468 MulAmt2 = MulAmt / 3;
24471 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
24474 if (isPowerOf2_64(MulAmt2) &&
24475 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
24476 // If second multiplifer is pow2, issue it first. We want the multiply by
24477 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
24479 std::swap(MulAmt1, MulAmt2);
24482 if (isPowerOf2_64(MulAmt1))
24483 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
24484 DAG.getConstant(Log2_64(MulAmt1), DL, MVT::i8));
24486 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
24487 DAG.getConstant(MulAmt1, DL, VT));
24489 if (isPowerOf2_64(MulAmt2))
24490 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
24491 DAG.getConstant(Log2_64(MulAmt2), DL, MVT::i8));
24493 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
24494 DAG.getConstant(MulAmt2, DL, VT));
24496 // Do not add new nodes to DAG combiner worklist.
24497 DCI.CombineTo(N, NewMul, false);
24502 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
24503 SDValue N0 = N->getOperand(0);
24504 SDValue N1 = N->getOperand(1);
24505 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
24506 EVT VT = N0.getValueType();
24508 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
24509 // since the result of setcc_c is all zero's or all ones.
24510 if (VT.isInteger() && !VT.isVector() &&
24511 N1C && N0.getOpcode() == ISD::AND &&
24512 N0.getOperand(1).getOpcode() == ISD::Constant) {
24513 SDValue N00 = N0.getOperand(0);
24514 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
24515 APInt ShAmt = N1C->getAPIntValue();
24516 Mask = Mask.shl(ShAmt);
24517 bool MaskOK = false;
24518 // We can handle cases concerning bit-widening nodes containing setcc_c if
24519 // we carefully interrogate the mask to make sure we are semantics
24521 // The transform is not safe if the result of C1 << C2 exceeds the bitwidth
24522 // of the underlying setcc_c operation if the setcc_c was zero extended.
24523 // Consider the following example:
24524 // zext(setcc_c) -> i32 0x0000FFFF
24525 // c1 -> i32 0x0000FFFF
24526 // c2 -> i32 0x00000001
24527 // (shl (and (setcc_c), c1), c2) -> i32 0x0001FFFE
24528 // (and setcc_c, (c1 << c2)) -> i32 0x0000FFFE
24529 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
24531 } else if (N00.getOpcode() == ISD::SIGN_EXTEND &&
24532 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
24534 } else if ((N00.getOpcode() == ISD::ZERO_EXTEND ||
24535 N00.getOpcode() == ISD::ANY_EXTEND) &&
24536 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
24537 MaskOK = Mask.isIntN(N00.getOperand(0).getValueSizeInBits());
24539 if (MaskOK && Mask != 0) {
24541 return DAG.getNode(ISD::AND, DL, VT, N00, DAG.getConstant(Mask, DL, VT));
24545 // Hardware support for vector shifts is sparse which makes us scalarize the
24546 // vector operations in many cases. Also, on sandybridge ADD is faster than
24548 // (shl V, 1) -> add V,V
24549 if (auto *N1BV = dyn_cast<BuildVectorSDNode>(N1))
24550 if (auto *N1SplatC = N1BV->getConstantSplatNode()) {
24551 assert(N0.getValueType().isVector() && "Invalid vector shift type");
24552 // We shift all of the values by one. In many cases we do not have
24553 // hardware support for this operation. This is better expressed as an ADD
24555 if (N1SplatC->getAPIntValue() == 1)
24556 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N0);
24562 /// \brief Returns a vector of 0s if the node in input is a vector logical
24563 /// shift by a constant amount which is known to be bigger than or equal
24564 /// to the vector element size in bits.
24565 static SDValue performShiftToAllZeros(SDNode *N, SelectionDAG &DAG,
24566 const X86Subtarget *Subtarget) {
24567 EVT VT = N->getValueType(0);
24569 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16 &&
24570 (!Subtarget->hasInt256() ||
24571 (VT != MVT::v4i64 && VT != MVT::v8i32 && VT != MVT::v16i16)))
24574 SDValue Amt = N->getOperand(1);
24576 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Amt))
24577 if (auto *AmtSplat = AmtBV->getConstantSplatNode()) {
24578 APInt ShiftAmt = AmtSplat->getAPIntValue();
24579 unsigned MaxAmount =
24580 VT.getSimpleVT().getVectorElementType().getSizeInBits();
24582 // SSE2/AVX2 logical shifts always return a vector of 0s
24583 // if the shift amount is bigger than or equal to
24584 // the element size. The constant shift amount will be
24585 // encoded as a 8-bit immediate.
24586 if (ShiftAmt.trunc(8).uge(MaxAmount))
24587 return getZeroVector(VT, Subtarget, DAG, DL);
24593 /// PerformShiftCombine - Combine shifts.
24594 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
24595 TargetLowering::DAGCombinerInfo &DCI,
24596 const X86Subtarget *Subtarget) {
24597 if (N->getOpcode() == ISD::SHL)
24598 if (SDValue V = PerformSHLCombine(N, DAG))
24601 // Try to fold this logical shift into a zero vector.
24602 if (N->getOpcode() != ISD::SRA)
24603 if (SDValue V = performShiftToAllZeros(N, DAG, Subtarget))
24609 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
24610 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
24611 // and friends. Likewise for OR -> CMPNEQSS.
24612 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
24613 TargetLowering::DAGCombinerInfo &DCI,
24614 const X86Subtarget *Subtarget) {
24617 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
24618 // we're requiring SSE2 for both.
24619 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
24620 SDValue N0 = N->getOperand(0);
24621 SDValue N1 = N->getOperand(1);
24622 SDValue CMP0 = N0->getOperand(1);
24623 SDValue CMP1 = N1->getOperand(1);
24626 // The SETCCs should both refer to the same CMP.
24627 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
24630 SDValue CMP00 = CMP0->getOperand(0);
24631 SDValue CMP01 = CMP0->getOperand(1);
24632 EVT VT = CMP00.getValueType();
24634 if (VT == MVT::f32 || VT == MVT::f64) {
24635 bool ExpectingFlags = false;
24636 // Check for any users that want flags:
24637 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
24638 !ExpectingFlags && UI != UE; ++UI)
24639 switch (UI->getOpcode()) {
24644 ExpectingFlags = true;
24646 case ISD::CopyToReg:
24647 case ISD::SIGN_EXTEND:
24648 case ISD::ZERO_EXTEND:
24649 case ISD::ANY_EXTEND:
24653 if (!ExpectingFlags) {
24654 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
24655 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
24657 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
24658 X86::CondCode tmp = cc0;
24663 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
24664 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
24665 // FIXME: need symbolic constants for these magic numbers.
24666 // See X86ATTInstPrinter.cpp:printSSECC().
24667 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
24668 if (Subtarget->hasAVX512()) {
24669 SDValue FSetCC = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CMP00,
24671 DAG.getConstant(x86cc, DL, MVT::i8));
24672 if (N->getValueType(0) != MVT::i1)
24673 return DAG.getNode(ISD::ZERO_EXTEND, DL, N->getValueType(0),
24677 SDValue OnesOrZeroesF = DAG.getNode(X86ISD::FSETCC, DL,
24678 CMP00.getValueType(), CMP00, CMP01,
24679 DAG.getConstant(x86cc, DL,
24682 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
24683 MVT IntVT = is64BitFP ? MVT::i64 : MVT::i32;
24685 if (is64BitFP && !Subtarget->is64Bit()) {
24686 // On a 32-bit target, we cannot bitcast the 64-bit float to a
24687 // 64-bit integer, since that's not a legal type. Since
24688 // OnesOrZeroesF is all ones of all zeroes, we don't need all the
24689 // bits, but can do this little dance to extract the lowest 32 bits
24690 // and work with those going forward.
24691 SDValue Vector64 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64,
24693 SDValue Vector32 = DAG.getBitcast(MVT::v4f32, Vector64);
24694 OnesOrZeroesF = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32,
24695 Vector32, DAG.getIntPtrConstant(0, DL));
24699 SDValue OnesOrZeroesI = DAG.getBitcast(IntVT, OnesOrZeroesF);
24700 SDValue ANDed = DAG.getNode(ISD::AND, DL, IntVT, OnesOrZeroesI,
24701 DAG.getConstant(1, DL, IntVT));
24702 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8,
24704 return OneBitOfTruth;
24712 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
24713 /// so it can be folded inside ANDNP.
24714 static bool CanFoldXORWithAllOnes(const SDNode *N) {
24715 EVT VT = N->getValueType(0);
24717 // Match direct AllOnes for 128 and 256-bit vectors
24718 if (ISD::isBuildVectorAllOnes(N))
24721 // Look through a bit convert.
24722 if (N->getOpcode() == ISD::BITCAST)
24723 N = N->getOperand(0).getNode();
24725 // Sometimes the operand may come from a insert_subvector building a 256-bit
24727 if (VT.is256BitVector() &&
24728 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
24729 SDValue V1 = N->getOperand(0);
24730 SDValue V2 = N->getOperand(1);
24732 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
24733 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
24734 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
24735 ISD::isBuildVectorAllOnes(V2.getNode()))
24742 // On AVX/AVX2 the type v8i1 is legalized to v8i16, which is an XMM sized
24743 // register. In most cases we actually compare or select YMM-sized registers
24744 // and mixing the two types creates horrible code. This method optimizes
24745 // some of the transition sequences.
24746 static SDValue WidenMaskArithmetic(SDNode *N, SelectionDAG &DAG,
24747 TargetLowering::DAGCombinerInfo &DCI,
24748 const X86Subtarget *Subtarget) {
24749 EVT VT = N->getValueType(0);
24750 if (!VT.is256BitVector())
24753 assert((N->getOpcode() == ISD::ANY_EXTEND ||
24754 N->getOpcode() == ISD::ZERO_EXTEND ||
24755 N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node");
24757 SDValue Narrow = N->getOperand(0);
24758 EVT NarrowVT = Narrow->getValueType(0);
24759 if (!NarrowVT.is128BitVector())
24762 if (Narrow->getOpcode() != ISD::XOR &&
24763 Narrow->getOpcode() != ISD::AND &&
24764 Narrow->getOpcode() != ISD::OR)
24767 SDValue N0 = Narrow->getOperand(0);
24768 SDValue N1 = Narrow->getOperand(1);
24771 // The Left side has to be a trunc.
24772 if (N0.getOpcode() != ISD::TRUNCATE)
24775 // The type of the truncated inputs.
24776 EVT WideVT = N0->getOperand(0)->getValueType(0);
24780 // The right side has to be a 'trunc' or a constant vector.
24781 bool RHSTrunc = N1.getOpcode() == ISD::TRUNCATE;
24782 ConstantSDNode *RHSConstSplat = nullptr;
24783 if (auto *RHSBV = dyn_cast<BuildVectorSDNode>(N1))
24784 RHSConstSplat = RHSBV->getConstantSplatNode();
24785 if (!RHSTrunc && !RHSConstSplat)
24788 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24790 if (!TLI.isOperationLegalOrPromote(Narrow->getOpcode(), WideVT))
24793 // Set N0 and N1 to hold the inputs to the new wide operation.
24794 N0 = N0->getOperand(0);
24795 if (RHSConstSplat) {
24796 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT.getVectorElementType(),
24797 SDValue(RHSConstSplat, 0));
24798 SmallVector<SDValue, 8> C(WideVT.getVectorNumElements(), N1);
24799 N1 = DAG.getNode(ISD::BUILD_VECTOR, DL, WideVT, C);
24800 } else if (RHSTrunc) {
24801 N1 = N1->getOperand(0);
24804 // Generate the wide operation.
24805 SDValue Op = DAG.getNode(Narrow->getOpcode(), DL, WideVT, N0, N1);
24806 unsigned Opcode = N->getOpcode();
24808 case ISD::ANY_EXTEND:
24810 case ISD::ZERO_EXTEND: {
24811 unsigned InBits = NarrowVT.getScalarSizeInBits();
24812 APInt Mask = APInt::getAllOnesValue(InBits);
24813 Mask = Mask.zext(VT.getScalarSizeInBits());
24814 return DAG.getNode(ISD::AND, DL, VT,
24815 Op, DAG.getConstant(Mask, DL, VT));
24817 case ISD::SIGN_EXTEND:
24818 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT,
24819 Op, DAG.getValueType(NarrowVT));
24821 llvm_unreachable("Unexpected opcode");
24825 static SDValue VectorZextCombine(SDNode *N, SelectionDAG &DAG,
24826 TargetLowering::DAGCombinerInfo &DCI,
24827 const X86Subtarget *Subtarget) {
24828 SDValue N0 = N->getOperand(0);
24829 SDValue N1 = N->getOperand(1);
24832 // A vector zext_in_reg may be represented as a shuffle,
24833 // feeding into a bitcast (this represents anyext) feeding into
24834 // an and with a mask.
24835 // We'd like to try to combine that into a shuffle with zero
24836 // plus a bitcast, removing the and.
24837 if (N0.getOpcode() != ISD::BITCAST ||
24838 N0.getOperand(0).getOpcode() != ISD::VECTOR_SHUFFLE)
24841 // The other side of the AND should be a splat of 2^C, where C
24842 // is the number of bits in the source type.
24843 if (N1.getOpcode() == ISD::BITCAST)
24844 N1 = N1.getOperand(0);
24845 if (N1.getOpcode() != ISD::BUILD_VECTOR)
24847 BuildVectorSDNode *Vector = cast<BuildVectorSDNode>(N1);
24849 ShuffleVectorSDNode *Shuffle = cast<ShuffleVectorSDNode>(N0.getOperand(0));
24850 EVT SrcType = Shuffle->getValueType(0);
24852 // We expect a single-source shuffle
24853 if (Shuffle->getOperand(1)->getOpcode() != ISD::UNDEF)
24856 unsigned SrcSize = SrcType.getScalarSizeInBits();
24858 APInt SplatValue, SplatUndef;
24859 unsigned SplatBitSize;
24861 if (!Vector->isConstantSplat(SplatValue, SplatUndef,
24862 SplatBitSize, HasAnyUndefs))
24865 unsigned ResSize = N1.getValueType().getScalarSizeInBits();
24866 // Make sure the splat matches the mask we expect
24867 if (SplatBitSize > ResSize ||
24868 (SplatValue + 1).exactLogBase2() != (int)SrcSize)
24871 // Make sure the input and output size make sense
24872 if (SrcSize >= ResSize || ResSize % SrcSize)
24875 // We expect a shuffle of the form <0, u, u, u, 1, u, u, u...>
24876 // The number of u's between each two values depends on the ratio between
24877 // the source and dest type.
