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 // The extracted elements must start at a valid index and all mask
7355 // elements must be in the lower half.
7356 if (i > M || M >= HalfSize)
7359 if (Idx < 0 || (Src == V && Idx == (M - i))) {
7370 assert((Idx + Len) <= HalfSize && "Illegal extraction mask");
7371 int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
7372 int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
7373 return DAG.getNode(X86ISD::EXTRQI, DL, VT, Src,
7374 DAG.getConstant(BitLen, DL, MVT::i8),
7375 DAG.getConstant(BitIdx, DL, MVT::i8));
7378 if (SDValue ExtrQ = LowerAsEXTRQ())
7381 // INSERTQ: Extract lowest Len elements from lower half of second source and
7382 // insert over first source, starting at Idx.
7383 // { A[0], .., A[Idx-1], B[0], .., B[Len-1], A[Idx+Len], .., UNDEF, ... }
7384 auto LowerAsInsertQ = [&]() {
7385 for (int Idx = 0; Idx != HalfSize; ++Idx) {
7388 // Attempt to match first source from mask before insertion point.
7389 if (isUndefInRange(Mask, 0, Idx)) {
7391 } else if (isSequentialOrUndefInRange(Mask, 0, Idx, 0)) {
7393 } else if (isSequentialOrUndefInRange(Mask, 0, Idx, Size)) {
7399 // Extend the extraction length looking to match both the insertion of
7400 // the second source and the remaining elements of the first.
7401 for (int Hi = Idx + 1; Hi <= HalfSize; ++Hi) {
7406 if (isSequentialOrUndefInRange(Mask, Idx, Len, 0)) {
7408 } else if (isSequentialOrUndefInRange(Mask, Idx, Len, Size)) {
7414 // Match the remaining elements of the lower half.
7415 if (isUndefInRange(Mask, Hi, HalfSize - Hi)) {
7417 } else if ((!Base || (Base == V1)) &&
7418 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi, Hi)) {
7420 } else if ((!Base || (Base == V2)) &&
7421 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi,
7428 // We may not have a base (first source) - this can safely be undefined.
7430 Base = DAG.getUNDEF(VT);
7432 int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
7433 int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
7434 return DAG.getNode(X86ISD::INSERTQI, DL, VT, Base, Insert,
7435 DAG.getConstant(BitLen, DL, MVT::i8),
7436 DAG.getConstant(BitIdx, DL, MVT::i8));
7443 if (SDValue InsertQ = LowerAsInsertQ())
7449 /// \brief Lower a vector shuffle as a zero or any extension.
7451 /// Given a specific number of elements, element bit width, and extension
7452 /// stride, produce either a zero or any extension based on the available
7453 /// features of the subtarget. The extended elements are consecutive and
7454 /// begin and can start from an offseted element index in the input; to
7455 /// avoid excess shuffling the offset must either being in the bottom lane
7456 /// or at the start of a higher lane. All extended elements must be from
7458 static SDValue lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7459 SDLoc DL, MVT VT, int Scale, int Offset, bool AnyExt, SDValue InputV,
7460 ArrayRef<int> Mask, const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7461 assert(Scale > 1 && "Need a scale to extend.");
7462 int EltBits = VT.getScalarSizeInBits();
7463 int NumElements = VT.getVectorNumElements();
7464 int NumEltsPerLane = 128 / EltBits;
7465 int OffsetLane = Offset / NumEltsPerLane;
7466 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
7467 "Only 8, 16, and 32 bit elements can be extended.");
7468 assert(Scale * EltBits <= 64 && "Cannot zero extend past 64 bits.");
7469 assert(0 <= Offset && "Extension offset must be positive.");
7470 assert((Offset < NumEltsPerLane || Offset % NumEltsPerLane == 0) &&
7471 "Extension offset must be in the first lane or start an upper lane.");
7473 // Check that an index is in same lane as the base offset.
7474 auto SafeOffset = [&](int Idx) {
7475 return OffsetLane == (Idx / NumEltsPerLane);
7478 // Shift along an input so that the offset base moves to the first element.
7479 auto ShuffleOffset = [&](SDValue V) {
7483 SmallVector<int, 8> ShMask((unsigned)NumElements, -1);
7484 for (int i = 0; i * Scale < NumElements; ++i) {
7485 int SrcIdx = i + Offset;
7486 ShMask[i] = SafeOffset(SrcIdx) ? SrcIdx : -1;
7488 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), ShMask);
7491 // Found a valid zext mask! Try various lowering strategies based on the
7492 // input type and available ISA extensions.
7493 if (Subtarget->hasSSE41()) {
7494 // Not worth offseting 128-bit vectors if scale == 2, a pattern using
7495 // PUNPCK will catch this in a later shuffle match.
7496 if (Offset && Scale == 2 && VT.is128BitVector())
7498 MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits * Scale),
7499 NumElements / Scale);
7500 InputV = DAG.getNode(X86ISD::VZEXT, DL, ExtVT, ShuffleOffset(InputV));
7501 return DAG.getBitcast(VT, InputV);
7504 assert(VT.is128BitVector() && "Only 128-bit vectors can be extended.");
7506 // For any extends we can cheat for larger element sizes and use shuffle
7507 // instructions that can fold with a load and/or copy.
7508 if (AnyExt && EltBits == 32) {
7509 int PSHUFDMask[4] = {Offset, -1, SafeOffset(Offset + 1) ? Offset + 1 : -1,
7511 return DAG.getBitcast(
7512 VT, DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7513 DAG.getBitcast(MVT::v4i32, InputV),
7514 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
7516 if (AnyExt && EltBits == 16 && Scale > 2) {
7517 int PSHUFDMask[4] = {Offset / 2, -1,
7518 SafeOffset(Offset + 1) ? (Offset + 1) / 2 : -1, -1};
7519 InputV = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7520 DAG.getBitcast(MVT::v4i32, InputV),
7521 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG));
7522 int PSHUFWMask[4] = {1, -1, -1, -1};
7523 unsigned OddEvenOp = (Offset & 1 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW);
7524 return DAG.getBitcast(
7525 VT, DAG.getNode(OddEvenOp, DL, MVT::v8i16,
7526 DAG.getBitcast(MVT::v8i16, InputV),
7527 getV4X86ShuffleImm8ForMask(PSHUFWMask, DL, DAG)));
7530 // The SSE4A EXTRQ instruction can efficiently extend the first 2 lanes
7532 if ((Scale * EltBits) == 64 && EltBits < 32 && Subtarget->hasSSE4A()) {
7533 assert(NumElements == (int)Mask.size() && "Unexpected shuffle mask size!");
7534 assert(VT.is128BitVector() && "Unexpected vector width!");
7536 int LoIdx = Offset * EltBits;
7537 SDValue Lo = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7538 DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
7539 DAG.getConstant(EltBits, DL, MVT::i8),
7540 DAG.getConstant(LoIdx, DL, MVT::i8)));
7542 if (isUndefInRange(Mask, NumElements / 2, NumElements / 2) ||
7543 !SafeOffset(Offset + 1))
7544 return DAG.getNode(ISD::BITCAST, DL, VT, Lo);
7546 int HiIdx = (Offset + 1) * EltBits;
7547 SDValue Hi = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7548 DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
7549 DAG.getConstant(EltBits, DL, MVT::i8),
7550 DAG.getConstant(HiIdx, DL, MVT::i8)));
7551 return DAG.getNode(ISD::BITCAST, DL, VT,
7552 DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, Lo, Hi));
7555 // If this would require more than 2 unpack instructions to expand, use
7556 // pshufb when available. We can only use more than 2 unpack instructions
7557 // when zero extending i8 elements which also makes it easier to use pshufb.
7558 if (Scale > 4 && EltBits == 8 && Subtarget->hasSSSE3()) {
7559 assert(NumElements == 16 && "Unexpected byte vector width!");
7560 SDValue PSHUFBMask[16];
7561 for (int i = 0; i < 16; ++i) {
7562 int Idx = Offset + (i / Scale);
7563 PSHUFBMask[i] = DAG.getConstant(
7564 (i % Scale == 0 && SafeOffset(Idx)) ? Idx : 0x80, DL, MVT::i8);
7566 InputV = DAG.getBitcast(MVT::v16i8, InputV);
7567 return DAG.getBitcast(VT,
7568 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, InputV,
7569 DAG.getNode(ISD::BUILD_VECTOR, DL,
7570 MVT::v16i8, PSHUFBMask)));
7573 // If we are extending from an offset, ensure we start on a boundary that
7574 // we can unpack from.
7575 int AlignToUnpack = Offset % (NumElements / Scale);
7576 if (AlignToUnpack) {
7577 SmallVector<int, 8> ShMask((unsigned)NumElements, -1);
7578 for (int i = AlignToUnpack; i < NumElements; ++i)
7579 ShMask[i - AlignToUnpack] = i;
7580 InputV = DAG.getVectorShuffle(VT, DL, InputV, DAG.getUNDEF(VT), ShMask);
7581 Offset -= AlignToUnpack;
7584 // Otherwise emit a sequence of unpacks.
7586 unsigned UnpackLoHi = X86ISD::UNPCKL;
7587 if (Offset >= (NumElements / 2)) {
7588 UnpackLoHi = X86ISD::UNPCKH;
7589 Offset -= (NumElements / 2);
7592 MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements);
7593 SDValue Ext = AnyExt ? DAG.getUNDEF(InputVT)
7594 : getZeroVector(InputVT, Subtarget, DAG, DL);
7595 InputV = DAG.getBitcast(InputVT, InputV);
7596 InputV = DAG.getNode(UnpackLoHi, DL, InputVT, InputV, Ext);
7600 } while (Scale > 1);
7601 return DAG.getBitcast(VT, InputV);
7604 /// \brief Try to lower a vector shuffle as a zero extension on any microarch.
7606 /// This routine will try to do everything in its power to cleverly lower
7607 /// a shuffle which happens to match the pattern of a zero extend. It doesn't
7608 /// check for the profitability of this lowering, it tries to aggressively
7609 /// match this pattern. It will use all of the micro-architectural details it
7610 /// can to emit an efficient lowering. It handles both blends with all-zero
7611 /// inputs to explicitly zero-extend and undef-lanes (sometimes undef due to
7612 /// masking out later).
7614 /// The reason we have dedicated lowering for zext-style shuffles is that they
7615 /// are both incredibly common and often quite performance sensitive.
7616 static SDValue lowerVectorShuffleAsZeroOrAnyExtend(
7617 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7618 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7619 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7621 int Bits = VT.getSizeInBits();
7622 int NumLanes = Bits / 128;
7623 int NumElements = VT.getVectorNumElements();
7624 int NumEltsPerLane = NumElements / NumLanes;
7625 assert(VT.getScalarSizeInBits() <= 32 &&
7626 "Exceeds 32-bit integer zero extension limit");
7627 assert((int)Mask.size() == NumElements && "Unexpected shuffle mask size");
7629 // Define a helper function to check a particular ext-scale and lower to it if
7631 auto Lower = [&](int Scale) -> SDValue {
7636 for (int i = 0; i < NumElements; ++i) {
7639 continue; // Valid anywhere but doesn't tell us anything.
7640 if (i % Scale != 0) {
7641 // Each of the extended elements need to be zeroable.
7645 // We no longer are in the anyext case.
7650 // Each of the base elements needs to be consecutive indices into the
7651 // same input vector.
7652 SDValue V = M < NumElements ? V1 : V2;
7653 M = M % NumElements;
7656 Offset = M - (i / Scale);
7657 } else if (InputV != V)
7658 return SDValue(); // Flip-flopping inputs.
7660 // Offset must start in the lowest 128-bit lane or at the start of an
7662 // FIXME: Is it ever worth allowing a negative base offset?
7663 if (!((0 <= Offset && Offset < NumEltsPerLane) ||
7664 (Offset % NumEltsPerLane) == 0))
7667 // If we are offsetting, all referenced entries must come from the same
7669 if (Offset && (Offset / NumEltsPerLane) != (M / NumEltsPerLane))
7672 if ((M % NumElements) != (Offset + (i / Scale)))
7673 return SDValue(); // Non-consecutive strided elements.
7677 // If we fail to find an input, we have a zero-shuffle which should always
7678 // have already been handled.
7679 // FIXME: Maybe handle this here in case during blending we end up with one?
7683 // If we are offsetting, don't extend if we only match a single input, we
7684 // can always do better by using a basic PSHUF or PUNPCK.
7685 if (Offset != 0 && Matches < 2)
7688 return lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7689 DL, VT, Scale, Offset, AnyExt, InputV, Mask, Subtarget, DAG);
7692 // The widest scale possible for extending is to a 64-bit integer.
7693 assert(Bits % 64 == 0 &&
7694 "The number of bits in a vector must be divisible by 64 on x86!");
7695 int NumExtElements = Bits / 64;
7697 // Each iteration, try extending the elements half as much, but into twice as
7699 for (; NumExtElements < NumElements; NumExtElements *= 2) {
7700 assert(NumElements % NumExtElements == 0 &&
7701 "The input vector size must be divisible by the extended size.");
7702 if (SDValue V = Lower(NumElements / NumExtElements))
7706 // General extends failed, but 128-bit vectors may be able to use MOVQ.
7710 // Returns one of the source operands if the shuffle can be reduced to a
7711 // MOVQ, copying the lower 64-bits and zero-extending to the upper 64-bits.
7712 auto CanZExtLowHalf = [&]() {
7713 for (int i = NumElements / 2; i != NumElements; ++i)
7716 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, 0))
7718 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, NumElements))
7723 if (SDValue V = CanZExtLowHalf()) {
7724 V = DAG.getBitcast(MVT::v2i64, V);
7725 V = DAG.getNode(X86ISD::VZEXT_MOVL, DL, MVT::v2i64, V);
7726 return DAG.getBitcast(VT, V);
7729 // No viable ext lowering found.
7733 /// \brief Try to get a scalar value for a specific element of a vector.
7735 /// Looks through BUILD_VECTOR and SCALAR_TO_VECTOR nodes to find a scalar.
7736 static SDValue getScalarValueForVectorElement(SDValue V, int Idx,
7737 SelectionDAG &DAG) {
7738 MVT VT = V.getSimpleValueType();
7739 MVT EltVT = VT.getVectorElementType();
7740 while (V.getOpcode() == ISD::BITCAST)
7741 V = V.getOperand(0);
7742 // If the bitcasts shift the element size, we can't extract an equivalent
7744 MVT NewVT = V.getSimpleValueType();
7745 if (!NewVT.isVector() || NewVT.getScalarSizeInBits() != VT.getScalarSizeInBits())
7748 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7749 (Idx == 0 && V.getOpcode() == ISD::SCALAR_TO_VECTOR)) {
7750 // Ensure the scalar operand is the same size as the destination.
7751 // FIXME: Add support for scalar truncation where possible.
7752 SDValue S = V.getOperand(Idx);
7753 if (EltVT.getSizeInBits() == S.getSimpleValueType().getSizeInBits())
7754 return DAG.getNode(ISD::BITCAST, SDLoc(V), EltVT, S);
7760 /// \brief Helper to test for a load that can be folded with x86 shuffles.
7762 /// This is particularly important because the set of instructions varies
7763 /// significantly based on whether the operand is a load or not.
7764 static bool isShuffleFoldableLoad(SDValue V) {
7765 while (V.getOpcode() == ISD::BITCAST)
7766 V = V.getOperand(0);
7768 return ISD::isNON_EXTLoad(V.getNode());
7771 /// \brief Try to lower insertion of a single element into a zero vector.
7773 /// This is a common pattern that we have especially efficient patterns to lower
7774 /// across all subtarget feature sets.
7775 static SDValue lowerVectorShuffleAsElementInsertion(
7776 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7777 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7778 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7780 MVT EltVT = VT.getVectorElementType();
7782 int V2Index = std::find_if(Mask.begin(), Mask.end(),
7783 [&Mask](int M) { return M >= (int)Mask.size(); }) -
7785 bool IsV1Zeroable = true;
7786 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7787 if (i != V2Index && !Zeroable[i]) {
7788 IsV1Zeroable = false;
7792 // Check for a single input from a SCALAR_TO_VECTOR node.
7793 // FIXME: All of this should be canonicalized into INSERT_VECTOR_ELT and
7794 // all the smarts here sunk into that routine. However, the current
7795 // lowering of BUILD_VECTOR makes that nearly impossible until the old
7796 // vector shuffle lowering is dead.
7797 SDValue V2S = getScalarValueForVectorElement(V2, Mask[V2Index] - Mask.size(),
7799 if (V2S && DAG.getTargetLoweringInfo().isTypeLegal(V2S.getValueType())) {
7800 // We need to zext the scalar if it is smaller than an i32.
7801 V2S = DAG.getBitcast(EltVT, V2S);
7802 if (EltVT == MVT::i8 || EltVT == MVT::i16) {
7803 // Using zext to expand a narrow element won't work for non-zero
7808 // Zero-extend directly to i32.
7810 V2S = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, V2S);
7812 V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtVT, V2S);
7813 } else if (Mask[V2Index] != (int)Mask.size() || EltVT == MVT::i8 ||
7814 EltVT == MVT::i16) {
7815 // Either not inserting from the low element of the input or the input
7816 // element size is too small to use VZEXT_MOVL to clear the high bits.
7820 if (!IsV1Zeroable) {
7821 // If V1 can't be treated as a zero vector we have fewer options to lower
7822 // this. We can't support integer vectors or non-zero targets cheaply, and
7823 // the V1 elements can't be permuted in any way.
7824 assert(VT == ExtVT && "Cannot change extended type when non-zeroable!");
7825 if (!VT.isFloatingPoint() || V2Index != 0)
7827 SmallVector<int, 8> V1Mask(Mask.begin(), Mask.end());
7828 V1Mask[V2Index] = -1;
7829 if (!isNoopShuffleMask(V1Mask))
7831 // This is essentially a special case blend operation, but if we have
7832 // general purpose blend operations, they are always faster. Bail and let
7833 // the rest of the lowering handle these as blends.
7834 if (Subtarget->hasSSE41())
7837 // Otherwise, use MOVSD or MOVSS.
7838 assert((EltVT == MVT::f32 || EltVT == MVT::f64) &&
7839 "Only two types of floating point element types to handle!");
7840 return DAG.getNode(EltVT == MVT::f32 ? X86ISD::MOVSS : X86ISD::MOVSD, DL,
7844 // This lowering only works for the low element with floating point vectors.
7845 if (VT.isFloatingPoint() && V2Index != 0)
7848 V2 = DAG.getNode(X86ISD::VZEXT_MOVL, DL, ExtVT, V2);
7850 V2 = DAG.getBitcast(VT, V2);
7853 // If we have 4 or fewer lanes we can cheaply shuffle the element into
7854 // the desired position. Otherwise it is more efficient to do a vector
7855 // shift left. We know that we can do a vector shift left because all
7856 // the inputs are zero.
7857 if (VT.isFloatingPoint() || VT.getVectorNumElements() <= 4) {
7858 SmallVector<int, 4> V2Shuffle(Mask.size(), 1);
7859 V2Shuffle[V2Index] = 0;
7860 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Shuffle);
7862 V2 = DAG.getBitcast(MVT::v2i64, V2);
7864 X86ISD::VSHLDQ, DL, MVT::v2i64, V2,
7865 DAG.getConstant(V2Index * EltVT.getSizeInBits() / 8, DL,
7866 DAG.getTargetLoweringInfo().getScalarShiftAmountTy(
7867 DAG.getDataLayout(), VT)));
7868 V2 = DAG.getBitcast(VT, V2);
7874 /// \brief Try to lower broadcast of a single - truncated - integer element,
7875 /// coming from a scalar_to_vector/build_vector node \p V0 with larger elements.
7877 /// This assumes we have AVX2.
7878 static SDValue lowerVectorShuffleAsTruncBroadcast(SDLoc DL, MVT VT, SDValue V0,
7880 const X86Subtarget *Subtarget,
7881 SelectionDAG &DAG) {
7882 assert(Subtarget->hasAVX2() &&
7883 "We can only lower integer broadcasts with AVX2!");
7885 EVT EltVT = VT.getVectorElementType();
7886 EVT V0VT = V0.getValueType();
7888 assert(VT.isInteger() && "Unexpected non-integer trunc broadcast!");
7889 assert(V0VT.isVector() && "Unexpected non-vector vector-sized value!");
7891 EVT V0EltVT = V0VT.getVectorElementType();
7892 if (!V0EltVT.isInteger())
7895 const unsigned EltSize = EltVT.getSizeInBits();
7896 const unsigned V0EltSize = V0EltVT.getSizeInBits();
7898 // This is only a truncation if the original element type is larger.
7899 if (V0EltSize <= EltSize)
7902 assert(((V0EltSize % EltSize) == 0) &&
7903 "Scalar type sizes must all be powers of 2 on x86!");
7905 const unsigned V0Opc = V0.getOpcode();
7906 const unsigned Scale = V0EltSize / EltSize;
7907 const unsigned V0BroadcastIdx = BroadcastIdx / Scale;
7909 if ((V0Opc != ISD::SCALAR_TO_VECTOR || V0BroadcastIdx != 0) &&
7910 V0Opc != ISD::BUILD_VECTOR)
7913 SDValue Scalar = V0.getOperand(V0BroadcastIdx);
7915 // If we're extracting non-least-significant bits, shift so we can truncate.
7916 // Hopefully, we can fold away the trunc/srl/load into the broadcast.
7917 // Even if we can't (and !isShuffleFoldableLoad(Scalar)), prefer
7918 // vpbroadcast+vmovd+shr to vpshufb(m)+vmovd.
7919 if (const int OffsetIdx = BroadcastIdx % Scale)
7920 Scalar = DAG.getNode(ISD::SRL, DL, Scalar.getValueType(), Scalar,
7921 DAG.getConstant(OffsetIdx * EltSize, DL, Scalar.getValueType()));
7923 return DAG.getNode(X86ISD::VBROADCAST, DL, VT,
7924 DAG.getNode(ISD::TRUNCATE, DL, EltVT, Scalar));
7927 /// \brief Try to lower broadcast of a single element.
7929 /// For convenience, this code also bundles all of the subtarget feature set
7930 /// filtering. While a little annoying to re-dispatch on type here, there isn't
7931 /// a convenient way to factor it out.
7932 static SDValue lowerVectorShuffleAsBroadcast(SDLoc DL, MVT VT, SDValue V,
7934 const X86Subtarget *Subtarget,
7935 SelectionDAG &DAG) {
7936 if (!Subtarget->hasAVX())
7938 if (VT.isInteger() && !Subtarget->hasAVX2())
7941 // Check that the mask is a broadcast.
7942 int BroadcastIdx = -1;
7944 if (M >= 0 && BroadcastIdx == -1)
7946 else if (M >= 0 && M != BroadcastIdx)
7949 assert(BroadcastIdx < (int)Mask.size() && "We only expect to be called with "
7950 "a sorted mask where the broadcast "
7953 // Go up the chain of (vector) values to find a scalar load that we can
7954 // combine with the broadcast.
7956 switch (V.getOpcode()) {
7957 case ISD::CONCAT_VECTORS: {
7958 int OperandSize = Mask.size() / V.getNumOperands();
7959 V = V.getOperand(BroadcastIdx / OperandSize);
7960 BroadcastIdx %= OperandSize;
7964 case ISD::INSERT_SUBVECTOR: {
7965 SDValue VOuter = V.getOperand(0), VInner = V.getOperand(1);
7966 auto ConstantIdx = dyn_cast<ConstantSDNode>(V.getOperand(2));
7970 int BeginIdx = (int)ConstantIdx->getZExtValue();
7972 BeginIdx + (int)VInner.getSimpleValueType().getVectorNumElements();
7973 if (BroadcastIdx >= BeginIdx && BroadcastIdx < EndIdx) {
7974 BroadcastIdx -= BeginIdx;
7985 // Check if this is a broadcast of a scalar. We special case lowering
7986 // for scalars so that we can more effectively fold with loads.
7987 // First, look through bitcast: if the original value has a larger element
7988 // type than the shuffle, the broadcast element is in essence truncated.
7989 // Make that explicit to ease folding.
7990 if (V.getOpcode() == ISD::BITCAST && VT.isInteger())
7991 if (SDValue TruncBroadcast = lowerVectorShuffleAsTruncBroadcast(
7992 DL, VT, V.getOperand(0), BroadcastIdx, Subtarget, DAG))
7993 return TruncBroadcast;
7995 // Also check the simpler case, where we can directly reuse the scalar.
7996 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7997 (V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)) {
7998 V = V.getOperand(BroadcastIdx);
8000 // If the scalar isn't a load, we can't broadcast from it in AVX1.
8001 // Only AVX2 has register broadcasts.
8002 if (!Subtarget->hasAVX2() && !isShuffleFoldableLoad(V))
8004 } else if (BroadcastIdx != 0 || !Subtarget->hasAVX2()) {
8005 // We can't broadcast from a vector register without AVX2, and we can only
8006 // broadcast from the zero-element of a vector register.
8010 return DAG.getNode(X86ISD::VBROADCAST, DL, VT, V);
8013 // Check for whether we can use INSERTPS to perform the shuffle. We only use
8014 // INSERTPS when the V1 elements are already in the correct locations
8015 // because otherwise we can just always use two SHUFPS instructions which
8016 // are much smaller to encode than a SHUFPS and an INSERTPS. We can also
8017 // perform INSERTPS if a single V1 element is out of place and all V2
8018 // elements are zeroable.
8019 static SDValue lowerVectorShuffleAsInsertPS(SDValue Op, SDValue V1, SDValue V2,
8021 SelectionDAG &DAG) {
8022 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
8023 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8024 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8025 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8027 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
8030 int V1DstIndex = -1;
8031 int V2DstIndex = -1;
8032 bool V1UsedInPlace = false;
8034 for (int i = 0; i < 4; ++i) {
8035 // Synthesize a zero mask from the zeroable elements (includes undefs).
8041 // Flag if we use any V1 inputs in place.
8043 V1UsedInPlace = true;
8047 // We can only insert a single non-zeroable element.
8048 if (V1DstIndex != -1 || V2DstIndex != -1)
8052 // V1 input out of place for insertion.
8055 // V2 input for insertion.
8060 // Don't bother if we have no (non-zeroable) element for insertion.
8061 if (V1DstIndex == -1 && V2DstIndex == -1)
8064 // Determine element insertion src/dst indices. The src index is from the
8065 // start of the inserted vector, not the start of the concatenated vector.
8066 unsigned V2SrcIndex = 0;
8067 if (V1DstIndex != -1) {
8068 // If we have a V1 input out of place, we use V1 as the V2 element insertion
8069 // and don't use the original V2 at all.
8070 V2SrcIndex = Mask[V1DstIndex];
8071 V2DstIndex = V1DstIndex;
8074 V2SrcIndex = Mask[V2DstIndex] - 4;
8077 // If no V1 inputs are used in place, then the result is created only from
8078 // the zero mask and the V2 insertion - so remove V1 dependency.
8080 V1 = DAG.getUNDEF(MVT::v4f32);
8082 unsigned InsertPSMask = V2SrcIndex << 6 | V2DstIndex << 4 | ZMask;
8083 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
8085 // Insert the V2 element into the desired position.
8087 return DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
8088 DAG.getConstant(InsertPSMask, DL, MVT::i8));
8091 /// \brief Try to lower a shuffle as a permute of the inputs followed by an
8092 /// UNPCK instruction.
8094 /// This specifically targets cases where we end up with alternating between
8095 /// the two inputs, and so can permute them into something that feeds a single
8096 /// UNPCK instruction. Note that this routine only targets integer vectors
8097 /// because for floating point vectors we have a generalized SHUFPS lowering
8098 /// strategy that handles everything that doesn't *exactly* match an unpack,
8099 /// making this clever lowering unnecessary.
8100 static SDValue lowerVectorShuffleAsPermuteAndUnpack(SDLoc DL, MVT VT,
8101 SDValue V1, SDValue V2,
8103 SelectionDAG &DAG) {
8104 assert(!VT.isFloatingPoint() &&
8105 "This routine only supports integer vectors.");
8106 assert(!isSingleInputShuffleMask(Mask) &&
8107 "This routine should only be used when blending two inputs.");
8108 assert(Mask.size() >= 2 && "Single element masks are invalid.");
8110 int Size = Mask.size();
8112 int NumLoInputs = std::count_if(Mask.begin(), Mask.end(), [Size](int M) {
8113 return M >= 0 && M % Size < Size / 2;
8115 int NumHiInputs = std::count_if(
8116 Mask.begin(), Mask.end(), [Size](int M) { return M % Size >= Size / 2; });
8118 bool UnpackLo = NumLoInputs >= NumHiInputs;
8120 auto TryUnpack = [&](MVT UnpackVT, int Scale) {
8121 SmallVector<int, 32> V1Mask(Mask.size(), -1);
8122 SmallVector<int, 32> V2Mask(Mask.size(), -1);
8124 for (int i = 0; i < Size; ++i) {
8128 // Each element of the unpack contains Scale elements from this mask.