24878 unsigned ZextRatio = ResSize / SrcSize;
24879 bool IsZext = true;
24880 for (unsigned i = 0; i < SrcType.getVectorNumElements(); ++i) {
24881 if (i % ZextRatio) {
24882 if (Shuffle->getMaskElt(i) > 0) {
24888 if (Shuffle->getMaskElt(i) != (int)(i / ZextRatio)) {
24889 // Expected element number
24899 // Ok, perform the transformation - replace the shuffle with
24900 // a shuffle of the form <0, k, k, k, 1, k, k, k> with zero
24901 // (instead of undef) where the k elements come from the zero vector.
24902 SmallVector<int, 8> Mask;
24903 unsigned NumElems = SrcType.getVectorNumElements();
24904 for (unsigned i = 0; i < NumElems; ++i)
24906 Mask.push_back(NumElems);
24908 Mask.push_back(i / ZextRatio);
24910 SDValue NewShuffle = DAG.getVectorShuffle(Shuffle->getValueType(0), DL,
24911 Shuffle->getOperand(0), DAG.getConstant(0, DL, SrcType), Mask);
24912 return DAG.getBitcast(N0.getValueType(), NewShuffle);
24915 /// If both input operands of a logic op are being cast from floating point
24916 /// types, try to convert this into a floating point logic node to avoid
24917 /// unnecessary moves from SSE to integer registers.
24918 static SDValue convertIntLogicToFPLogic(SDNode *N, SelectionDAG &DAG,
24919 const X86Subtarget *Subtarget) {
24920 unsigned FPOpcode = ISD::DELETED_NODE;
24921 if (N->getOpcode() == ISD::AND)
24922 FPOpcode = X86ISD::FAND;
24923 else if (N->getOpcode() == ISD::OR)
24924 FPOpcode = X86ISD::FOR;
24925 else if (N->getOpcode() == ISD::XOR)
24926 FPOpcode = X86ISD::FXOR;
24928 assert(FPOpcode != ISD::DELETED_NODE &&
24929 "Unexpected input node for FP logic conversion");
24931 EVT VT = N->getValueType(0);
24932 SDValue N0 = N->getOperand(0);
24933 SDValue N1 = N->getOperand(1);
24935 if (N0.getOpcode() == ISD::BITCAST && N1.getOpcode() == ISD::BITCAST &&
24936 ((Subtarget->hasSSE1() && VT == MVT::i32) ||
24937 (Subtarget->hasSSE2() && VT == MVT::i64))) {
24938 SDValue N00 = N0.getOperand(0);
24939 SDValue N10 = N1.getOperand(0);
24940 EVT N00Type = N00.getValueType();
24941 EVT N10Type = N10.getValueType();
24942 if (N00Type.isFloatingPoint() && N10Type.isFloatingPoint()) {
24943 SDValue FPLogic = DAG.getNode(FPOpcode, DL, N00Type, N00, N10);
24944 return DAG.getBitcast(VT, FPLogic);
24950 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
24951 TargetLowering::DAGCombinerInfo &DCI,
24952 const X86Subtarget *Subtarget) {
24953 if (DCI.isBeforeLegalizeOps())
24956 if (SDValue Zext = VectorZextCombine(N, DAG, DCI, Subtarget))
24959 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
24962 if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
24965 EVT VT = N->getValueType(0);
24966 SDValue N0 = N->getOperand(0);
24967 SDValue N1 = N->getOperand(1);
24970 // Create BEXTR instructions
24971 // BEXTR is ((X >> imm) & (2**size-1))
24972 if (VT == MVT::i32 || VT == MVT::i64) {
24973 // Check for BEXTR.
24974 if ((Subtarget->hasBMI() || Subtarget->hasTBM()) &&
24975 (N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::SRL)) {
24976 ConstantSDNode *MaskNode = dyn_cast<ConstantSDNode>(N1);
24977 ConstantSDNode *ShiftNode = dyn_cast<ConstantSDNode>(N0.getOperand(1));
24978 if (MaskNode && ShiftNode) {
24979 uint64_t Mask = MaskNode->getZExtValue();
24980 uint64_t Shift = ShiftNode->getZExtValue();
24981 if (isMask_64(Mask)) {
24982 uint64_t MaskSize = countPopulation(Mask);
24983 if (Shift + MaskSize <= VT.getSizeInBits())
24984 return DAG.getNode(X86ISD::BEXTR, DL, VT, N0.getOperand(0),
24985 DAG.getConstant(Shift | (MaskSize << 8), DL,
24994 // Want to form ANDNP nodes:
24995 // 1) In the hopes of then easily combining them with OR and AND nodes
24996 // to form PBLEND/PSIGN.
24997 // 2) To match ANDN packed intrinsics
24998 if (VT != MVT::v2i64 && VT != MVT::v4i64)
25001 // Check LHS for vnot
25002 if (N0.getOpcode() == ISD::XOR &&
25003 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
25004 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
25005 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
25007 // Check RHS for vnot
25008 if (N1.getOpcode() == ISD::XOR &&
25009 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
25010 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
25011 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
25016 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
25017 TargetLowering::DAGCombinerInfo &DCI,
25018 const X86Subtarget *Subtarget) {
25019 if (DCI.isBeforeLegalizeOps())
25022 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
25025 if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
25028 SDValue N0 = N->getOperand(0);
25029 SDValue N1 = N->getOperand(1);
25030 EVT VT = N->getValueType(0);
25032 // look for psign/blend
25033 if (VT == MVT::v2i64 || VT == MVT::v4i64) {
25034 if (!Subtarget->hasSSSE3() ||
25035 (VT == MVT::v4i64 && !Subtarget->hasInt256()))
25038 // Canonicalize pandn to RHS
25039 if (N0.getOpcode() == X86ISD::ANDNP)
25041 // or (and (m, y), (pandn m, x))
25042 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
25043 SDValue Mask = N1.getOperand(0);
25044 SDValue X = N1.getOperand(1);
25046 if (N0.getOperand(0) == Mask)
25047 Y = N0.getOperand(1);
25048 if (N0.getOperand(1) == Mask)
25049 Y = N0.getOperand(0);
25051 // Check to see if the mask appeared in both the AND and ANDNP and
25055 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
25056 // Look through mask bitcast.
25057 if (Mask.getOpcode() == ISD::BITCAST)
25058 Mask = Mask.getOperand(0);
25059 if (X.getOpcode() == ISD::BITCAST)
25060 X = X.getOperand(0);
25061 if (Y.getOpcode() == ISD::BITCAST)
25062 Y = Y.getOperand(0);
25064 EVT MaskVT = Mask.getValueType();
25066 // Validate that the Mask operand is a vector sra node.
25067 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
25068 // there is no psrai.b
25069 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
25070 unsigned SraAmt = ~0;
25071 if (Mask.getOpcode() == ISD::SRA) {
25072 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Mask.getOperand(1)))
25073 if (auto *AmtConst = AmtBV->getConstantSplatNode())
25074 SraAmt = AmtConst->getZExtValue();
25075 } else if (Mask.getOpcode() == X86ISD::VSRAI) {
25076 SDValue SraC = Mask.getOperand(1);
25077 SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
25079 if ((SraAmt + 1) != EltBits)
25084 // Now we know we at least have a plendvb with the mask val. See if
25085 // we can form a psignb/w/d.
25086 // psign = x.type == y.type == mask.type && y = sub(0, x);
25087 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
25088 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
25089 X.getValueType() == MaskVT && Y.getValueType() == MaskVT) {
25090 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
25091 "Unsupported VT for PSIGN");
25092 Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask.getOperand(0));
25093 return DAG.getBitcast(VT, Mask);
25095 // PBLENDVB only available on SSE 4.1
25096 if (!Subtarget->hasSSE41())
25099 MVT BlendVT = (VT == MVT::v4i64) ? MVT::v32i8 : MVT::v16i8;
25101 X = DAG.getBitcast(BlendVT, X);
25102 Y = DAG.getBitcast(BlendVT, Y);
25103 Mask = DAG.getBitcast(BlendVT, Mask);
25104 Mask = DAG.getNode(ISD::VSELECT, DL, BlendVT, Mask, Y, X);
25105 return DAG.getBitcast(VT, Mask);
25109 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
25112 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
25113 bool OptForSize = DAG.getMachineFunction().getFunction()->optForSize();
25115 // SHLD/SHRD instructions have lower register pressure, but on some
25116 // platforms they have higher latency than the equivalent
25117 // series of shifts/or that would otherwise be generated.
25118 // Don't fold (or (x << c) | (y >> (64 - c))) if SHLD/SHRD instructions
25119 // have higher latencies and we are not optimizing for size.
25120 if (!OptForSize && Subtarget->isSHLDSlow())
25123 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
25125 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
25127 if (!N0.hasOneUse() || !N1.hasOneUse())
25130 SDValue ShAmt0 = N0.getOperand(1);
25131 if (ShAmt0.getValueType() != MVT::i8)
25133 SDValue ShAmt1 = N1.getOperand(1);
25134 if (ShAmt1.getValueType() != MVT::i8)
25136 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
25137 ShAmt0 = ShAmt0.getOperand(0);
25138 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
25139 ShAmt1 = ShAmt1.getOperand(0);
25142 unsigned Opc = X86ISD::SHLD;
25143 SDValue Op0 = N0.getOperand(0);
25144 SDValue Op1 = N1.getOperand(0);
25145 if (ShAmt0.getOpcode() == ISD::SUB) {
25146 Opc = X86ISD::SHRD;
25147 std::swap(Op0, Op1);
25148 std::swap(ShAmt0, ShAmt1);
25151 unsigned Bits = VT.getSizeInBits();
25152 if (ShAmt1.getOpcode() == ISD::SUB) {
25153 SDValue Sum = ShAmt1.getOperand(0);
25154 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
25155 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
25156 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
25157 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
25158 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
25159 return DAG.getNode(Opc, DL, VT,
25161 DAG.getNode(ISD::TRUNCATE, DL,
25164 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
25165 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
25167 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
25168 return DAG.getNode(Opc, DL, VT,
25169 N0.getOperand(0), N1.getOperand(0),
25170 DAG.getNode(ISD::TRUNCATE, DL,
25177 // Generate NEG and CMOV for integer abs.
25178 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
25179 EVT VT = N->getValueType(0);
25181 // Since X86 does not have CMOV for 8-bit integer, we don't convert
25182 // 8-bit integer abs to NEG and CMOV.
25183 if (VT.isInteger() && VT.getSizeInBits() == 8)
25186 SDValue N0 = N->getOperand(0);
25187 SDValue N1 = N->getOperand(1);
25190 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
25191 // and change it to SUB and CMOV.
25192 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
25193 N0.getOpcode() == ISD::ADD &&
25194 N0.getOperand(1) == N1 &&
25195 N1.getOpcode() == ISD::SRA &&
25196 N1.getOperand(0) == N0.getOperand(0))
25197 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
25198 if (Y1C->getAPIntValue() == VT.getSizeInBits()-1) {
25199 // Generate SUB & CMOV.
25200 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32),
25201 DAG.getConstant(0, DL, VT), N0.getOperand(0));
25203 SDValue Ops[] = { N0.getOperand(0), Neg,
25204 DAG.getConstant(X86::COND_GE, DL, MVT::i8),
25205 SDValue(Neg.getNode(), 1) };
25206 return DAG.getNode(X86ISD::CMOV, DL, DAG.getVTList(VT, MVT::Glue), Ops);
25211 // Try to turn tests against the signbit in the form of:
25212 // XOR(TRUNCATE(SRL(X, size(X)-1)), 1)
25215 static SDValue foldXorTruncShiftIntoCmp(SDNode *N, SelectionDAG &DAG) {
25216 // This is only worth doing if the output type is i8.
25217 if (N->getValueType(0) != MVT::i8)
25220 SDValue N0 = N->getOperand(0);
25221 SDValue N1 = N->getOperand(1);
25223 // We should be performing an xor against a truncated shift.
25224 if (N0.getOpcode() != ISD::TRUNCATE || !N0.hasOneUse())
25227 // Make sure we are performing an xor against one.
25228 if (!isa<ConstantSDNode>(N1) || !cast<ConstantSDNode>(N1)->isOne())
25231 // SetCC on x86 zero extends so only act on this if it's a logical shift.
25232 SDValue Shift = N0.getOperand(0);
25233 if (Shift.getOpcode() != ISD::SRL || !Shift.hasOneUse())
25236 // Make sure we are truncating from one of i16, i32 or i64.
25237 EVT ShiftTy = Shift.getValueType();
25238 if (ShiftTy != MVT::i16 && ShiftTy != MVT::i32 && ShiftTy != MVT::i64)
25241 // Make sure the shift amount extracts the sign bit.
25242 if (!isa<ConstantSDNode>(Shift.getOperand(1)) ||
25243 Shift.getConstantOperandVal(1) != ShiftTy.getSizeInBits() - 1)
25246 // Create a greater-than comparison against -1.
25247 // N.B. Using SETGE against 0 works but we want a canonical looking
25248 // comparison, using SETGT matches up with what TranslateX86CC.
25250 SDValue ShiftOp = Shift.getOperand(0);
25251 EVT ShiftOpTy = ShiftOp.getValueType();
25252 SDValue Cond = DAG.getSetCC(DL, MVT::i8, ShiftOp,
25253 DAG.getConstant(-1, DL, ShiftOpTy), ISD::SETGT);
25257 static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
25258 TargetLowering::DAGCombinerInfo &DCI,
25259 const X86Subtarget *Subtarget) {
25260 if (DCI.isBeforeLegalizeOps())
25263 if (SDValue RV = foldXorTruncShiftIntoCmp(N, DAG))
25266 if (Subtarget->hasCMov())
25267 if (SDValue RV = performIntegerAbsCombine(N, DAG))
25270 if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
25276 /// PerformLOADCombine - Do target-specific dag combines on LOAD nodes.
25277 static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
25278 TargetLowering::DAGCombinerInfo &DCI,
25279 const X86Subtarget *Subtarget) {
25280 LoadSDNode *Ld = cast<LoadSDNode>(N);
25281 EVT RegVT = Ld->getValueType(0);
25282 EVT MemVT = Ld->getMemoryVT();
25284 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25286 // For chips with slow 32-byte unaligned loads, break the 32-byte operation
25287 // into two 16-byte operations.
25288 ISD::LoadExtType Ext = Ld->getExtensionType();
25290 unsigned AddressSpace = Ld->getAddressSpace();
25291 unsigned Alignment = Ld->getAlignment();
25292 if (RegVT.is256BitVector() && !DCI.isBeforeLegalizeOps() &&
25293 Ext == ISD::NON_EXTLOAD &&
25294 TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), RegVT,
25295 AddressSpace, Alignment, &Fast) && !Fast) {
25296 unsigned NumElems = RegVT.getVectorNumElements();
25300 SDValue Ptr = Ld->getBasePtr();
25301 SDValue Increment =
25302 DAG.getConstant(16, dl, TLI.getPointerTy(DAG.getDataLayout()));
25304 EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
25306 SDValue Load1 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
25307 Ld->getPointerInfo(), Ld->isVolatile(),
25308 Ld->isNonTemporal(), Ld->isInvariant(),
25310 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
25311 SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
25312 Ld->getPointerInfo(), Ld->isVolatile(),
25313 Ld->isNonTemporal(), Ld->isInvariant(),
25314 std::min(16U, Alignment));
25315 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
25317 Load2.getValue(1));
25319 SDValue NewVec = DAG.getUNDEF(RegVT);
25320 NewVec = Insert128BitVector(NewVec, Load1, 0, DAG, dl);
25321 NewVec = Insert128BitVector(NewVec, Load2, NumElems/2, DAG, dl);
25322 return DCI.CombineTo(N, NewVec, TF, true);
25328 /// PerformMLOADCombine - Resolve extending loads
25329 static SDValue PerformMLOADCombine(SDNode *N, SelectionDAG &DAG,
25330 TargetLowering::DAGCombinerInfo &DCI,
25331 const X86Subtarget *Subtarget) {
25332 MaskedLoadSDNode *Mld = cast<MaskedLoadSDNode>(N);
25333 if (Mld->getExtensionType() != ISD::SEXTLOAD)
25336 EVT VT = Mld->getValueType(0);
25337 unsigned NumElems = VT.getVectorNumElements();
25338 EVT LdVT = Mld->getMemoryVT();
25341 assert(LdVT != VT && "Cannot extend to the same type");
25342 unsigned ToSz = VT.getVectorElementType().getSizeInBits();
25343 unsigned FromSz = LdVT.getVectorElementType().getSizeInBits();
25344 // From, To sizes and ElemCount must be pow of two
25345 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
25346 "Unexpected size for extending masked load");
25348 unsigned SizeRatio = ToSz / FromSz;
25349 assert(SizeRatio * NumElems * FromSz == VT.getSizeInBits());
25351 // Create a type on which we perform the shuffle
25352 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
25353 LdVT.getScalarType(), NumElems*SizeRatio);
25354 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
25356 // Convert Src0 value
25357 SDValue WideSrc0 = DAG.getBitcast(WideVecVT, Mld->getSrc0());
25358 if (Mld->getSrc0().getOpcode() != ISD::UNDEF) {
25359 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
25360 for (unsigned i = 0; i != NumElems; ++i)
25361 ShuffleVec[i] = i * SizeRatio;
25363 // Can't shuffle using an illegal type.
25364 assert(DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) &&
25365 "WideVecVT should be legal");
25366 WideSrc0 = DAG.getVectorShuffle(WideVecVT, dl, WideSrc0,
25367 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
25369 // Prepare the new mask
25371 SDValue Mask = Mld->getMask();
25372 if (Mask.getValueType() == VT) {
25373 // Mask and original value have the same type
25374 NewMask = DAG.getBitcast(WideVecVT, Mask);
25375 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
25376 for (unsigned i = 0; i != NumElems; ++i)
25377 ShuffleVec[i] = i * SizeRatio;
25378 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
25379 ShuffleVec[i] = NumElems*SizeRatio;
25380 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
25381 DAG.getConstant(0, dl, WideVecVT),
25385 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
25386 unsigned WidenNumElts = NumElems*SizeRatio;
25387 unsigned MaskNumElts = VT.getVectorNumElements();
25388 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
25391 unsigned NumConcat = WidenNumElts / MaskNumElts;
25392 SmallVector<SDValue, 16> Ops(NumConcat);
25393 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
25395 for (unsigned i = 1; i != NumConcat; ++i)
25398 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
25401 SDValue WideLd = DAG.getMaskedLoad(WideVecVT, dl, Mld->getChain(),
25402 Mld->getBasePtr(), NewMask, WideSrc0,
25403 Mld->getMemoryVT(), Mld->getMemOperand(),
25405 SDValue NewVec = DAG.getNode(X86ISD::VSEXT, dl, VT, WideLd);
25406 return DCI.CombineTo(N, NewVec, WideLd.getValue(1), true);
25408 /// PerformMSTORECombine - Resolve truncating stores
25409 static SDValue PerformMSTORECombine(SDNode *N, SelectionDAG &DAG,
25410 const X86Subtarget *Subtarget) {
25411 MaskedStoreSDNode *Mst = cast<MaskedStoreSDNode>(N);
25412 if (!Mst->isTruncatingStore())
25415 EVT VT = Mst->getValue().getValueType();
25416 unsigned NumElems = VT.getVectorNumElements();
25417 EVT StVT = Mst->getMemoryVT();
25420 assert(StVT != VT && "Cannot truncate to the same type");
25421 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
25422 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
25424 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25426 // The truncating store is legal in some cases. For example
25427 // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
25428 // are designated for truncate store.
25429 // In this case we don't need any further transformations.
25430 if (TLI.isTruncStoreLegal(VT, StVT))
25433 // From, To sizes and ElemCount must be pow of two
25434 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
25435 "Unexpected size for truncating masked store");
25436 // We are going to use the original vector elt for storing.
25437 // Accumulated smaller vector elements must be a multiple of the store size.
25438 assert (((NumElems * FromSz) % ToSz) == 0 &&
25439 "Unexpected ratio for truncating masked store");
25441 unsigned SizeRatio = FromSz / ToSz;
25442 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
25444 // Create a type on which we perform the shuffle
25445 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
25446 StVT.getScalarType(), NumElems*SizeRatio);
25448 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
25450 SDValue WideVec = DAG.getBitcast(WideVecVT, Mst->getValue());
25451 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
25452 for (unsigned i = 0; i != NumElems; ++i)
25453 ShuffleVec[i] = i * SizeRatio;
25455 // Can't shuffle using an illegal type.
25456 assert(DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) &&
25457 "WideVecVT should be legal");
25459 SDValue TruncatedVal = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
25460 DAG.getUNDEF(WideVecVT),
25464 SDValue Mask = Mst->getMask();
25465 if (Mask.getValueType() == VT) {
25466 // Mask and original value have the same type
25467 NewMask = DAG.getBitcast(WideVecVT, Mask);
25468 for (unsigned i = 0; i != NumElems; ++i)
25469 ShuffleVec[i] = i * SizeRatio;
25470 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
25471 ShuffleVec[i] = NumElems*SizeRatio;
25472 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
25473 DAG.getConstant(0, dl, WideVecVT),
25477 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
25478 unsigned WidenNumElts = NumElems*SizeRatio;
25479 unsigned MaskNumElts = VT.getVectorNumElements();
25480 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
25483 unsigned NumConcat = WidenNumElts / MaskNumElts;
25484 SmallVector<SDValue, 16> Ops(NumConcat);
25485 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
25487 for (unsigned i = 1; i != NumConcat; ++i)
25490 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
25493 return DAG.getMaskedStore(Mst->getChain(), dl, TruncatedVal, Mst->getBasePtr(),
25494 NewMask, StVT, Mst->getMemOperand(), false);
25496 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
25497 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
25498 const X86Subtarget *Subtarget) {
25499 StoreSDNode *St = cast<StoreSDNode>(N);
25500 EVT VT = St->getValue().getValueType();
25501 EVT StVT = St->getMemoryVT();
25503 SDValue StoredVal = St->getOperand(1);
25504 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25506 // If we are saving a concatenation of two XMM registers and 32-byte stores
25507 // are slow, such as on Sandy Bridge, perform two 16-byte stores.
25509 unsigned AddressSpace = St->getAddressSpace();
25510 unsigned Alignment = St->getAlignment();
25511 if (VT.is256BitVector() && StVT == VT &&
25512 TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), VT,
25513 AddressSpace, Alignment, &Fast) && !Fast) {
25514 unsigned NumElems = VT.getVectorNumElements();
25518 SDValue Value0 = Extract128BitVector(StoredVal, 0, DAG, dl);
25519 SDValue Value1 = Extract128BitVector(StoredVal, NumElems/2, DAG, dl);
25522 DAG.getConstant(16, dl, TLI.getPointerTy(DAG.getDataLayout()));
25523 SDValue Ptr0 = St->getBasePtr();
25524 SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
25526 SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
25527 St->getPointerInfo(), St->isVolatile(),
25528 St->isNonTemporal(), Alignment);
25529 SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
25530 St->getPointerInfo(), St->isVolatile(),
25531 St->isNonTemporal(),
25532 std::min(16U, Alignment));
25533 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
25536 // Optimize trunc store (of multiple scalars) to shuffle and store.
25537 // First, pack all of the elements in one place. Next, store to memory
25538 // in fewer chunks.
25539 if (St->isTruncatingStore() && VT.isVector()) {
25540 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25541 unsigned NumElems = VT.getVectorNumElements();
25542 assert(StVT != VT && "Cannot truncate to the same type");
25543 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
25544 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
25546 // The truncating store is legal in some cases. For example
25547 // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
25548 // are designated for truncate store.
25549 // In this case we don't need any further transformations.
25550 if (TLI.isTruncStoreLegal(VT, StVT))
25553 // From, To sizes and ElemCount must be pow of two
25554 if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
25555 // We are going to use the original vector elt for storing.
25556 // Accumulated smaller vector elements must be a multiple of the store size.
25557 if (0 != (NumElems * FromSz) % ToSz) return SDValue();
25559 unsigned SizeRatio = FromSz / ToSz;
25561 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
25563 // Create a type on which we perform the shuffle
25564 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
25565 StVT.getScalarType(), NumElems*SizeRatio);
25567 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
25569 SDValue WideVec = DAG.getBitcast(WideVecVT, St->getValue());
25570 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
25571 for (unsigned i = 0; i != NumElems; ++i)
25572 ShuffleVec[i] = i * SizeRatio;
25574 // Can't shuffle using an illegal type.
25575 if (!TLI.isTypeLegal(WideVecVT))
25578 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
25579 DAG.getUNDEF(WideVecVT),
25581 // At this point all of the data is stored at the bottom of the
25582 // register. We now need to save it to mem.
25584 // Find the largest store unit
25585 MVT StoreType = MVT::i8;
25586 for (MVT Tp : MVT::integer_valuetypes()) {
25587 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToSz)
25591 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
25592 if (TLI.isTypeLegal(MVT::f64) && StoreType.getSizeInBits() < 64 &&
25593 (64 <= NumElems * ToSz))
25594 StoreType = MVT::f64;
25596 // Bitcast the original vector into a vector of store-size units
25597 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
25598 StoreType, VT.getSizeInBits()/StoreType.getSizeInBits());
25599 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
25600 SDValue ShuffWide = DAG.getBitcast(StoreVecVT, Shuff);
25601 SmallVector<SDValue, 8> Chains;
25602 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits() / 8, dl,
25603 TLI.getPointerTy(DAG.getDataLayout()));
25604 SDValue Ptr = St->getBasePtr();
25606 // Perform one or more big stores into memory.
25607 for (unsigned i=0, e=(ToSz*NumElems)/StoreType.getSizeInBits(); i!=e; ++i) {
25608 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
25609 StoreType, ShuffWide,
25610 DAG.getIntPtrConstant(i, dl));
25611 SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr,
25612 St->getPointerInfo(), St->isVolatile(),
25613 St->isNonTemporal(), St->getAlignment());
25614 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
25615 Chains.push_back(Ch);
25618 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
25621 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
25622 // the FP state in cases where an emms may be missing.
25623 // A preferable solution to the general problem is to figure out the right
25624 // places to insert EMMS. This qualifies as a quick hack.
25626 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
25627 if (VT.getSizeInBits() != 64)
25630 const Function *F = DAG.getMachineFunction().getFunction();
25631 bool NoImplicitFloatOps = F->hasFnAttribute(Attribute::NoImplicitFloat);
25633 !Subtarget->useSoftFloat() && !NoImplicitFloatOps && Subtarget->hasSSE2();
25634 if ((VT.isVector() ||
25635 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
25636 isa<LoadSDNode>(St->getValue()) &&
25637 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
25638 St->getChain().hasOneUse() && !St->isVolatile()) {
25639 SDNode* LdVal = St->getValue().getNode();
25640 LoadSDNode *Ld = nullptr;
25641 int TokenFactorIndex = -1;
25642 SmallVector<SDValue, 8> Ops;
25643 SDNode* ChainVal = St->getChain().getNode();
25644 // Must be a store of a load. We currently handle two cases: the load
25645 // is a direct child, and it's under an intervening TokenFactor. It is
25646 // possible to dig deeper under nested TokenFactors.