8129 int UnpackIdx = i / Scale;
8131 // We only handle the case where V1 feeds the first slots of the unpack.
8132 // We rely on canonicalization to ensure this is the case.
8133 if ((UnpackIdx % 2 == 0) != (Mask[i] < Size))
8136 // Setup the mask for this input. The indexing is tricky as we have to
8137 // handle the unpack stride.
8138 SmallVectorImpl<int> &VMask = (UnpackIdx % 2 == 0) ? V1Mask : V2Mask;
8139 VMask[(UnpackIdx / 2) * Scale + i % Scale + (UnpackLo ? 0 : Size / 2)] =
8143 // If we will have to shuffle both inputs to use the unpack, check whether
8144 // we can just unpack first and shuffle the result. If so, skip this unpack.
8145 if ((NumLoInputs == 0 || NumHiInputs == 0) && !isNoopShuffleMask(V1Mask) &&
8146 !isNoopShuffleMask(V2Mask))
8149 // Shuffle the inputs into place.
8150 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
8151 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
8153 // Cast the inputs to the type we will use to unpack them.
8154 V1 = DAG.getBitcast(UnpackVT, V1);
8155 V2 = DAG.getBitcast(UnpackVT, V2);
8157 // Unpack the inputs and cast the result back to the desired type.
8158 return DAG.getBitcast(
8159 VT, DAG.getNode(UnpackLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
8163 // We try each unpack from the largest to the smallest to try and find one
8164 // that fits this mask.
8165 int OrigNumElements = VT.getVectorNumElements();
8166 int OrigScalarSize = VT.getScalarSizeInBits();
8167 for (int ScalarSize = 64; ScalarSize >= OrigScalarSize; ScalarSize /= 2) {
8168 int Scale = ScalarSize / OrigScalarSize;
8169 int NumElements = OrigNumElements / Scale;
8170 MVT UnpackVT = MVT::getVectorVT(MVT::getIntegerVT(ScalarSize), NumElements);
8171 if (SDValue Unpack = TryUnpack(UnpackVT, Scale))
8175 // If none of the unpack-rooted lowerings worked (or were profitable) try an
8177 if (NumLoInputs == 0 || NumHiInputs == 0) {
8178 assert((NumLoInputs > 0 || NumHiInputs > 0) &&
8179 "We have to have *some* inputs!");
8180 int HalfOffset = NumLoInputs == 0 ? Size / 2 : 0;
8182 // FIXME: We could consider the total complexity of the permute of each
8183 // possible unpacking. Or at the least we should consider how many
8184 // half-crossings are created.
8185 // FIXME: We could consider commuting the unpacks.
8187 SmallVector<int, 32> PermMask;
8188 PermMask.assign(Size, -1);
8189 for (int i = 0; i < Size; ++i) {
8193 assert(Mask[i] % Size >= HalfOffset && "Found input from wrong half!");
8196 2 * ((Mask[i] % Size) - HalfOffset) + (Mask[i] < Size ? 0 : 1);
8198 return DAG.getVectorShuffle(
8199 VT, DL, DAG.getNode(NumLoInputs == 0 ? X86ISD::UNPCKH : X86ISD::UNPCKL,
8201 DAG.getUNDEF(VT), PermMask);
8207 /// \brief Handle lowering of 2-lane 64-bit floating point shuffles.
8209 /// This is the basis function for the 2-lane 64-bit shuffles as we have full
8210 /// support for floating point shuffles but not integer shuffles. These
8211 /// instructions will incur a domain crossing penalty on some chips though so
8212 /// it is better to avoid lowering through this for integer vectors where
8214 static SDValue lowerV2F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8215 const X86Subtarget *Subtarget,
8216 SelectionDAG &DAG) {
8218 assert(Op.getSimpleValueType() == MVT::v2f64 && "Bad shuffle type!");
8219 assert(V1.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
8220 assert(V2.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
8221 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8222 ArrayRef<int> Mask = SVOp->getMask();
8223 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
8225 if (isSingleInputShuffleMask(Mask)) {
8226 // Use low duplicate instructions for masks that match their pattern.
8227 if (Subtarget->hasSSE3())
8228 if (isShuffleEquivalent(V1, V2, Mask, {0, 0}))
8229 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v2f64, V1);
8231 // Straight shuffle of a single input vector. Simulate this by using the
8232 // single input as both of the "inputs" to this instruction..
8233 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1);
8235 if (Subtarget->hasAVX()) {
8236 // If we have AVX, we can use VPERMILPS which will allow folding a load
8237 // into the shuffle.
8238 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v2f64, V1,
8239 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8242 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V1,
8243 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8245 assert(Mask[0] >= 0 && Mask[0] < 2 && "Non-canonicalized blend!");
8246 assert(Mask[1] >= 2 && "Non-canonicalized blend!");
8248 // If we have a single input, insert that into V1 if we can do so cheaply.
8249 if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1) {
8250 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8251 DL, MVT::v2f64, V1, V2, Mask, Subtarget, DAG))
8253 // Try inverting the insertion since for v2 masks it is easy to do and we
8254 // can't reliably sort the mask one way or the other.
8255 int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2),
8256 Mask[1] < 0 ? -1 : (Mask[1] ^ 2)};
8257 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8258 DL, MVT::v2f64, V2, V1, InverseMask, Subtarget, DAG))
8262 // Try to use one of the special instruction patterns to handle two common
8263 // blend patterns if a zero-blend above didn't work.
8264 if (isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
8265 isShuffleEquivalent(V1, V2, Mask, {1, 3}))
8266 if (SDValue V1S = getScalarValueForVectorElement(V1, Mask[0], DAG))
8267 // We can either use a special instruction to load over the low double or
8268 // to move just the low double.
8270 isShuffleFoldableLoad(V1S) ? X86ISD::MOVLPD : X86ISD::MOVSD,
8272 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, V1S));
8274 if (Subtarget->hasSSE41())
8275 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2f64, V1, V2, Mask,
8279 // Use dedicated unpack instructions for masks that match their pattern.
8281 lowerVectorShuffleWithUNPCK(DL, MVT::v2f64, Mask, V1, V2, DAG))
8284 unsigned SHUFPDMask = (Mask[0] == 1) | (((Mask[1] - 2) == 1) << 1);
8285 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V2,
8286 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8289 /// \brief Handle lowering of 2-lane 64-bit integer shuffles.
8291 /// Tries to lower a 2-lane 64-bit shuffle using shuffle operations provided by
8292 /// the integer unit to minimize domain crossing penalties. However, for blends
8293 /// it falls back to the floating point shuffle operation with appropriate bit
8295 static SDValue lowerV2I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8296 const X86Subtarget *Subtarget,
8297 SelectionDAG &DAG) {
8299 assert(Op.getSimpleValueType() == MVT::v2i64 && "Bad shuffle type!");
8300 assert(V1.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
8301 assert(V2.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
8302 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8303 ArrayRef<int> Mask = SVOp->getMask();
8304 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
8306 if (isSingleInputShuffleMask(Mask)) {
8307 // Check for being able to broadcast a single element.
8308 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v2i64, V1,
8309 Mask, Subtarget, DAG))
8312 // Straight shuffle of a single input vector. For everything from SSE2
8313 // onward this has a single fast instruction with no scary immediates.
8314 // We have to map the mask as it is actually a v4i32 shuffle instruction.
8315 V1 = DAG.getBitcast(MVT::v4i32, V1);
8316 int WidenedMask[4] = {
8317 std::max(Mask[0], 0) * 2, std::max(Mask[0], 0) * 2 + 1,
8318 std::max(Mask[1], 0) * 2, std::max(Mask[1], 0) * 2 + 1};
8319 return DAG.getBitcast(
8321 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
8322 getV4X86ShuffleImm8ForMask(WidenedMask, DL, DAG)));
8324 assert(Mask[0] != -1 && "No undef lanes in multi-input v2 shuffles!");
8325 assert(Mask[1] != -1 && "No undef lanes in multi-input v2 shuffles!");
8326 assert(Mask[0] < 2 && "We sort V1 to be the first input.");
8327 assert(Mask[1] >= 2 && "We sort V2 to be the second input.");
8329 // If we have a blend of two PACKUS operations an the blend aligns with the
8330 // low and half halves, we can just merge the PACKUS operations. This is
8331 // particularly important as it lets us merge shuffles that this routine itself
8333 auto GetPackNode = [](SDValue V) {
8334 while (V.getOpcode() == ISD::BITCAST)
8335 V = V.getOperand(0);
8337 return V.getOpcode() == X86ISD::PACKUS ? V : SDValue();
8339 if (SDValue V1Pack = GetPackNode(V1))
8340 if (SDValue V2Pack = GetPackNode(V2))
8341 return DAG.getBitcast(MVT::v2i64,
8342 DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8,
8343 Mask[0] == 0 ? V1Pack.getOperand(0)
8344 : V1Pack.getOperand(1),
8345 Mask[1] == 2 ? V2Pack.getOperand(0)
8346 : V2Pack.getOperand(1)));
8348 // Try to use shift instructions.
8350 lowerVectorShuffleAsShift(DL, MVT::v2i64, V1, V2, Mask, DAG))
8353 // When loading a scalar and then shuffling it into a vector we can often do
8354 // the insertion cheaply.
8355 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8356 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
8358 // Try inverting the insertion since for v2 masks it is easy to do and we
8359 // can't reliably sort the mask one way or the other.
8360 int InverseMask[2] = {Mask[0] ^ 2, Mask[1] ^ 2};
8361 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8362 DL, MVT::v2i64, V2, V1, InverseMask, Subtarget, DAG))
8365 // We have different paths for blend lowering, but they all must use the
8366 // *exact* same predicate.
8367 bool IsBlendSupported = Subtarget->hasSSE41();
8368 if (IsBlendSupported)
8369 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2i64, V1, V2, Mask,
8373 // Use dedicated unpack instructions for masks that match their pattern.
8375 lowerVectorShuffleWithUNPCK(DL, MVT::v2i64, Mask, V1, V2, DAG))
8378 // Try to use byte rotation instructions.
8379 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8380 if (Subtarget->hasSSSE3())
8381 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8382 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
8385 // If we have direct support for blends, we should lower by decomposing into
8386 // a permute. That will be faster than the domain cross.
8387 if (IsBlendSupported)
8388 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v2i64, V1, V2,
8391 // We implement this with SHUFPD which is pretty lame because it will likely
8392 // incur 2 cycles of stall for integer vectors on Nehalem and older chips.
8393 // However, all the alternatives are still more cycles and newer chips don't
8394 // have this problem. It would be really nice if x86 had better shuffles here.
8395 V1 = DAG.getBitcast(MVT::v2f64, V1);
8396 V2 = DAG.getBitcast(MVT::v2f64, V2);
8397 return DAG.getBitcast(MVT::v2i64,
8398 DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask));
8401 /// \brief Test whether this can be lowered with a single SHUFPS instruction.
8403 /// This is used to disable more specialized lowerings when the shufps lowering
8404 /// will happen to be efficient.
8405 static bool isSingleSHUFPSMask(ArrayRef<int> Mask) {
8406 // This routine only handles 128-bit shufps.
8407 assert(Mask.size() == 4 && "Unsupported mask size!");
8409 // To lower with a single SHUFPS we need to have the low half and high half
8410 // each requiring a single input.
8411 if (Mask[0] != -1 && Mask[1] != -1 && (Mask[0] < 4) != (Mask[1] < 4))
8413 if (Mask[2] != -1 && Mask[3] != -1 && (Mask[2] < 4) != (Mask[3] < 4))
8419 /// \brief Lower a vector shuffle using the SHUFPS instruction.
8421 /// This is a helper routine dedicated to lowering vector shuffles using SHUFPS.
8422 /// It makes no assumptions about whether this is the *best* lowering, it simply
8424 static SDValue lowerVectorShuffleWithSHUFPS(SDLoc DL, MVT VT,
8425 ArrayRef<int> Mask, SDValue V1,
8426 SDValue V2, SelectionDAG &DAG) {
8427 SDValue LowV = V1, HighV = V2;
8428 int NewMask[4] = {Mask[0], Mask[1], Mask[2], Mask[3]};
8431 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8433 if (NumV2Elements == 1) {
8435 std::find_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }) -
8438 // Compute the index adjacent to V2Index and in the same half by toggling
8440 int V2AdjIndex = V2Index ^ 1;
8442 if (Mask[V2AdjIndex] == -1) {
8443 // Handles all the cases where we have a single V2 element and an undef.
8444 // This will only ever happen in the high lanes because we commute the
8445 // vector otherwise.
8447 std::swap(LowV, HighV);
8448 NewMask[V2Index] -= 4;
8450 // Handle the case where the V2 element ends up adjacent to a V1 element.
8451 // To make this work, blend them together as the first step.
8452 int V1Index = V2AdjIndex;
8453 int BlendMask[4] = {Mask[V2Index] - 4, 0, Mask[V1Index], 0};
8454 V2 = DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
8455 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
8457 // Now proceed to reconstruct the final blend as we have the necessary
8458 // high or low half formed.
8465 NewMask[V1Index] = 2; // We put the V1 element in V2[2].
8466 NewMask[V2Index] = 0; // We shifted the V2 element into V2[0].
8468 } else if (NumV2Elements == 2) {
8469 if (Mask[0] < 4 && Mask[1] < 4) {
8470 // Handle the easy case where we have V1 in the low lanes and V2 in the
8474 } else if (Mask[2] < 4 && Mask[3] < 4) {
8475 // We also handle the reversed case because this utility may get called
8476 // when we detect a SHUFPS pattern but can't easily commute the shuffle to
8477 // arrange things in the right direction.
8483 // We have a mixture of V1 and V2 in both low and high lanes. Rather than
8484 // trying to place elements directly, just blend them and set up the final
8485 // shuffle to place them.
8487 // The first two blend mask elements are for V1, the second two are for
8489 int BlendMask[4] = {Mask[0] < 4 ? Mask[0] : Mask[1],
8490 Mask[2] < 4 ? Mask[2] : Mask[3],
8491 (Mask[0] >= 4 ? Mask[0] : Mask[1]) - 4,
8492 (Mask[2] >= 4 ? Mask[2] : Mask[3]) - 4};
8493 V1 = DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
8494 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
8496 // Now we do a normal shuffle of V1 by giving V1 as both operands to
8499 NewMask[0] = Mask[0] < 4 ? 0 : 2;
8500 NewMask[1] = Mask[0] < 4 ? 2 : 0;
8501 NewMask[2] = Mask[2] < 4 ? 1 : 3;
8502 NewMask[3] = Mask[2] < 4 ? 3 : 1;
8505 return DAG.getNode(X86ISD::SHUFP, DL, VT, LowV, HighV,
8506 getV4X86ShuffleImm8ForMask(NewMask, DL, DAG));
8509 /// \brief Lower 4-lane 32-bit floating point shuffles.
8511 /// Uses instructions exclusively from the floating point unit to minimize
8512 /// domain crossing penalties, as these are sufficient to implement all v4f32
8514 static SDValue lowerV4F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8515 const X86Subtarget *Subtarget,
8516 SelectionDAG &DAG) {
8518 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
8519 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8520 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8521 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8522 ArrayRef<int> Mask = SVOp->getMask();
8523 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8526 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8528 if (NumV2Elements == 0) {
8529 // Check for being able to broadcast a single element.
8530 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f32, V1,
8531 Mask, Subtarget, DAG))
8534 // Use even/odd duplicate instructions for masks that match their pattern.
8535 if (Subtarget->hasSSE3()) {
8536 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
8537 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v4f32, V1);
8538 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3}))
8539 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v4f32, V1);
8542 if (Subtarget->hasAVX()) {
8543 // If we have AVX, we can use VPERMILPS which will allow folding a load
8544 // into the shuffle.
8545 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f32, V1,
8546 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8549 // Otherwise, use a straight shuffle of a single input vector. We pass the
8550 // input vector to both operands to simulate this with a SHUFPS.
8551 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1,
8552 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8555 // There are special ways we can lower some single-element blends. However, we
8556 // have custom ways we can lower more complex single-element blends below that
8557 // we defer to if both this and BLENDPS fail to match, so restrict this to
8558 // when the V2 input is targeting element 0 of the mask -- that is the fast
8560 if (NumV2Elements == 1 && Mask[0] >= 4)
8561 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4f32, V1, V2,
8562 Mask, Subtarget, DAG))
8565 if (Subtarget->hasSSE41()) {
8566 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f32, V1, V2, Mask,
8570 // Use INSERTPS if we can complete the shuffle efficiently.
8571 if (SDValue V = lowerVectorShuffleAsInsertPS(Op, V1, V2, Mask, DAG))
8574 if (!isSingleSHUFPSMask(Mask))
8575 if (SDValue BlendPerm = lowerVectorShuffleAsBlendAndPermute(
8576 DL, MVT::v4f32, V1, V2, Mask, DAG))
8580 // Use dedicated unpack instructions for masks that match their pattern.
8582 lowerVectorShuffleWithUNPCK(DL, MVT::v4f32, Mask, V1, V2, DAG))
8585 // Otherwise fall back to a SHUFPS lowering strategy.
8586 return lowerVectorShuffleWithSHUFPS(DL, MVT::v4f32, Mask, V1, V2, DAG);
8589 /// \brief Lower 4-lane i32 vector shuffles.
8591 /// We try to handle these with integer-domain shuffles where we can, but for
8592 /// blends we use the floating point domain blend instructions.
8593 static SDValue lowerV4I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8594 const X86Subtarget *Subtarget,
8595 SelectionDAG &DAG) {
8597 assert(Op.getSimpleValueType() == MVT::v4i32 && "Bad shuffle type!");
8598 assert(V1.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8599 assert(V2.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8600 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8601 ArrayRef<int> Mask = SVOp->getMask();
8602 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8604 // Whenever we can lower this as a zext, that instruction is strictly faster
8605 // than any alternative. It also allows us to fold memory operands into the
8606 // shuffle in many cases.
8607 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v4i32, V1, V2,
8608 Mask, Subtarget, DAG))
8612 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8614 if (NumV2Elements == 0) {
8615 // Check for being able to broadcast a single element.
8616 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i32, V1,
8617 Mask, Subtarget, DAG))
8620 // Straight shuffle of a single input vector. For everything from SSE2
8621 // onward this has a single fast instruction with no scary immediates.
8622 // We coerce the shuffle pattern to be compatible with UNPCK instructions
8623 // but we aren't actually going to use the UNPCK instruction because doing
8624 // so prevents folding a load into this instruction or making a copy.
8625 const int UnpackLoMask[] = {0, 0, 1, 1};
8626 const int UnpackHiMask[] = {2, 2, 3, 3};
8627 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 1, 1}))
8628 Mask = UnpackLoMask;
8629 else if (isShuffleEquivalent(V1, V2, Mask, {2, 2, 3, 3}))
8630 Mask = UnpackHiMask;
8632 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
8633 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8636 // Try to use shift instructions.
8638 lowerVectorShuffleAsShift(DL, MVT::v4i32, V1, V2, Mask, DAG))
8641 // There are special ways we can lower some single-element blends.
8642 if (NumV2Elements == 1)
8643 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4i32, V1, V2,
8644 Mask, Subtarget, DAG))
8647 // We have different paths for blend lowering, but they all must use the
8648 // *exact* same predicate.
8649 bool IsBlendSupported = Subtarget->hasSSE41();
8650 if (IsBlendSupported)
8651 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i32, V1, V2, Mask,
8655 if (SDValue Masked =
8656 lowerVectorShuffleAsBitMask(DL, MVT::v4i32, V1, V2, Mask, DAG))
8659 // Use dedicated unpack instructions for masks that match their pattern.
8661 lowerVectorShuffleWithUNPCK(DL, MVT::v4i32, Mask, V1, V2, DAG))
8664 // Try to use byte rotation instructions.
8665 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8666 if (Subtarget->hasSSSE3())
8667 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8668 DL, MVT::v4i32, V1, V2, Mask, Subtarget, DAG))
8671 // If we have direct support for blends, we should lower by decomposing into
8672 // a permute. That will be faster than the domain cross.
8673 if (IsBlendSupported)
8674 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i32, V1, V2,
8677 // Try to lower by permuting the inputs into an unpack instruction.
8678 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(DL, MVT::v4i32, V1,
8682 // We implement this with SHUFPS because it can blend from two vectors.
8683 // Because we're going to eventually use SHUFPS, we use SHUFPS even to build
8684 // up the inputs, bypassing domain shift penalties that we would encur if we
8685 // directly used PSHUFD on Nehalem and older. For newer chips, this isn't
8687 return DAG.getBitcast(
8689 DAG.getVectorShuffle(MVT::v4f32, DL, DAG.getBitcast(MVT::v4f32, V1),
8690 DAG.getBitcast(MVT::v4f32, V2), Mask));
8693 /// \brief Lowering of single-input v8i16 shuffles is the cornerstone of SSE2
8694 /// shuffle lowering, and the most complex part.
8696 /// The lowering strategy is to try to form pairs of input lanes which are
8697 /// targeted at the same half of the final vector, and then use a dword shuffle
8698 /// to place them onto the right half, and finally unpack the paired lanes into
8699 /// their final position.
8701 /// The exact breakdown of how to form these dword pairs and align them on the
8702 /// correct sides is really tricky. See the comments within the function for
8703 /// more of the details.
8705 /// This code also handles repeated 128-bit lanes of v8i16 shuffles, but each
8706 /// lane must shuffle the *exact* same way. In fact, you must pass a v8 Mask to
8707 /// this routine for it to work correctly. To shuffle a 256-bit or 512-bit i16
8708 /// vector, form the analogous 128-bit 8-element Mask.
8709 static SDValue lowerV8I16GeneralSingleInputVectorShuffle(
8710 SDLoc DL, MVT VT, SDValue V, MutableArrayRef<int> Mask,
8711 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
8712 assert(VT.getVectorElementType() == MVT::i16 && "Bad input type!");
8713 MVT PSHUFDVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
8715 assert(Mask.size() == 8 && "Shuffle mask length doen't match!");
8716 MutableArrayRef<int> LoMask = Mask.slice(0, 4);
8717 MutableArrayRef<int> HiMask = Mask.slice(4, 4);
8719 SmallVector<int, 4> LoInputs;
8720 std::copy_if(LoMask.begin(), LoMask.end(), std::back_inserter(LoInputs),
8721 [](int M) { return M >= 0; });
8722 std::sort(LoInputs.begin(), LoInputs.end());
8723 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), LoInputs.end());
8724 SmallVector<int, 4> HiInputs;
8725 std::copy_if(HiMask.begin(), HiMask.end(), std::back_inserter(HiInputs),
8726 [](int M) { return M >= 0; });
8727 std::sort(HiInputs.begin(), HiInputs.end());
8728 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), HiInputs.end());
8730 std::lower_bound(LoInputs.begin(), LoInputs.end(), 4) - LoInputs.begin();
8731 int NumHToL = LoInputs.size() - NumLToL;
8733 std::lower_bound(HiInputs.begin(), HiInputs.end(), 4) - HiInputs.begin();
8734 int NumHToH = HiInputs.size() - NumLToH;
8735 MutableArrayRef<int> LToLInputs(LoInputs.data(), NumLToL);
8736 MutableArrayRef<int> LToHInputs(HiInputs.data(), NumLToH);
8737 MutableArrayRef<int> HToLInputs(LoInputs.data() + NumLToL, NumHToL);
8738 MutableArrayRef<int> HToHInputs(HiInputs.data() + NumLToH, NumHToH);
8740 // Simplify the 1-into-3 and 3-into-1 cases with a single pshufd. For all
8741 // such inputs we can swap two of the dwords across the half mark and end up
8742 // with <=2 inputs to each half in each half. Once there, we can fall through
8743 // to the generic code below. For example:
8745 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8746 // Mask: [0, 1, 2, 7, 4, 5, 6, 3] -----------------> [0, 1, 4, 7, 2, 3, 6, 5]
8748 // However in some very rare cases we have a 1-into-3 or 3-into-1 on one half
8749 // and an existing 2-into-2 on the other half. In this case we may have to
8750 // pre-shuffle the 2-into-2 half to avoid turning it into a 3-into-1 or
8751 // 1-into-3 which could cause us to cycle endlessly fixing each side in turn.
8752 // Fortunately, we don't have to handle anything but a 2-into-2 pattern
8753 // because any other situation (including a 3-into-1 or 1-into-3 in the other
8754 // half than the one we target for fixing) will be fixed when we re-enter this
8755 // path. We will also combine away any sequence of PSHUFD instructions that
8756 // result into a single instruction. Here is an example of the tricky case:
8758 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8759 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -THIS-IS-BAD!!!!-> [5, 7, 1, 0, 4, 7, 5, 3]
8761 // This now has a 1-into-3 in the high half! Instead, we do two shuffles:
8763 // Input: [a, b, c, d, e, f, g, h] PSHUFHW[0,2,1,3]-> [a, b, c, d, e, g, f, h]
8764 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -----------------> [3, 7, 1, 0, 2, 7, 3, 6]
8766 // Input: [a, b, c, d, e, g, f, h] -PSHUFD[0,2,1,3]-> [a, b, e, g, c, d, f, h]
8767 // Mask: [3, 7, 1, 0, 2, 7, 3, 6] -----------------> [5, 7, 1, 0, 4, 7, 5, 6]
8769 // The result is fine to be handled by the generic logic.
8770 auto balanceSides = [&](ArrayRef<int> AToAInputs, ArrayRef<int> BToAInputs,
8771 ArrayRef<int> BToBInputs, ArrayRef<int> AToBInputs,
8772 int AOffset, int BOffset) {
8773 assert((AToAInputs.size() == 3 || AToAInputs.size() == 1) &&
8774 "Must call this with A having 3 or 1 inputs from the A half.");
8775 assert((BToAInputs.size() == 1 || BToAInputs.size() == 3) &&
8776 "Must call this with B having 1 or 3 inputs from the B half.");
8777 assert(AToAInputs.size() + BToAInputs.size() == 4 &&
8778 "Must call this with either 3:1 or 1:3 inputs (summing to 4).");
8780 bool ThreeAInputs = AToAInputs.size() == 3;
8782 // Compute the index of dword with only one word among the three inputs in
8783 // a half by taking the sum of the half with three inputs and subtracting
8784 // the sum of the actual three inputs. The difference is the remaining
8787 int &TripleDWord = ThreeAInputs ? ADWord : BDWord;
8788 int &OneInputDWord = ThreeAInputs ? BDWord : ADWord;
8789 int TripleInputOffset = ThreeAInputs ? AOffset : BOffset;
8790 ArrayRef<int> TripleInputs = ThreeAInputs ? AToAInputs : BToAInputs;
8791 int OneInput = ThreeAInputs ? BToAInputs[0] : AToAInputs[0];
8792 int TripleInputSum = 0 + 1 + 2 + 3 + (4 * TripleInputOffset);
8793 int TripleNonInputIdx =
8794 TripleInputSum - std::accumulate(TripleInputs.begin(), TripleInputs.end(), 0);
8795 TripleDWord = TripleNonInputIdx / 2;
8797 // We use xor with one to compute the adjacent DWord to whichever one the
8799 OneInputDWord = (OneInput / 2) ^ 1;
8801 // Check for one tricky case: We're fixing a 3<-1 or a 1<-3 shuffle for AToA
8802 // and BToA inputs. If there is also such a problem with the BToB and AToB
8803 // inputs, we don't try to fix it necessarily -- we'll recurse and see it in
8804 // the next pass. However, if we have a 2<-2 in the BToB and AToB inputs, it
8805 // is essential that we don't *create* a 3<-1 as then we might oscillate.
8806 if (BToBInputs.size() == 2 && AToBInputs.size() == 2) {
8807 // Compute how many inputs will be flipped by swapping these DWords. We
8809 // to balance this to ensure we don't form a 3-1 shuffle in the other
8811 int NumFlippedAToBInputs =
8812 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord) +
8813 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord + 1);
8814 int NumFlippedBToBInputs =
8815 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord) +
8816 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord + 1);
8817 if ((NumFlippedAToBInputs == 1 &&
8818 (NumFlippedBToBInputs == 0 || NumFlippedBToBInputs == 2)) ||
8819 (NumFlippedBToBInputs == 1 &&
8820 (NumFlippedAToBInputs == 0 || NumFlippedAToBInputs == 2))) {
8821 // We choose whether to fix the A half or B half based on whether that
8822 // half has zero flipped inputs. At zero, we may not be able to fix it
8823 // with that half. We also bias towards fixing the B half because that
8824 // will more commonly be the high half, and we have to bias one way.
8825 auto FixFlippedInputs = [&V, &DL, &Mask, &DAG](int PinnedIdx, int DWord,
8826 ArrayRef<int> Inputs) {
8827 int FixIdx = PinnedIdx ^ 1; // The adjacent slot to the pinned slot.