25647 if (ChainVal == LdVal)
25648 Ld = cast<LoadSDNode>(St->getChain());
25649 else if (St->getValue().hasOneUse() &&
25650 ChainVal->getOpcode() == ISD::TokenFactor) {
25651 for (unsigned i = 0, e = ChainVal->getNumOperands(); i != e; ++i) {
25652 if (ChainVal->getOperand(i).getNode() == LdVal) {
25653 TokenFactorIndex = i;
25654 Ld = cast<LoadSDNode>(St->getValue());
25656 Ops.push_back(ChainVal->getOperand(i));
25660 if (!Ld || !ISD::isNormalLoad(Ld))
25663 // If this is not the MMX case, i.e. we are just turning i64 load/store
25664 // into f64 load/store, avoid the transformation if there are multiple
25665 // uses of the loaded value.
25666 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
25671 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
25672 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
25674 if (Subtarget->is64Bit() || F64IsLegal) {
25675 MVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
25676 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
25677 Ld->getPointerInfo(), Ld->isVolatile(),
25678 Ld->isNonTemporal(), Ld->isInvariant(),
25679 Ld->getAlignment());
25680 SDValue NewChain = NewLd.getValue(1);
25681 if (TokenFactorIndex != -1) {
25682 Ops.push_back(NewChain);
25683 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
25685 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
25686 St->getPointerInfo(),
25687 St->isVolatile(), St->isNonTemporal(),
25688 St->getAlignment());
25691 // Otherwise, lower to two pairs of 32-bit loads / stores.
25692 SDValue LoAddr = Ld->getBasePtr();
25693 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
25694 DAG.getConstant(4, LdDL, MVT::i32));
25696 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
25697 Ld->getPointerInfo(),
25698 Ld->isVolatile(), Ld->isNonTemporal(),
25699 Ld->isInvariant(), Ld->getAlignment());
25700 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
25701 Ld->getPointerInfo().getWithOffset(4),
25702 Ld->isVolatile(), Ld->isNonTemporal(),
25704 MinAlign(Ld->getAlignment(), 4));
25706 SDValue NewChain = LoLd.getValue(1);
25707 if (TokenFactorIndex != -1) {
25708 Ops.push_back(LoLd);
25709 Ops.push_back(HiLd);
25710 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
25713 LoAddr = St->getBasePtr();
25714 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
25715 DAG.getConstant(4, StDL, MVT::i32));
25717 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
25718 St->getPointerInfo(),
25719 St->isVolatile(), St->isNonTemporal(),
25720 St->getAlignment());
25721 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
25722 St->getPointerInfo().getWithOffset(4),
25724 St->isNonTemporal(),
25725 MinAlign(St->getAlignment(), 4));
25726 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
25729 // This is similar to the above case, but here we handle a scalar 64-bit
25730 // integer store that is extracted from a vector on a 32-bit target.
25731 // If we have SSE2, then we can treat it like a floating-point double
25732 // to get past legalization. The execution dependencies fixup pass will
25733 // choose the optimal machine instruction for the store if this really is
25734 // an integer or v2f32 rather than an f64.
25735 if (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit() &&
25736 St->getOperand(1).getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
25737 SDValue OldExtract = St->getOperand(1);
25738 SDValue ExtOp0 = OldExtract.getOperand(0);
25739 unsigned VecSize = ExtOp0.getValueSizeInBits();
25740 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, VecSize / 64);
25741 SDValue BitCast = DAG.getBitcast(VecVT, ExtOp0);
25742 SDValue NewExtract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
25743 BitCast, OldExtract.getOperand(1));
25744 return DAG.getStore(St->getChain(), dl, NewExtract, St->getBasePtr(),
25745 St->getPointerInfo(), St->isVolatile(),
25746 St->isNonTemporal(), St->getAlignment());
25752 /// Return 'true' if this vector operation is "horizontal"
25753 /// and return the operands for the horizontal operation in LHS and RHS. A
25754 /// horizontal operation performs the binary operation on successive elements
25755 /// of its first operand, then on successive elements of its second operand,
25756 /// returning the resulting values in a vector. For example, if
25757 /// A = < float a0, float a1, float a2, float a3 >
25759 /// B = < float b0, float b1, float b2, float b3 >
25760 /// then the result of doing a horizontal operation on A and B is
25761 /// A horizontal-op B = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >.
25762 /// In short, LHS and RHS are inspected to see if LHS op RHS is of the form
25763 /// A horizontal-op B, for some already available A and B, and if so then LHS is
25764 /// set to A, RHS to B, and the routine returns 'true'.
25765 /// Note that the binary operation should have the property that if one of the
25766 /// operands is UNDEF then the result is UNDEF.
25767 static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool IsCommutative) {
25768 // Look for the following pattern: if
25769 // A = < float a0, float a1, float a2, float a3 >
25770 // B = < float b0, float b1, float b2, float b3 >
25772 // LHS = VECTOR_SHUFFLE A, B, <0, 2, 4, 6>
25773 // RHS = VECTOR_SHUFFLE A, B, <1, 3, 5, 7>
25774 // then LHS op RHS = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >
25775 // which is A horizontal-op B.
25777 // At least one of the operands should be a vector shuffle.
25778 if (LHS.getOpcode() != ISD::VECTOR_SHUFFLE &&
25779 RHS.getOpcode() != ISD::VECTOR_SHUFFLE)
25782 MVT VT = LHS.getSimpleValueType();
25784 assert((VT.is128BitVector() || VT.is256BitVector()) &&
25785 "Unsupported vector type for horizontal add/sub");
25787 // Handle 128 and 256-bit vector lengths. AVX defines horizontal add/sub to
25788 // operate independently on 128-bit lanes.
25789 unsigned NumElts = VT.getVectorNumElements();
25790 unsigned NumLanes = VT.getSizeInBits()/128;
25791 unsigned NumLaneElts = NumElts / NumLanes;
25792 assert((NumLaneElts % 2 == 0) &&
25793 "Vector type should have an even number of elements in each lane");
25794 unsigned HalfLaneElts = NumLaneElts/2;
25796 // View LHS in the form
25797 // LHS = VECTOR_SHUFFLE A, B, LMask
25798 // If LHS is not a shuffle then pretend it is the shuffle
25799 // LHS = VECTOR_SHUFFLE LHS, undef, <0, 1, ..., N-1>
25800 // NOTE: in what follows a default initialized SDValue represents an UNDEF of
25803 SmallVector<int, 16> LMask(NumElts);
25804 if (LHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
25805 if (LHS.getOperand(0).getOpcode() != ISD::UNDEF)
25806 A = LHS.getOperand(0);
25807 if (LHS.getOperand(1).getOpcode() != ISD::UNDEF)
25808 B = LHS.getOperand(1);
25809 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(LHS.getNode())->getMask();
25810 std::copy(Mask.begin(), Mask.end(), LMask.begin());
25812 if (LHS.getOpcode() != ISD::UNDEF)
25814 for (unsigned i = 0; i != NumElts; ++i)
25818 // Likewise, view RHS in the form
25819 // RHS = VECTOR_SHUFFLE C, D, RMask
25821 SmallVector<int, 16> RMask(NumElts);
25822 if (RHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
25823 if (RHS.getOperand(0).getOpcode() != ISD::UNDEF)
25824 C = RHS.getOperand(0);
25825 if (RHS.getOperand(1).getOpcode() != ISD::UNDEF)
25826 D = RHS.getOperand(1);
25827 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(RHS.getNode())->getMask();
25828 std::copy(Mask.begin(), Mask.end(), RMask.begin());
25830 if (RHS.getOpcode() != ISD::UNDEF)
25832 for (unsigned i = 0; i != NumElts; ++i)
25836 // Check that the shuffles are both shuffling the same vectors.
25837 if (!(A == C && B == D) && !(A == D && B == C))
25840 // If everything is UNDEF then bail out: it would be better to fold to UNDEF.
25841 if (!A.getNode() && !B.getNode())
25844 // If A and B occur in reverse order in RHS, then "swap" them (which means
25845 // rewriting the mask).
25847 ShuffleVectorSDNode::commuteMask(RMask);
25849 // At this point LHS and RHS are equivalent to
25850 // LHS = VECTOR_SHUFFLE A, B, LMask
25851 // RHS = VECTOR_SHUFFLE A, B, RMask
25852 // Check that the masks correspond to performing a horizontal operation.
25853 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
25854 for (unsigned i = 0; i != NumLaneElts; ++i) {
25855 int LIdx = LMask[i+l], RIdx = RMask[i+l];
25857 // Ignore any UNDEF components.
25858 if (LIdx < 0 || RIdx < 0 ||
25859 (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) ||
25860 (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts)))
25863 // Check that successive elements are being operated on. If not, this is
25864 // not a horizontal operation.
25865 unsigned Src = (i/HalfLaneElts); // each lane is split between srcs
25866 int Index = 2*(i%HalfLaneElts) + NumElts*Src + l;
25867 if (!(LIdx == Index && RIdx == Index + 1) &&
25868 !(IsCommutative && LIdx == Index + 1 && RIdx == Index))
25873 LHS = A.getNode() ? A : B; // If A is 'UNDEF', use B for it.
25874 RHS = B.getNode() ? B : A; // If B is 'UNDEF', use A for it.
25878 /// Do target-specific dag combines on floating point adds.
25879 static SDValue PerformFADDCombine(SDNode *N, SelectionDAG &DAG,
25880 const X86Subtarget *Subtarget) {
25881 EVT VT = N->getValueType(0);
25882 SDValue LHS = N->getOperand(0);
25883 SDValue RHS = N->getOperand(1);
25885 // Try to synthesize horizontal adds from adds of shuffles.
25886 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
25887 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
25888 isHorizontalBinOp(LHS, RHS, true))
25889 return DAG.getNode(X86ISD::FHADD, SDLoc(N), VT, LHS, RHS);
25893 /// Do target-specific dag combines on floating point subs.
25894 static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG,
25895 const X86Subtarget *Subtarget) {
25896 EVT VT = N->getValueType(0);
25897 SDValue LHS = N->getOperand(0);
25898 SDValue RHS = N->getOperand(1);
25900 // Try to synthesize horizontal subs from subs of shuffles.
25901 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
25902 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
25903 isHorizontalBinOp(LHS, RHS, false))
25904 return DAG.getNode(X86ISD::FHSUB, SDLoc(N), VT, LHS, RHS);
25908 /// Do target-specific dag combines on X86ISD::FOR and X86ISD::FXOR nodes.
25909 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG,
25910 const X86Subtarget *Subtarget) {
25911 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
25913 // F[X]OR(0.0, x) -> x
25914 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25915 if (C->getValueAPF().isPosZero())
25916 return N->getOperand(1);
25918 // F[X]OR(x, 0.0) -> x
25919 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25920 if (C->getValueAPF().isPosZero())
25921 return N->getOperand(0);
25923 EVT VT = N->getValueType(0);
25924 if (VT.is512BitVector() && !Subtarget->hasDQI()) {
25926 MVT IntScalar = MVT::getIntegerVT(VT.getScalarSizeInBits());
25927 MVT IntVT = MVT::getVectorVT(IntScalar, VT.getVectorNumElements());
25929 SDValue Op0 = DAG.getNode(ISD::BITCAST, dl, IntVT, N->getOperand(0));
25930 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, IntVT, N->getOperand(1));
25931 unsigned IntOpcode = (N->getOpcode() == X86ISD::FOR) ? ISD::OR : ISD::XOR;
25932 SDValue IntOp = DAG.getNode(IntOpcode, dl, IntVT, Op0, Op1);
25933 return DAG.getNode(ISD::BITCAST, dl, VT, IntOp);
25938 /// Do target-specific dag combines on X86ISD::FMIN and X86ISD::FMAX nodes.
25939 static SDValue PerformFMinFMaxCombine(SDNode *N, SelectionDAG &DAG) {
25940 assert(N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD::FMAX);
25942 // Only perform optimizations if UnsafeMath is used.
25943 if (!DAG.getTarget().Options.UnsafeFPMath)
25946 // If we run in unsafe-math mode, then convert the FMAX and FMIN nodes
25947 // into FMINC and FMAXC, which are Commutative operations.
25948 unsigned NewOp = 0;
25949 switch (N->getOpcode()) {
25950 default: llvm_unreachable("unknown opcode");
25951 case X86ISD::FMIN: NewOp = X86ISD::FMINC; break;
25952 case X86ISD::FMAX: NewOp = X86ISD::FMAXC; break;
25955 return DAG.getNode(NewOp, SDLoc(N), N->getValueType(0),
25956 N->getOperand(0), N->getOperand(1));
25959 /// Do target-specific dag combines on X86ISD::FAND nodes.
25960 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
25961 // FAND(0.0, x) -> 0.0
25962 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25963 if (C->getValueAPF().isPosZero())
25964 return N->getOperand(0);
25966 // FAND(x, 0.0) -> 0.0
25967 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25968 if (C->getValueAPF().isPosZero())
25969 return N->getOperand(1);
25974 /// Do target-specific dag combines on X86ISD::FANDN nodes
25975 static SDValue PerformFANDNCombine(SDNode *N, SelectionDAG &DAG) {
25976 // FANDN(0.0, x) -> x
25977 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25978 if (C->getValueAPF().isPosZero())
25979 return N->getOperand(1);
25981 // FANDN(x, 0.0) -> 0.0
25982 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25983 if (C->getValueAPF().isPosZero())
25984 return N->getOperand(1);
25989 static SDValue PerformBTCombine(SDNode *N,
25991 TargetLowering::DAGCombinerInfo &DCI) {
25992 // BT ignores high bits in the bit index operand.
25993 SDValue Op1 = N->getOperand(1);
25994 if (Op1.hasOneUse()) {
25995 unsigned BitWidth = Op1.getValueSizeInBits();
25996 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
25997 APInt KnownZero, KnownOne;
25998 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
25999 !DCI.isBeforeLegalizeOps());
26000 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
26001 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
26002 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
26003 DCI.CommitTargetLoweringOpt(TLO);
26008 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
26009 SDValue Op = N->getOperand(0);
26010 if (Op.getOpcode() == ISD::BITCAST)
26011 Op = Op.getOperand(0);
26012 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
26013 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
26014 VT.getVectorElementType().getSizeInBits() ==
26015 OpVT.getVectorElementType().getSizeInBits()) {
26016 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
26021 static SDValue PerformSIGN_EXTEND_INREGCombine(SDNode *N, SelectionDAG &DAG,
26022 const X86Subtarget *Subtarget) {
26023 EVT VT = N->getValueType(0);
26024 if (!VT.isVector())
26027 SDValue N0 = N->getOperand(0);
26028 SDValue N1 = N->getOperand(1);
26029 EVT ExtraVT = cast<VTSDNode>(N1)->getVT();
26032 // The SIGN_EXTEND_INREG to v4i64 is expensive operation on the
26033 // both SSE and AVX2 since there is no sign-extended shift right
26034 // operation on a vector with 64-bit elements.
26035 //(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) ->
26036 // (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT)))
26037 if (VT == MVT::v4i64 && (N0.getOpcode() == ISD::ANY_EXTEND ||
26038 N0.getOpcode() == ISD::SIGN_EXTEND)) {
26039 SDValue N00 = N0.getOperand(0);
26041 // EXTLOAD has a better solution on AVX2,
26042 // it may be replaced with X86ISD::VSEXT node.
26043 if (N00.getOpcode() == ISD::LOAD && Subtarget->hasInt256())
26044 if (!ISD::isNormalLoad(N00.getNode()))
26047 if (N00.getValueType() == MVT::v4i32 && ExtraVT.getSizeInBits() < 128) {
26048 SDValue Tmp = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32,
26050 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i64, Tmp);
26056 /// sext(add_nsw(x, C)) --> add(sext(x), C_sext)
26057 /// Promoting a sign extension ahead of an 'add nsw' exposes opportunities
26058 /// to combine math ops, use an LEA, or use a complex addressing mode. This can
26059 /// eliminate extend, add, and shift instructions.
26060 static SDValue promoteSextBeforeAddNSW(SDNode *Sext, SelectionDAG &DAG,
26061 const X86Subtarget *Subtarget) {
26062 // TODO: This should be valid for other integer types.
26063 EVT VT = Sext->getValueType(0);
26064 if (VT != MVT::i64)
26067 // We need an 'add nsw' feeding into the 'sext'.
26068 SDValue Add = Sext->getOperand(0);
26069 if (Add.getOpcode() != ISD::ADD || !Add->getFlags()->hasNoSignedWrap())
26072 // Having a constant operand to the 'add' ensures that we are not increasing
26073 // the instruction count because the constant is extended for free below.
26074 // A constant operand can also become the displacement field of an LEA.
26075 auto *AddOp1 = dyn_cast<ConstantSDNode>(Add.getOperand(1));
26079 // Don't make the 'add' bigger if there's no hope of combining it with some
26080 // other 'add' or 'shl' instruction.
26081 // TODO: It may be profitable to generate simpler LEA instructions in place
26082 // of single 'add' instructions, but the cost model for selecting an LEA
26083 // currently has a high threshold.
26084 bool HasLEAPotential = false;
26085 for (auto *User : Sext->uses()) {
26086 if (User->getOpcode() == ISD::ADD || User->getOpcode() == ISD::SHL) {
26087 HasLEAPotential = true;
26091 if (!HasLEAPotential)
26094 // Everything looks good, so pull the 'sext' ahead of the 'add'.
26095 int64_t AddConstant = AddOp1->getSExtValue();
26096 SDValue AddOp0 = Add.getOperand(0);
26097 SDValue NewSext = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(Sext), VT, AddOp0);
26098 SDValue NewConstant = DAG.getConstant(AddConstant, SDLoc(Add), VT);
26100 // The wider add is guaranteed to not wrap because both operands are
26103 Flags.setNoSignedWrap(true);
26104 return DAG.getNode(ISD::ADD, SDLoc(Add), VT, NewSext, NewConstant, &Flags);
26107 static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
26108 TargetLowering::DAGCombinerInfo &DCI,
26109 const X86Subtarget *Subtarget) {
26110 SDValue N0 = N->getOperand(0);
26111 EVT VT = N->getValueType(0);
26112 EVT SVT = VT.getScalarType();
26113 EVT InVT = N0.getValueType();
26114 EVT InSVT = InVT.getScalarType();
26117 // (i8,i32 sext (sdivrem (i8 x, i8 y)) ->
26118 // (i8,i32 (sdivrem_sext_hreg (i8 x, i8 y)
26119 // This exposes the sext to the sdivrem lowering, so that it directly extends
26120 // from AH (which we otherwise need to do contortions to access).
26121 if (N0.getOpcode() == ISD::SDIVREM && N0.getResNo() == 1 &&
26122 InVT == MVT::i8 && VT == MVT::i32) {
26123 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
26124 SDValue R = DAG.getNode(X86ISD::SDIVREM8_SEXT_HREG, DL, NodeTys,
26125 N0.getOperand(0), N0.getOperand(1));
26126 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
26127 return R.getValue(1);
26130 if (!DCI.isBeforeLegalizeOps()) {
26131 if (InVT == MVT::i1) {
26132 SDValue Zero = DAG.getConstant(0, DL, VT);
26134 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), DL, VT);
26135 return DAG.getNode(ISD::SELECT, DL, VT, N0, AllOnes, Zero);
26140 if (VT.isVector() && Subtarget->hasSSE2()) {
26141 auto ExtendVecSize = [&DAG](SDLoc DL, SDValue N, unsigned Size) {
26142 EVT InVT = N.getValueType();
26143 EVT OutVT = EVT::getVectorVT(*DAG.getContext(), InVT.getScalarType(),
26144 Size / InVT.getScalarSizeInBits());
26145 SmallVector<SDValue, 8> Opnds(Size / InVT.getSizeInBits(),
26146 DAG.getUNDEF(InVT));
26148 return DAG.getNode(ISD::CONCAT_VECTORS, DL, OutVT, Opnds);
26151 // If target-size is less than 128-bits, extend to a type that would extend
26152 // to 128 bits, extend that and extract the original target vector.
26153 if (VT.getSizeInBits() < 128 && !(128 % VT.getSizeInBits()) &&
26154 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
26155 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
26156 unsigned Scale = 128 / VT.getSizeInBits();
26158 EVT::getVectorVT(*DAG.getContext(), SVT, 128 / SVT.getSizeInBits());
26159 SDValue Ex = ExtendVecSize(DL, N0, Scale * InVT.getSizeInBits());
26160 SDValue SExt = DAG.getNode(ISD::SIGN_EXTEND, DL, ExVT, Ex);
26161 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, SExt,
26162 DAG.getIntPtrConstant(0, DL));
26165 // If target-size is 128-bits, then convert to ISD::SIGN_EXTEND_VECTOR_INREG
26166 // which ensures lowering to X86ISD::VSEXT (pmovsx*).
26167 if (VT.getSizeInBits() == 128 &&
26168 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
26169 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
26170 SDValue ExOp = ExtendVecSize(DL, N0, 128);
26171 return DAG.getSignExtendVectorInReg(ExOp, DL, VT);
26174 // On pre-AVX2 targets, split into 128-bit nodes of
26175 // ISD::SIGN_EXTEND_VECTOR_INREG.
26176 if (!Subtarget->hasInt256() && !(VT.getSizeInBits() % 128) &&
26177 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
26178 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
26179 unsigned NumVecs = VT.getSizeInBits() / 128;
26180 unsigned NumSubElts = 128 / SVT.getSizeInBits();
26181 EVT SubVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumSubElts);
26182 EVT InSubVT = EVT::getVectorVT(*DAG.getContext(), InSVT, NumSubElts);
26184 SmallVector<SDValue, 8> Opnds;
26185 for (unsigned i = 0, Offset = 0; i != NumVecs;
26186 ++i, Offset += NumSubElts) {
26187 SDValue SrcVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InSubVT, N0,
26188 DAG.getIntPtrConstant(Offset, DL));
26189 SrcVec = ExtendVecSize(DL, SrcVec, 128);
26190 SrcVec = DAG.getSignExtendVectorInReg(SrcVec, DL, SubVT);
26191 Opnds.push_back(SrcVec);
26193 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Opnds);
26197 if (Subtarget->hasAVX() && VT.is256BitVector())
26198 if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
26201 if (SDValue NewAdd = promoteSextBeforeAddNSW(N, DAG, Subtarget))
26207 static SDValue PerformFMACombine(SDNode *N, SelectionDAG &DAG,
26208 const X86Subtarget* Subtarget) {
26210 EVT VT = N->getValueType(0);
26212 // Let legalize expand this if it isn't a legal type yet.
26213 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
26216 EVT ScalarVT = VT.getScalarType();
26217 if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) ||
26218 (!Subtarget->hasFMA() && !Subtarget->hasFMA4() &&
26219 !Subtarget->hasAVX512()))
26222 SDValue A = N->getOperand(0);
26223 SDValue B = N->getOperand(1);
26224 SDValue C = N->getOperand(2);
26226 bool NegA = (A.getOpcode() == ISD::FNEG);
26227 bool NegB = (B.getOpcode() == ISD::FNEG);
26228 bool NegC = (C.getOpcode() == ISD::FNEG);
26230 // Negative multiplication when NegA xor NegB
26231 bool NegMul = (NegA != NegB);
26233 A = A.getOperand(0);
26235 B = B.getOperand(0);
26237 C = C.getOperand(0);
26241 Opcode = (!NegC) ? X86ISD::FMADD : X86ISD::FMSUB;
26243 Opcode = (!NegC) ? X86ISD::FNMADD : X86ISD::FNMSUB;
26245 return DAG.getNode(Opcode, dl, VT, A, B, C);
26248 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG,
26249 TargetLowering::DAGCombinerInfo &DCI,
26250 const X86Subtarget *Subtarget) {
26251 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
26252 // (and (i32 x86isd::setcc_carry), 1)
26253 // This eliminates the zext. This transformation is necessary because
26254 // ISD::SETCC is always legalized to i8.
26256 SDValue N0 = N->getOperand(0);
26257 EVT VT = N->getValueType(0);
26259 if (N0.getOpcode() == ISD::AND &&
26261 N0.getOperand(0).hasOneUse()) {
26262 SDValue N00 = N0.getOperand(0);
26263 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
26264 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
26265 if (!C || C->getZExtValue() != 1)
26267 return DAG.getNode(ISD::AND, dl, VT,
26268 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
26269 N00.getOperand(0), N00.getOperand(1)),
26270 DAG.getConstant(1, dl, VT));
26274 if (N0.getOpcode() == ISD::TRUNCATE &&
26276 N0.getOperand(0).hasOneUse()) {
26277 SDValue N00 = N0.getOperand(0);
26278 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
26279 return DAG.getNode(ISD::AND, dl, VT,
26280 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
26281 N00.getOperand(0), N00.getOperand(1)),
26282 DAG.getConstant(1, dl, VT));
26286 if (VT.is256BitVector())
26287 if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
26290 // (i8,i32 zext (udivrem (i8 x, i8 y)) ->
26291 // (i8,i32 (udivrem_zext_hreg (i8 x, i8 y)
26292 // This exposes the zext to the udivrem lowering, so that it directly extends
26293 // from AH (which we otherwise need to do contortions to access).
26294 if (N0.getOpcode() == ISD::UDIVREM &&
26295 N0.getResNo() == 1 && N0.getValueType() == MVT::i8 &&
26296 (VT == MVT::i32 || VT == MVT::i64)) {
26297 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
26298 SDValue R = DAG.getNode(X86ISD::UDIVREM8_ZEXT_HREG, dl, NodeTys,
26299 N0.getOperand(0), N0.getOperand(1));
26300 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
26301 return R.getValue(1);
26307 // Optimize x == -y --> x+y == 0
26308 // x != -y --> x+y != 0
26309 static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG,
26310 const X86Subtarget* Subtarget) {
26311 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
26312 SDValue LHS = N->getOperand(0);
26313 SDValue RHS = N->getOperand(1);
26314 EVT VT = N->getValueType(0);
26317 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB)
26318 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(LHS.getOperand(0)))
26319 if (C->getAPIntValue() == 0 && LHS.hasOneUse()) {
26320 SDValue addV = DAG.getNode(ISD::ADD, DL, LHS.getValueType(), RHS,
26321 LHS.getOperand(1));
26322 return DAG.getSetCC(DL, N->getValueType(0), addV,
26323 DAG.getConstant(0, DL, addV.getValueType()), CC);
26325 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && RHS.getOpcode() == ISD::SUB)
26326 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS.getOperand(0)))
26327 if (C->getAPIntValue() == 0 && RHS.hasOneUse()) {
26328 SDValue addV = DAG.getNode(ISD::ADD, DL, RHS.getValueType(), LHS,
26329 RHS.getOperand(1));
26330 return DAG.getSetCC(DL, N->getValueType(0), addV,
26331 DAG.getConstant(0, DL, addV.getValueType()), CC);
26334 if (VT.getScalarType() == MVT::i1 &&
26335 (CC == ISD::SETNE || CC == ISD::SETEQ || ISD::isSignedIntSetCC(CC))) {
26337 (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
26338 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
26339 bool IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
26341 if (!IsSEXT0 || !IsVZero1) {
26342 // Swap the operands and update the condition code.