8828 bool IsFixIdxInput = std::find(Inputs.begin(), Inputs.end(),
8829 PinnedIdx ^ 1) != Inputs.end();
8830 // Determine whether the free index is in the flipped dword or the
8831 // unflipped dword based on where the pinned index is. We use this bit
8832 // in an xor to conditionally select the adjacent dword.
8833 int FixFreeIdx = 2 * (DWord ^ (PinnedIdx / 2 == DWord));
8834 bool IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8835 FixFreeIdx) != Inputs.end();
8836 if (IsFixIdxInput == IsFixFreeIdxInput)
8838 IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8839 FixFreeIdx) != Inputs.end();
8840 assert(IsFixIdxInput != IsFixFreeIdxInput &&
8841 "We need to be changing the number of flipped inputs!");
8842 int PSHUFHalfMask[] = {0, 1, 2, 3};
8843 std::swap(PSHUFHalfMask[FixFreeIdx % 4], PSHUFHalfMask[FixIdx % 4]);
8844 V = DAG.getNode(FixIdx < 4 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW, DL,
8846 getV4X86ShuffleImm8ForMask(PSHUFHalfMask, DL, DAG));
8849 if (M != -1 && M == FixIdx)
8851 else if (M != -1 && M == FixFreeIdx)
8854 if (NumFlippedBToBInputs != 0) {
8856 BToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
8857 FixFlippedInputs(BPinnedIdx, BDWord, BToBInputs);
8859 assert(NumFlippedAToBInputs != 0 && "Impossible given predicates!");
8860 int APinnedIdx = ThreeAInputs ? TripleNonInputIdx : OneInput;
8861 FixFlippedInputs(APinnedIdx, ADWord, AToBInputs);
8866 int PSHUFDMask[] = {0, 1, 2, 3};
8867 PSHUFDMask[ADWord] = BDWord;
8868 PSHUFDMask[BDWord] = ADWord;
8871 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
8872 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
8874 // Adjust the mask to match the new locations of A and B.
8876 if (M != -1 && M/2 == ADWord)
8877 M = 2 * BDWord + M % 2;
8878 else if (M != -1 && M/2 == BDWord)
8879 M = 2 * ADWord + M % 2;
8881 // Recurse back into this routine to re-compute state now that this isn't
8882 // a 3 and 1 problem.
8883 return lowerV8I16GeneralSingleInputVectorShuffle(DL, VT, V, Mask, Subtarget,
8886 if ((NumLToL == 3 && NumHToL == 1) || (NumLToL == 1 && NumHToL == 3))
8887 return balanceSides(LToLInputs, HToLInputs, HToHInputs, LToHInputs, 0, 4);
8888 else if ((NumHToH == 3 && NumLToH == 1) || (NumHToH == 1 && NumLToH == 3))
8889 return balanceSides(HToHInputs, LToHInputs, LToLInputs, HToLInputs, 4, 0);
8891 // At this point there are at most two inputs to the low and high halves from
8892 // each half. That means the inputs can always be grouped into dwords and
8893 // those dwords can then be moved to the correct half with a dword shuffle.
8894 // We use at most one low and one high word shuffle to collect these paired
8895 // inputs into dwords, and finally a dword shuffle to place them.
8896 int PSHUFLMask[4] = {-1, -1, -1, -1};
8897 int PSHUFHMask[4] = {-1, -1, -1, -1};
8898 int PSHUFDMask[4] = {-1, -1, -1, -1};
8900 // First fix the masks for all the inputs that are staying in their
8901 // original halves. This will then dictate the targets of the cross-half
8903 auto fixInPlaceInputs =
8904 [&PSHUFDMask](ArrayRef<int> InPlaceInputs, ArrayRef<int> IncomingInputs,
8905 MutableArrayRef<int> SourceHalfMask,
8906 MutableArrayRef<int> HalfMask, int HalfOffset) {
8907 if (InPlaceInputs.empty())
8909 if (InPlaceInputs.size() == 1) {
8910 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8911 InPlaceInputs[0] - HalfOffset;
8912 PSHUFDMask[InPlaceInputs[0] / 2] = InPlaceInputs[0] / 2;
8915 if (IncomingInputs.empty()) {
8916 // Just fix all of the in place inputs.
8917 for (int Input : InPlaceInputs) {
8918 SourceHalfMask[Input - HalfOffset] = Input - HalfOffset;
8919 PSHUFDMask[Input / 2] = Input / 2;
8924 assert(InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!");
8925 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8926 InPlaceInputs[0] - HalfOffset;
8927 // Put the second input next to the first so that they are packed into
8928 // a dword. We find the adjacent index by toggling the low bit.
8929 int AdjIndex = InPlaceInputs[0] ^ 1;
8930 SourceHalfMask[AdjIndex - HalfOffset] = InPlaceInputs[1] - HalfOffset;
8931 std::replace(HalfMask.begin(), HalfMask.end(), InPlaceInputs[1], AdjIndex);
8932 PSHUFDMask[AdjIndex / 2] = AdjIndex / 2;
8934 fixInPlaceInputs(LToLInputs, HToLInputs, PSHUFLMask, LoMask, 0);
8935 fixInPlaceInputs(HToHInputs, LToHInputs, PSHUFHMask, HiMask, 4);
8937 // Now gather the cross-half inputs and place them into a free dword of
8938 // their target half.
8939 // FIXME: This operation could almost certainly be simplified dramatically to
8940 // look more like the 3-1 fixing operation.
8941 auto moveInputsToRightHalf = [&PSHUFDMask](
8942 MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs,
8943 MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask,
8944 MutableArrayRef<int> FinalSourceHalfMask, int SourceOffset,
8946 auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) {
8947 return SourceHalfMask[Word] != -1 && SourceHalfMask[Word] != Word;
8949 auto isDWordClobbered = [&isWordClobbered](ArrayRef<int> SourceHalfMask,
8951 int LowWord = Word & ~1;
8952 int HighWord = Word | 1;
8953 return isWordClobbered(SourceHalfMask, LowWord) ||
8954 isWordClobbered(SourceHalfMask, HighWord);
8957 if (IncomingInputs.empty())
8960 if (ExistingInputs.empty()) {
8961 // Map any dwords with inputs from them into the right half.
8962 for (int Input : IncomingInputs) {
8963 // If the source half mask maps over the inputs, turn those into
8964 // swaps and use the swapped lane.
8965 if (isWordClobbered(SourceHalfMask, Input - SourceOffset)) {
8966 if (SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == -1) {
8967 SourceHalfMask[SourceHalfMask[Input - SourceOffset]] =
8968 Input - SourceOffset;
8969 // We have to swap the uses in our half mask in one sweep.
8970 for (int &M : HalfMask)
8971 if (M == SourceHalfMask[Input - SourceOffset] + SourceOffset)
8973 else if (M == Input)
8974 M = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8976 assert(SourceHalfMask[SourceHalfMask[Input - SourceOffset]] ==
8977 Input - SourceOffset &&
8978 "Previous placement doesn't match!");
8980 // Note that this correctly re-maps both when we do a swap and when
8981 // we observe the other side of the swap above. We rely on that to
8982 // avoid swapping the members of the input list directly.
8983 Input = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8986 // Map the input's dword into the correct half.
8987 if (PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == -1)
8988 PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] = Input / 2;
8990 assert(PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] ==
8992 "Previous placement doesn't match!");
8995 // And just directly shift any other-half mask elements to be same-half
8996 // as we will have mirrored the dword containing the element into the
8997 // same position within that half.
8998 for (int &M : HalfMask)
8999 if (M >= SourceOffset && M < SourceOffset + 4) {
9000 M = M - SourceOffset + DestOffset;
9001 assert(M >= 0 && "This should never wrap below zero!");
9006 // Ensure we have the input in a viable dword of its current half. This
9007 // is particularly tricky because the original position may be clobbered
9008 // by inputs being moved and *staying* in that half.
9009 if (IncomingInputs.size() == 1) {
9010 if (isWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
9011 int InputFixed = std::find(std::begin(SourceHalfMask),
9012 std::end(SourceHalfMask), -1) -
9013 std::begin(SourceHalfMask) + SourceOffset;
9014 SourceHalfMask[InputFixed - SourceOffset] =
9015 IncomingInputs[0] - SourceOffset;
9016 std::replace(HalfMask.begin(), HalfMask.end(), IncomingInputs[0],
9018 IncomingInputs[0] = InputFixed;
9020 } else if (IncomingInputs.size() == 2) {
9021 if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 ||
9022 isDWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
9023 // We have two non-adjacent or clobbered inputs we need to extract from
9024 // the source half. To do this, we need to map them into some adjacent
9025 // dword slot in the source mask.
9026 int InputsFixed[2] = {IncomingInputs[0] - SourceOffset,
9027 IncomingInputs[1] - SourceOffset};
9029 // If there is a free slot in the source half mask adjacent to one of
9030 // the inputs, place the other input in it. We use (Index XOR 1) to
9031 // compute an adjacent index.
9032 if (!isWordClobbered(SourceHalfMask, InputsFixed[0]) &&
9033 SourceHalfMask[InputsFixed[0] ^ 1] == -1) {
9034 SourceHalfMask[InputsFixed[0]] = InputsFixed[0];
9035 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
9036 InputsFixed[1] = InputsFixed[0] ^ 1;
9037 } else if (!isWordClobbered(SourceHalfMask, InputsFixed[1]) &&
9038 SourceHalfMask[InputsFixed[1] ^ 1] == -1) {
9039 SourceHalfMask[InputsFixed[1]] = InputsFixed[1];
9040 SourceHalfMask[InputsFixed[1] ^ 1] = InputsFixed[0];
9041 InputsFixed[0] = InputsFixed[1] ^ 1;
9042 } else if (SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] == -1 &&
9043 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] == -1) {
9044 // The two inputs are in the same DWord but it is clobbered and the
9045 // adjacent DWord isn't used at all. Move both inputs to the free
9047 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] = InputsFixed[0];
9048 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] = InputsFixed[1];
9049 InputsFixed[0] = 2 * ((InputsFixed[0] / 2) ^ 1);
9050 InputsFixed[1] = 2 * ((InputsFixed[0] / 2) ^ 1) + 1;
9052 // The only way we hit this point is if there is no clobbering
9053 // (because there are no off-half inputs to this half) and there is no
9054 // free slot adjacent to one of the inputs. In this case, we have to
9055 // swap an input with a non-input.
9056 for (int i = 0; i < 4; ++i)
9057 assert((SourceHalfMask[i] == -1 || SourceHalfMask[i] == i) &&
9058 "We can't handle any clobbers here!");
9059 assert(InputsFixed[1] != (InputsFixed[0] ^ 1) &&
9060 "Cannot have adjacent inputs here!");
9062 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
9063 SourceHalfMask[InputsFixed[1]] = InputsFixed[0] ^ 1;
9065 // We also have to update the final source mask in this case because
9066 // it may need to undo the above swap.
9067 for (int &M : FinalSourceHalfMask)
9068 if (M == (InputsFixed[0] ^ 1) + SourceOffset)
9069 M = InputsFixed[1] + SourceOffset;
9070 else if (M == InputsFixed[1] + SourceOffset)
9071 M = (InputsFixed[0] ^ 1) + SourceOffset;
9073 InputsFixed[1] = InputsFixed[0] ^ 1;
9076 // Point everything at the fixed inputs.
9077 for (int &M : HalfMask)
9078 if (M == IncomingInputs[0])
9079 M = InputsFixed[0] + SourceOffset;
9080 else if (M == IncomingInputs[1])
9081 M = InputsFixed[1] + SourceOffset;
9083 IncomingInputs[0] = InputsFixed[0] + SourceOffset;
9084 IncomingInputs[1] = InputsFixed[1] + SourceOffset;
9087 llvm_unreachable("Unhandled input size!");
9090 // Now hoist the DWord down to the right half.
9091 int FreeDWord = (PSHUFDMask[DestOffset / 2] == -1 ? 0 : 1) + DestOffset / 2;
9092 assert(PSHUFDMask[FreeDWord] == -1 && "DWord not free");
9093 PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2;
9094 for (int &M : HalfMask)
9095 for (int Input : IncomingInputs)
9097 M = FreeDWord * 2 + Input % 2;
9099 moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask, HiMask,
9100 /*SourceOffset*/ 4, /*DestOffset*/ 0);
9101 moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask, LoMask,
9102 /*SourceOffset*/ 0, /*DestOffset*/ 4);
9104 // Now enact all the shuffles we've computed to move the inputs into their
9106 if (!isNoopShuffleMask(PSHUFLMask))
9107 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
9108 getV4X86ShuffleImm8ForMask(PSHUFLMask, DL, DAG));
9109 if (!isNoopShuffleMask(PSHUFHMask))
9110 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
9111 getV4X86ShuffleImm8ForMask(PSHUFHMask, DL, DAG));
9112 if (!isNoopShuffleMask(PSHUFDMask))
9115 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
9116 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
9118 // At this point, each half should contain all its inputs, and we can then
9119 // just shuffle them into their final position.
9120 assert(std::count_if(LoMask.begin(), LoMask.end(),
9121 [](int M) { return M >= 4; }) == 0 &&
9122 "Failed to lift all the high half inputs to the low mask!");
9123 assert(std::count_if(HiMask.begin(), HiMask.end(),
9124 [](int M) { return M >= 0 && M < 4; }) == 0 &&
9125 "Failed to lift all the low half inputs to the high mask!");
9127 // Do a half shuffle for the low mask.
9128 if (!isNoopShuffleMask(LoMask))
9129 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
9130 getV4X86ShuffleImm8ForMask(LoMask, DL, DAG));
9132 // Do a half shuffle with the high mask after shifting its values down.
9133 for (int &M : HiMask)
9136 if (!isNoopShuffleMask(HiMask))
9137 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
9138 getV4X86ShuffleImm8ForMask(HiMask, DL, DAG));
9143 /// \brief Helper to form a PSHUFB-based shuffle+blend.
9144 static SDValue lowerVectorShuffleAsPSHUFB(SDLoc DL, MVT VT, SDValue V1,
9145 SDValue V2, ArrayRef<int> Mask,
9146 SelectionDAG &DAG, bool &V1InUse,
9148 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
9154 int Size = Mask.size();
9155 int Scale = 16 / Size;
9156 for (int i = 0; i < 16; ++i) {
9157 if (Mask[i / Scale] == -1) {
9158 V1Mask[i] = V2Mask[i] = DAG.getUNDEF(MVT::i8);
9160 const int ZeroMask = 0x80;
9161 int V1Idx = Mask[i / Scale] < Size ? Mask[i / Scale] * Scale + i % Scale
9163 int V2Idx = Mask[i / Scale] < Size
9165 : (Mask[i / Scale] - Size) * Scale + i % Scale;
9166 if (Zeroable[i / Scale])
9167 V1Idx = V2Idx = ZeroMask;
9168 V1Mask[i] = DAG.getConstant(V1Idx, DL, MVT::i8);
9169 V2Mask[i] = DAG.getConstant(V2Idx, DL, MVT::i8);
9170 V1InUse |= (ZeroMask != V1Idx);
9171 V2InUse |= (ZeroMask != V2Idx);
9176 V1 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
9177 DAG.getBitcast(MVT::v16i8, V1),
9178 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V1Mask));
9180 V2 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
9181 DAG.getBitcast(MVT::v16i8, V2),
9182 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V2Mask));
9184 // If we need shuffled inputs from both, blend the two.
9186 if (V1InUse && V2InUse)
9187 V = DAG.getNode(ISD::OR, DL, MVT::v16i8, V1, V2);
9189 V = V1InUse ? V1 : V2;
9191 // Cast the result back to the correct type.
9192 return DAG.getBitcast(VT, V);
9195 /// \brief Generic lowering of 8-lane i16 shuffles.
9197 /// This handles both single-input shuffles and combined shuffle/blends with
9198 /// two inputs. The single input shuffles are immediately delegated to
9199 /// a dedicated lowering routine.
9201 /// The blends are lowered in one of three fundamental ways. If there are few
9202 /// enough inputs, it delegates to a basic UNPCK-based strategy. If the shuffle
9203 /// of the input is significantly cheaper when lowered as an interleaving of
9204 /// the two inputs, try to interleave them. Otherwise, blend the low and high
9205 /// halves of the inputs separately (making them have relatively few inputs)
9206 /// and then concatenate them.
9207 static SDValue lowerV8I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9208 const X86Subtarget *Subtarget,
9209 SelectionDAG &DAG) {
9211 assert(Op.getSimpleValueType() == MVT::v8i16 && "Bad shuffle type!");
9212 assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
9213 assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
9214 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9215 ArrayRef<int> OrigMask = SVOp->getMask();
9216 int MaskStorage[8] = {OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
9217 OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7]};
9218 MutableArrayRef<int> Mask(MaskStorage);
9220 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9222 // Whenever we can lower this as a zext, that instruction is strictly faster
9223 // than any alternative.
9224 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
9225 DL, MVT::v8i16, V1, V2, OrigMask, Subtarget, DAG))
9228 auto isV1 = [](int M) { return M >= 0 && M < 8; };
9230 auto isV2 = [](int M) { return M >= 8; };
9232 int NumV2Inputs = std::count_if(Mask.begin(), Mask.end(), isV2);
9234 if (NumV2Inputs == 0) {
9235 // Check for being able to broadcast a single element.
9236 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i16, V1,
9237 Mask, Subtarget, DAG))
9240 // Try to use shift instructions.
9242 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V1, Mask, DAG))
9245 // Use dedicated unpack instructions for masks that match their pattern.
9247 lowerVectorShuffleWithUNPCK(DL, MVT::v8i16, Mask, V1, V2, DAG))
9250 // Try to use byte rotation instructions.
9251 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(DL, MVT::v8i16, V1, V1,
9252 Mask, Subtarget, DAG))
9255 return lowerV8I16GeneralSingleInputVectorShuffle(DL, MVT::v8i16, V1, Mask,
9259 assert(std::any_of(Mask.begin(), Mask.end(), isV1) &&
9260 "All single-input shuffles should be canonicalized to be V1-input "
9263 // Try to use shift instructions.
9265 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V2, Mask, DAG))
9268 // See if we can use SSE4A Extraction / Insertion.
9269 if (Subtarget->hasSSE4A())
9270 if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v8i16, V1, V2, Mask, DAG))
9273 // There are special ways we can lower some single-element blends.
9274 if (NumV2Inputs == 1)
9275 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v8i16, V1, V2,
9276 Mask, Subtarget, DAG))
9279 // We have different paths for blend lowering, but they all must use the
9280 // *exact* same predicate.
9281 bool IsBlendSupported = Subtarget->hasSSE41();
9282 if (IsBlendSupported)
9283 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i16, V1, V2, Mask,
9287 if (SDValue Masked =
9288 lowerVectorShuffleAsBitMask(DL, MVT::v8i16, V1, V2, Mask, DAG))
9291 // Use dedicated unpack instructions for masks that match their pattern.
9293 lowerVectorShuffleWithUNPCK(DL, MVT::v8i16, Mask, V1, V2, DAG))
9296 // Try to use byte rotation instructions.
9297 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9298 DL, MVT::v8i16, V1, V2, Mask, Subtarget, DAG))
9301 if (SDValue BitBlend =
9302 lowerVectorShuffleAsBitBlend(DL, MVT::v8i16, V1, V2, Mask, DAG))
9305 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(DL, MVT::v8i16, V1,
9309 // If we can't directly blend but can use PSHUFB, that will be better as it
9310 // can both shuffle and set up the inefficient blend.
9311 if (!IsBlendSupported && Subtarget->hasSSSE3()) {
9312 bool V1InUse, V2InUse;
9313 return lowerVectorShuffleAsPSHUFB(DL, MVT::v8i16, V1, V2, Mask, DAG,
9317 // We can always bit-blend if we have to so the fallback strategy is to
9318 // decompose into single-input permutes and blends.
9319 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i16, V1, V2,
9323 /// \brief Check whether a compaction lowering can be done by dropping even
9324 /// elements and compute how many times even elements must be dropped.
9326 /// This handles shuffles which take every Nth element where N is a power of
9327 /// two. Example shuffle masks:
9329 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 0, 2, 4, 6, 8, 10, 12, 14
9330 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30
9331 /// N = 2: 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12
9332 /// N = 2: 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28
9333 /// N = 3: 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8
9334 /// N = 3: 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24
9336 /// Any of these lanes can of course be undef.
9338 /// This routine only supports N <= 3.
9339 /// FIXME: Evaluate whether either AVX or AVX-512 have any opportunities here
9342 /// \returns N above, or the number of times even elements must be dropped if
9343 /// there is such a number. Otherwise returns zero.
9344 static int canLowerByDroppingEvenElements(ArrayRef<int> Mask) {
9345 // Figure out whether we're looping over two inputs or just one.
9346 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9348 // The modulus for the shuffle vector entries is based on whether this is
9349 // a single input or not.
9350 int ShuffleModulus = Mask.size() * (IsSingleInput ? 1 : 2);
9351 assert(isPowerOf2_32((uint32_t)ShuffleModulus) &&
9352 "We should only be called with masks with a power-of-2 size!");
9354 uint64_t ModMask = (uint64_t)ShuffleModulus - 1;
9356 // We track whether the input is viable for all power-of-2 strides 2^1, 2^2,
9357 // and 2^3 simultaneously. This is because we may have ambiguity with
9358 // partially undef inputs.
9359 bool ViableForN[3] = {true, true, true};
9361 for (int i = 0, e = Mask.size(); i < e; ++i) {
9362 // Ignore undef lanes, we'll optimistically collapse them to the pattern we
9367 bool IsAnyViable = false;
9368 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
9369 if (ViableForN[j]) {
9372 // The shuffle mask must be equal to (i * 2^N) % M.
9373 if ((uint64_t)Mask[i] == (((uint64_t)i << N) & ModMask))
9376 ViableForN[j] = false;
9378 // Early exit if we exhaust the possible powers of two.
9383 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
9387 // Return 0 as there is no viable power of two.
9391 /// \brief Generic lowering of v16i8 shuffles.
9393 /// This is a hybrid strategy to lower v16i8 vectors. It first attempts to
9394 /// detect any complexity reducing interleaving. If that doesn't help, it uses
9395 /// UNPCK to spread the i8 elements across two i16-element vectors, and uses
9396 /// the existing lowering for v8i16 blends on each half, finally PACK-ing them
9398 static SDValue lowerV16I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9399 const X86Subtarget *Subtarget,
9400 SelectionDAG &DAG) {
9402 assert(Op.getSimpleValueType() == MVT::v16i8 && "Bad shuffle type!");
9403 assert(V1.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
9404 assert(V2.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
9405 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9406 ArrayRef<int> Mask = SVOp->getMask();
9407 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
9409 // Try to use shift instructions.
9411 lowerVectorShuffleAsShift(DL, MVT::v16i8, V1, V2, Mask, DAG))
9414 // Try to use byte rotation instructions.
9415 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9416 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
9419 // Try to use a zext lowering.
9420 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
9421 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
9424 // See if we can use SSE4A Extraction / Insertion.
9425 if (Subtarget->hasSSE4A())
9426 if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v16i8, V1, V2, Mask, DAG))
9430 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 16; });
9432 // For single-input shuffles, there are some nicer lowering tricks we can use.
9433 if (NumV2Elements == 0) {
9434 // Check for being able to broadcast a single element.
9435 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i8, V1,
9436 Mask, Subtarget, DAG))
9439 // Check whether we can widen this to an i16 shuffle by duplicating bytes.
9440 // Notably, this handles splat and partial-splat shuffles more efficiently.
9441 // However, it only makes sense if the pre-duplication shuffle simplifies
9442 // things significantly. Currently, this means we need to be able to
9443 // express the pre-duplication shuffle as an i16 shuffle.
9445 // FIXME: We should check for other patterns which can be widened into an
9446 // i16 shuffle as well.
9447 auto canWidenViaDuplication = [](ArrayRef<int> Mask) {
9448 for (int i = 0; i < 16; i += 2)
9449 if (Mask[i] != -1 && Mask[i + 1] != -1 && Mask[i] != Mask[i + 1])
9454 auto tryToWidenViaDuplication = [&]() -> SDValue {
9455 if (!canWidenViaDuplication(Mask))
9457 SmallVector<int, 4> LoInputs;
9458 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(LoInputs),
9459 [](int M) { return M >= 0 && M < 8; });
9460 std::sort(LoInputs.begin(), LoInputs.end());
9461 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()),
9463 SmallVector<int, 4> HiInputs;
9464 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(HiInputs),
9465 [](int M) { return M >= 8; });
9466 std::sort(HiInputs.begin(), HiInputs.end());
9467 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()),
9470 bool TargetLo = LoInputs.size() >= HiInputs.size();
9471 ArrayRef<int> InPlaceInputs = TargetLo ? LoInputs : HiInputs;
9472 ArrayRef<int> MovingInputs = TargetLo ? HiInputs : LoInputs;
9474 int PreDupI16Shuffle[] = {-1, -1, -1, -1, -1, -1, -1, -1};
9475 SmallDenseMap<int, int, 8> LaneMap;
9476 for (int I : InPlaceInputs) {
9477 PreDupI16Shuffle[I/2] = I/2;
9480 int j = TargetLo ? 0 : 4, je = j + 4;
9481 for (int i = 0, ie = MovingInputs.size(); i < ie; ++i) {
9482 // Check if j is already a shuffle of this input. This happens when
9483 // there are two adjacent bytes after we move the low one.
9484 if (PreDupI16Shuffle[j] != MovingInputs[i] / 2) {
9485 // If we haven't yet mapped the input, search for a slot into which
9487 while (j < je && PreDupI16Shuffle[j] != -1)
9491 // We can't place the inputs into a single half with a simple i16 shuffle, so bail.
9494 // Map this input with the i16 shuffle.
9495 PreDupI16Shuffle[j] = MovingInputs[i] / 2;
9498 // Update the lane map based on the mapping we ended up with.
9499 LaneMap[MovingInputs[i]] = 2 * j + MovingInputs[i] % 2;
9501 V1 = DAG.getBitcast(
9503 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
9504 DAG.getUNDEF(MVT::v8i16), PreDupI16Shuffle));
9506 // Unpack the bytes to form the i16s that will be shuffled into place.
9507 V1 = DAG.getNode(TargetLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
9508 MVT::v16i8, V1, V1);
9510 int PostDupI16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9511 for (int i = 0; i < 16; ++i)
9512 if (Mask[i] != -1) {
9513 int MappedMask = LaneMap[Mask[i]] - (TargetLo ? 0 : 8);
9514 assert(MappedMask < 8 && "Invalid v8 shuffle mask!");
9515 if (PostDupI16Shuffle[i / 2] == -1)
9516 PostDupI16Shuffle[i / 2] = MappedMask;
9518 assert(PostDupI16Shuffle[i / 2] == MappedMask &&
9519 "Conflicting entrties in the original shuffle!");
9521 return DAG.getBitcast(
9523 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
9524 DAG.getUNDEF(MVT::v8i16), PostDupI16Shuffle));
9526 if (SDValue V = tryToWidenViaDuplication())
9530 if (SDValue Masked =
9531 lowerVectorShuffleAsBitMask(DL, MVT::v16i8, V1, V2, Mask, DAG))
9534 // Use dedicated unpack instructions for masks that match their pattern.
9536 lowerVectorShuffleWithUNPCK(DL, MVT::v16i8, Mask, V1, V2, DAG))
9539 // Check for SSSE3 which lets us lower all v16i8 shuffles much more directly
9540 // with PSHUFB. It is important to do this before we attempt to generate any
9541 // blends but after all of the single-input lowerings. If the single input
9542 // lowerings can find an instruction sequence that is faster than a PSHUFB, we
9543 // want to preserve that and we can DAG combine any longer sequences into
9544 // a PSHUFB in the end. But once we start blending from multiple inputs,
9545 // the complexity of DAG combining bad patterns back into PSHUFB is too high,
9546 // and there are *very* few patterns that would actually be faster than the
9547 // PSHUFB approach because of its ability to zero lanes.
9549 // FIXME: The only exceptions to the above are blends which are exact
9550 // interleavings with direct instructions supporting them. We currently don't
9551 // handle those well here.
9552 if (Subtarget->hasSSSE3()) {
9553 bool V1InUse = false;
9554 bool V2InUse = false;
9556 SDValue PSHUFB = lowerVectorShuffleAsPSHUFB(DL, MVT::v16i8, V1, V2, Mask,
9557 DAG, V1InUse, V2InUse);
9559 // If both V1 and V2 are in use and we can use a direct blend or an unpack,
9560 // do so. This avoids using them to handle blends-with-zero which is
9561 // important as a single pshufb is significantly faster for that.
9562 if (V1InUse && V2InUse) {
9563 if (Subtarget->hasSSE41())
9564 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i8, V1, V2,
9565 Mask, Subtarget, DAG))
9568 // We can use an unpack to do the blending rather than an or in some
9569 // cases. Even though the or may be (very minorly) more efficient, we
9570 // preference this lowering because there are common cases where part of
9571 // the complexity of the shuffles goes away when we do the final blend as
9573 // FIXME: It might be worth trying to detect if the unpack-feeding
9574 // shuffles will both be pshufb, in which case we shouldn't bother with
9576 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(
9577 DL, MVT::v16i8, V1, V2, Mask, DAG))
9584 // There are special ways we can lower some single-element blends.
9585 if (NumV2Elements == 1)
9586 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v16i8, V1, V2,
9587 Mask, Subtarget, DAG))
9590 if (SDValue BitBlend =
9591 lowerVectorShuffleAsBitBlend(DL, MVT::v16i8, V1, V2, Mask, DAG))
9594 // Check whether a compaction lowering can be done. This handles shuffles
9595 // which take every Nth element for some even N. See the helper function for
9598 // We special case these as they can be particularly efficiently handled with
9599 // the PACKUSB instruction on x86 and they show up in common patterns of
9600 // rearranging bytes to truncate wide elements.
9601 if (int NumEvenDrops = canLowerByDroppingEvenElements(Mask)) {
9602 // NumEvenDrops is the power of two stride of the elements. Another way of
9603 // thinking about it is that we need to drop the even elements this many
9604 // times to get the original input.
9605 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9607 // First we need to zero all the dropped bytes.
9608 assert(NumEvenDrops <= 3 &&
9609 "No support for dropping even elements more than 3 times.");
9610 // We use the mask type to pick which bytes are preserved based on how many
9611 // elements are dropped.
9612 MVT MaskVTs[] = { MVT::v8i16, MVT::v4i32, MVT::v2i64 };
9613 SDValue ByteClearMask = DAG.getBitcast(
9614 MVT::v16i8, DAG.getConstant(0xFF, DL, MaskVTs[NumEvenDrops - 1]));
9615 V1 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V1, ByteClearMask);
9617 V2 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V2, ByteClearMask);
9619 // Now pack things back together.
9620 V1 = DAG.getBitcast(MVT::v8i16, V1);
9621 V2 = IsSingleInput ? V1 : DAG.getBitcast(MVT::v8i16, V2);
9622 SDValue Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, V1, V2);
9623 for (int i = 1; i < NumEvenDrops; ++i) {
9624 Result = DAG.getBitcast(MVT::v8i16, Result);
9625 Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, Result, Result);
9631 // Handle multi-input cases by blending single-input shuffles.
9632 if (NumV2Elements > 0)
9633 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v16i8, V1, V2,
9636 // The fallback path for single-input shuffles widens this into two v8i16
9637 // vectors with unpacks, shuffles those, and then pulls them back together
9641 int LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9642 int HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9643 for (int i = 0; i < 16; ++i)
9645 (i < 8 ? LoBlendMask[i] : HiBlendMask[i % 8]) = Mask[i];
9647 SDValue Zero = getZeroVector(MVT::v8i16, Subtarget, DAG, DL);
9649 SDValue VLoHalf, VHiHalf;
9650 // Check if any of the odd lanes in the v16i8 are used. If not, we can mask
9651 // them out and avoid using UNPCK{L,H} to extract the elements of V as
9653 if (std::none_of(std::begin(LoBlendMask), std::end(LoBlendMask),
9654 [](int M) { return M >= 0 && M % 2 == 1; }) &&
9655 std::none_of(std::begin(HiBlendMask), std::end(HiBlendMask),
9656 [](int M) { return M >= 0 && M % 2 == 1; })) {
9657 // Use a mask to drop the high bytes.
9658 VLoHalf = DAG.getBitcast(MVT::v8i16, V);
9659 VLoHalf = DAG.getNode(ISD::AND, DL, MVT::v8i16, VLoHalf,
9660 DAG.getConstant(0x00FF, DL, MVT::v8i16));
9662 // This will be a single vector shuffle instead of a blend so nuke VHiHalf.
9663 VHiHalf = DAG.getUNDEF(MVT::v8i16);
9665 // Squash the masks to point directly into VLoHalf.
9666 for (int &M : LoBlendMask)
9669 for (int &M : HiBlendMask)
9673 // Otherwise just unpack the low half of V into VLoHalf and the high half into
9674 // VHiHalf so that we can blend them as i16s.
9675 VLoHalf = DAG.getBitcast(
9676 MVT::v8i16, DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V, Zero));
9677 VHiHalf = DAG.getBitcast(
9678 MVT::v8i16, DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V, Zero));
9681 SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, LoBlendMask);
9682 SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, HiBlendMask);
9684 return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, LoV, HiV);
9687 /// \brief Dispatching routine to lower various 128-bit x86 vector shuffles.
9689 /// This routine breaks down the specific type of 128-bit shuffle and
9690 /// dispatches to the lowering routines accordingly.
9691 static SDValue lower128BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9692 MVT VT, const X86Subtarget *Subtarget,
9693 SelectionDAG &DAG) {
9694 switch (VT.SimpleTy) {
9696 return lowerV2I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9698 return lowerV2F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9700 return lowerV4I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9702 return lowerV4F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9704 return lowerV8I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
9706 return lowerV16I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
9709 llvm_unreachable("Unimplemented!");
9713 /// \brief Helper function to test whether a shuffle mask could be
9714 /// simplified by widening the elements being shuffled.
9716 /// Appends the mask for wider elements in WidenedMask if valid. Otherwise
9717 /// leaves it in an unspecified state.
9719 /// NOTE: This must handle normal vector shuffle masks and *target* vector
9720 /// shuffle masks. The latter have the special property of a '-2' representing
9721 /// a zero-ed lane of a vector.
9722 static bool canWidenShuffleElements(ArrayRef<int> Mask,
9723 SmallVectorImpl<int> &WidenedMask) {
9724 for (int i = 0, Size = Mask.size(); i < Size; i += 2) {
9725 // If both elements are undef, its trivial.
9726 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] == SM_SentinelUndef) {
9727 WidenedMask.push_back(SM_SentinelUndef);
9731 // Check for an undef mask and a mask value properly aligned to fit with
9732 // a pair of values. If we find such a case, use the non-undef mask's value.
9733 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] >= 0 && Mask[i + 1] % 2 == 1) {
9734 WidenedMask.push_back(Mask[i + 1] / 2);
9737 if (Mask[i + 1] == SM_SentinelUndef && Mask[i] >= 0 && Mask[i] % 2 == 0) {
9738 WidenedMask.push_back(Mask[i] / 2);
9742 // When zeroing, we need to spread the zeroing across both lanes to widen.
9743 if (Mask[i] == SM_SentinelZero || Mask[i + 1] == SM_SentinelZero) {
9744 if ((Mask[i] == SM_SentinelZero || Mask[i] == SM_SentinelUndef) &&
9745 (Mask[i + 1] == SM_SentinelZero || Mask[i + 1] == SM_SentinelUndef)) {
9746 WidenedMask.push_back(SM_SentinelZero);
9752 // Finally check if the two mask values are adjacent and aligned with
9754 if (Mask[i] != SM_SentinelUndef && Mask[i] % 2 == 0 && Mask[i] + 1 == Mask[i + 1]) {
9755 WidenedMask.push_back(Mask[i] / 2);
9759 // Otherwise we can't safely widen the elements used in this shuffle.
9762 assert(WidenedMask.size() == Mask.size() / 2 &&
9763 "Incorrect size of mask after widening the elements!");
9768 /// \brief Generic routine to split vector shuffle into half-sized shuffles.
9770 /// This routine just extracts two subvectors, shuffles them independently, and
9771 /// then concatenates them back together. This should work effectively with all
9772 /// AVX vector shuffle types.
9773 static SDValue splitAndLowerVectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9774 SDValue V2, ArrayRef<int> Mask,
9775 SelectionDAG &DAG) {
9776 assert(VT.getSizeInBits() >= 256 &&
9777 "Only for 256-bit or wider vector shuffles!");
9778 assert(V1.getSimpleValueType() == VT && "Bad operand type!");
9779 assert(V2.getSimpleValueType() == VT && "Bad operand type!");
9781 ArrayRef<int> LoMask = Mask.slice(0, Mask.size() / 2);
9782 ArrayRef<int> HiMask = Mask.slice(Mask.size() / 2);
9784 int NumElements = VT.getVectorNumElements();
9785 int SplitNumElements = NumElements / 2;
9786 MVT ScalarVT = VT.getVectorElementType();
9787 MVT SplitVT = MVT::getVectorVT(ScalarVT, NumElements / 2);
9789 // Rather than splitting build-vectors, just build two narrower build
9790 // vectors. This helps shuffling with splats and zeros.
9791 auto SplitVector = [&](SDValue V) {
9792 while (V.getOpcode() == ISD::BITCAST)
9793 V = V->getOperand(0);
9795 MVT OrigVT = V.getSimpleValueType();
9796 int OrigNumElements = OrigVT.getVectorNumElements();
9797 int OrigSplitNumElements = OrigNumElements / 2;
9798 MVT OrigScalarVT = OrigVT.getVectorElementType();
9799 MVT OrigSplitVT = MVT::getVectorVT(OrigScalarVT, OrigNumElements / 2);
9803 auto *BV = dyn_cast<BuildVectorSDNode>(V);
9805 LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9806 DAG.getIntPtrConstant(0, DL));
9807 HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9808 DAG.getIntPtrConstant(OrigSplitNumElements, DL));
9811 SmallVector<SDValue, 16> LoOps, HiOps;
9812 for (int i = 0; i < OrigSplitNumElements; ++i) {
9813 LoOps.push_back(BV->getOperand(i));
9814 HiOps.push_back(BV->getOperand(i + OrigSplitNumElements));
9816 LoV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, LoOps);
9817 HiV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, HiOps);
9819 return std::make_pair(DAG.getBitcast(SplitVT, LoV),
9820 DAG.getBitcast(SplitVT, HiV));
9823 SDValue LoV1, HiV1, LoV2, HiV2;
9824 std::tie(LoV1, HiV1) = SplitVector(V1);
9825 std::tie(LoV2, HiV2) = SplitVector(V2);
9827 // Now create two 4-way blends of these half-width vectors.
9828 auto HalfBlend = [&](ArrayRef<int> HalfMask) {
9829 bool UseLoV1 = false, UseHiV1 = false, UseLoV2 = false, UseHiV2 = false;
9830 SmallVector<int, 32> V1BlendMask, V2BlendMask, BlendMask;
9831 for (int i = 0; i < SplitNumElements; ++i) {
9832 int M = HalfMask[i];
9833 if (M >= NumElements) {
9834 if (M >= NumElements + SplitNumElements)
9838 V2BlendMask.push_back(M - NumElements);
9839 V1BlendMask.push_back(-1);
9840 BlendMask.push_back(SplitNumElements + i);
9841 } else if (M >= 0) {
9842 if (M >= SplitNumElements)
9846 V2BlendMask.push_back(-1);
9847 V1BlendMask.push_back(M);
9848 BlendMask.push_back(i);
9850 V2BlendMask.push_back(-1);
9851 V1BlendMask.push_back(-1);
9852 BlendMask.push_back(-1);
9856 // Because the lowering happens after all combining takes place, we need to
9857 // manually combine these blend masks as much as possible so that we create
9858 // a minimal number of high-level vector shuffle nodes.
9860 // First try just blending the halves of V1 or V2.
9861 if (!UseLoV1 && !UseHiV1 && !UseLoV2 && !UseHiV2)
9862 return DAG.getUNDEF(SplitVT);
9863 if (!UseLoV2 && !UseHiV2)
9864 return DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9865 if (!UseLoV1 && !UseHiV1)
9866 return DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9868 SDValue V1Blend, V2Blend;
9869 if (UseLoV1 && UseHiV1) {
9871 DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9873 // We only use half of V1 so map the usage down into the final blend mask.
9874 V1Blend = UseLoV1 ? LoV1 : HiV1;
9875 for (int i = 0; i < SplitNumElements; ++i)
9876 if (BlendMask[i] >= 0 && BlendMask[i] < SplitNumElements)
9877 BlendMask[i] = V1BlendMask[i] - (UseLoV1 ? 0 : SplitNumElements);
9879 if (UseLoV2 && UseHiV2) {
9881 DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9883 // We only use half of V2 so map the usage down into the final blend mask.
9884 V2Blend = UseLoV2 ? LoV2 : HiV2;
9885 for (int i = 0; i < SplitNumElements; ++i)
9886 if (BlendMask[i] >= SplitNumElements)
9887 BlendMask[i] = V2BlendMask[i] + (UseLoV2 ? SplitNumElements : 0);
9889 return DAG.getVectorShuffle(SplitVT, DL, V1Blend, V2Blend, BlendMask);
9891 SDValue Lo = HalfBlend(LoMask);
9892 SDValue Hi = HalfBlend(HiMask);
9893 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
9896 /// \brief Either split a vector in halves or decompose the shuffles and the
9899 /// This is provided as a good fallback for many lowerings of non-single-input
9900 /// shuffles with more than one 128-bit lane. In those cases, we want to select
9901 /// between splitting the shuffle into 128-bit components and stitching those
9902 /// back together vs. extracting the single-input shuffles and blending those
9904 static SDValue lowerVectorShuffleAsSplitOrBlend(SDLoc DL, MVT VT, SDValue V1,
9905 SDValue V2, ArrayRef<int> Mask,
9906 SelectionDAG &DAG) {
9907 assert(!isSingleInputShuffleMask(Mask) && "This routine must not be used to "
9908 "lower single-input shuffles as it "
9909 "could then recurse on itself.");
9910 int Size = Mask.size();
9912 // If this can be modeled as a broadcast of two elements followed by a blend,
9913 // prefer that lowering. This is especially important because broadcasts can
9914 // often fold with memory operands.
9915 auto DoBothBroadcast = [&] {
9916 int V1BroadcastIdx = -1, V2BroadcastIdx = -1;
9919 if (V2BroadcastIdx == -1)
9920 V2BroadcastIdx = M - Size;
9921 else if (M - Size != V2BroadcastIdx)
9923 } else if (M >= 0) {
9924 if (V1BroadcastIdx == -1)
9926 else if (M != V1BroadcastIdx)
9931 if (DoBothBroadcast())
9932 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask,
9935 // If the inputs all stem from a single 128-bit lane of each input, then we
9936 // split them rather than blending because the split will decompose to
9937 // unusually few instructions.
9938 int LaneCount = VT.getSizeInBits() / 128;
9939 int LaneSize = Size / LaneCount;
9940 SmallBitVector LaneInputs[2];
9941 LaneInputs[0].resize(LaneCount, false);
9942 LaneInputs[1].resize(LaneCount, false);
9943 for (int i = 0; i < Size; ++i)
9945 LaneInputs[Mask[i] / Size][(Mask[i] % Size) / LaneSize] = true;
9946 if (LaneInputs[0].count() <= 1 && LaneInputs[1].count() <= 1)
9947 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9949 // Otherwise, just fall back to decomposed shuffles and a blend. This requires
9950 // that the decomposed single-input shuffles don't end up here.
9951 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
9954 /// \brief Lower a vector shuffle crossing multiple 128-bit lanes as
9955 /// a permutation and blend of those lanes.
9957 /// This essentially blends the out-of-lane inputs to each lane into the lane
9958 /// from a permuted copy of the vector. This lowering strategy results in four
9959 /// instructions in the worst case for a single-input cross lane shuffle which
9960 /// is lower than any other fully general cross-lane shuffle strategy I'm aware
9961 /// of. Special cases for each particular shuffle pattern should be handled
9962 /// prior to trying this lowering.
9963 static SDValue lowerVectorShuffleAsLanePermuteAndBlend(SDLoc DL, MVT VT,
9964 SDValue V1, SDValue V2,
9966 SelectionDAG &DAG) {
9967 // FIXME: This should probably be generalized for 512-bit vectors as well.
9968 assert(VT.is256BitVector() && "Only for 256-bit vector shuffles!");
9969 int LaneSize = Mask.size() / 2;
9971 // If there are only inputs from one 128-bit lane, splitting will in fact be
9972 // less expensive. The flags track whether the given lane contains an element
9973 // that crosses to another lane.
9974 bool LaneCrossing[2] = {false, false};
9975 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9976 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
9977 LaneCrossing[(Mask[i] % Size) / LaneSize] = true;
9978 if (!LaneCrossing[0] || !LaneCrossing[1])
9979 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9981 if (isSingleInputShuffleMask(Mask)) {
9982 SmallVector<int, 32> FlippedBlendMask;
9983 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9984 FlippedBlendMask.push_back(
9985 Mask[i] < 0 ? -1 : (((Mask[i] % Size) / LaneSize == i / LaneSize)
9987 : Mask[i] % LaneSize +
9988 (i / LaneSize) * LaneSize + Size));
9990 // Flip the vector, and blend the results which should now be in-lane. The
9991 // VPERM2X128 mask uses the low 2 bits for the low source and bits 4 and
9992 // 5 for the high source. The value 3 selects the high half of source 2 and
9993 // the value 2 selects the low half of source 2. We only use source 2 to
9994 // allow folding it into a memory operand.
9995 unsigned PERMMask = 3 | 2 << 4;
9996 SDValue Flipped = DAG.getNode(X86ISD::VPERM2X128, DL, VT, DAG.getUNDEF(VT),
9997 V1, DAG.getConstant(PERMMask, DL, MVT::i8));
9998 return DAG.getVectorShuffle(VT, DL, V1, Flipped, FlippedBlendMask);
10001 // This now reduces to two single-input shuffles of V1 and V2 which at worst
10002 // will be handled by the above logic and a blend of the results, much like
10003 // other patterns in AVX.
10004 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
10007 /// \brief Handle lowering 2-lane 128-bit shuffles.
10008 static SDValue lowerV2X128VectorShuffle(SDLoc DL, MVT VT, SDValue V1,
10009 SDValue V2, ArrayRef<int> Mask,
10010 const X86Subtarget *Subtarget,
10011 SelectionDAG &DAG) {
10012 // TODO: If minimizing size and one of the inputs is a zero vector and the
10013 // the zero vector has only one use, we could use a VPERM2X128 to save the
10014 // instruction bytes needed to explicitly generate the zero vector.
10016 // Blends are faster and handle all the non-lane-crossing cases.
10017 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, VT, V1, V2, Mask,
10021 bool IsV1Zero = ISD::isBuildVectorAllZeros(V1.getNode());
10022 bool IsV2Zero = ISD::isBuildVectorAllZeros(V2.getNode());
10024 // If either input operand is a zero vector, use VPERM2X128 because its mask
10025 // allows us to replace the zero input with an implicit zero.
10026 if (!IsV1Zero && !IsV2Zero) {
10027 // Check for patterns which can be matched with a single insert of a 128-bit
10029 bool OnlyUsesV1 = isShuffleEquivalent(V1, V2, Mask, {0, 1, 0, 1});
10030 if (OnlyUsesV1 || isShuffleEquivalent(V1, V2, Mask, {0, 1, 4, 5})) {
10031 MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(),
10032 VT.getVectorNumElements() / 2);
10033 SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
10034 DAG.getIntPtrConstant(0, DL));
10035 SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT,
10036 OnlyUsesV1 ? V1 : V2,
10037 DAG.getIntPtrConstant(0, DL));
10038 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
10042 // Otherwise form a 128-bit permutation. After accounting for undefs,
10043 // convert the 64-bit shuffle mask selection values into 128-bit
10044 // selection bits by dividing the indexes by 2 and shifting into positions
10045 // defined by a vperm2*128 instruction's immediate control byte.
10047 // The immediate permute control byte looks like this:
10048 // [1:0] - select 128 bits from sources for low half of destination
10050 // [3] - zero low half of destination
10051 // [5:4] - select 128 bits from sources for high half of destination
10053 // [7] - zero high half of destination
10055 int MaskLO = Mask[0];
10056 if (MaskLO == SM_SentinelUndef)
10057 MaskLO = Mask[1] == SM_SentinelUndef ? 0 : Mask[1];
10059 int MaskHI = Mask[2];
10060 if (MaskHI == SM_SentinelUndef)
10061 MaskHI = Mask[3] == SM_SentinelUndef ? 0 : Mask[3];
10063 unsigned PermMask = MaskLO / 2 | (MaskHI / 2) << 4;
10065 // If either input is a zero vector, replace it with an undef input.
10066 // Shuffle mask values < 4 are selecting elements of V1.
10067 // Shuffle mask values >= 4 are selecting elements of V2.
10068 // Adjust each half of the permute mask by clearing the half that was
10069 // selecting the zero vector and setting the zero mask bit.
10071 V1 = DAG.getUNDEF(VT);
10073 PermMask = (PermMask & 0xf0) | 0x08;
10075 PermMask = (PermMask & 0x0f) | 0x80;
10078 V2 = DAG.getUNDEF(VT);
10080 PermMask = (PermMask & 0xf0) | 0x08;
10082 PermMask = (PermMask & 0x0f) | 0x80;
10085 return DAG.getNode(X86ISD::VPERM2X128, DL, VT, V1, V2,
10086 DAG.getConstant(PermMask, DL, MVT::i8));
10089 /// \brief Lower a vector shuffle by first fixing the 128-bit lanes and then
10090 /// shuffling each lane.
10092 /// This will only succeed when the result of fixing the 128-bit lanes results
10093 /// in a single-input non-lane-crossing shuffle with a repeating shuffle mask in
10094 /// each 128-bit lanes. This handles many cases where we can quickly blend away
10095 /// the lane crosses early and then use simpler shuffles within each lane.
10097 /// FIXME: It might be worthwhile at some point to support this without
10098 /// requiring the 128-bit lane-relative shuffles to be repeating, but currently
10099 /// in x86 only floating point has interesting non-repeating shuffles, and even
10100 /// those are still *marginally* more expensive.
10101 static SDValue lowerVectorShuffleByMerging128BitLanes(
10102 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
10103 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
10104 assert(!isSingleInputShuffleMask(Mask) &&
10105 "This is only useful with multiple inputs.");
10107 int Size = Mask.size();
10108 int LaneSize = 128 / VT.getScalarSizeInBits();
10109 int NumLanes = Size / LaneSize;
10110 assert(NumLanes > 1 && "Only handles 256-bit and wider shuffles.");
10112 // See if we can build a hypothetical 128-bit lane-fixing shuffle mask. Also
10113 // check whether the in-128-bit lane shuffles share a repeating pattern.
10114 SmallVector<int, 4> Lanes;
10115 Lanes.resize(NumLanes, -1);
10116 SmallVector<int, 4> InLaneMask;
10117 InLaneMask.resize(LaneSize, -1);
10118 for (int i = 0; i < Size; ++i) {
10122 int j = i / LaneSize;
10124 if (Lanes[j] < 0) {
10125 // First entry we've seen for this lane.
10126 Lanes[j] = Mask[i] / LaneSize;
10127 } else if (Lanes[j] != Mask[i] / LaneSize) {
10128 // This doesn't match the lane selected previously!
10132 // Check that within each lane we have a consistent shuffle mask.
10133 int k = i % LaneSize;
10134 if (InLaneMask[k] < 0) {
10135 InLaneMask[k] = Mask[i] % LaneSize;
10136 } else if (InLaneMask[k] != Mask[i] % LaneSize) {
10137 // This doesn't fit a repeating in-lane mask.
10142 // First shuffle the lanes into place.
10143 MVT LaneVT = MVT::getVectorVT(VT.isFloatingPoint() ? MVT::f64 : MVT::i64,
10144 VT.getSizeInBits() / 64);
10145 SmallVector<int, 8> LaneMask;
10146 LaneMask.resize(NumLanes * 2, -1);
10147 for (int i = 0; i < NumLanes; ++i)
10148 if (Lanes[i] >= 0) {
10149 LaneMask[2 * i + 0] = 2*Lanes[i] + 0;
10150 LaneMask[2 * i + 1] = 2*Lanes[i] + 1;
10153 V1 = DAG.getBitcast(LaneVT, V1);
10154 V2 = DAG.getBitcast(LaneVT, V2);
10155 SDValue LaneShuffle = DAG.getVectorShuffle(LaneVT, DL, V1, V2, LaneMask);
10157 // Cast it back to the type we actually want.
10158 LaneShuffle = DAG.getBitcast(VT, LaneShuffle);
10160 // Now do a simple shuffle that isn't lane crossing.
10161 SmallVector<int, 8> NewMask;
10162 NewMask.resize(Size, -1);
10163 for (int i = 0; i < Size; ++i)
10165 NewMask[i] = (i / LaneSize) * LaneSize + Mask[i] % LaneSize;
10166 assert(!is128BitLaneCrossingShuffleMask(VT, NewMask) &&
10167 "Must not introduce lane crosses at this point!");
10169 return DAG.getVectorShuffle(VT, DL, LaneShuffle, DAG.getUNDEF(VT), NewMask);
10172 /// \brief Test whether the specified input (0 or 1) is in-place blended by the
10175 /// This returns true if the elements from a particular input are already in the
10176 /// slot required by the given mask and require no permutation.
10177 static bool isShuffleMaskInputInPlace(int Input, ArrayRef<int> Mask) {
10178 assert((Input == 0 || Input == 1) && "Only two inputs to shuffles.");
10179 int Size = Mask.size();
10180 for (int i = 0; i < Size; ++i)
10181 if (Mask[i] >= 0 && Mask[i] / Size == Input && Mask[i] % Size != i)
10187 static SDValue lowerVectorShuffleWithSHUFPD(SDLoc DL, MVT VT,
10188 ArrayRef<int> Mask, SDValue V1,
10189 SDValue V2, SelectionDAG &DAG) {
10191 // Mask for V8F64: 0/1, 8/9, 2/3, 10/11, 4/5, ..
10192 // Mask for V4F64; 0/1, 4/5, 2/3, 6/7..
10193 assert(VT.getScalarSizeInBits() == 64 && "Unexpected data type for VSHUFPD");
10194 int NumElts = VT.getVectorNumElements();
10195 bool ShufpdMask = true;
10196 bool CommutableMask = true;
10197 unsigned Immediate = 0;
10198 for (int i = 0; i < NumElts; ++i) {
10201 int Val = (i & 6) + NumElts * (i & 1);
10202 int CommutVal = (i & 0xe) + NumElts * ((i & 1)^1);
10203 if (Mask[i] < Val || Mask[i] > Val + 1)
10204 ShufpdMask = false;
10205 if (Mask[i] < CommutVal || Mask[i] > CommutVal + 1)
10206 CommutableMask = false;
10207 Immediate |= (Mask[i] % 2) << i;
10210 return DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
10211 DAG.getConstant(Immediate, DL, MVT::i8));
10212 if (CommutableMask)
10213 return DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
10214 DAG.getConstant(Immediate, DL, MVT::i8));
10218 /// \brief Handle lowering of 4-lane 64-bit floating point shuffles.
10220 /// Also ends up handling lowering of 4-lane 64-bit integer shuffles when AVX2
10221 /// isn't available.
10222 static SDValue lowerV4F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10223 const X86Subtarget *Subtarget,
10224 SelectionDAG &DAG) {
10226 assert(V1.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
10227 assert(V2.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
10228 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10229 ArrayRef<int> Mask = SVOp->getMask();
10230 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
10232 SmallVector<int, 4> WidenedMask;
10233 if (canWidenShuffleElements(Mask, WidenedMask))
10234 return lowerV2X128VectorShuffle(DL, MVT::v4f64, V1, V2, Mask, Subtarget,
10237 if (isSingleInputShuffleMask(Mask)) {
10238 // Check for being able to broadcast a single element.
10239 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f64, V1,
10240 Mask, Subtarget, DAG))
10243 // Use low duplicate instructions for masks that match their pattern.
10244 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
10245 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v4f64, V1);
10247 if (!is128BitLaneCrossingShuffleMask(MVT::v4f64, Mask)) {
10248 // Non-half-crossing single input shuffles can be lowerid with an
10249 // interleaved permutation.
10250 unsigned VPERMILPMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1) |
10251 ((Mask[2] == 3) << 2) | ((Mask[3] == 3) << 3);
10252 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f64, V1,
10253 DAG.getConstant(VPERMILPMask, DL, MVT::i8));
10256 // With AVX2 we have direct support for this permutation.
10257 if (Subtarget->hasAVX2())
10258 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4f64, V1,
10259 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
10261 // Otherwise, fall back.
10262 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v4f64, V1, V2, Mask,
10266 // Use dedicated unpack instructions for masks that match their pattern.
10268 lowerVectorShuffleWithUNPCK(DL, MVT::v4f64, Mask, V1, V2, DAG))
10271 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f64, V1, V2, Mask,
10275 // Check if the blend happens to exactly fit that of SHUFPD.
10277 lowerVectorShuffleWithSHUFPD(DL, MVT::v4f64, Mask, V1, V2, DAG))
10280 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10281 // shuffle. However, if we have AVX2 and either inputs are already in place,
10282 // we will be able to shuffle even across lanes the other input in a single
10283 // instruction so skip this pattern.