26343 std::swap(LHS, RHS);
26344 CC = ISD::getSetCCSwappedOperands(CC);
26346 IsSEXT0 = (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
26347 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
26348 IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
26351 if (IsSEXT0 && IsVZero1) {
26352 assert(VT == LHS.getOperand(0).getValueType() &&
26353 "Uexpected operand type");
26354 if (CC == ISD::SETGT)
26355 return DAG.getConstant(0, DL, VT);
26356 if (CC == ISD::SETLE)
26357 return DAG.getConstant(1, DL, VT);
26358 if (CC == ISD::SETEQ || CC == ISD::SETGE)
26359 return DAG.getNOT(DL, LHS.getOperand(0), VT);
26361 assert((CC == ISD::SETNE || CC == ISD::SETLT) &&
26362 "Unexpected condition code!");
26363 return LHS.getOperand(0);
26370 static SDValue PerformBLENDICombine(SDNode *N, SelectionDAG &DAG) {
26371 SDValue V0 = N->getOperand(0);
26372 SDValue V1 = N->getOperand(1);
26374 EVT VT = N->getValueType(0);
26376 // Canonicalize a v2f64 blend with a mask of 2 by swapping the vector
26377 // operands and changing the mask to 1. This saves us a bunch of
26378 // pattern-matching possibilities related to scalar math ops in SSE/AVX.
26379 // x86InstrInfo knows how to commute this back after instruction selection
26380 // if it would help register allocation.
26382 // TODO: If optimizing for size or a processor that doesn't suffer from
26383 // partial register update stalls, this should be transformed into a MOVSD
26384 // instruction because a MOVSD is 1-2 bytes smaller than a BLENDPD.
26386 if (VT == MVT::v2f64)
26387 if (auto *Mask = dyn_cast<ConstantSDNode>(N->getOperand(2)))
26388 if (Mask->getZExtValue() == 2 && !isShuffleFoldableLoad(V0)) {
26389 SDValue NewMask = DAG.getConstant(1, DL, MVT::i8);
26390 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V0, NewMask);
26396 // Helper function of PerformSETCCCombine. It is to materialize "setb reg"
26397 // as "sbb reg,reg", since it can be extended without zext and produces
26398 // an all-ones bit which is more useful than 0/1 in some cases.
26399 static SDValue MaterializeSETB(SDLoc DL, SDValue EFLAGS, SelectionDAG &DAG,
26402 return DAG.getNode(ISD::AND, DL, VT,
26403 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
26404 DAG.getConstant(X86::COND_B, DL, MVT::i8),
26406 DAG.getConstant(1, DL, VT));
26407 assert (VT == MVT::i1 && "Unexpected type for SECCC node");
26408 return DAG.getNode(ISD::TRUNCATE, DL, MVT::i1,
26409 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
26410 DAG.getConstant(X86::COND_B, DL, MVT::i8),
26414 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
26415 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG,
26416 TargetLowering::DAGCombinerInfo &DCI,
26417 const X86Subtarget *Subtarget) {
26419 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0));
26420 SDValue EFLAGS = N->getOperand(1);
26422 if (CC == X86::COND_A) {
26423 // Try to convert COND_A into COND_B in an attempt to facilitate
26424 // materializing "setb reg".
26426 // Do not flip "e > c", where "c" is a constant, because Cmp instruction
26427 // cannot take an immediate as its first operand.
26429 if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() &&
26430 EFLAGS.getValueType().isInteger() &&
26431 !isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
26432 SDValue NewSub = DAG.getNode(X86ISD::SUB, SDLoc(EFLAGS),
26433 EFLAGS.getNode()->getVTList(),
26434 EFLAGS.getOperand(1), EFLAGS.getOperand(0));
26435 SDValue NewEFLAGS = SDValue(NewSub.getNode(), EFLAGS.getResNo());
26436 return MaterializeSETB(DL, NewEFLAGS, DAG, N->getSimpleValueType(0));
26440 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
26441 // a zext and produces an all-ones bit which is more useful than 0/1 in some
26443 if (CC == X86::COND_B)
26444 return MaterializeSETB(DL, EFLAGS, DAG, N->getSimpleValueType(0));
26446 if (SDValue Flags = checkBoolTestSetCCCombine(EFLAGS, CC)) {
26447 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
26448 return DAG.getNode(X86ISD::SETCC, DL, N->getVTList(), Cond, Flags);
26454 // Optimize branch condition evaluation.
26456 static SDValue PerformBrCondCombine(SDNode *N, SelectionDAG &DAG,
26457 TargetLowering::DAGCombinerInfo &DCI,
26458 const X86Subtarget *Subtarget) {
26460 SDValue Chain = N->getOperand(0);
26461 SDValue Dest = N->getOperand(1);
26462 SDValue EFLAGS = N->getOperand(3);
26463 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(2));
26465 if (SDValue Flags = checkBoolTestSetCCCombine(EFLAGS, CC)) {
26466 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
26467 return DAG.getNode(X86ISD::BRCOND, DL, N->getVTList(), Chain, Dest, Cond,
26474 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
26475 SelectionDAG &DAG) {
26476 // Take advantage of vector comparisons producing 0 or -1 in each lane to
26477 // optimize away operation when it's from a constant.
26479 // The general transformation is:
26480 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
26481 // AND(VECTOR_CMP(x,y), constant2)
26482 // constant2 = UNARYOP(constant)
26484 // Early exit if this isn't a vector operation, the operand of the
26485 // unary operation isn't a bitwise AND, or if the sizes of the operations
26486 // aren't the same.
26487 EVT VT = N->getValueType(0);
26488 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
26489 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
26490 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
26493 // Now check that the other operand of the AND is a constant. We could
26494 // make the transformation for non-constant splats as well, but it's unclear
26495 // that would be a benefit as it would not eliminate any operations, just
26496 // perform one more step in scalar code before moving to the vector unit.
26497 if (BuildVectorSDNode *BV =
26498 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
26499 // Bail out if the vector isn't a constant.
26500 if (!BV->isConstant())
26503 // Everything checks out. Build up the new and improved node.
26505 EVT IntVT = BV->getValueType(0);
26506 // Create a new constant of the appropriate type for the transformed
26508 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
26509 // The AND node needs bitcasts to/from an integer vector type around it.
26510 SDValue MaskConst = DAG.getBitcast(IntVT, SourceConst);
26511 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
26512 N->getOperand(0)->getOperand(0), MaskConst);
26513 SDValue Res = DAG.getBitcast(VT, NewAnd);
26520 static SDValue PerformUINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
26521 const X86Subtarget *Subtarget) {
26522 SDValue Op0 = N->getOperand(0);
26523 EVT VT = N->getValueType(0);
26524 EVT InVT = Op0.getValueType();
26525 EVT InSVT = InVT.getScalarType();
26526 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
26528 // UINT_TO_FP(vXi8) -> SINT_TO_FP(ZEXT(vXi8 to vXi32))
26529 // UINT_TO_FP(vXi16) -> SINT_TO_FP(ZEXT(vXi16 to vXi32))
26530 if (InVT.isVector() && (InSVT == MVT::i8 || InSVT == MVT::i16)) {
26532 EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
26533 InVT.getVectorNumElements());
26534 SDValue P = DAG.getNode(ISD::ZERO_EXTEND, dl, DstVT, Op0);
26536 if (TLI.isOperationLegal(ISD::UINT_TO_FP, DstVT))
26537 return DAG.getNode(ISD::UINT_TO_FP, dl, VT, P);
26539 return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P);
26545 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
26546 const X86Subtarget *Subtarget) {
26547 // First try to optimize away the conversion entirely when it's
26548 // conditionally from a constant. Vectors only.
26549 if (SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG))
26552 // Now move on to more general possibilities.
26553 SDValue Op0 = N->getOperand(0);
26554 EVT VT = N->getValueType(0);
26555 EVT InVT = Op0.getValueType();
26556 EVT InSVT = InVT.getScalarType();
26558 // SINT_TO_FP(vXi8) -> SINT_TO_FP(SEXT(vXi8 to vXi32))
26559 // SINT_TO_FP(vXi16) -> SINT_TO_FP(SEXT(vXi16 to vXi32))
26560 if (InVT.isVector() && (InSVT == MVT::i8 || InSVT == MVT::i16)) {
26562 EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
26563 InVT.getVectorNumElements());
26564 SDValue P = DAG.getNode(ISD::SIGN_EXTEND, dl, DstVT, Op0);
26565 return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P);
26568 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
26569 // a 32-bit target where SSE doesn't support i64->FP operations.
26570 if (!Subtarget->useSoftFloat() && Op0.getOpcode() == ISD::LOAD) {
26571 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
26572 EVT LdVT = Ld->getValueType(0);
26574 // This transformation is not supported if the result type is f16
26575 if (VT == MVT::f16)
26578 if (!Ld->isVolatile() && !VT.isVector() &&
26579 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
26580 !Subtarget->is64Bit() && LdVT == MVT::i64) {
26581 SDValue FILDChain = Subtarget->getTargetLowering()->BuildFILD(
26582 SDValue(N, 0), LdVT, Ld->getChain(), Op0, DAG);
26583 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
26590 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
26591 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
26592 X86TargetLowering::DAGCombinerInfo &DCI) {
26593 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
26594 // the result is either zero or one (depending on the input carry bit).
26595 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
26596 if (X86::isZeroNode(N->getOperand(0)) &&
26597 X86::isZeroNode(N->getOperand(1)) &&
26598 // We don't have a good way to replace an EFLAGS use, so only do this when
26600 SDValue(N, 1).use_empty()) {
26602 EVT VT = N->getValueType(0);
26603 SDValue CarryOut = DAG.getConstant(0, DL, N->getValueType(1));
26604 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
26605 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
26606 DAG.getConstant(X86::COND_B, DL,
26609 DAG.getConstant(1, DL, VT));
26610 return DCI.CombineTo(N, Res1, CarryOut);
26616 // fold (add Y, (sete X, 0)) -> adc 0, Y
26617 // (add Y, (setne X, 0)) -> sbb -1, Y
26618 // (sub (sete X, 0), Y) -> sbb 0, Y
26619 // (sub (setne X, 0), Y) -> adc -1, Y
26620 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
26623 // Look through ZExts.
26624 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
26625 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
26628 SDValue SetCC = Ext.getOperand(0);
26629 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
26632 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
26633 if (CC != X86::COND_E && CC != X86::COND_NE)
26636 SDValue Cmp = SetCC.getOperand(1);
26637 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
26638 !X86::isZeroNode(Cmp.getOperand(1)) ||
26639 !Cmp.getOperand(0).getValueType().isInteger())
26642 SDValue CmpOp0 = Cmp.getOperand(0);
26643 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
26644 DAG.getConstant(1, DL, CmpOp0.getValueType()));
26646 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
26647 if (CC == X86::COND_NE)
26648 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
26649 DL, OtherVal.getValueType(), OtherVal,
26650 DAG.getConstant(-1ULL, DL, OtherVal.getValueType()),
26652 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
26653 DL, OtherVal.getValueType(), OtherVal,
26654 DAG.getConstant(0, DL, OtherVal.getValueType()), NewCmp);
26657 /// PerformADDCombine - Do target-specific dag combines on integer adds.
26658 static SDValue PerformAddCombine(SDNode *N, SelectionDAG &DAG,
26659 const X86Subtarget *Subtarget) {
26660 EVT VT = N->getValueType(0);
26661 SDValue Op0 = N->getOperand(0);
26662 SDValue Op1 = N->getOperand(1);
26664 // Try to synthesize horizontal adds from adds of shuffles.
26665 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
26666 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
26667 isHorizontalBinOp(Op0, Op1, true))
26668 return DAG.getNode(X86ISD::HADD, SDLoc(N), VT, Op0, Op1);
26670 return OptimizeConditionalInDecrement(N, DAG);
26673 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG,
26674 const X86Subtarget *Subtarget) {
26675 SDValue Op0 = N->getOperand(0);
26676 SDValue Op1 = N->getOperand(1);
26678 // X86 can't encode an immediate LHS of a sub. See if we can push the
26679 // negation into a preceding instruction.
26680 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
26681 // If the RHS of the sub is a XOR with one use and a constant, invert the
26682 // immediate. Then add one to the LHS of the sub so we can turn
26683 // X-Y -> X+~Y+1, saving one register.
26684 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
26685 isa<ConstantSDNode>(Op1.getOperand(1))) {
26686 APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue();
26687 EVT VT = Op0.getValueType();
26688 SDValue NewXor = DAG.getNode(ISD::XOR, SDLoc(Op1), VT,
26690 DAG.getConstant(~XorC, SDLoc(Op1), VT));
26691 return DAG.getNode(ISD::ADD, SDLoc(N), VT, NewXor,
26692 DAG.getConstant(C->getAPIntValue() + 1, SDLoc(N), VT));
26696 // Try to synthesize horizontal adds from adds of shuffles.
26697 EVT VT = N->getValueType(0);
26698 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
26699 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
26700 isHorizontalBinOp(Op0, Op1, true))
26701 return DAG.getNode(X86ISD::HSUB, SDLoc(N), VT, Op0, Op1);
26703 return OptimizeConditionalInDecrement(N, DAG);
26706 /// performVZEXTCombine - Performs build vector combines
26707 static SDValue performVZEXTCombine(SDNode *N, SelectionDAG &DAG,
26708 TargetLowering::DAGCombinerInfo &DCI,
26709 const X86Subtarget *Subtarget) {
26711 MVT VT = N->getSimpleValueType(0);
26712 SDValue Op = N->getOperand(0);
26713 MVT OpVT = Op.getSimpleValueType();
26714 MVT OpEltVT = OpVT.getVectorElementType();
26715 unsigned InputBits = OpEltVT.getSizeInBits() * VT.getVectorNumElements();
26717 // (vzext (bitcast (vzext (x)) -> (vzext x)
26719 while (V.getOpcode() == ISD::BITCAST)
26720 V = V.getOperand(0);
26722 if (V != Op && V.getOpcode() == X86ISD::VZEXT) {
26723 MVT InnerVT = V.getSimpleValueType();
26724 MVT InnerEltVT = InnerVT.getVectorElementType();
26726 // If the element sizes match exactly, we can just do one larger vzext. This
26727 // is always an exact type match as vzext operates on integer types.