10284 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
10285 isShuffleMaskInputInPlace(1, Mask))))
10286 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10287 DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG))
10290 // If we have AVX2 then we always want to lower with a blend because an v4 we
10291 // can fully permute the elements.
10292 if (Subtarget->hasAVX2())
10293 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4f64, V1, V2,
10296 // Otherwise fall back on generic lowering.
10297 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v4f64, V1, V2, Mask, DAG);
10300 /// \brief Handle lowering of 4-lane 64-bit integer shuffles.
10302 /// This routine is only called when we have AVX2 and thus a reasonable
10303 /// instruction set for v4i64 shuffling..
10304 static SDValue lowerV4I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10305 const X86Subtarget *Subtarget,
10306 SelectionDAG &DAG) {
10308 assert(V1.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
10309 assert(V2.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
10310 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10311 ArrayRef<int> Mask = SVOp->getMask();
10312 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
10313 assert(Subtarget->hasAVX2() && "We can only lower v4i64 with AVX2!");
10315 SmallVector<int, 4> WidenedMask;
10316 if (canWidenShuffleElements(Mask, WidenedMask))
10317 return lowerV2X128VectorShuffle(DL, MVT::v4i64, V1, V2, Mask, Subtarget,
10320 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i64, V1, V2, Mask,
10324 // Check for being able to broadcast a single element.
10325 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i64, V1,
10326 Mask, Subtarget, DAG))
10329 // When the shuffle is mirrored between the 128-bit lanes of the unit, we can
10330 // use lower latency instructions that will operate on both 128-bit lanes.
10331 SmallVector<int, 2> RepeatedMask;
10332 if (is128BitLaneRepeatedShuffleMask(MVT::v4i64, Mask, RepeatedMask)) {
10333 if (isSingleInputShuffleMask(Mask)) {
10334 int PSHUFDMask[] = {-1, -1, -1, -1};
10335 for (int i = 0; i < 2; ++i)
10336 if (RepeatedMask[i] >= 0) {
10337 PSHUFDMask[2 * i] = 2 * RepeatedMask[i];
10338 PSHUFDMask[2 * i + 1] = 2 * RepeatedMask[i] + 1;
10340 return DAG.getBitcast(
10342 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32,
10343 DAG.getBitcast(MVT::v8i32, V1),
10344 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
10348 // AVX2 provides a direct instruction for permuting a single input across
10350 if (isSingleInputShuffleMask(Mask))
10351 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4i64, V1,
10352 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
10354 // Try to use shift instructions.
10355 if (SDValue Shift =
10356 lowerVectorShuffleAsShift(DL, MVT::v4i64, V1, V2, Mask, DAG))
10359 // Use dedicated unpack instructions for masks that match their pattern.
10361 lowerVectorShuffleWithUNPCK(DL, MVT::v4i64, Mask, V1, V2, DAG))
10364 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10365 // shuffle. However, if we have AVX2 and either inputs are already in place,
10366 // we will be able to shuffle even across lanes the other input in a single
10367 // instruction so skip this pattern.
10368 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
10369 isShuffleMaskInputInPlace(1, Mask))))
10370 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10371 DL, MVT::v4i64, V1, V2, Mask, Subtarget, DAG))
10374 // Otherwise fall back on generic blend lowering.
10375 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i64, V1, V2,
10379 /// \brief Handle lowering of 8-lane 32-bit floating point shuffles.
10381 /// Also ends up handling lowering of 8-lane 32-bit integer shuffles when AVX2
10382 /// isn't available.
10383 static SDValue lowerV8F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10384 const X86Subtarget *Subtarget,
10385 SelectionDAG &DAG) {
10387 assert(V1.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
10388 assert(V2.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
10389 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10390 ArrayRef<int> Mask = SVOp->getMask();
10391 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10393 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8f32, V1, V2, Mask,
10397 // Check for being able to broadcast a single element.
10398 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8f32, V1,
10399 Mask, Subtarget, DAG))
10402 // If the shuffle mask is repeated in each 128-bit lane, we have many more
10403 // options to efficiently lower the shuffle.
10404 SmallVector<int, 4> RepeatedMask;
10405 if (is128BitLaneRepeatedShuffleMask(MVT::v8f32, Mask, RepeatedMask)) {
10406 assert(RepeatedMask.size() == 4 &&
10407 "Repeated masks must be half the mask width!");
10409 // Use even/odd duplicate instructions for masks that match their pattern.
10410 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2, 4, 4, 6, 6}))
10411 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v8f32, V1);
10412 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3, 5, 5, 7, 7}))
10413 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v8f32, V1);
10415 if (isSingleInputShuffleMask(Mask))
10416 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f32, V1,
10417 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
10419 // Use dedicated unpack instructions for masks that match their pattern.
10421 lowerVectorShuffleWithUNPCK(DL, MVT::v8f32, Mask, V1, V2, DAG))
10424 // Otherwise, fall back to a SHUFPS sequence. Here it is important that we
10425 // have already handled any direct blends. We also need to squash the
10426 // repeated mask into a simulated v4f32 mask.
10427 for (int i = 0; i < 4; ++i)
10428 if (RepeatedMask[i] >= 8)
10429 RepeatedMask[i] -= 4;
10430 return lowerVectorShuffleWithSHUFPS(DL, MVT::v8f32, RepeatedMask, V1, V2, DAG);
10433 // If we have a single input shuffle with different shuffle patterns in the
10434 // two 128-bit lanes use the variable mask to VPERMILPS.
10435 if (isSingleInputShuffleMask(Mask)) {
10436 SDValue VPermMask[8];
10437 for (int i = 0; i < 8; ++i)
10438 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10439 : DAG.getConstant(Mask[i], DL, MVT::i32);
10440 if (!is128BitLaneCrossingShuffleMask(MVT::v8f32, Mask))
10441 return DAG.getNode(
10442 X86ISD::VPERMILPV, DL, MVT::v8f32, V1,
10443 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask));
10445 if (Subtarget->hasAVX2())
10446 return DAG.getNode(
10447 X86ISD::VPERMV, DL, MVT::v8f32,
10448 DAG.getBitcast(MVT::v8f32, DAG.getNode(ISD::BUILD_VECTOR, DL,
10449 MVT::v8i32, VPermMask)),
10452 // Otherwise, fall back.
10453 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v8f32, V1, V2, Mask,
10457 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10459 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10460 DL, MVT::v8f32, V1, V2, Mask, Subtarget, DAG))
10463 // If we have AVX2 then we always want to lower with a blend because at v8 we
10464 // can fully permute the elements.
10465 if (Subtarget->hasAVX2())
10466 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8f32, V1, V2,
10469 // Otherwise fall back on generic lowering.
10470 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v8f32, V1, V2, Mask, DAG);
10473 /// \brief Handle lowering of 8-lane 32-bit integer shuffles.
10475 /// This routine is only called when we have AVX2 and thus a reasonable
10476 /// instruction set for v8i32 shuffling..
10477 static SDValue lowerV8I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10478 const X86Subtarget *Subtarget,
10479 SelectionDAG &DAG) {
10481 assert(V1.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10482 assert(V2.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10483 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10484 ArrayRef<int> Mask = SVOp->getMask();
10485 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10486 assert(Subtarget->hasAVX2() && "We can only lower v8i32 with AVX2!");
10488 // Whenever we can lower this as a zext, that instruction is strictly faster
10489 // than any alternative. It also allows us to fold memory operands into the
10490 // shuffle in many cases.
10491 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v8i32, V1, V2,
10492 Mask, Subtarget, DAG))
10495 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i32, V1, V2, Mask,
10499 // Check for being able to broadcast a single element.
10500 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i32, V1,
10501 Mask, Subtarget, DAG))
10504 // If the shuffle mask is repeated in each 128-bit lane we can use more
10505 // efficient instructions that mirror the shuffles across the two 128-bit
10507 SmallVector<int, 4> RepeatedMask;
10508 if (is128BitLaneRepeatedShuffleMask(MVT::v8i32, Mask, RepeatedMask)) {
10509 assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!");
10510 if (isSingleInputShuffleMask(Mask))
10511 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, V1,
10512 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
10514 // Use dedicated unpack instructions for masks that match their pattern.
10516 lowerVectorShuffleWithUNPCK(DL, MVT::v8i32, Mask, V1, V2, DAG))
10520 // Try to use shift instructions.
10521 if (SDValue Shift =
10522 lowerVectorShuffleAsShift(DL, MVT::v8i32, V1, V2, Mask, DAG))
10525 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10526 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10529 // If the shuffle patterns aren't repeated but it is a single input, directly
10530 // generate a cross-lane VPERMD instruction.
10531 if (isSingleInputShuffleMask(Mask)) {
10532 SDValue VPermMask[8];
10533 for (int i = 0; i < 8; ++i)
10534 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10535 : DAG.getConstant(Mask[i], DL, MVT::i32);
10536 return DAG.getNode(
10537 X86ISD::VPERMV, DL, MVT::v8i32,
10538 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1);
10541 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10543 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10544 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10547 // Otherwise fall back on generic blend lowering.
10548 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i32, V1, V2,
10552 /// \brief Handle lowering of 16-lane 16-bit integer shuffles.
10554 /// This routine is only called when we have AVX2 and thus a reasonable
10555 /// instruction set for v16i16 shuffling..
10556 static SDValue lowerV16I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10557 const X86Subtarget *Subtarget,
10558 SelectionDAG &DAG) {
10560 assert(V1.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10561 assert(V2.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10562 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10563 ArrayRef<int> Mask = SVOp->getMask();
10564 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10565 assert(Subtarget->hasAVX2() && "We can only lower v16i16 with AVX2!");
10567 // Whenever we can lower this as a zext, that instruction is strictly faster
10568 // than any alternative. It also allows us to fold memory operands into the
10569 // shuffle in many cases.
10570 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v16i16, V1, V2,
10571 Mask, Subtarget, DAG))
10574 // Check for being able to broadcast a single element.
10575 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i16, V1,
10576 Mask, Subtarget, DAG))
10579 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i16, V1, V2, Mask,
10583 // Use dedicated unpack instructions for masks that match their pattern.
10585 lowerVectorShuffleWithUNPCK(DL, MVT::v16i16, Mask, V1, V2, DAG))
10588 // Try to use shift instructions.
10589 if (SDValue Shift =
10590 lowerVectorShuffleAsShift(DL, MVT::v16i16, V1, V2, Mask, DAG))
10593 // Try to use byte rotation instructions.
10594 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10595 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10598 if (isSingleInputShuffleMask(Mask)) {
10599 // There are no generalized cross-lane shuffle operations available on i16
10601 if (is128BitLaneCrossingShuffleMask(MVT::v16i16, Mask))
10602 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v16i16, V1, V2,
10605 SmallVector<int, 8> RepeatedMask;
10606 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
10607 // As this is a single-input shuffle, the repeated mask should be
10608 // a strictly valid v8i16 mask that we can pass through to the v8i16
10609 // lowering to handle even the v16 case.
10610 return lowerV8I16GeneralSingleInputVectorShuffle(
10611 DL, MVT::v16i16, V1, RepeatedMask, Subtarget, DAG);
10614 SDValue PSHUFBMask[32];
10615 for (int i = 0; i < 16; ++i) {
10616 if (Mask[i] == -1) {
10617 PSHUFBMask[2 * i] = PSHUFBMask[2 * i + 1] = DAG.getUNDEF(MVT::i8);
10621 int M = i < 8 ? Mask[i] : Mask[i] - 8;
10622 assert(M >= 0 && M < 8 && "Invalid single-input mask!");
10623 PSHUFBMask[2 * i] = DAG.getConstant(2 * M, DL, MVT::i8);
10624 PSHUFBMask[2 * i + 1] = DAG.getConstant(2 * M + 1, DL, MVT::i8);
10626 return DAG.getBitcast(MVT::v16i16,
10627 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8,
10628 DAG.getBitcast(MVT::v32i8, V1),
10629 DAG.getNode(ISD::BUILD_VECTOR, DL,
10630 MVT::v32i8, PSHUFBMask)));
10633 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10635 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10636 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10639 // Otherwise fall back on generic lowering.
10640 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v16i16, V1, V2, Mask, DAG);
10643 /// \brief Handle lowering of 32-lane 8-bit integer shuffles.
10645 /// This routine is only called when we have AVX2 and thus a reasonable
10646 /// instruction set for v32i8 shuffling..
10647 static SDValue lowerV32I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10648 const X86Subtarget *Subtarget,
10649 SelectionDAG &DAG) {
10651 assert(V1.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10652 assert(V2.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10653 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10654 ArrayRef<int> Mask = SVOp->getMask();
10655 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10656 assert(Subtarget->hasAVX2() && "We can only lower v32i8 with AVX2!");
10658 // Whenever we can lower this as a zext, that instruction is strictly faster
10659 // than any alternative. It also allows us to fold memory operands into the
10660 // shuffle in many cases.
10661 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v32i8, V1, V2,
10662 Mask, Subtarget, DAG))
10665 // Check for being able to broadcast a single element.
10666 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v32i8, V1,
10667 Mask, Subtarget, DAG))
10670 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v32i8, V1, V2, Mask,
10674 // Use dedicated unpack instructions for masks that match their pattern.
10676 lowerVectorShuffleWithUNPCK(DL, MVT::v32i8, Mask, V1, V2, DAG))
10679 // Try to use shift instructions.
10680 if (SDValue Shift =
10681 lowerVectorShuffleAsShift(DL, MVT::v32i8, V1, V2, Mask, DAG))
10684 // Try to use byte rotation instructions.
10685 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10686 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10689 if (isSingleInputShuffleMask(Mask)) {
10690 // There are no generalized cross-lane shuffle operations available on i8
10692 if (is128BitLaneCrossingShuffleMask(MVT::v32i8, Mask))
10693 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v32i8, V1, V2,
10696 SDValue PSHUFBMask[32];
10697 for (int i = 0; i < 32; ++i)
10700 ? DAG.getUNDEF(MVT::i8)
10701 : DAG.getConstant(Mask[i] < 16 ? Mask[i] : Mask[i] - 16, DL,
10704 return DAG.getNode(
10705 X86ISD::PSHUFB, DL, MVT::v32i8, V1,
10706 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask));
10709 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10711 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10712 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10715 // Otherwise fall back on generic lowering.
10716 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v32i8, V1, V2, Mask, DAG);
10719 /// \brief High-level routine to lower various 256-bit x86 vector shuffles.
10721 /// This routine either breaks down the specific type of a 256-bit x86 vector
10722 /// shuffle or splits it into two 128-bit shuffles and fuses the results back
10723 /// together based on the available instructions.
10724 static SDValue lower256BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10725 MVT VT, const X86Subtarget *Subtarget,
10726 SelectionDAG &DAG) {
10728 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10729 ArrayRef<int> Mask = SVOp->getMask();
10731 // If we have a single input to the zero element, insert that into V1 if we
10732 // can do so cheaply.
10733 int NumElts = VT.getVectorNumElements();
10734 int NumV2Elements = std::count_if(Mask.begin(), Mask.end(), [NumElts](int M) {
10735 return M >= NumElts;
10738 if (NumV2Elements == 1 && Mask[0] >= NumElts)
10739 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
10740 DL, VT, V1, V2, Mask, Subtarget, DAG))
10743 // There is a really nice hard cut-over between AVX1 and AVX2 that means we
10744 // can check for those subtargets here and avoid much of the subtarget
10745 // querying in the per-vector-type lowering routines. With AVX1 we have
10746 // essentially *zero* ability to manipulate a 256-bit vector with integer
10747 // types. Since we'll use floating point types there eventually, just
10748 // immediately cast everything to a float and operate entirely in that domain.
10749 if (VT.isInteger() && !Subtarget->hasAVX2()) {
10750 int ElementBits = VT.getScalarSizeInBits();
10751 if (ElementBits < 32)
10752 // No floating point type available, decompose into 128-bit vectors.
10753 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10755 MVT FpVT = MVT::getVectorVT(MVT::getFloatingPointVT(ElementBits),
10756 VT.getVectorNumElements());
10757 V1 = DAG.getBitcast(FpVT, V1);
10758 V2 = DAG.getBitcast(FpVT, V2);
10759 return DAG.getBitcast(VT, DAG.getVectorShuffle(FpVT, DL, V1, V2, Mask));
10762 switch (VT.SimpleTy) {
10764 return lowerV4F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10766 return lowerV4I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10768 return lowerV8F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10770 return lowerV8I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10772 return lowerV16I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10774 return lowerV32I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10777 llvm_unreachable("Not a valid 256-bit x86 vector type!");
10781 /// \brief Try to lower a vector shuffle as a 128-bit shuffles.
10782 static SDValue lowerV4X128VectorShuffle(SDLoc DL, MVT VT,
10783 ArrayRef<int> Mask,
10784 SDValue V1, SDValue V2,
10785 SelectionDAG &DAG) {
10786 assert(VT.getScalarSizeInBits() == 64 &&
10787 "Unexpected element type size for 128bit shuffle.");
10789 // To handle 256 bit vector requires VLX and most probably
10790 // function lowerV2X128VectorShuffle() is better solution.
10791 assert(VT.is512BitVector() && "Unexpected vector size for 128bit shuffle.");
10793 SmallVector<int, 4> WidenedMask;
10794 if (!canWidenShuffleElements(Mask, WidenedMask))
10797 // Form a 128-bit permutation.
10798 // Convert the 64-bit shuffle mask selection values into 128-bit selection
10799 // bits defined by a vshuf64x2 instruction's immediate control byte.
10800 unsigned PermMask = 0, Imm = 0;
10801 unsigned ControlBitsNum = WidenedMask.size() / 2;
10803 for (int i = 0, Size = WidenedMask.size(); i < Size; ++i) {
10804 if (WidenedMask[i] == SM_SentinelZero)
10807 // Use first element in place of undef mask.
10808 Imm = (WidenedMask[i] == SM_SentinelUndef) ? 0 : WidenedMask[i];
10809 PermMask |= (Imm % WidenedMask.size()) << (i * ControlBitsNum);
10812 return DAG.getNode(X86ISD::SHUF128, DL, VT, V1, V2,
10813 DAG.getConstant(PermMask, DL, MVT::i8));
10816 static SDValue lowerVectorShuffleWithPERMV(SDLoc DL, MVT VT,
10817 ArrayRef<int> Mask, SDValue V1,
10818 SDValue V2, SelectionDAG &DAG) {
10820 assert(VT.getScalarSizeInBits() >= 16 && "Unexpected data type for PERMV");
10822 MVT MaskEltVT = MVT::getIntegerVT(VT.getScalarSizeInBits());
10823 MVT MaskVecVT = MVT::getVectorVT(MaskEltVT, VT.getVectorNumElements());
10825 SDValue MaskNode = getConstVector(Mask, MaskVecVT, DAG, DL, true);
10826 if (isSingleInputShuffleMask(Mask))
10827 return DAG.getNode(X86ISD::VPERMV, DL, VT, MaskNode, V1);
10829 return DAG.getNode(X86ISD::VPERMV3, DL, VT, V1, MaskNode, V2);
10832 /// \brief Handle lowering of 8-lane 64-bit floating point shuffles.
10833 static SDValue lowerV8F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10834 const X86Subtarget *Subtarget,
10835 SelectionDAG &DAG) {
10837 assert(V1.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10838 assert(V2.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10839 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10840 ArrayRef<int> Mask = SVOp->getMask();
10841 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10843 if (SDValue Shuf128 =
10844 lowerV4X128VectorShuffle(DL, MVT::v8f64, Mask, V1, V2, DAG))
10847 if (SDValue Unpck =
10848 lowerVectorShuffleWithUNPCK(DL, MVT::v8f64, Mask, V1, V2, DAG))
10851 return lowerVectorShuffleWithPERMV(DL, MVT::v8f64, Mask, V1, V2, DAG);
10854 /// \brief Handle lowering of 16-lane 32-bit floating point shuffles.
10855 static SDValue lowerV16F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10856 const X86Subtarget *Subtarget,
10857 SelectionDAG &DAG) {
10859 assert(V1.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10860 assert(V2.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10861 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10862 ArrayRef<int> Mask = SVOp->getMask();
10863 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10865 if (SDValue Unpck =
10866 lowerVectorShuffleWithUNPCK(DL, MVT::v16f32, Mask, V1, V2, DAG))
10869 return lowerVectorShuffleWithPERMV(DL, MVT::v16f32, Mask, V1, V2, DAG);
10872 /// \brief Handle lowering of 8-lane 64-bit integer shuffles.
10873 static SDValue lowerV8I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10874 const X86Subtarget *Subtarget,
10875 SelectionDAG &DAG) {
10877 assert(V1.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
10878 assert(V2.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
10879 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10880 ArrayRef<int> Mask = SVOp->getMask();
10881 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10883 if (SDValue Shuf128 =
10884 lowerV4X128VectorShuffle(DL, MVT::v8i64, Mask, V1, V2, DAG))
10887 if (SDValue Unpck =
10888 lowerVectorShuffleWithUNPCK(DL, MVT::v8i64, Mask, V1, V2, DAG))
10891 return lowerVectorShuffleWithPERMV(DL, MVT::v8i64, Mask, V1, V2, DAG);
10894 /// \brief Handle lowering of 16-lane 32-bit integer shuffles.
10895 static SDValue lowerV16I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10896 const X86Subtarget *Subtarget,
10897 SelectionDAG &DAG) {
10899 assert(V1.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
10900 assert(V2.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
10901 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10902 ArrayRef<int> Mask = SVOp->getMask();
10903 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10905 if (SDValue Unpck =
10906 lowerVectorShuffleWithUNPCK(DL, MVT::v16i32, Mask, V1, V2, DAG))
10909 return lowerVectorShuffleWithPERMV(DL, MVT::v16i32, Mask, V1, V2, DAG);
10912 /// \brief Handle lowering of 32-lane 16-bit integer shuffles.
10913 static SDValue lowerV32I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10914 const X86Subtarget *Subtarget,
10915 SelectionDAG &DAG) {
10917 assert(V1.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10918 assert(V2.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10919 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10920 ArrayRef<int> Mask = SVOp->getMask();
10921 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10922 assert(Subtarget->hasBWI() && "We can only lower v32i16 with AVX-512-BWI!");
10924 return lowerVectorShuffleWithPERMV(DL, MVT::v32i16, Mask, V1, V2, DAG);
10927 /// \brief Handle lowering of 64-lane 8-bit integer shuffles.
10928 static SDValue lowerV64I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10929 const X86Subtarget *Subtarget,
10930 SelectionDAG &DAG) {
10932 assert(V1.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10933 assert(V2.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10934 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10935 ArrayRef<int> Mask = SVOp->getMask();
10936 assert(Mask.size() == 64 && "Unexpected mask size for v64 shuffle!");
10937 assert(Subtarget->hasBWI() && "We can only lower v64i8 with AVX-512-BWI!");
10939 // FIXME: Implement direct support for this type!
10940 return splitAndLowerVectorShuffle(DL, MVT::v64i8, V1, V2, Mask, DAG);
10943 /// \brief High-level routine to lower various 512-bit x86 vector shuffles.
10945 /// This routine either breaks down the specific type of a 512-bit x86 vector
10946 /// shuffle or splits it into two 256-bit shuffles and fuses the results back
10947 /// together based on the available instructions.
10948 static SDValue lower512BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10949 MVT VT, const X86Subtarget *Subtarget,
10950 SelectionDAG &DAG) {
10952 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10953 ArrayRef<int> Mask = SVOp->getMask();
10954 assert(Subtarget->hasAVX512() &&
10955 "Cannot lower 512-bit vectors w/ basic ISA!");
10957 // Check for being able to broadcast a single element.
10958 if (SDValue Broadcast =
10959 lowerVectorShuffleAsBroadcast(DL, VT, V1, Mask, Subtarget, DAG))
10962 // Dispatch to each element type for lowering. If we don't have supprot for
10963 // specific element type shuffles at 512 bits, immediately split them and
10964 // lower them. Each lowering routine of a given type is allowed to assume that
10965 // the requisite ISA extensions for that element type are available.
10966 switch (VT.SimpleTy) {
10968 return lowerV8F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10970 return lowerV16F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10972 return lowerV8I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10974 return lowerV16I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10976 if (Subtarget->hasBWI())
10977 return lowerV32I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10980 if (Subtarget->hasBWI())
10981 return lowerV64I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10985 llvm_unreachable("Not a valid 512-bit x86 vector type!");
10988 // Otherwise fall back on splitting.
10989 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10992 // Lower vXi1 vector shuffles.
10993 // There is no a dedicated instruction on AVX-512 that shuffles the masks.
10994 // The only way to shuffle bits is to sign-extend the mask vector to SIMD
10995 // vector, shuffle and then truncate it back.
10996 static SDValue lower1BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10997 MVT VT, const X86Subtarget *Subtarget,
10998 SelectionDAG &DAG) {
11000 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11001 ArrayRef<int> Mask = SVOp->getMask();
11002 assert(Subtarget->hasAVX512() &&
11003 "Cannot lower 512-bit vectors w/o basic ISA!");
11005 switch (VT.SimpleTy) {
11007 llvm_unreachable("Expected a vector of i1 elements");
11009 ExtVT = MVT::v2i64;
11012 ExtVT = MVT::v4i32;
11015 ExtVT = MVT::v8i64; // Take 512-bit type, more shuffles on KNL
11018 ExtVT = MVT::v16i32;
11021 ExtVT = MVT::v32i16;
11024 ExtVT = MVT::v64i8;
11028 if (ISD::isBuildVectorAllZeros(V1.getNode()))
11029 V1 = getZeroVector(ExtVT, Subtarget, DAG, DL);
11030 else if (ISD::isBuildVectorAllOnes(V1.getNode()))
11031 V1 = getOnesVector(ExtVT, Subtarget, DAG, DL);
11033 V1 = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, V1);
11036 V2 = DAG.getUNDEF(ExtVT);
11037 else if (ISD::isBuildVectorAllZeros(V2.getNode()))
11038 V2 = getZeroVector(ExtVT, Subtarget, DAG, DL);
11039 else if (ISD::isBuildVectorAllOnes(V2.getNode()))
11040 V2 = getOnesVector(ExtVT, Subtarget, DAG, DL);
11042 V2 = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, V2);
11043 return DAG.getNode(ISD::TRUNCATE, DL, VT,
11044 DAG.getVectorShuffle(ExtVT, DL, V1, V2, Mask));
11046 /// \brief Top-level lowering for x86 vector shuffles.
11048 /// This handles decomposition, canonicalization, and lowering of all x86
11049 /// vector shuffles. Most of the specific lowering strategies are encapsulated
11050 /// above in helper routines. The canonicalization attempts to widen shuffles
11051 /// to involve fewer lanes of wider elements, consolidate symmetric patterns
11052 /// s.t. only one of the two inputs needs to be tested, etc.
11053 static SDValue lowerVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
11054 SelectionDAG &DAG) {
11055 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11056 ArrayRef<int> Mask = SVOp->getMask();
11057 SDValue V1 = Op.getOperand(0);
11058 SDValue V2 = Op.getOperand(1);
11059 MVT VT = Op.getSimpleValueType();
11060 int NumElements = VT.getVectorNumElements();
11062 bool Is1BitVector = (VT.getVectorElementType() == MVT::i1);
11064 assert((VT.getSizeInBits() != 64 || Is1BitVector) &&
11065 "Can't lower MMX shuffles");
11067 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
11068 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
11069 if (V1IsUndef && V2IsUndef)
11070 return DAG.getUNDEF(VT);
11072 // When we create a shuffle node we put the UNDEF node to second operand,
11073 // but in some cases the first operand may be transformed to UNDEF.
11074 // In this case we should just commute the node.
11076 return DAG.getCommutedVectorShuffle(*SVOp);
11078 // Check for non-undef masks pointing at an undef vector and make the masks
11079 // undef as well. This makes it easier to match the shuffle based solely on
11083 if (M >= NumElements) {
11084 SmallVector<int, 8> NewMask(Mask.begin(), Mask.end());
11085 for (int &M : NewMask)
11086 if (M >= NumElements)
11088 return DAG.getVectorShuffle(VT, dl, V1, V2, NewMask);
11091 // We actually see shuffles that are entirely re-arrangements of a set of
11092 // zero inputs. This mostly happens while decomposing complex shuffles into
11093 // simple ones. Directly lower these as a buildvector of zeros.
11094 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
11095 if (Zeroable.all())
11096 return getZeroVector(VT, Subtarget, DAG, dl);
11098 // Try to collapse shuffles into using a vector type with fewer elements but
11099 // wider element types. We cap this to not form integers or floating point
11100 // elements wider than 64 bits, but it might be interesting to form i128
11101 // integers to handle flipping the low and high halves of AVX 256-bit vectors.
11102 SmallVector<int, 16> WidenedMask;
11103 if (VT.getScalarSizeInBits() < 64 && !Is1BitVector &&
11104 canWidenShuffleElements(Mask, WidenedMask)) {
11105 MVT NewEltVT = VT.isFloatingPoint()
11106 ? MVT::getFloatingPointVT(VT.getScalarSizeInBits() * 2)
11107 : MVT::getIntegerVT(VT.getScalarSizeInBits() * 2);
11108 MVT NewVT = MVT::getVectorVT(NewEltVT, VT.getVectorNumElements() / 2);
11109 // Make sure that the new vector type is legal. For example, v2f64 isn't
11111 if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) {
11112 V1 = DAG.getBitcast(NewVT, V1);
11113 V2 = DAG.getBitcast(NewVT, V2);
11114 return DAG.getBitcast(
11115 VT, DAG.getVectorShuffle(NewVT, dl, V1, V2, WidenedMask));
11119 int NumV1Elements = 0, NumUndefElements = 0, NumV2Elements = 0;
11120 for (int M : SVOp->getMask())
11122 ++NumUndefElements;
11123 else if (M < NumElements)
11128 // Commute the shuffle as needed such that more elements come from V1 than
11129 // V2. This allows us to match the shuffle pattern strictly on how many
11130 // elements come from V1 without handling the symmetric cases.
11131 if (NumV2Elements > NumV1Elements)
11132 return DAG.getCommutedVectorShuffle(*SVOp);
11134 // When the number of V1 and V2 elements are the same, try to minimize the
11135 // number of uses of V2 in the low half of the vector. When that is tied,
11136 // ensure that the sum of indices for V1 is equal to or lower than the sum
11137 // indices for V2. When those are equal, try to ensure that the number of odd
11138 // indices for V1 is lower than the number of odd indices for V2.
11139 if (NumV1Elements == NumV2Elements) {
11140 int LowV1Elements = 0, LowV2Elements = 0;
11141 for (int M : SVOp->getMask().slice(0, NumElements / 2))
11142 if (M >= NumElements)
11146 if (LowV2Elements > LowV1Elements) {
11147 return DAG.getCommutedVectorShuffle(*SVOp);
11148 } else if (LowV2Elements == LowV1Elements) {
11149 int SumV1Indices = 0, SumV2Indices = 0;
11150 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
11151 if (SVOp->getMask()[i] >= NumElements)
11153 else if (SVOp->getMask()[i] >= 0)
11155 if (SumV2Indices < SumV1Indices) {
11156 return DAG.getCommutedVectorShuffle(*SVOp);
11157 } else if (SumV2Indices == SumV1Indices) {
11158 int NumV1OddIndices = 0, NumV2OddIndices = 0;
11159 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
11160 if (SVOp->getMask()[i] >= NumElements)
11161 NumV2OddIndices += i % 2;
11162 else if (SVOp->getMask()[i] >= 0)
11163 NumV1OddIndices += i % 2;
11164 if (NumV2OddIndices < NumV1OddIndices)
11165 return DAG.getCommutedVectorShuffle(*SVOp);
11170 // For each vector width, delegate to a specialized lowering routine.
11171 if (VT.is128BitVector())
11172 return lower128BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11174 if (VT.is256BitVector())
11175 return lower256BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11177 if (VT.is512BitVector())
11178 return lower512BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11181 return lower1BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11182 llvm_unreachable("Unimplemented!");
11185 // This function assumes its argument is a BUILD_VECTOR of constants or
11186 // undef SDNodes. i.e: ISD::isBuildVectorOfConstantSDNodes(BuildVector) is
11188 static bool BUILD_VECTORtoBlendMask(BuildVectorSDNode *BuildVector,
11189 unsigned &MaskValue) {
11191 unsigned NumElems = BuildVector->getNumOperands();
11193 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
11194 // We don't handle the >2 lanes case right now.
11195 unsigned NumLanes = (NumElems - 1) / 8 + 1;
11199 unsigned NumElemsInLane = NumElems / NumLanes;
11201 // Blend for v16i16 should be symmetric for the both lanes.
11202 for (unsigned i = 0; i < NumElemsInLane; ++i) {
11203 SDValue EltCond = BuildVector->getOperand(i);
11204 SDValue SndLaneEltCond =
11205 (NumLanes == 2) ? BuildVector->getOperand(i + NumElemsInLane) : EltCond;
11207 int Lane1Cond = -1, Lane2Cond = -1;
11208 if (isa<ConstantSDNode>(EltCond))
11209 Lane1Cond = !isZero(EltCond);
11210 if (isa<ConstantSDNode>(SndLaneEltCond))
11211 Lane2Cond = !isZero(SndLaneEltCond);
11213 unsigned LaneMask = 0;
11214 if (Lane1Cond == Lane2Cond || Lane2Cond < 0)
11215 // Lane1Cond != 0, means we want the first argument.
11216 // Lane1Cond == 0, means we want the second argument.
11217 // The encoding of this argument is 0 for the first argument, 1
11218 // for the second. Therefore, invert the condition.
11219 LaneMask = !Lane1Cond << i;
11220 else if (Lane1Cond < 0)
11221 LaneMask = !Lane2Cond << i;
11225 MaskValue |= LaneMask;
11227 MaskValue |= LaneMask << NumElemsInLane;
11232 /// \brief Try to lower a VSELECT instruction to a vector shuffle.
11233 static SDValue lowerVSELECTtoVectorShuffle(SDValue Op,
11234 const X86Subtarget *Subtarget,
11235 SelectionDAG &DAG) {
11236 SDValue Cond = Op.getOperand(0);
11237 SDValue LHS = Op.getOperand(1);
11238 SDValue RHS = Op.getOperand(2);
11240 MVT VT = Op.getSimpleValueType();
11242 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
11244 auto *CondBV = cast<BuildVectorSDNode>(Cond);
11246 // Only non-legal VSELECTs reach this lowering, convert those into generic
11247 // shuffles and re-use the shuffle lowering path for blends.
11248 SmallVector<int, 32> Mask;
11249 for (int i = 0, Size = VT.getVectorNumElements(); i < Size; ++i) {
11250 SDValue CondElt = CondBV->getOperand(i);
11252 isa<ConstantSDNode>(CondElt) ? i + (isZero(CondElt) ? Size : 0) : -1);
11254 return DAG.getVectorShuffle(VT, dl, LHS, RHS, Mask);
11257 SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const {
11258 // A vselect where all conditions and data are constants can be optimized into
11259 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
11260 if (ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(0).getNode()) &&
11261 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(1).getNode()) &&
11262 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(2).getNode()))
11265 // Try to lower this to a blend-style vector shuffle. This can handle all
11266 // constant condition cases.
11267 if (SDValue BlendOp = lowerVSELECTtoVectorShuffle(Op, Subtarget, DAG))
11270 // Variable blends are only legal from SSE4.1 onward.
11271 if (!Subtarget->hasSSE41())
11274 // Only some types will be legal on some subtargets. If we can emit a legal
11275 // VSELECT-matching blend, return Op, and but if we need to expand, return
11277 switch (Op.getSimpleValueType().SimpleTy) {
11279 // Most of the vector types have blends past SSE4.1.
11283 // The byte blends for AVX vectors were introduced only in AVX2.
11284 if (Subtarget->hasAVX2())
11291 // AVX-512 BWI and VLX features support VSELECT with i16 elements.
11292 if (Subtarget->hasBWI() && Subtarget->hasVLX())
11295 // FIXME: We should custom lower this by fixing the condition and using i8
11301 static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
11302 MVT VT = Op.getSimpleValueType();
11305 if (!Op.getOperand(0).getSimpleValueType().is128BitVector())
11308 if (VT.getSizeInBits() == 8) {
11309 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
11310 Op.getOperand(0), Op.getOperand(1));
11311 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
11312 DAG.getValueType(VT));
11313 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11316 if (VT.getSizeInBits() == 16) {
11317 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11318 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
11320 return DAG.getNode(
11321 ISD::TRUNCATE, dl, MVT::i16,
11322 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11323 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
11324 Op.getOperand(1)));
11325 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
11326 Op.getOperand(0), Op.getOperand(1));
11327 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
11328 DAG.getValueType(VT));
11329 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11332 if (VT == MVT::f32) {
11333 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
11334 // the result back to FR32 register. It's only worth matching if the
11335 // result has a single use which is a store or a bitcast to i32. And in
11336 // the case of a store, it's not worth it if the index is a constant 0,
11337 // because a MOVSSmr can be used instead, which is smaller and faster.
11338 if (!Op.hasOneUse())
11340 SDNode *User = *Op.getNode()->use_begin();
11341 if ((User->getOpcode() != ISD::STORE ||
11342 (isa<ConstantSDNode>(Op.getOperand(1)) &&
11343 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
11344 (User->getOpcode() != ISD::BITCAST ||
11345 User->getValueType(0) != MVT::i32))
11347 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11348 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
11350 return DAG.getBitcast(MVT::f32, Extract);
11353 if (VT == MVT::i32 || VT == MVT::i64) {
11354 // ExtractPS/pextrq works with constant index.
11355 if (isa<ConstantSDNode>(Op.getOperand(1)))
11361 /// Extract one bit from mask vector, like v16i1 or v8i1.
11362 /// AVX-512 feature.
11364 X86TargetLowering::ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const {
11365 SDValue Vec = Op.getOperand(0);
11367 MVT VecVT = Vec.getSimpleValueType();
11368 SDValue Idx = Op.getOperand(1);
11369 MVT EltVT = Op.getSimpleValueType();
11371 assert((EltVT == MVT::i1) && "Unexpected operands in ExtractBitFromMaskVector");
11372 assert((VecVT.getVectorNumElements() <= 16 || Subtarget->hasBWI()) &&
11373 "Unexpected vector type in ExtractBitFromMaskVector");
11375 // variable index can't be handled in mask registers,
11376 // extend vector to VR512
11377 if (!isa<ConstantSDNode>(Idx)) {
11378 MVT ExtVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
11379 SDValue Ext = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVT, Vec);
11380 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
11381 ExtVT.getVectorElementType(), Ext, Idx);
11382 return DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
11385 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11386 const TargetRegisterClass* rc = getRegClassFor(VecVT);
11387 if (!Subtarget->hasDQI() && (VecVT.getVectorNumElements() <= 8))
11388 rc = getRegClassFor(MVT::v16i1);
11389 unsigned MaxSift = rc->getSize()*8 - 1;
11390 Vec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, Vec,
11391 DAG.getConstant(MaxSift - IdxVal, dl, MVT::i8));
11392 Vec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, Vec,
11393 DAG.getConstant(MaxSift, dl, MVT::i8));
11394 return DAG.getNode(X86ISD::VEXTRACT, dl, MVT::i1, Vec,
11395 DAG.getIntPtrConstant(0, dl));
11399 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
11400 SelectionDAG &DAG) const {
11402 SDValue Vec = Op.getOperand(0);
11403 MVT VecVT = Vec.getSimpleValueType();
11404 SDValue Idx = Op.getOperand(1);
11406 if (Op.getSimpleValueType() == MVT::i1)
11407 return ExtractBitFromMaskVector(Op, DAG);
11409 if (!isa<ConstantSDNode>(Idx)) {
11410 if (VecVT.is512BitVector() ||
11411 (VecVT.is256BitVector() && Subtarget->hasInt256() &&
11412 VecVT.getVectorElementType().getSizeInBits() == 32)) {
11415 MVT::getIntegerVT(VecVT.getVectorElementType().getSizeInBits());
11416 MVT MaskVT = MVT::getVectorVT(MaskEltVT, VecVT.getSizeInBits() /
11417 MaskEltVT.getSizeInBits());
11419 Idx = DAG.getZExtOrTrunc(Idx, dl, MaskEltVT);
11420 auto PtrVT = getPointerTy(DAG.getDataLayout());
11421 SDValue Mask = DAG.getNode(X86ISD::VINSERT, dl, MaskVT,
11422 getZeroVector(MaskVT, Subtarget, DAG, dl), Idx,
11423 DAG.getConstant(0, dl, PtrVT));
11424 SDValue Perm = DAG.getNode(X86ISD::VPERMV, dl, VecVT, Mask, Vec);
11425 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Perm,
11426 DAG.getConstant(0, dl, PtrVT));
11431 // If this is a 256-bit vector result, first extract the 128-bit vector and
11432 // then extract the element from the 128-bit vector.
11433 if (VecVT.is256BitVector() || VecVT.is512BitVector()) {
11435 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11436 // Get the 128-bit vector.
11437 Vec = Extract128BitVector(Vec, IdxVal, DAG, dl);
11438 MVT EltVT = VecVT.getVectorElementType();
11440 unsigned ElemsPerChunk = 128 / EltVT.getSizeInBits();
11441 assert(isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2");
11443 // Find IdxVal modulo ElemsPerChunk. Since ElemsPerChunk is a power of 2
11444 // this can be done with a mask.
11445 IdxVal &= ElemsPerChunk - 1;
11446 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
11447 DAG.getConstant(IdxVal, dl, MVT::i32));
11450 assert(VecVT.is128BitVector() && "Unexpected vector length");
11452 if (Subtarget->hasSSE41())
11453 if (SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG))
11456 MVT VT = Op.getSimpleValueType();
11457 // TODO: handle v16i8.
11458 if (VT.getSizeInBits() == 16) {
11459 SDValue Vec = Op.getOperand(0);
11460 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11462 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
11463 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11464 DAG.getBitcast(MVT::v4i32, Vec),
11465 Op.getOperand(1)));
11466 // Transform it so it match pextrw which produces a 32-bit result.
11467 MVT EltVT = MVT::i32;
11468 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
11469 Op.getOperand(0), Op.getOperand(1));
11470 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
11471 DAG.getValueType(VT));
11472 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11475 if (VT.getSizeInBits() == 32) {
11476 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11480 // SHUFPS the element to the lowest double word, then movss.
11481 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
11482 MVT VVT = Op.getOperand(0).getSimpleValueType();
11483 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
11484 DAG.getUNDEF(VVT), Mask);
11485 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
11486 DAG.getIntPtrConstant(0, dl));
11489 if (VT.getSizeInBits() == 64) {
11490 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
11491 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
11492 // to match extract_elt for f64.
11493 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11497 // UNPCKHPD the element to the lowest double word, then movsd.
11498 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
11499 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
11500 int Mask[2] = { 1, -1 };
11501 MVT VVT = Op.getOperand(0).getSimpleValueType();
11502 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
11503 DAG.getUNDEF(VVT), Mask);
11504 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
11505 DAG.getIntPtrConstant(0, dl));
11511 /// Insert one bit to mask vector, like v16i1 or v8i1.
11512 /// AVX-512 feature.
11514 X86TargetLowering::InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const {
11516 SDValue Vec = Op.getOperand(0);
11517 SDValue Elt = Op.getOperand(1);
11518 SDValue Idx = Op.getOperand(2);
11519 MVT VecVT = Vec.getSimpleValueType();
11521 if (!isa<ConstantSDNode>(Idx)) {
11522 // Non constant index. Extend source and destination,
11523 // insert element and then truncate the result.
11524 MVT ExtVecVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
11525 MVT ExtEltVT = (VecVT == MVT::v8i1 ? MVT::i64 : MVT::i32);
11526 SDValue ExtOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ExtVecVT,
11527 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVecVT, Vec),
11528 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtEltVT, Elt), Idx);
11529 return DAG.getNode(ISD::TRUNCATE, dl, VecVT, ExtOp);
11532 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11533 SDValue EltInVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Elt);
11535 EltInVec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
11536 DAG.getConstant(IdxVal, dl, MVT::i8));
11537 if (Vec.getOpcode() == ISD::UNDEF)
11539 return DAG.getNode(ISD::OR, dl, VecVT, Vec, EltInVec);
11542 SDValue X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
11543 SelectionDAG &DAG) const {
11544 MVT VT = Op.getSimpleValueType();
11545 MVT EltVT = VT.getVectorElementType();
11547 if (EltVT == MVT::i1)
11548 return InsertBitToMaskVector(Op, DAG);
11551 SDValue N0 = Op.getOperand(0);
11552 SDValue N1 = Op.getOperand(1);
11553 SDValue N2 = Op.getOperand(2);
11554 if (!isa<ConstantSDNode>(N2))
11556 auto *N2C = cast<ConstantSDNode>(N2);
11557 unsigned IdxVal = N2C->getZExtValue();
11559 // If the vector is wider than 128 bits, extract the 128-bit subvector, insert
11560 // into that, and then insert the subvector back into the result.
11561 if (VT.is256BitVector() || VT.is512BitVector()) {
11562 // With a 256-bit vector, we can insert into the zero element efficiently
11563 // using a blend if we have AVX or AVX2 and the right data type.
11564 if (VT.is256BitVector() && IdxVal == 0) {
11565 // TODO: It is worthwhile to cast integer to floating point and back
11566 // and incur a domain crossing penalty if that's what we'll end up
11567 // doing anyway after extracting to a 128-bit vector.
11568 if ((Subtarget->hasAVX() && (EltVT == MVT::f64 || EltVT == MVT::f32)) ||
11569 (Subtarget->hasAVX2() && EltVT == MVT::i32)) {
11570 SDValue N1Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, N1);
11571 N2 = DAG.getIntPtrConstant(1, dl);
11572 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1Vec, N2);
11576 // Get the desired 128-bit vector chunk.
11577 SDValue V = Extract128BitVector(N0, IdxVal, DAG, dl);
11579 // Insert the element into the desired chunk.
11580 unsigned NumEltsIn128 = 128 / EltVT.getSizeInBits();
11581 assert(isPowerOf2_32(NumEltsIn128));
11582 // Since NumEltsIn128 is a power of 2 we can use mask instead of modulo.
11583 unsigned IdxIn128 = IdxVal & (NumEltsIn128 - 1);
11585 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1,
11586 DAG.getConstant(IdxIn128, dl, MVT::i32));
11588 // Insert the changed part back into the bigger vector
11589 return Insert128BitVector(N0, V, IdxVal, DAG, dl);
11591 assert(VT.is128BitVector() && "Only 128-bit vector types should be left!");
11593 if (Subtarget->hasSSE41()) {
11594 if (EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) {
11596 if (VT == MVT::v8i16) {
11597 Opc = X86ISD::PINSRW;
11599 assert(VT == MVT::v16i8);
11600 Opc = X86ISD::PINSRB;
11603 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
11605 if (N1.getValueType() != MVT::i32)
11606 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11607 if (N2.getValueType() != MVT::i32)
11608 N2 = DAG.getIntPtrConstant(IdxVal, dl);
11609 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
11612 if (EltVT == MVT::f32) {
11613 // Bits [7:6] of the constant are the source select. This will always be
11614 // zero here. The DAG Combiner may combine an extract_elt index into
11615 // these bits. For example (insert (extract, 3), 2) could be matched by
11616 // putting the '3' into bits [7:6] of X86ISD::INSERTPS.
11617 // Bits [5:4] of the constant are the destination select. This is the
11618 // value of the incoming immediate.
11619 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
11620 // combine either bitwise AND or insert of float 0.0 to set these bits.
11622 bool MinSize = DAG.getMachineFunction().getFunction()->optForMinSize();
11623 if (IdxVal == 0 && (!MinSize || !MayFoldLoad(N1))) {
11624 // If this is an insertion of 32-bits into the low 32-bits of
11625 // a vector, we prefer to generate a blend with immediate rather
11626 // than an insertps. Blends are simpler operations in hardware and so
11627 // will always have equal or better performance than insertps.
11628 // But if optimizing for size and there's a load folding opportunity,
11629 // generate insertps because blendps does not have a 32-bit memory
11631 N2 = DAG.getIntPtrConstant(1, dl);
11632 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
11633 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1, N2);
11635 N2 = DAG.getIntPtrConstant(IdxVal << 4, dl);
11636 // Create this as a scalar to vector..
11637 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
11638 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
11641 if (EltVT == MVT::i32 || EltVT == MVT::i64) {
11642 // PINSR* works with constant index.
11647 if (EltVT == MVT::i8)
11650 if (EltVT.getSizeInBits() == 16) {
11651 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
11652 // as its second argument.
11653 if (N1.getValueType() != MVT::i32)
11654 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11655 if (N2.getValueType() != MVT::i32)
11656 N2 = DAG.getIntPtrConstant(IdxVal, dl);
11657 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
11662 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
11664 MVT OpVT = Op.getSimpleValueType();
11666 // If this is a 256-bit vector result, first insert into a 128-bit
11667 // vector and then insert into the 256-bit vector.
11668 if (!OpVT.is128BitVector()) {
11669 // Insert into a 128-bit vector.
11670 unsigned SizeFactor = OpVT.getSizeInBits()/128;
11671 MVT VT128 = MVT::getVectorVT(OpVT.getVectorElementType(),
11672 OpVT.getVectorNumElements() / SizeFactor);
11674 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
11676 // Insert the 128-bit vector.
11677 return Insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl);
11680 if (OpVT == MVT::v1i64 &&
11681 Op.getOperand(0).getValueType() == MVT::i64)
11682 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
11684 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
11685 assert(OpVT.is128BitVector() && "Expected an SSE type!");
11686 return DAG.getBitcast(
11687 OpVT, DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, AnyExt));
11690 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
11691 // a simple subregister reference or explicit instructions to grab
11692 // upper bits of a vector.
11693 static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
11694 SelectionDAG &DAG) {
11696 SDValue In = Op.getOperand(0);
11697 SDValue Idx = Op.getOperand(1);
11698 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11699 MVT ResVT = Op.getSimpleValueType();
11700 MVT InVT = In.getSimpleValueType();
11702 if (Subtarget->hasFp256()) {
11703 if (ResVT.is128BitVector() &&
11704 (InVT.is256BitVector() || InVT.is512BitVector()) &&
11705 isa<ConstantSDNode>(Idx)) {
11706 return Extract128BitVector(In, IdxVal, DAG, dl);
11708 if (ResVT.is256BitVector() && InVT.is512BitVector() &&
11709 isa<ConstantSDNode>(Idx)) {
11710 return Extract256BitVector(In, IdxVal, DAG, dl);
11716 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
11717 // simple superregister reference or explicit instructions to insert
11718 // the upper bits of a vector.
11719 static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
11720 SelectionDAG &DAG) {
11721 if (!Subtarget->hasAVX())
11725 SDValue Vec = Op.getOperand(0);
11726 SDValue SubVec = Op.getOperand(1);
11727 SDValue Idx = Op.getOperand(2);
11729 if (!isa<ConstantSDNode>(Idx))
11732 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11733 MVT OpVT = Op.getSimpleValueType();
11734 MVT SubVecVT = SubVec.getSimpleValueType();
11736 // Fold two 16-byte subvector loads into one 32-byte load:
11737 // (insert_subvector (insert_subvector undef, (load addr), 0),
11738 // (load addr + 16), Elts/2)
11740 if ((IdxVal == OpVT.getVectorNumElements() / 2) &&
11741 Vec.getOpcode() == ISD::INSERT_SUBVECTOR &&
11742 OpVT.is256BitVector() && SubVecVT.is128BitVector()) {
11743 auto *Idx2 = dyn_cast<ConstantSDNode>(Vec.getOperand(2));
11744 if (Idx2 && Idx2->getZExtValue() == 0) {
11745 SDValue SubVec2 = Vec.getOperand(1);
11746 // If needed, look through a bitcast to get to the load.
11747 if (SubVec2.getNode() && SubVec2.getOpcode() == ISD::BITCAST)
11748 SubVec2 = SubVec2.getOperand(0);
11750 if (auto *FirstLd = dyn_cast<LoadSDNode>(SubVec2)) {
11752 unsigned Alignment = FirstLd->getAlignment();
11753 unsigned AS = FirstLd->getAddressSpace();
11754 const X86TargetLowering *TLI = Subtarget->getTargetLowering();
11755 if (TLI->allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(),
11756 OpVT, AS, Alignment, &Fast) && Fast) {
11757 SDValue Ops[] = { SubVec2, SubVec };
11758 if (SDValue Ld = EltsFromConsecutiveLoads(OpVT, Ops, dl, DAG, false))
11765 if ((OpVT.is256BitVector() || OpVT.is512BitVector()) &&
11766 SubVecVT.is128BitVector())
11767 return Insert128BitVector(Vec, SubVec, IdxVal, DAG, dl);
11769 if (OpVT.is512BitVector() && SubVecVT.is256BitVector())
11770 return Insert256BitVector(Vec, SubVec, IdxVal, DAG, dl);
11772 if (OpVT.getVectorElementType() == MVT::i1) {
11773 if (IdxVal == 0 && Vec.getOpcode() == ISD::UNDEF) // the operation is legal
11775 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
11776 SDValue Undef = DAG.getUNDEF(OpVT);
11777 unsigned NumElems = OpVT.getVectorNumElements();
11778 SDValue ShiftBits = DAG.getConstant(NumElems/2, dl, MVT::i8);
11780 if (IdxVal == OpVT.getVectorNumElements() / 2) {
11781 // Zero upper bits of the Vec
11782 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
11783 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
11785 SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
11787 Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
11788 return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
11791 SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
11793 // Zero upper bits of the Vec2
11794 Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
11795 Vec2 = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec2, ShiftBits);
11796 // Zero lower bits of the Vec
11797 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
11798 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
11799 // Merge them together
11800 return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
11806 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
11807 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
11808 // one of the above mentioned nodes. It has to be wrapped because otherwise
11809 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
11810 // be used to form addressing mode. These wrapped nodes will be selected
11813 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
11814 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
11816 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11817 // global base reg.
11818 unsigned char OpFlag = 0;
11819 unsigned WrapperKind = X86ISD::Wrapper;
11820 CodeModel::Model M = DAG.getTarget().getCodeModel();
11822 if (Subtarget->isPICStyleRIPRel() &&
11823 (M == CodeModel::Small || M == CodeModel::Kernel))
11824 WrapperKind = X86ISD::WrapperRIP;
11825 else if (Subtarget->isPICStyleGOT())
11826 OpFlag = X86II::MO_GOTOFF;
11827 else if (Subtarget->isPICStyleStubPIC())
11828 OpFlag = X86II::MO_PIC_BASE_OFFSET;
11830 auto PtrVT = getPointerTy(DAG.getDataLayout());
11831 SDValue Result = DAG.getTargetConstantPool(
11832 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(), OpFlag);
11834 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11835 // With PIC, the address is actually $g + Offset.
11838 DAG.getNode(ISD::ADD, DL, PtrVT,
11839 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11845 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
11846 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
11848 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11849 // global base reg.
11850 unsigned char OpFlag = 0;
11851 unsigned WrapperKind = X86ISD::Wrapper;
11852 CodeModel::Model M = DAG.getTarget().getCodeModel();
11854 if (Subtarget->isPICStyleRIPRel() &&
11855 (M == CodeModel::Small || M == CodeModel::Kernel))
11856 WrapperKind = X86ISD::WrapperRIP;
11857 else if (Subtarget->isPICStyleGOT())
11858 OpFlag = X86II::MO_GOTOFF;
11859 else if (Subtarget->isPICStyleStubPIC())
11860 OpFlag = X86II::MO_PIC_BASE_OFFSET;
11862 auto PtrVT = getPointerTy(DAG.getDataLayout());
11863 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, OpFlag);
11865 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11867 // With PIC, the address is actually $g + Offset.
11870 DAG.getNode(ISD::ADD, DL, PtrVT,
11871 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11877 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
11878 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
11880 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11881 // global base reg.
11882 unsigned char OpFlag = 0;
11883 unsigned WrapperKind = X86ISD::Wrapper;
11884 CodeModel::Model M = DAG.getTarget().getCodeModel();
11886 if (Subtarget->isPICStyleRIPRel() &&
11887 (M == CodeModel::Small || M == CodeModel::Kernel)) {
11888 if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF())
11889 OpFlag = X86II::MO_GOTPCREL;
11890 WrapperKind = X86ISD::WrapperRIP;
11891 } else if (Subtarget->isPICStyleGOT()) {
11892 OpFlag = X86II::MO_GOT;
11893 } else if (Subtarget->isPICStyleStubPIC()) {
11894 OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE;
11895 } else if (Subtarget->isPICStyleStubNoDynamic()) {
11896 OpFlag = X86II::MO_DARWIN_NONLAZY;
11899 auto PtrVT = getPointerTy(DAG.getDataLayout());
11900 SDValue Result = DAG.getTargetExternalSymbol(Sym, PtrVT, OpFlag);
11903 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11905 // With PIC, the address is actually $g + Offset.
11906 if (DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
11907 !Subtarget->is64Bit()) {
11909 DAG.getNode(ISD::ADD, DL, PtrVT,
11910 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11913 // For symbols that require a load from a stub to get the address, emit the
11915 if (isGlobalStubReference(OpFlag))
11916 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
11917 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
11918 false, false, false, 0);
11924 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
11925 // Create the TargetBlockAddressAddress node.