26728 if (OpEltVT == InnerEltVT) {
26729 assert(OpVT == InnerVT && "Types must match for vzext!");
26730 return DAG.getNode(X86ISD::VZEXT, DL, VT, V.getOperand(0));
26733 // The only other way we can combine them is if only a single element of the
26734 // inner vzext is used in the input to the outer vzext.
26735 if (InnerEltVT.getSizeInBits() < InputBits)
26738 // In this case, the inner vzext is completely dead because we're going to
26739 // only look at bits inside of the low element. Just do the outer vzext on
26740 // a bitcast of the input to the inner.
26741 return DAG.getNode(X86ISD::VZEXT, DL, VT, DAG.getBitcast(OpVT, V));
26744 // Check if we can bypass extracting and re-inserting an element of an input
26745 // vector. Essentially:
26746 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast x)
26747 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR &&
26748 V.getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
26749 V.getOperand(0).getSimpleValueType().getSizeInBits() == InputBits) {
26750 SDValue ExtractedV = V.getOperand(0);
26751 SDValue OrigV = ExtractedV.getOperand(0);
26752 if (auto *ExtractIdx = dyn_cast<ConstantSDNode>(ExtractedV.getOperand(1)))
26753 if (ExtractIdx->getZExtValue() == 0) {
26754 MVT OrigVT = OrigV.getSimpleValueType();
26755 // Extract a subvector if necessary...
26756 if (OrigVT.getSizeInBits() > OpVT.getSizeInBits()) {
26757 int Ratio = OrigVT.getSizeInBits() / OpVT.getSizeInBits();
26758 OrigVT = MVT::getVectorVT(OrigVT.getVectorElementType(),
26759 OrigVT.getVectorNumElements() / Ratio);
26760 OrigV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigVT, OrigV,
26761 DAG.getIntPtrConstant(0, DL));
26763 Op = DAG.getBitcast(OpVT, OrigV);
26764 return DAG.getNode(X86ISD::VZEXT, DL, VT, Op);
26771 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
26772 DAGCombinerInfo &DCI) const {
26773 SelectionDAG &DAG = DCI.DAG;
26774 switch (N->getOpcode()) {
26776 case ISD::EXTRACT_VECTOR_ELT:
26777 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, DCI);
26780 case X86ISD::SHRUNKBLEND:
26781 return PerformSELECTCombine(N, DAG, DCI, Subtarget);
26782 case ISD::BITCAST: return PerformBITCASTCombine(N, DAG, Subtarget);
26783 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI, Subtarget);
26784 case ISD::ADD: return PerformAddCombine(N, DAG, Subtarget);
26785 case ISD::SUB: return PerformSubCombine(N, DAG, Subtarget);
26786 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
26787 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
26790 case ISD::SRL: return PerformShiftCombine(N, DAG, DCI, Subtarget);
26791 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
26792 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
26793 case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget);
26794 case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget);
26795 case ISD::MLOAD: return PerformMLOADCombine(N, DAG, DCI, Subtarget);
26796 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
26797 case ISD::MSTORE: return PerformMSTORECombine(N, DAG, Subtarget);
26798 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, Subtarget);
26799 case ISD::UINT_TO_FP: return PerformUINT_TO_FPCombine(N, DAG, Subtarget);
26800 case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
26801 case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
26803 case X86ISD::FOR: return PerformFORCombine(N, DAG, Subtarget);
26805 case X86ISD::FMAX: return PerformFMinFMaxCombine(N, DAG);
26806 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
26807 case X86ISD::FANDN: return PerformFANDNCombine(N, DAG);
26808 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
26809 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
26810 case ISD::ANY_EXTEND:
26811 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG, DCI, Subtarget);
26812 case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
26813 case ISD::SIGN_EXTEND_INREG:
26814 return PerformSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
26815 case ISD::SETCC: return PerformISDSETCCCombine(N, DAG, Subtarget);
26816 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
26817 case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
26818 case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);
26819 case X86ISD::SHUFP: // Handle all target specific shuffles
26820 case X86ISD::PALIGNR:
26821 case X86ISD::UNPCKH:
26822 case X86ISD::UNPCKL:
26823 case X86ISD::MOVHLPS:
26824 case X86ISD::MOVLHPS:
26825 case X86ISD::PSHUFB:
26826 case X86ISD::PSHUFD:
26827 case X86ISD::PSHUFHW:
26828 case X86ISD::PSHUFLW:
26829 case X86ISD::MOVSS:
26830 case X86ISD::MOVSD:
26831 case X86ISD::VPERMILPI:
26832 case X86ISD::VPERM2X128:
26833 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
26834 case ISD::FMA: return PerformFMACombine(N, DAG, Subtarget);
26835 case X86ISD::BLENDI: return PerformBLENDICombine(N, DAG);
26841 /// isTypeDesirableForOp - Return true if the target has native support for
26842 /// the specified value type and it is 'desirable' to use the type for the
26843 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
26844 /// instruction encodings are longer and some i16 instructions are slow.
26845 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
26846 if (!isTypeLegal(VT))
26848 if (VT != MVT::i16)
26855 case ISD::SIGN_EXTEND:
26856 case ISD::ZERO_EXTEND:
26857 case ISD::ANY_EXTEND:
26870 /// IsDesirableToPromoteOp - This method query the target whether it is
26871 /// beneficial for dag combiner to promote the specified node. If true, it
26872 /// should return the desired promotion type by reference.
26873 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
26874 EVT VT = Op.getValueType();
26875 if (VT != MVT::i16)
26878 bool Promote = false;
26879 bool Commute = false;
26880 switch (Op.getOpcode()) {
26883 LoadSDNode *LD = cast<LoadSDNode>(Op);
26884 // If the non-extending load has a single use and it's not live out, then it
26885 // might be folded.
26886 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
26887 Op.hasOneUse()*/) {
26888 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
26889 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
26890 // The only case where we'd want to promote LOAD (rather then it being
26891 // promoted as an operand is when it's only use is liveout.
26892 if (UI->getOpcode() != ISD::CopyToReg)
26899 case ISD::SIGN_EXTEND:
26900 case ISD::ZERO_EXTEND:
26901 case ISD::ANY_EXTEND:
26906 SDValue N0 = Op.getOperand(0);
26907 // Look out for (store (shl (load), x)).
26908 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
26921 SDValue N0 = Op.getOperand(0);
26922 SDValue N1 = Op.getOperand(1);
26923 if (!Commute && MayFoldLoad(N1))
26925 // Avoid disabling potential load folding opportunities.
26926 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
26928 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
26938 //===----------------------------------------------------------------------===//
26939 // X86 Inline Assembly Support
26940 //===----------------------------------------------------------------------===//
26942 // Helper to match a string separated by whitespace.
26943 static bool matchAsm(StringRef S, ArrayRef<const char *> Pieces) {
26944 S = S.substr(S.find_first_not_of(" \t")); // Skip leading whitespace.
26946 for (StringRef Piece : Pieces) {
26947 if (!S.startswith(Piece)) // Check if the piece matches.
26950 S = S.substr(Piece.size());
26951 StringRef::size_type Pos = S.find_first_not_of(" \t");
26952 if (Pos == 0) // We matched a prefix.
26961 static bool clobbersFlagRegisters(const SmallVector<StringRef, 4> &AsmPieces) {
26963 if (AsmPieces.size() == 3 || AsmPieces.size() == 4) {
26964 if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{cc}") &&
26965 std::count(AsmPieces.begin(), AsmPieces.end(), "~{flags}") &&
26966 std::count(AsmPieces.begin(), AsmPieces.end(), "~{fpsr}")) {
26968 if (AsmPieces.size() == 3)
26970 else if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{dirflag}"))
26977 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
26978 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
26980 std::string AsmStr = IA->getAsmString();
26982 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
26983 if (!Ty || Ty->getBitWidth() % 16 != 0)
26986 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
26987 SmallVector<StringRef, 4> AsmPieces;
26988 SplitString(AsmStr, AsmPieces, ";\n");
26990 switch (AsmPieces.size()) {
26991 default: return false;
26993 // FIXME: this should verify that we are targeting a 486 or better. If not,
26994 // we will turn this bswap into something that will be lowered to logical
26995 // ops instead of emitting the bswap asm. For now, we don't support 486 or
26996 // lower so don't worry about this.
26998 if (matchAsm(AsmPieces[0], {"bswap", "$0"}) ||
26999 matchAsm(AsmPieces[0], {"bswapl", "$0"}) ||
27000 matchAsm(AsmPieces[0], {"bswapq", "$0"}) ||
27001 matchAsm(AsmPieces[0], {"bswap", "${0:q}"}) ||
27002 matchAsm(AsmPieces[0], {"bswapl", "${0:q}"}) ||
27003 matchAsm(AsmPieces[0], {"bswapq", "${0:q}"})) {
27004 // No need to check constraints, nothing other than the equivalent of
27005 // "=r,0" would be valid here.
27006 return IntrinsicLowering::LowerToByteSwap(CI);
27009 // rorw $$8, ${0:w} --> llvm.bswap.i16
27010 if (CI->getType()->isIntegerTy(16) &&
27011 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
27012 (matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) ||
27013 matchAsm(AsmPieces[0], {"rolw", "$$8,", "${0:w}"}))) {
27015 StringRef ConstraintsStr = IA->getConstraintString();
27016 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
27017 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
27018 if (clobbersFlagRegisters(AsmPieces))
27019 return IntrinsicLowering::LowerToByteSwap(CI);
27023 if (CI->getType()->isIntegerTy(32) &&
27024 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
27025 matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) &&
27026 matchAsm(AsmPieces[1], {"rorl", "$$16,", "$0"}) &&
27027 matchAsm(AsmPieces[2], {"rorw", "$$8,", "${0:w}"})) {
27029 StringRef ConstraintsStr = IA->getConstraintString();
27030 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
27031 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
27032 if (clobbersFlagRegisters(AsmPieces))
27033 return IntrinsicLowering::LowerToByteSwap(CI);
27036 if (CI->getType()->isIntegerTy(64)) {
27037 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
27038 if (Constraints.size() >= 2 &&
27039 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
27040 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
27041 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
27042 if (matchAsm(AsmPieces[0], {"bswap", "%eax"}) &&
27043 matchAsm(AsmPieces[1], {"bswap", "%edx"}) &&
27044 matchAsm(AsmPieces[2], {"xchgl", "%eax,", "%edx"}))
27045 return IntrinsicLowering::LowerToByteSwap(CI);
27053 /// getConstraintType - Given a constraint letter, return the type of
27054 /// constraint it is for this target.
27055 X86TargetLowering::ConstraintType
27056 X86TargetLowering::getConstraintType(StringRef Constraint) const {
27057 if (Constraint.size() == 1) {
27058 switch (Constraint[0]) {
27069 return C_RegisterClass;
27093 return TargetLowering::getConstraintType(Constraint);
27096 /// Examine constraint type and operand type and determine a weight value.
27097 /// This object must already have been set up with the operand type
27098 /// and the current alternative constraint selected.
27099 TargetLowering::ConstraintWeight
27100 X86TargetLowering::getSingleConstraintMatchWeight(
27101 AsmOperandInfo &info, const char *constraint) const {
27102 ConstraintWeight weight = CW_Invalid;
27103 Value *CallOperandVal = info.CallOperandVal;
27104 // If we don't have a value, we can't do a match,
27105 // but allow it at the lowest weight.
27106 if (!CallOperandVal)
27108 Type *type = CallOperandVal->getType();
27109 // Look at the constraint type.
27110 switch (*constraint) {
27112 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
27123 if (CallOperandVal->getType()->isIntegerTy())
27124 weight = CW_SpecificReg;
27129 if (type->isFloatingPointTy())
27130 weight = CW_SpecificReg;
27133 if (type->isX86_MMXTy() && Subtarget->hasMMX())
27134 weight = CW_SpecificReg;
27138 if (((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasSSE1()) ||
27139 ((type->getPrimitiveSizeInBits() == 256) && Subtarget->hasFp256()))
27140 weight = CW_Register;
27143 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
27144 if (C->getZExtValue() <= 31)
27145 weight = CW_Constant;
27149 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27150 if (C->getZExtValue() <= 63)
27151 weight = CW_Constant;
27155 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27156 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
27157 weight = CW_Constant;
27161 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27162 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
27163 weight = CW_Constant;
27167 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27168 if (C->getZExtValue() <= 3)
27169 weight = CW_Constant;
27173 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27174 if (C->getZExtValue() <= 0xff)
27175 weight = CW_Constant;
27180 if (isa<ConstantFP>(CallOperandVal)) {
27181 weight = CW_Constant;
27185 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27186 if ((C->getSExtValue() >= -0x80000000LL) &&
27187 (C->getSExtValue() <= 0x7fffffffLL))
27188 weight = CW_Constant;
27192 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27193 if (C->getZExtValue() <= 0xffffffff)
27194 weight = CW_Constant;
27201 /// LowerXConstraint - try to replace an X constraint, which matches anything,
27202 /// with another that has more specific requirements based on the type of the
27203 /// corresponding operand.
27204 const char *X86TargetLowering::
27205 LowerXConstraint(EVT ConstraintVT) const {
27206 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
27207 // 'f' like normal targets.
27208 if (ConstraintVT.isFloatingPoint()) {
27209 if (Subtarget->hasSSE2())
27211 if (Subtarget->hasSSE1())
27215 return TargetLowering::LowerXConstraint(ConstraintVT);
27218 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
27219 /// vector. If it is invalid, don't add anything to Ops.
27220 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
27221 std::string &Constraint,
27222 std::vector<SDValue>&Ops,
27223 SelectionDAG &DAG) const {
27226 // Only support length 1 constraints for now.