11926 unsigned char OpFlags =
11927 Subtarget->ClassifyBlockAddressReference();
11928 CodeModel::Model M = DAG.getTarget().getCodeModel();
11929 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
11930 int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset();
11932 auto PtrVT = getPointerTy(DAG.getDataLayout());
11933 SDValue Result = DAG.getTargetBlockAddress(BA, PtrVT, Offset, OpFlags);
11935 if (Subtarget->isPICStyleRIPRel() &&
11936 (M == CodeModel::Small || M == CodeModel::Kernel))
11937 Result = DAG.getNode(X86ISD::WrapperRIP, dl, PtrVT, Result);
11939 Result = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, Result);
11941 // With PIC, the address is actually $g + Offset.
11942 if (isGlobalRelativeToPICBase(OpFlags)) {
11943 Result = DAG.getNode(ISD::ADD, dl, PtrVT,
11944 DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
11951 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
11952 int64_t Offset, SelectionDAG &DAG) const {
11953 // Create the TargetGlobalAddress node, folding in the constant
11954 // offset if it is legal.
11955 unsigned char OpFlags =
11956 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget());
11957 CodeModel::Model M = DAG.getTarget().getCodeModel();
11958 auto PtrVT = getPointerTy(DAG.getDataLayout());
11960 if (OpFlags == X86II::MO_NO_FLAG &&
11961 X86::isOffsetSuitableForCodeModel(Offset, M)) {
11962 // A direct static reference to a global.
11963 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, Offset);
11966 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, OpFlags);
11969 if (Subtarget->isPICStyleRIPRel() &&
11970 (M == CodeModel::Small || M == CodeModel::Kernel))
11971 Result = DAG.getNode(X86ISD::WrapperRIP, dl, PtrVT, Result);
11973 Result = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, Result);
11975 // With PIC, the address is actually $g + Offset.
11976 if (isGlobalRelativeToPICBase(OpFlags)) {
11977 Result = DAG.getNode(ISD::ADD, dl, PtrVT,
11978 DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
11981 // For globals that require a load from a stub to get the address, emit the
11983 if (isGlobalStubReference(OpFlags))
11984 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
11985 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
11986 false, false, false, 0);
11988 // If there was a non-zero offset that we didn't fold, create an explicit
11989 // addition for it.
11991 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result,
11992 DAG.getConstant(Offset, dl, PtrVT));
11998 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
11999 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
12000 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
12001 return LowerGlobalAddress(GV, SDLoc(Op), Offset, DAG);
12005 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
12006 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
12007 unsigned char OperandFlags, bool LocalDynamic = false) {
12008 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
12009 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
12011 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
12012 GA->getValueType(0),
12016 X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR
12020 SDValue Ops[] = { Chain, TGA, *InFlag };
12021 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
12023 SDValue Ops[] = { Chain, TGA };
12024 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
12027 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
12028 MFI->setAdjustsStack(true);
12029 MFI->setHasCalls(true);
12031 SDValue Flag = Chain.getValue(1);
12032 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
12035 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
12037 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
12040 SDLoc dl(GA); // ? function entry point might be better
12041 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
12042 DAG.getNode(X86ISD::GlobalBaseReg,
12043 SDLoc(), PtrVT), InFlag);
12044 InFlag = Chain.getValue(1);
12046 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
12049 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
12051 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
12053 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT,
12054 X86::RAX, X86II::MO_TLSGD);
12057 static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA,
12063 // Get the start address of the TLS block for this module.
12064 X86MachineFunctionInfo* MFI = DAG.getMachineFunction()
12065 .getInfo<X86MachineFunctionInfo>();
12066 MFI->incNumLocalDynamicTLSAccesses();
12070 Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, X86::RAX,
12071 X86II::MO_TLSLD, /*LocalDynamic=*/true);
12074 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
12075 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), InFlag);
12076 InFlag = Chain.getValue(1);
12077 Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX,
12078 X86II::MO_TLSLDM, /*LocalDynamic=*/true);
12081 // Note: the CleanupLocalDynamicTLSPass will remove redundant computations
12085 unsigned char OperandFlags = X86II::MO_DTPOFF;
12086 unsigned WrapperKind = X86ISD::Wrapper;
12087 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
12088 GA->getValueType(0),
12089 GA->getOffset(), OperandFlags);
12090 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
12092 // Add x@dtpoff with the base.
12093 return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base);
12096 // Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model.
12097 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
12098 const EVT PtrVT, TLSModel::Model model,
12099 bool is64Bit, bool isPIC) {
12102 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
12103 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
12104 is64Bit ? 257 : 256));
12106 SDValue ThreadPointer =
12107 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), DAG.getIntPtrConstant(0, dl),
12108 MachinePointerInfo(Ptr), false, false, false, 0);
12110 unsigned char OperandFlags = 0;
12111 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
12113 unsigned WrapperKind = X86ISD::Wrapper;
12114 if (model == TLSModel::LocalExec) {
12115 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
12116 } else if (model == TLSModel::InitialExec) {
12118 OperandFlags = X86II::MO_GOTTPOFF;
12119 WrapperKind = X86ISD::WrapperRIP;
12121 OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF;
12124 llvm_unreachable("Unexpected model");
12127 // emit "addl x@ntpoff,%eax" (local exec)
12128 // or "addl x@indntpoff,%eax" (initial exec)
12129 // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic)
12131 DAG.getTargetGlobalAddress(GA->getGlobal(), dl, GA->getValueType(0),
12132 GA->getOffset(), OperandFlags);
12133 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
12135 if (model == TLSModel::InitialExec) {
12136 if (isPIC && !is64Bit) {
12137 Offset = DAG.getNode(ISD::ADD, dl, PtrVT,
12138 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
12142 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
12143 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
12144 false, false, false, 0);
12147 // The address of the thread local variable is the add of the thread
12148 // pointer with the offset of the variable.
12149 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
12153 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
12155 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
12156 const GlobalValue *GV = GA->getGlobal();
12157 auto PtrVT = getPointerTy(DAG.getDataLayout());
12159 if (Subtarget->isTargetELF()) {
12160 if (DAG.getTarget().Options.EmulatedTLS)
12161 return LowerToTLSEmulatedModel(GA, DAG);
12162 TLSModel::Model model = DAG.getTarget().getTLSModel(GV);
12164 case TLSModel::GeneralDynamic:
12165 if (Subtarget->is64Bit())
12166 return LowerToTLSGeneralDynamicModel64(GA, DAG, PtrVT);
12167 return LowerToTLSGeneralDynamicModel32(GA, DAG, PtrVT);
12168 case TLSModel::LocalDynamic:
12169 return LowerToTLSLocalDynamicModel(GA, DAG, PtrVT,
12170 Subtarget->is64Bit());
12171 case TLSModel::InitialExec:
12172 case TLSModel::LocalExec:
12173 return LowerToTLSExecModel(GA, DAG, PtrVT, model, Subtarget->is64Bit(),
12174 DAG.getTarget().getRelocationModel() ==
12177 llvm_unreachable("Unknown TLS model.");
12180 if (Subtarget->isTargetDarwin()) {
12181 // Darwin only has one model of TLS. Lower to that.
12182 unsigned char OpFlag = 0;
12183 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
12184 X86ISD::WrapperRIP : X86ISD::Wrapper;
12186 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
12187 // global base reg.
12188 bool PIC32 = (DAG.getTarget().getRelocationModel() == Reloc::PIC_) &&
12189 !Subtarget->is64Bit();
12191 OpFlag = X86II::MO_TLVP_PIC_BASE;
12193 OpFlag = X86II::MO_TLVP;
12195 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
12196 GA->getValueType(0),
12197 GA->getOffset(), OpFlag);
12198 SDValue Offset = DAG.getNode(WrapperKind, DL, PtrVT, Result);
12200 // With PIC32, the address is actually $g + Offset.
12202 Offset = DAG.getNode(ISD::ADD, DL, PtrVT,
12203 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
12206 // Lowering the machine isd will make sure everything is in the right
12208 SDValue Chain = DAG.getEntryNode();
12209 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
12210 SDValue Args[] = { Chain, Offset };
12211 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args);
12213 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
12214 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
12215 MFI->setAdjustsStack(true);
12217 // And our return value (tls address) is in the standard call return value
12219 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
12220 return DAG.getCopyFromReg(Chain, DL, Reg, PtrVT, Chain.getValue(1));
12223 if (Subtarget->isTargetKnownWindowsMSVC() ||
12224 Subtarget->isTargetWindowsGNU()) {
12225 // Just use the implicit TLS architecture
12226 // Need to generate someting similar to:
12227 // mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage
12229 // mov ecx, dword [rel _tls_index]: Load index (from C runtime)
12230 // mov rcx, qword [rdx+rcx*8]
12231 // mov eax, .tls$:tlsvar
12232 // [rax+rcx] contains the address
12233 // Windows 64bit: gs:0x58
12234 // Windows 32bit: fs:__tls_array
12237 SDValue Chain = DAG.getEntryNode();
12239 // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or
12240 // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly
12241 // use its literal value of 0x2C.
12242 Value *Ptr = Constant::getNullValue(Subtarget->is64Bit()
12243 ? Type::getInt8PtrTy(*DAG.getContext(),
12245 : Type::getInt32PtrTy(*DAG.getContext(),
12248 SDValue TlsArray = Subtarget->is64Bit()
12249 ? DAG.getIntPtrConstant(0x58, dl)
12250 : (Subtarget->isTargetWindowsGNU()
12251 ? DAG.getIntPtrConstant(0x2C, dl)
12252 : DAG.getExternalSymbol("_tls_array", PtrVT));
12254 SDValue ThreadPointer =
12255 DAG.getLoad(PtrVT, dl, Chain, TlsArray, MachinePointerInfo(Ptr), false,
12259 if (GV->getThreadLocalMode() == GlobalVariable::LocalExecTLSModel) {
12260 res = ThreadPointer;
12262 // Load the _tls_index variable
12263 SDValue IDX = DAG.getExternalSymbol("_tls_index", PtrVT);
12264 if (Subtarget->is64Bit())
12265 IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, PtrVT, Chain, IDX,
12266 MachinePointerInfo(), MVT::i32, false, false,
12269 IDX = DAG.getLoad(PtrVT, dl, Chain, IDX, MachinePointerInfo(), false,
12272 auto &DL = DAG.getDataLayout();
12274 DAG.getConstant(Log2_64_Ceil(DL.getPointerSize()), dl, PtrVT);
12275 IDX = DAG.getNode(ISD::SHL, dl, PtrVT, IDX, Scale);
12277 res = DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, IDX);
12280 res = DAG.getLoad(PtrVT, dl, Chain, res, MachinePointerInfo(), false, false,
12283 // Get the offset of start of .tls section
12284 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
12285 GA->getValueType(0),
12286 GA->getOffset(), X86II::MO_SECREL);
12287 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, TGA);
12289 // The address of the thread local variable is the add of the thread
12290 // pointer with the offset of the variable.
12291 return DAG.getNode(ISD::ADD, dl, PtrVT, res, Offset);
12294 llvm_unreachable("TLS not implemented for this target.");
12297 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values
12298 /// and take a 2 x i32 value to shift plus a shift amount.
12299 static SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) {
12300 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
12301 MVT VT = Op.getSimpleValueType();
12302 unsigned VTBits = VT.getSizeInBits();
12304 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
12305 SDValue ShOpLo = Op.getOperand(0);
12306 SDValue ShOpHi = Op.getOperand(1);
12307 SDValue ShAmt = Op.getOperand(2);
12308 // X86ISD::SHLD and X86ISD::SHRD have defined overflow behavior but the
12309 // generic ISD nodes haven't. Insert an AND to be safe, it's optimized away
12311 SDValue SafeShAmt = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
12312 DAG.getConstant(VTBits - 1, dl, MVT::i8));
12313 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
12314 DAG.getConstant(VTBits - 1, dl, MVT::i8))
12315 : DAG.getConstant(0, dl, VT);
12317 SDValue Tmp2, Tmp3;
12318 if (Op.getOpcode() == ISD::SHL_PARTS) {
12319 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
12320 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, SafeShAmt);
12322 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
12323 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, SafeShAmt);
12326 // If the shift amount is larger or equal than the width of a part we can't
12327 // rely on the results of shld/shrd. Insert a test and select the appropriate
12328 // values for large shift amounts.
12329 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
12330 DAG.getConstant(VTBits, dl, MVT::i8));
12331 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
12332 AndNode, DAG.getConstant(0, dl, MVT::i8));
12335 SDValue CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
12336 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
12337 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
12339 if (Op.getOpcode() == ISD::SHL_PARTS) {
12340 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
12341 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
12343 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
12344 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
12347 SDValue Ops[2] = { Lo, Hi };
12348 return DAG.getMergeValues(Ops, dl);
12351 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
12352 SelectionDAG &DAG) const {
12353 SDValue Src = Op.getOperand(0);
12354 MVT SrcVT = Src.getSimpleValueType();
12355 MVT VT = Op.getSimpleValueType();
12358 if (SrcVT.isVector()) {
12359 if (SrcVT == MVT::v2i32 && VT == MVT::v2f64) {
12360 return DAG.getNode(X86ISD::CVTDQ2PD, dl, VT,
12361 DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4i32, Src,
12362 DAG.getUNDEF(SrcVT)));
12364 if (SrcVT.getVectorElementType() == MVT::i1) {
12365 MVT IntegerVT = MVT::getVectorVT(MVT::i32, SrcVT.getVectorNumElements());
12366 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
12367 DAG.getNode(ISD::SIGN_EXTEND, dl, IntegerVT, Src));
12372 assert(SrcVT <= MVT::i64 && SrcVT >= MVT::i16 &&
12373 "Unknown SINT_TO_FP to lower!");
12375 // These are really Legal; return the operand so the caller accepts it as
12377 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
12379 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
12380 Subtarget->is64Bit()) {
12384 unsigned Size = SrcVT.getSizeInBits()/8;
12385 MachineFunction &MF = DAG.getMachineFunction();
12386 auto PtrVT = getPointerTy(MF.getDataLayout());
12387 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
12388 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12389 SDValue Chain = DAG.getStore(
12390 DAG.getEntryNode(), dl, Op.getOperand(0), StackSlot,
12391 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI), false,
12393 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
12396 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
12398 SelectionDAG &DAG) const {
12402 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
12404 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
12406 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
12408 unsigned ByteSize = SrcVT.getSizeInBits()/8;
12410 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
12411 MachineMemOperand *MMO;
12413 int SSFI = FI->getIndex();
12414 MMO = DAG.getMachineFunction().getMachineMemOperand(
12415 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12416 MachineMemOperand::MOLoad, ByteSize, ByteSize);
12418 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
12419 StackSlot = StackSlot.getOperand(1);
12421 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
12422 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
12424 Tys, Ops, SrcVT, MMO);
12427 Chain = Result.getValue(1);
12428 SDValue InFlag = Result.getValue(2);
12430 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
12431 // shouldn't be necessary except that RFP cannot be live across
12432 // multiple blocks. When stackifier is fixed, they can be uncoupled.
12433 MachineFunction &MF = DAG.getMachineFunction();
12434 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
12435 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
12436 auto PtrVT = getPointerTy(MF.getDataLayout());
12437 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12438 Tys = DAG.getVTList(MVT::Other);
12440 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
12442 MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand(
12443 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12444 MachineMemOperand::MOStore, SSFISize, SSFISize);
12446 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
12447 Ops, Op.getValueType(), MMO);
12448 Result = DAG.getLoad(
12449 Op.getValueType(), DL, Chain, StackSlot,
12450 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12451 false, false, false, 0);
12457 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
12458 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
12459 SelectionDAG &DAG) const {
12460 // This algorithm is not obvious. Here it is what we're trying to output:
12463 punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U }
12464 subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 }
12466 haddpd %xmm0, %xmm0
12468 pshufd $0x4e, %xmm0, %xmm1
12474 LLVMContext *Context = DAG.getContext();
12476 // Build some magic constants.
12477 static const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 };
12478 Constant *C0 = ConstantDataVector::get(*Context, CV0);
12479 auto PtrVT = getPointerTy(DAG.getDataLayout());
12480 SDValue CPIdx0 = DAG.getConstantPool(C0, PtrVT, 16);
12482 SmallVector<Constant*,2> CV1;
12484 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
12485 APInt(64, 0x4330000000000000ULL))));
12487 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
12488 APInt(64, 0x4530000000000000ULL))));
12489 Constant *C1 = ConstantVector::get(CV1);
12490 SDValue CPIdx1 = DAG.getConstantPool(C1, PtrVT, 16);
12492 // Load the 64-bit value into an XMM register.
12493 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
12496 DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
12497 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12498 false, false, false, 16);
12500 getUnpackl(DAG, dl, MVT::v4i32, DAG.getBitcast(MVT::v4i32, XR1), CLod0);
12503 DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
12504 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12505 false, false, false, 16);
12506 SDValue XR2F = DAG.getBitcast(MVT::v2f64, Unpck1);
12507 // TODO: Are there any fast-math-flags to propagate here?
12508 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
12511 if (Subtarget->hasSSE3()) {
12512 // FIXME: The 'haddpd' instruction may be slower than 'movhlps + addsd'.
12513 Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub);
12515 SDValue S2F = DAG.getBitcast(MVT::v4i32, Sub);
12516 SDValue Shuffle = getTargetShuffleNode(X86ISD::PSHUFD, dl, MVT::v4i32,
12518 Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64,
12519 DAG.getBitcast(MVT::v2f64, Shuffle), Sub);
12522 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result,
12523 DAG.getIntPtrConstant(0, dl));
12526 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
12527 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
12528 SelectionDAG &DAG) const {
12530 // FP constant to bias correct the final result.
12531 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
12534 // Load the 32-bit value into an XMM register.
12535 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
12538 // Zero out the upper parts of the register.
12539 Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG);
12541 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12542 DAG.getBitcast(MVT::v2f64, Load),
12543 DAG.getIntPtrConstant(0, dl));
12545 // Or the load with the bias.
12546 SDValue Or = DAG.getNode(
12547 ISD::OR, dl, MVT::v2i64,
12548 DAG.getBitcast(MVT::v2i64,
12549 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Load)),
12550 DAG.getBitcast(MVT::v2i64,
12551 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Bias)));
12553 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12554 DAG.getBitcast(MVT::v2f64, Or), DAG.getIntPtrConstant(0, dl));
12556 // Subtract the bias.
12557 // TODO: Are there any fast-math-flags to propagate here?
12558 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
12560 // Handle final rounding.
12561 MVT DestVT = Op.getSimpleValueType();
12563 if (DestVT.bitsLT(MVT::f64))
12564 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
12565 DAG.getIntPtrConstant(0, dl));
12566 if (DestVT.bitsGT(MVT::f64))
12567 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
12569 // Handle final rounding.
12573 static SDValue lowerUINT_TO_FP_vXi32(SDValue Op, SelectionDAG &DAG,
12574 const X86Subtarget &Subtarget) {
12575 // The algorithm is the following:
12576 // #ifdef __SSE4_1__
12577 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
12578 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
12579 // (uint4) 0x53000000, 0xaa);
12581 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
12582 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
12584 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
12585 // return (float4) lo + fhi;
12587 // We shouldn't use it when unsafe-fp-math is enabled though: we might later
12588 // reassociate the two FADDs, and if we do that, the algorithm fails
12589 // spectacularly (PR24512).
12590 // FIXME: If we ever have some kind of Machine FMF, this should be marked
12591 // as non-fast and always be enabled. Why isn't SDAG FMF enough? Because
12592 // there's also the MachineCombiner reassociations happening on Machine IR.
12593 if (DAG.getTarget().Options.UnsafeFPMath)
12597 SDValue V = Op->getOperand(0);
12598 MVT VecIntVT = V.getSimpleValueType();
12599 bool Is128 = VecIntVT == MVT::v4i32;
12600 MVT VecFloatVT = Is128 ? MVT::v4f32 : MVT::v8f32;
12601 // If we convert to something else than the supported type, e.g., to v4f64,
12603 if (VecFloatVT != Op->getSimpleValueType(0))
12606 unsigned NumElts = VecIntVT.getVectorNumElements();
12607 assert((VecIntVT == MVT::v4i32 || VecIntVT == MVT::v8i32) &&
12608 "Unsupported custom type");
12609 assert(NumElts <= 8 && "The size of the constant array must be fixed");
12611 // In the #idef/#else code, we have in common:
12612 // - The vector of constants:
12618 // Create the splat vector for 0x4b000000.
12619 SDValue CstLow = DAG.getConstant(0x4b000000, DL, MVT::i32);
12620 SDValue CstLowArray[] = {CstLow, CstLow, CstLow, CstLow,
12621 CstLow, CstLow, CstLow, CstLow};
12622 SDValue VecCstLow = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12623 makeArrayRef(&CstLowArray[0], NumElts));
12624 // Create the splat vector for 0x53000000.
12625 SDValue CstHigh = DAG.getConstant(0x53000000, DL, MVT::i32);
12626 SDValue CstHighArray[] = {CstHigh, CstHigh, CstHigh, CstHigh,
12627 CstHigh, CstHigh, CstHigh, CstHigh};
12628 SDValue VecCstHigh = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12629 makeArrayRef(&CstHighArray[0], NumElts));
12631 // Create the right shift.
12632 SDValue CstShift = DAG.getConstant(16, DL, MVT::i32);
12633 SDValue CstShiftArray[] = {CstShift, CstShift, CstShift, CstShift,
12634 CstShift, CstShift, CstShift, CstShift};
12635 SDValue VecCstShift = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12636 makeArrayRef(&CstShiftArray[0], NumElts));
12637 SDValue HighShift = DAG.getNode(ISD::SRL, DL, VecIntVT, V, VecCstShift);
12640 if (Subtarget.hasSSE41()) {
12641 MVT VecI16VT = Is128 ? MVT::v8i16 : MVT::v16i16;
12642 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
12643 SDValue VecCstLowBitcast = DAG.getBitcast(VecI16VT, VecCstLow);
12644 SDValue VecBitcast = DAG.getBitcast(VecI16VT, V);
12645 // Low will be bitcasted right away, so do not bother bitcasting back to its
12647 Low = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecBitcast,
12648 VecCstLowBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
12649 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
12650 // (uint4) 0x53000000, 0xaa);
12651 SDValue VecCstHighBitcast = DAG.getBitcast(VecI16VT, VecCstHigh);
12652 SDValue VecShiftBitcast = DAG.getBitcast(VecI16VT, HighShift);
12653 // High will be bitcasted right away, so do not bother bitcasting back to
12654 // its original type.
12655 High = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecShiftBitcast,
12656 VecCstHighBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
12658 SDValue CstMask = DAG.getConstant(0xffff, DL, MVT::i32);
12659 SDValue VecCstMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT, CstMask,
12660 CstMask, CstMask, CstMask);
12661 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
12662 SDValue LowAnd = DAG.getNode(ISD::AND, DL, VecIntVT, V, VecCstMask);
12663 Low = DAG.getNode(ISD::OR, DL, VecIntVT, LowAnd, VecCstLow);
12665 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
12666 High = DAG.getNode(ISD::OR, DL, VecIntVT, HighShift, VecCstHigh);
12669 // Create the vector constant for -(0x1.0p39f + 0x1.0p23f).
12670 SDValue CstFAdd = DAG.getConstantFP(
12671 APFloat(APFloat::IEEEsingle, APInt(32, 0xD3000080)), DL, MVT::f32);
12672 SDValue CstFAddArray[] = {CstFAdd, CstFAdd, CstFAdd, CstFAdd,
12673 CstFAdd, CstFAdd, CstFAdd, CstFAdd};
12674 SDValue VecCstFAdd = DAG.getNode(ISD::BUILD_VECTOR, DL, VecFloatVT,
12675 makeArrayRef(&CstFAddArray[0], NumElts));
12677 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
12678 SDValue HighBitcast = DAG.getBitcast(VecFloatVT, High);
12679 // TODO: Are there any fast-math-flags to propagate here?
12681 DAG.getNode(ISD::FADD, DL, VecFloatVT, HighBitcast, VecCstFAdd);
12682 // return (float4) lo + fhi;
12683 SDValue LowBitcast = DAG.getBitcast(VecFloatVT, Low);
12684 return DAG.getNode(ISD::FADD, DL, VecFloatVT, LowBitcast, FHigh);
12687 SDValue X86TargetLowering::lowerUINT_TO_FP_vec(SDValue Op,
12688 SelectionDAG &DAG) const {
12689 SDValue N0 = Op.getOperand(0);
12690 MVT SVT = N0.getSimpleValueType();
12693 switch (SVT.SimpleTy) {
12695 llvm_unreachable("Custom UINT_TO_FP is not supported!");
12700 MVT NVT = MVT::getVectorVT(MVT::i32, SVT.getVectorNumElements());
12701 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
12702 DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0));
12706 return lowerUINT_TO_FP_vXi32(Op, DAG, *Subtarget);
12709 assert(Subtarget->hasAVX512());
12710 return DAG.getNode(ISD::UINT_TO_FP, dl, Op.getValueType(),
12711 DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v16i32, N0));
12715 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
12716 SelectionDAG &DAG) const {
12717 SDValue N0 = Op.getOperand(0);
12719 auto PtrVT = getPointerTy(DAG.getDataLayout());
12721 if (Op.getSimpleValueType().isVector())
12722 return lowerUINT_TO_FP_vec(Op, DAG);
12724 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
12725 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
12726 // the optimization here.
12727 if (DAG.SignBitIsZero(N0))
12728 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
12730 MVT SrcVT = N0.getSimpleValueType();
12731 MVT DstVT = Op.getSimpleValueType();
12733 if (Subtarget->hasAVX512() && isScalarFPTypeInSSEReg(DstVT) &&
12734 (SrcVT == MVT::i32 || (SrcVT == MVT::i64 && Subtarget->is64Bit()))) {
12735 // Conversions from unsigned i32 to f32/f64 are legal,
12736 // using VCVTUSI2SS/SD. Same for i64 in 64-bit mode.
12740 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
12741 return LowerUINT_TO_FP_i64(Op, DAG);
12742 if (SrcVT == MVT::i32 && X86ScalarSSEf64)
12743 return LowerUINT_TO_FP_i32(Op, DAG);
12744 if (Subtarget->is64Bit() && SrcVT == MVT::i64 && DstVT == MVT::f32)
12747 // Make a 64-bit buffer, and use it to build an FILD.
12748 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
12749 if (SrcVT == MVT::i32) {
12750 SDValue WordOff = DAG.getConstant(4, dl, PtrVT);
12751 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, WordOff);
12752 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
12753 StackSlot, MachinePointerInfo(),
12755 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, dl, MVT::i32),
12756 OffsetSlot, MachinePointerInfo(),
12758 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
12762 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
12763 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
12764 StackSlot, MachinePointerInfo(),
12766 // For i64 source, we need to add the appropriate power of 2 if the input
12767 // was negative. This is the same as the optimization in
12768 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
12769 // we must be careful to do the computation in x87 extended precision, not
12770 // in SSE. (The generic code can't know it's OK to do this, or how to.)
12771 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
12772 MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand(
12773 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12774 MachineMemOperand::MOLoad, 8, 8);
12776 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
12777 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
12778 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops,
12781 APInt FF(32, 0x5F800000ULL);
12783 // Check whether the sign bit is set.
12784 SDValue SignSet = DAG.getSetCC(
12785 dl, getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i64),
12786 Op.getOperand(0), DAG.getConstant(0, dl, MVT::i64), ISD::SETLT);
12788 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
12789 SDValue FudgePtr = DAG.getConstantPool(
12790 ConstantInt::get(*DAG.getContext(), FF.zext(64)), PtrVT);
12792 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
12793 SDValue Zero = DAG.getIntPtrConstant(0, dl);
12794 SDValue Four = DAG.getIntPtrConstant(4, dl);
12795 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
12797 FudgePtr = DAG.getNode(ISD::ADD, dl, PtrVT, FudgePtr, Offset);
12799 // Load the value out, extending it from f32 to f80.
12800 // FIXME: Avoid the extend by constructing the right constant pool?
12801 SDValue Fudge = DAG.getExtLoad(
12802 ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(), FudgePtr,
12803 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), MVT::f32,
12804 false, false, false, 4);
12805 // Extend everything to 80 bits to force it to be done on x87.
12806 // TODO: Are there any fast-math-flags to propagate here?
12807 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
12808 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add,
12809 DAG.getIntPtrConstant(0, dl));
12812 // If the given FP_TO_SINT (IsSigned) or FP_TO_UINT (!IsSigned) operation
12813 // is legal, or has an fp128 or f16 source (which needs to be promoted to f32),
12814 // just return an <SDValue(), SDValue()> pair.
12815 // Otherwise it is assumed to be a conversion from one of f32, f64 or f80
12816 // to i16, i32 or i64, and we lower it to a legal sequence.
12817 // If lowered to the final integer result we return a <result, SDValue()> pair.
12818 // Otherwise we lower it to a sequence ending with a FIST, return a
12819 // <FIST, StackSlot> pair, and the caller is responsible for loading
12820 // the final integer result from StackSlot.