27227 if (Constraint.length() > 1) return;
27229 char ConstraintLetter = Constraint[0];
27230 switch (ConstraintLetter) {
27233 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27234 if (C->getZExtValue() <= 31) {
27235 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27236 Op.getValueType());
27242 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27243 if (C->getZExtValue() <= 63) {
27244 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27245 Op.getValueType());
27251 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27252 if (isInt<8>(C->getSExtValue())) {
27253 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27254 Op.getValueType());
27260 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27261 if (C->getZExtValue() == 0xff || C->getZExtValue() == 0xffff ||
27262 (Subtarget->is64Bit() && C->getZExtValue() == 0xffffffff)) {
27263 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op),
27264 Op.getValueType());
27270 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27271 if (C->getZExtValue() <= 3) {
27272 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27273 Op.getValueType());
27279 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27280 if (C->getZExtValue() <= 255) {
27281 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27282 Op.getValueType());
27288 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27289 if (C->getZExtValue() <= 127) {
27290 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27291 Op.getValueType());
27297 // 32-bit signed value
27298 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27299 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
27300 C->getSExtValue())) {
27301 // Widen to 64 bits here to get it sign extended.
27302 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op), MVT::i64);
27305 // FIXME gcc accepts some relocatable values here too, but only in certain
27306 // memory models; it's complicated.
27311 // 32-bit unsigned value
27312 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27313 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
27314 C->getZExtValue())) {
27315 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27316 Op.getValueType());
27320 // FIXME gcc accepts some relocatable values here too, but only in certain
27321 // memory models; it's complicated.
27325 // Literal immediates are always ok.
27326 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
27327 // Widen to 64 bits here to get it sign extended.
27328 Result = DAG.getTargetConstant(CST->getSExtValue(), SDLoc(Op), MVT::i64);
27332 // In any sort of PIC mode addresses need to be computed at runtime by
27333 // adding in a register or some sort of table lookup. These can't
27334 // be used as immediates.
27335 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
27338 // If we are in non-pic codegen mode, we allow the address of a global (with
27339 // an optional displacement) to be used with 'i'.
27340 GlobalAddressSDNode *GA = nullptr;
27341 int64_t Offset = 0;
27343 // Match either (GA), (GA+C), (GA+C1+C2), etc.
27345 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
27346 Offset += GA->getOffset();
27348 } else if (Op.getOpcode() == ISD::ADD) {
27349 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
27350 Offset += C->getZExtValue();
27351 Op = Op.getOperand(0);
27354 } else if (Op.getOpcode() == ISD::SUB) {
27355 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
27356 Offset += -C->getZExtValue();
27357 Op = Op.getOperand(0);
27362 // Otherwise, this isn't something we can handle, reject it.
27366 const GlobalValue *GV = GA->getGlobal();
27367 // If we require an extra load to get this address, as in PIC mode, we
27368 // can't accept it.
27369 if (isGlobalStubReference(
27370 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget())))
27373 Result = DAG.getTargetGlobalAddress(GV, SDLoc(Op),
27374 GA->getValueType(0), Offset);
27379 if (Result.getNode()) {
27380 Ops.push_back(Result);
27383 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
27386 std::pair<unsigned, const TargetRegisterClass *>
27387 X86TargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
27388 StringRef Constraint,
27390 // First, see if this is a constraint that directly corresponds to an LLVM
27392 if (Constraint.size() == 1) {
27393 // GCC Constraint Letters
27394 switch (Constraint[0]) {
27396 // TODO: Slight differences here in allocation order and leaving
27397 // RIP in the class. Do they matter any more here than they do
27398 // in the normal allocation?
27399 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
27400 if (Subtarget->is64Bit()) {
27401 if (VT == MVT::i32 || VT == MVT::f32)
27402 return std::make_pair(0U, &X86::GR32RegClass);
27403 if (VT == MVT::i16)
27404 return std::make_pair(0U, &X86::GR16RegClass);
27405 if (VT == MVT::i8 || VT == MVT::i1)
27406 return std::make_pair(0U, &X86::GR8RegClass);
27407 if (VT == MVT::i64 || VT == MVT::f64)
27408 return std::make_pair(0U, &X86::GR64RegClass);
27411 // 32-bit fallthrough
27412 case 'Q': // Q_REGS
27413 if (VT == MVT::i32 || VT == MVT::f32)
27414 return std::make_pair(0U, &X86::GR32_ABCDRegClass);
27415 if (VT == MVT::i16)
27416 return std::make_pair(0U, &X86::GR16_ABCDRegClass);
27417 if (VT == MVT::i8 || VT == MVT::i1)
27418 return std::make_pair(0U, &X86::GR8_ABCD_LRegClass);
27419 if (VT == MVT::i64)
27420 return std::make_pair(0U, &X86::GR64_ABCDRegClass);
27422 case 'r': // GENERAL_REGS
27423 case 'l': // INDEX_REGS
27424 if (VT == MVT::i8 || VT == MVT::i1)
27425 return std::make_pair(0U, &X86::GR8RegClass);
27426 if (VT == MVT::i16)
27427 return std::make_pair(0U, &X86::GR16RegClass);
27428 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
27429 return std::make_pair(0U, &X86::GR32RegClass);
27430 return std::make_pair(0U, &X86::GR64RegClass);
27431 case 'R': // LEGACY_REGS
27432 if (VT == MVT::i8 || VT == MVT::i1)
27433 return std::make_pair(0U, &X86::GR8_NOREXRegClass);
27434 if (VT == MVT::i16)
27435 return std::make_pair(0U, &X86::GR16_NOREXRegClass);
27436 if (VT == MVT::i32 || !Subtarget->is64Bit())
27437 return std::make_pair(0U, &X86::GR32_NOREXRegClass);
27438 return std::make_pair(0U, &X86::GR64_NOREXRegClass);
27439 case 'f': // FP Stack registers.
27440 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
27441 // value to the correct fpstack register class.
27442 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
27443 return std::make_pair(0U, &X86::RFP32RegClass);
27444 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
27445 return std::make_pair(0U, &X86::RFP64RegClass);
27446 return std::make_pair(0U, &X86::RFP80RegClass);
27447 case 'y': // MMX_REGS if MMX allowed.
27448 if (!Subtarget->hasMMX()) break;
27449 return std::make_pair(0U, &X86::VR64RegClass);
27450 case 'Y': // SSE_REGS if SSE2 allowed
27451 if (!Subtarget->hasSSE2()) break;
27453 case 'x': // SSE_REGS if SSE1 allowed or AVX_REGS if AVX allowed
27454 if (!Subtarget->hasSSE1()) break;
27456 switch (VT.SimpleTy) {
27458 // Scalar SSE types.
27461 return std::make_pair(0U, &X86::FR32RegClass);
27464 return std::make_pair(0U, &X86::FR64RegClass);
27472 return std::make_pair(0U, &X86::VR128RegClass);
27480 return std::make_pair(0U, &X86::VR256RegClass);
27485 return std::make_pair(0U, &X86::VR512RegClass);
27491 // Use the default implementation in TargetLowering to convert the register
27492 // constraint into a member of a register class.
27493 std::pair<unsigned, const TargetRegisterClass*> Res;
27494 Res = TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
27496 // Not found as a standard register?
27498 // Map st(0) -> st(7) -> ST0
27499 if (Constraint.size() == 7 && Constraint[0] == '{' &&
27500 tolower(Constraint[1]) == 's' &&
27501 tolower(Constraint[2]) == 't' &&
27502 Constraint[3] == '(' &&
27503 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
27504 Constraint[5] == ')' &&
27505 Constraint[6] == '}') {
27507 Res.first = X86::FP0+Constraint[4]-'0';
27508 Res.second = &X86::RFP80RegClass;
27512 // GCC allows "st(0)" to be called just plain "st".
27513 if (StringRef("{st}").equals_lower(Constraint)) {
27514 Res.first = X86::FP0;
27515 Res.second = &X86::RFP80RegClass;
27520 if (StringRef("{flags}").equals_lower(Constraint)) {
27521 Res.first = X86::EFLAGS;
27522 Res.second = &X86::CCRRegClass;
27526 // 'A' means EAX + EDX.
27527 if (Constraint == "A") {
27528 Res.first = X86::EAX;
27529 Res.second = &X86::GR32_ADRegClass;
27535 // Otherwise, check to see if this is a register class of the wrong value
27536 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
27537 // turn into {ax},{dx}.
27538 // MVT::Other is used to specify clobber names.
27539 if (Res.second->hasType(VT) || VT == MVT::Other)
27540 return Res; // Correct type already, nothing to do.
27542 // Get a matching integer of the correct size. i.e. "ax" with MVT::32 should
27543 // return "eax". This should even work for things like getting 64bit integer
27544 // registers when given an f64 type.
27545 const TargetRegisterClass *Class = Res.second;
27546 if (Class == &X86::GR8RegClass || Class == &X86::GR16RegClass ||
27547 Class == &X86::GR32RegClass || Class == &X86::GR64RegClass) {
27548 unsigned Size = VT.getSizeInBits();
27549 MVT::SimpleValueType SimpleTy = Size == 1 || Size == 8 ? MVT::i8
27550 : Size == 16 ? MVT::i16
27551 : Size == 32 ? MVT::i32
27552 : Size == 64 ? MVT::i64
27554 unsigned DestReg = getX86SubSuperRegisterOrZero(Res.first, SimpleTy);
27556 Res.first = DestReg;
27557 Res.second = SimpleTy == MVT::i8 ? &X86::GR8RegClass
27558 : SimpleTy == MVT::i16 ? &X86::GR16RegClass
27559 : SimpleTy == MVT::i32 ? &X86::GR32RegClass
27560 : &X86::GR64RegClass;
27561 assert(Res.second->contains(Res.first) && "Register in register class");
27563 // No register found/type mismatch.
27565 Res.second = nullptr;
27567 } else if (Class == &X86::FR32RegClass || Class == &X86::FR64RegClass ||
27568 Class == &X86::VR128RegClass || Class == &X86::VR256RegClass ||
27569 Class == &X86::FR32XRegClass || Class == &X86::FR64XRegClass ||
27570 Class == &X86::VR128XRegClass || Class == &X86::VR256XRegClass ||
27571 Class == &X86::VR512RegClass) {
27572 // Handle references to XMM physical registers that got mapped into the
27573 // wrong class. This can happen with constraints like {xmm0} where the
27574 // target independent register mapper will just pick the first match it can
27575 // find, ignoring the required type.
27577 if (VT == MVT::f32 || VT == MVT::i32)
27578 Res.second = &X86::FR32RegClass;
27579 else if (VT == MVT::f64 || VT == MVT::i64)
27580 Res.second = &X86::FR64RegClass;
27581 else if (X86::VR128RegClass.hasType(VT))
27582 Res.second = &X86::VR128RegClass;
27583 else if (X86::VR256RegClass.hasType(VT))
27584 Res.second = &X86::VR256RegClass;
27585 else if (X86::VR512RegClass.hasType(VT))
27586 Res.second = &X86::VR512RegClass;
27588 // Type mismatch and not a clobber: Return an error;
27590 Res.second = nullptr;
27597 int X86TargetLowering::getScalingFactorCost(const DataLayout &DL,
27598 const AddrMode &AM, Type *Ty,
27599 unsigned AS) const {
27600 // Scaling factors are not free at all.
27601 // An indexed folded instruction, i.e., inst (reg1, reg2, scale),
27602 // will take 2 allocations in the out of order engine instead of 1
27603 // for plain addressing mode, i.e. inst (reg1).
27605 // vaddps (%rsi,%drx), %ymm0, %ymm1
27606 // Requires two allocations (one for the load, one for the computation)
27608 // vaddps (%rsi), %ymm0, %ymm1
27609 // Requires just 1 allocation, i.e., freeing allocations for other operations
27610 // and having less micro operations to execute.
27612 // For some X86 architectures, this is even worse because for instance for
27613 // stores, the complex addressing mode forces the instruction to use the
27614 // "load" ports instead of the dedicated "store" port.
27615 // E.g., on Haswell:
27616 // vmovaps %ymm1, (%r8, %rdi) can use port 2 or 3.
27617 // vmovaps %ymm1, (%r8) can use port 2, 3, or 7.
27618 if (isLegalAddressingMode(DL, AM, Ty, AS))
27619 // Scale represents reg2 * scale, thus account for 1
27620 // as soon as we use a second register.
27621 return AM.Scale != 0;
27625 bool X86TargetLowering::isIntDivCheap(EVT VT, AttributeSet Attr) const {
27626 // Integer division on x86 is expensive. However, when aggressively optimizing
27627 // for code size, we prefer to use a div instruction, as it is usually smaller
27628 // than the alternative sequence.
27629 // The exception to this is vector division. Since x86 doesn't have vector
27630 // integer division, leaving the division as-is is a loss even in terms of
27631 // size, because it will have to be scalarized, while the alternative code
27632 // sequence can be performed in vector form.
27633 bool OptSize = Attr.hasAttribute(AttributeSet::FunctionIndex,
27634 Attribute::MinSize);
27635 return OptSize && !VT.isVector();
27638 void X86TargetLowering::markInRegArguments(SelectionDAG &DAG,
27639 TargetLowering::ArgListTy& Args) const {
27640 // The MCU psABI requires some arguments to be passed in-register.
27641 // For regular calls, the inreg arguments are marked by the front-end.
27642 // However, for compiler generated library calls, we have to patch this
27644 if (!Subtarget->isTargetMCU() || !Args.size())
27647 unsigned FreeRegs = 3;
27648 for (auto &Arg : Args) {
27649 // For library functions, we do not expect any fancy types.
27650 unsigned Size = DAG.getDataLayout().getTypeSizeInBits(Arg.Ty);
27651 unsigned SizeInRegs = (Size + 31) / 32;
27652 if (SizeInRegs > 2 || SizeInRegs > FreeRegs)
27655 Arg.isInReg = true;
27656 FreeRegs -= SizeInRegs;