12821 std::pair<SDValue,SDValue>
12822 X86TargetLowering::FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
12823 bool IsSigned, bool IsReplace) const {
12826 EVT DstTy = Op.getValueType();
12827 EVT TheVT = Op.getOperand(0).getValueType();
12828 auto PtrVT = getPointerTy(DAG.getDataLayout());
12830 if (TheVT != MVT::f32 && TheVT != MVT::f64 && TheVT != MVT::f80) {
12831 // f16 must be promoted before using the lowering in this routine.
12832 // fp128 does not use this lowering.
12833 return std::make_pair(SDValue(), SDValue());
12836 // If using FIST to compute an unsigned i64, we'll need some fixup
12837 // to handle values above the maximum signed i64. A FIST is always
12838 // used for the 32-bit subtarget, but also for f80 on a 64-bit target.
12839 bool UnsignedFixup = !IsSigned &&
12840 DstTy == MVT::i64 &&
12841 (!Subtarget->is64Bit() ||
12842 !isScalarFPTypeInSSEReg(TheVT));
12844 if (!IsSigned && DstTy != MVT::i64 && !Subtarget->hasAVX512()) {
12845 // Replace the fp-to-uint32 operation with an fp-to-sint64 FIST.
12846 // The low 32 bits of the fist result will have the correct uint32 result.
12847 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
12851 assert(DstTy.getSimpleVT() <= MVT::i64 &&
12852 DstTy.getSimpleVT() >= MVT::i16 &&
12853 "Unknown FP_TO_INT to lower!");
12855 // These are really Legal.
12856 if (DstTy == MVT::i32 &&
12857 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
12858 return std::make_pair(SDValue(), SDValue());
12859 if (Subtarget->is64Bit() &&
12860 DstTy == MVT::i64 &&
12861 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
12862 return std::make_pair(SDValue(), SDValue());
12864 // We lower FP->int64 into FISTP64 followed by a load from a temporary
12866 MachineFunction &MF = DAG.getMachineFunction();
12867 unsigned MemSize = DstTy.getSizeInBits()/8;
12868 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
12869 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12872 switch (DstTy.getSimpleVT().SimpleTy) {
12873 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
12874 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
12875 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
12876 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
12879 SDValue Chain = DAG.getEntryNode();
12880 SDValue Value = Op.getOperand(0);
12881 SDValue Adjust; // 0x0 or 0x80000000, for result sign bit adjustment.
12883 if (UnsignedFixup) {
12885 // Conversion to unsigned i64 is implemented with a select,
12886 // depending on whether the source value fits in the range
12887 // of a signed i64. Let Thresh be the FP equivalent of
12888 // 0x8000000000000000ULL.
12890 // Adjust i32 = (Value < Thresh) ? 0 : 0x80000000;
12891 // FistSrc = (Value < Thresh) ? Value : (Value - Thresh);
12892 // Fist-to-mem64 FistSrc
12893 // Add 0 or 0x800...0ULL to the 64-bit result, which is equivalent
12894 // to XOR'ing the high 32 bits with Adjust.
12896 // Being a power of 2, Thresh is exactly representable in all FP formats.
12897 // For X87 we'd like to use the smallest FP type for this constant, but
12898 // for DAG type consistency we have to match the FP operand type.
12900 APFloat Thresh(APFloat::IEEEsingle, APInt(32, 0x5f000000));
12901 LLVM_ATTRIBUTE_UNUSED APFloat::opStatus Status = APFloat::opOK;
12902 bool LosesInfo = false;
12903 if (TheVT == MVT::f64)
12904 // The rounding mode is irrelevant as the conversion should be exact.
12905 Status = Thresh.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven,
12907 else if (TheVT == MVT::f80)
12908 Status = Thresh.convert(APFloat::x87DoubleExtended,
12909 APFloat::rmNearestTiesToEven, &LosesInfo);
12911 assert(Status == APFloat::opOK && !LosesInfo &&
12912 "FP conversion should have been exact");
12914 SDValue ThreshVal = DAG.getConstantFP(Thresh, DL, TheVT);
12916 SDValue Cmp = DAG.getSetCC(DL,
12917 getSetCCResultType(DAG.getDataLayout(),
12918 *DAG.getContext(), TheVT),
12919 Value, ThreshVal, ISD::SETLT);
12920 Adjust = DAG.getSelect(DL, MVT::i32, Cmp,
12921 DAG.getConstant(0, DL, MVT::i32),
12922 DAG.getConstant(0x80000000, DL, MVT::i32));
12923 SDValue Sub = DAG.getNode(ISD::FSUB, DL, TheVT, Value, ThreshVal);
12924 Cmp = DAG.getSetCC(DL, getSetCCResultType(DAG.getDataLayout(),
12925 *DAG.getContext(), TheVT),
12926 Value, ThreshVal, ISD::SETLT);
12927 Value = DAG.getSelect(DL, TheVT, Cmp, Value, Sub);
12930 // FIXME This causes a redundant load/store if the SSE-class value is already
12931 // in memory, such as if it is on the callstack.
12932 if (isScalarFPTypeInSSEReg(TheVT)) {
12933 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
12934 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
12935 MachinePointerInfo::getFixedStack(MF, SSFI), false,
12937 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
12939 Chain, StackSlot, DAG.getValueType(TheVT)
12942 MachineMemOperand *MMO =
12943 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
12944 MachineMemOperand::MOLoad, MemSize, MemSize);
12945 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, DstTy, MMO);
12946 Chain = Value.getValue(1);
12947 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
12948 StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12951 MachineMemOperand *MMO =
12952 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
12953 MachineMemOperand::MOStore, MemSize, MemSize);
12955 if (UnsignedFixup) {
12957 // Insert the FIST, load its result as two i32's,
12958 // and XOR the high i32 with Adjust.
12960 SDValue FistOps[] = { Chain, Value, StackSlot };
12961 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
12962 FistOps, DstTy, MMO);
12964 SDValue Low32 = DAG.getLoad(MVT::i32, DL, FIST, StackSlot,
12965 MachinePointerInfo(),
12966 false, false, false, 0);
12967 SDValue HighAddr = DAG.getNode(ISD::ADD, DL, PtrVT, StackSlot,
12968 DAG.getConstant(4, DL, PtrVT));
12970 SDValue High32 = DAG.getLoad(MVT::i32, DL, FIST, HighAddr,
12971 MachinePointerInfo(),
12972 false, false, false, 0);
12973 High32 = DAG.getNode(ISD::XOR, DL, MVT::i32, High32, Adjust);
12975 if (Subtarget->is64Bit()) {
12976 // Join High32 and Low32 into a 64-bit result.
12977 // (High32 << 32) | Low32
12978 Low32 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, Low32);
12979 High32 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, High32);
12980 High32 = DAG.getNode(ISD::SHL, DL, MVT::i64, High32,
12981 DAG.getConstant(32, DL, MVT::i8));
12982 SDValue Result = DAG.getNode(ISD::OR, DL, MVT::i64, High32, Low32);
12983 return std::make_pair(Result, SDValue());
12986 SDValue ResultOps[] = { Low32, High32 };
12988 SDValue pair = IsReplace
12989 ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, ResultOps)
12990 : DAG.getMergeValues(ResultOps, DL);
12991 return std::make_pair(pair, SDValue());
12993 // Build the FP_TO_INT*_IN_MEM
12994 SDValue Ops[] = { Chain, Value, StackSlot };
12995 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
12997 return std::make_pair(FIST, StackSlot);
13001 static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
13002 const X86Subtarget *Subtarget) {
13003 MVT VT = Op->getSimpleValueType(0);
13004 SDValue In = Op->getOperand(0);
13005 MVT InVT = In.getSimpleValueType();
13008 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
13009 return DAG.getNode(ISD::ZERO_EXTEND, dl, VT, In);
13011 // Optimize vectors in AVX mode:
13014 // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32.
13015 // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32.
13016 // Concat upper and lower parts.
13019 // Use vpunpckldq for 4 lower elements v4i32 -> v2i64.
13020 // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64.
13021 // Concat upper and lower parts.
13024 if (((VT != MVT::v16i16) || (InVT != MVT::v16i8)) &&
13025 ((VT != MVT::v8i32) || (InVT != MVT::v8i16)) &&
13026 ((VT != MVT::v4i64) || (InVT != MVT::v4i32)))
13029 if (Subtarget->hasInt256())
13030 return DAG.getNode(X86ISD::VZEXT, dl, VT, In);
13032 SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl);
13033 SDValue Undef = DAG.getUNDEF(InVT);
13034 bool NeedZero = Op.getOpcode() == ISD::ZERO_EXTEND;
13035 SDValue OpLo = getUnpackl(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
13036 SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
13038 MVT HVT = MVT::getVectorVT(VT.getVectorElementType(),
13039 VT.getVectorNumElements()/2);
13041 OpLo = DAG.getBitcast(HVT, OpLo);
13042 OpHi = DAG.getBitcast(HVT, OpHi);
13044 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
13047 static SDValue LowerZERO_EXTEND_AVX512(SDValue Op,
13048 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
13049 MVT VT = Op->getSimpleValueType(0);
13050 SDValue In = Op->getOperand(0);
13051 MVT InVT = In.getSimpleValueType();
13053 unsigned int NumElts = VT.getVectorNumElements();
13054 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
13057 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1)
13058 return DAG.getNode(X86ISD::VZEXT, DL, VT, In);
13060 assert(InVT.getVectorElementType() == MVT::i1);
13061 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
13063 DAG.getConstant(APInt(ExtVT.getScalarSizeInBits(), 1), DL, ExtVT);
13065 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), DL, ExtVT);
13067 SDValue V = DAG.getNode(ISD::VSELECT, DL, ExtVT, In, One, Zero);
13068 if (VT.is512BitVector())
13070 return DAG.getNode(X86ISD::VTRUNC, DL, VT, V);
13073 static SDValue LowerANY_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
13074 SelectionDAG &DAG) {
13075 if (Subtarget->hasFp256())
13076 if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget))
13082 static SDValue LowerZERO_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
13083 SelectionDAG &DAG) {
13085 MVT VT = Op.getSimpleValueType();
13086 SDValue In = Op.getOperand(0);
13087 MVT SVT = In.getSimpleValueType();
13089 if (VT.is512BitVector() || SVT.getVectorElementType() == MVT::i1)
13090 return LowerZERO_EXTEND_AVX512(Op, Subtarget, DAG);
13092 if (Subtarget->hasFp256())
13093 if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget))
13096 assert(!VT.is256BitVector() || !SVT.is128BitVector() ||
13097 VT.getVectorNumElements() != SVT.getVectorNumElements());
13101 SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
13103 MVT VT = Op.getSimpleValueType();
13104 SDValue In = Op.getOperand(0);
13105 MVT InVT = In.getSimpleValueType();
13107 if (VT == MVT::i1) {
13108 assert((InVT.isInteger() && (InVT.getSizeInBits() <= 64)) &&
13109 "Invalid scalar TRUNCATE operation");
13110 if (InVT.getSizeInBits() >= 32)
13112 In = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, In);
13113 return DAG.getNode(ISD::TRUNCATE, DL, VT, In);
13115 assert(VT.getVectorNumElements() == InVT.getVectorNumElements() &&
13116 "Invalid TRUNCATE operation");
13118 // move vector to mask - truncate solution for SKX
13119 if (VT.getVectorElementType() == MVT::i1) {
13120 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() <= 16 &&
13121 Subtarget->hasBWI())
13122 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
13123 if ((InVT.is256BitVector() || InVT.is128BitVector())
13124 && InVT.getScalarSizeInBits() <= 16 &&
13125 Subtarget->hasBWI() && Subtarget->hasVLX())
13126 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
13127 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() >= 32 &&
13128 Subtarget->hasDQI())
13129 return Op; // legal, will go to VPMOVD2M, VPMOVQ2M
13130 if ((InVT.is256BitVector() || InVT.is128BitVector())
13131 && InVT.getScalarSizeInBits() >= 32 &&
13132 Subtarget->hasDQI() && Subtarget->hasVLX())
13133 return Op; // legal, will go to VPMOVB2M, VPMOVQ2M
13136 if (VT.getVectorElementType() == MVT::i1) {
13137 assert(VT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
13138 unsigned NumElts = InVT.getVectorNumElements();
13139 assert ((NumElts == 8 || NumElts == 16) && "Unexpected vector type");
13140 if (InVT.getSizeInBits() < 512) {
13141 MVT ExtVT = (NumElts == 16)? MVT::v16i32 : MVT::v8i64;
13142 In = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, In);
13147 DAG.getConstant(APInt::getSignBit(InVT.getScalarSizeInBits()), DL, InVT);
13148 SDValue And = DAG.getNode(ISD::AND, DL, InVT, OneV, In);
13149 return DAG.getNode(X86ISD::TESTM, DL, VT, And, And);
13152 // vpmovqb/w/d, vpmovdb/w, vpmovwb
13153 if (Subtarget->hasAVX512()) {
13154 // word to byte only under BWI
13155 if (InVT == MVT::v16i16 && !Subtarget->hasBWI()) // v16i16 -> v16i8
13156 return DAG.getNode(X86ISD::VTRUNC, DL, VT,
13157 DAG.getNode(X86ISD::VSEXT, DL, MVT::v16i32, In));
13158 return DAG.getNode(X86ISD::VTRUNC, DL, VT, In);
13160 if ((VT == MVT::v4i32) && (InVT == MVT::v4i64)) {
13161 // On AVX2, v4i64 -> v4i32 becomes VPERMD.
13162 if (Subtarget->hasInt256()) {
13163 static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1};
13164 In = DAG.getBitcast(MVT::v8i32, In);
13165 In = DAG.getVectorShuffle(MVT::v8i32, DL, In, DAG.getUNDEF(MVT::v8i32),
13167 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, In,
13168 DAG.getIntPtrConstant(0, DL));
13171 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13172 DAG.getIntPtrConstant(0, DL));
13173 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13174 DAG.getIntPtrConstant(2, DL));
13175 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
13176 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
13177 static const int ShufMask[] = {0, 2, 4, 6};
13178 return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask);
13181 if ((VT == MVT::v8i16) && (InVT == MVT::v8i32)) {
13182 // On AVX2, v8i32 -> v8i16 becomed PSHUFB.
13183 if (Subtarget->hasInt256()) {
13184 In = DAG.getBitcast(MVT::v32i8, In);
13186 SmallVector<SDValue,32> pshufbMask;
13187 for (unsigned i = 0; i < 2; ++i) {
13188 pshufbMask.push_back(DAG.getConstant(0x0, DL, MVT::i8));
13189 pshufbMask.push_back(DAG.getConstant(0x1, DL, MVT::i8));
13190 pshufbMask.push_back(DAG.getConstant(0x4, DL, MVT::i8));
13191 pshufbMask.push_back(DAG.getConstant(0x5, DL, MVT::i8));
13192 pshufbMask.push_back(DAG.getConstant(0x8, DL, MVT::i8));
13193 pshufbMask.push_back(DAG.getConstant(0x9, DL, MVT::i8));
13194 pshufbMask.push_back(DAG.getConstant(0xc, DL, MVT::i8));
13195 pshufbMask.push_back(DAG.getConstant(0xd, DL, MVT::i8));
13196 for (unsigned j = 0; j < 8; ++j)
13197 pshufbMask.push_back(DAG.getConstant(0x80, DL, MVT::i8));
13199 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, pshufbMask);
13200 In = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8, In, BV);
13201 In = DAG.getBitcast(MVT::v4i64, In);
13203 static const int ShufMask[] = {0, 2, -1, -1};
13204 In = DAG.getVectorShuffle(MVT::v4i64, DL, In, DAG.getUNDEF(MVT::v4i64),
13206 In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13207 DAG.getIntPtrConstant(0, DL));
13208 return DAG.getBitcast(VT, In);
13211 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
13212 DAG.getIntPtrConstant(0, DL));
13214 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
13215 DAG.getIntPtrConstant(4, DL));
13217 OpLo = DAG.getBitcast(MVT::v16i8, OpLo);
13218 OpHi = DAG.getBitcast(MVT::v16i8, OpHi);
13220 // The PSHUFB mask:
13221 static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13,
13222 -1, -1, -1, -1, -1, -1, -1, -1};
13224 SDValue Undef = DAG.getUNDEF(MVT::v16i8);
13225 OpLo = DAG.getVectorShuffle(MVT::v16i8, DL, OpLo, Undef, ShufMask1);
13226 OpHi = DAG.getVectorShuffle(MVT::v16i8, DL, OpHi, Undef, ShufMask1);
13228 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
13229 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
13231 // The MOVLHPS Mask:
13232 static const int ShufMask2[] = {0, 1, 4, 5};
13233 SDValue res = DAG.getVectorShuffle(MVT::v4i32, DL, OpLo, OpHi, ShufMask2);
13234 return DAG.getBitcast(MVT::v8i16, res);
13237 // Handle truncation of V256 to V128 using shuffles.
13238 if (!VT.is128BitVector() || !InVT.is256BitVector())
13241 assert(Subtarget->hasFp256() && "256-bit vector without AVX!");
13243 unsigned NumElems = VT.getVectorNumElements();
13244 MVT NVT = MVT::getVectorVT(VT.getVectorElementType(), NumElems * 2);
13246 SmallVector<int, 16> MaskVec(NumElems * 2, -1);
13247 // Prepare truncation shuffle mask
13248 for (unsigned i = 0; i != NumElems; ++i)
13249 MaskVec[i] = i * 2;
13250 SDValue V = DAG.getVectorShuffle(NVT, DL, DAG.getBitcast(NVT, In),
13251 DAG.getUNDEF(NVT), &MaskVec[0]);
13252 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V,
13253 DAG.getIntPtrConstant(0, DL));
13256 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
13257 SelectionDAG &DAG) const {
13258 assert(!Op.getSimpleValueType().isVector());
13260 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
13261 /*IsSigned=*/ true, /*IsReplace=*/ false);
13262 SDValue FIST = Vals.first, StackSlot = Vals.second;
13263 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
13264 if (!FIST.getNode())
13267 if (StackSlot.getNode())
13268 // Load the result.
13269 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
13270 FIST, StackSlot, MachinePointerInfo(),
13271 false, false, false, 0);
13273 // The node is the result.
13277 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
13278 SelectionDAG &DAG) const {
13279 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
13280 /*IsSigned=*/ false, /*IsReplace=*/ false);
13281 SDValue FIST = Vals.first, StackSlot = Vals.second;
13282 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
13283 if (!FIST.getNode())
13286 if (StackSlot.getNode())
13287 // Load the result.
13288 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
13289 FIST, StackSlot, MachinePointerInfo(),
13290 false, false, false, 0);
13292 // The node is the result.
13296 static SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) {
13298 MVT VT = Op.getSimpleValueType();
13299 SDValue In = Op.getOperand(0);
13300 MVT SVT = In.getSimpleValueType();
13302 assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!");
13304 return DAG.getNode(X86ISD::VFPEXT, DL, VT,
13305 DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32,
13306 In, DAG.getUNDEF(SVT)));
13309 /// The only differences between FABS and FNEG are the mask and the logic op.
13310 /// FNEG also has a folding opportunity for FNEG(FABS(x)).
13311 static SDValue LowerFABSorFNEG(SDValue Op, SelectionDAG &DAG) {
13312 assert((Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG) &&
13313 "Wrong opcode for lowering FABS or FNEG.");
13315 bool IsFABS = (Op.getOpcode() == ISD::FABS);
13317 // If this is a FABS and it has an FNEG user, bail out to fold the combination
13318 // into an FNABS. We'll lower the FABS after that if it is still in use.
13320 for (SDNode *User : Op->uses())
13321 if (User->getOpcode() == ISD::FNEG)
13325 MVT VT = Op.getSimpleValueType();
13327 // FIXME: Use function attribute "OptimizeForSize" and/or CodeGenOpt::Level to
13328 // decide if we should generate a 16-byte constant mask when we only need 4 or
13329 // 8 bytes for the scalar case.
13335 if (VT.isVector()) {
13337 EltVT = VT.getVectorElementType();
13338 NumElts = VT.getVectorNumElements();
13340 // There are no scalar bitwise logical SSE/AVX instructions, so we
13341 // generate a 16-byte vector constant and logic op even for the scalar case.
13342 // Using a 16-byte mask allows folding the load of the mask with
13343 // the logic op, so it can save (~4 bytes) on code size.
13344 LogicVT = (VT == MVT::f64) ? MVT::v2f64 : MVT::v4f32;
13346 NumElts = (VT == MVT::f64) ? 2 : 4;
13349 unsigned EltBits = EltVT.getSizeInBits();
13350 LLVMContext *Context = DAG.getContext();
13351 // For FABS, mask is 0x7f...; for FNEG, mask is 0x80...
13353 IsFABS ? APInt::getSignedMaxValue(EltBits) : APInt::getSignBit(EltBits);
13354 Constant *C = ConstantInt::get(*Context, MaskElt);
13355 C = ConstantVector::getSplat(NumElts, C);
13356 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13357 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout()));
13358 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
13360 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13361 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13362 false, false, false, Alignment);
13364 SDValue Op0 = Op.getOperand(0);
13365 bool IsFNABS = !IsFABS && (Op0.getOpcode() == ISD::FABS);
13367 IsFABS ? X86ISD::FAND : IsFNABS ? X86ISD::FOR : X86ISD::FXOR;
13368 SDValue Operand = IsFNABS ? Op0.getOperand(0) : Op0;
13371 return DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask);
13373 // For the scalar case extend to a 128-bit vector, perform the logic op,
13374 // and extract the scalar result back out.
13375 Operand = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Operand);
13376 SDValue LogicNode = DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask);
13377 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, LogicNode,
13378 DAG.getIntPtrConstant(0, dl));
13381 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
13382 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13383 LLVMContext *Context = DAG.getContext();
13384 SDValue Op0 = Op.getOperand(0);
13385 SDValue Op1 = Op.getOperand(1);
13387 MVT VT = Op.getSimpleValueType();
13388 MVT SrcVT = Op1.getSimpleValueType();
13390 // If second operand is smaller, extend it first.
13391 if (SrcVT.bitsLT(VT)) {
13392 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
13395 // And if it is bigger, shrink it first.
13396 if (SrcVT.bitsGT(VT)) {
13397 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1, dl));
13401 // At this point the operands and the result should have the same
13402 // type, and that won't be f80 since that is not custom lowered.
13404 const fltSemantics &Sem =
13405 VT == MVT::f64 ? APFloat::IEEEdouble : APFloat::IEEEsingle;
13406 const unsigned SizeInBits = VT.getSizeInBits();
13408 SmallVector<Constant *, 4> CV(
13409 VT == MVT::f64 ? 2 : 4,
13410 ConstantFP::get(*Context, APFloat(Sem, APInt(SizeInBits, 0))));
13412 // First, clear all bits but the sign bit from the second operand (sign).
13413 CV[0] = ConstantFP::get(*Context,
13414 APFloat(Sem, APInt::getHighBitsSet(SizeInBits, 1)));
13415 Constant *C = ConstantVector::get(CV);
13416 auto PtrVT = TLI.getPointerTy(DAG.getDataLayout());
13417 SDValue CPIdx = DAG.getConstantPool(C, PtrVT, 16);
13419 // Perform all logic operations as 16-byte vectors because there are no
13420 // scalar FP logic instructions in SSE. This allows load folding of the
13421 // constants into the logic instructions.
13422 MVT LogicVT = (VT == MVT::f64) ? MVT::v2f64 : MVT::v4f32;
13424 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13425 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13426 false, false, false, 16);
13427 Op1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Op1);
13428 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, LogicVT, Op1, Mask1);
13430 // Next, clear the sign bit from the first operand (magnitude).
13431 // If it's a constant, we can clear it here.
13432 if (ConstantFPSDNode *Op0CN = dyn_cast<ConstantFPSDNode>(Op0)) {
13433 APFloat APF = Op0CN->getValueAPF();
13434 // If the magnitude is a positive zero, the sign bit alone is enough.
13435 if (APF.isPosZero())
13436 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcVT, SignBit,
13437 DAG.getIntPtrConstant(0, dl));
13439 CV[0] = ConstantFP::get(*Context, APF);
13441 CV[0] = ConstantFP::get(
13443 APFloat(Sem, APInt::getLowBitsSet(SizeInBits, SizeInBits - 1)));
13445 C = ConstantVector::get(CV);
13446 CPIdx = DAG.getConstantPool(C, PtrVT, 16);
13448 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13449 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13450 false, false, false, 16);
13451 // If the magnitude operand wasn't a constant, we need to AND out the sign.
13452 if (!isa<ConstantFPSDNode>(Op0)) {
13453 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Op0);
13454 Val = DAG.getNode(X86ISD::FAND, dl, LogicVT, Op0, Val);
13456 // OR the magnitude value with the sign bit.
13457 Val = DAG.getNode(X86ISD::FOR, dl, LogicVT, Val, SignBit);
13458 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcVT, Val,
13459 DAG.getIntPtrConstant(0, dl));
13462 static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) {
13463 SDValue N0 = Op.getOperand(0);
13465 MVT VT = Op.getSimpleValueType();
13467 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
13468 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
13469 DAG.getConstant(1, dl, VT));
13470 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, dl, VT));
13473 // Check whether an OR'd tree is PTEST-able.
13474 static SDValue LowerVectorAllZeroTest(SDValue Op, const X86Subtarget *Subtarget,
13475 SelectionDAG &DAG) {
13476 assert(Op.getOpcode() == ISD::OR && "Only check OR'd tree.");
13478 if (!Subtarget->hasSSE41())
13481 if (!Op->hasOneUse())
13484 SDNode *N = Op.getNode();
13487 SmallVector<SDValue, 8> Opnds;
13488 DenseMap<SDValue, unsigned> VecInMap;
13489 SmallVector<SDValue, 8> VecIns;
13490 EVT VT = MVT::Other;
13492 // Recognize a special case where a vector is casted into wide integer to
13494 Opnds.push_back(N->getOperand(0));
13495 Opnds.push_back(N->getOperand(1));
13497 for (unsigned Slot = 0, e = Opnds.size(); Slot < e; ++Slot) {
13498 SmallVectorImpl<SDValue>::const_iterator I = Opnds.begin() + Slot;
13499 // BFS traverse all OR'd operands.
13500 if (I->getOpcode() == ISD::OR) {
13501 Opnds.push_back(I->getOperand(0));
13502 Opnds.push_back(I->getOperand(1));
13503 // Re-evaluate the number of nodes to be traversed.
13504 e += 2; // 2 more nodes (LHS and RHS) are pushed.
13508 // Quit if a non-EXTRACT_VECTOR_ELT
13509 if (I->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
13512 // Quit if without a constant index.
13513 SDValue Idx = I->getOperand(1);
13514 if (!isa<ConstantSDNode>(Idx))
13517 SDValue ExtractedFromVec = I->getOperand(0);
13518 DenseMap<SDValue, unsigned>::iterator M = VecInMap.find(ExtractedFromVec);
13519 if (M == VecInMap.end()) {
13520 VT = ExtractedFromVec.getValueType();
13521 // Quit if not 128/256-bit vector.
13522 if (!VT.is128BitVector() && !VT.is256BitVector())
13524 // Quit if not the same type.
13525 if (VecInMap.begin() != VecInMap.end() &&
13526 VT != VecInMap.begin()->first.getValueType())
13528 M = VecInMap.insert(std::make_pair(ExtractedFromVec, 0)).first;
13529 VecIns.push_back(ExtractedFromVec);
13531 M->second |= 1U << cast<ConstantSDNode>(Idx)->getZExtValue();
13534 assert((VT.is128BitVector() || VT.is256BitVector()) &&
13535 "Not extracted from 128-/256-bit vector.");
13537 unsigned FullMask = (1U << VT.getVectorNumElements()) - 1U;
13539 for (DenseMap<SDValue, unsigned>::const_iterator
13540 I = VecInMap.begin(), E = VecInMap.end(); I != E; ++I) {
13541 // Quit if not all elements are used.
13542 if (I->second != FullMask)
13546 MVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
13548 // Cast all vectors into TestVT for PTEST.
13549 for (unsigned i = 0, e = VecIns.size(); i < e; ++i)
13550 VecIns[i] = DAG.getBitcast(TestVT, VecIns[i]);
13552 // If more than one full vectors are evaluated, OR them first before PTEST.
13553 for (unsigned Slot = 0, e = VecIns.size(); e - Slot > 1; Slot += 2, e += 1) {
13554 // Each iteration will OR 2 nodes and append the result until there is only
13555 // 1 node left, i.e. the final OR'd value of all vectors.
13556 SDValue LHS = VecIns[Slot];
13557 SDValue RHS = VecIns[Slot + 1];
13558 VecIns.push_back(DAG.getNode(ISD::OR, DL, TestVT, LHS, RHS));
13561 return DAG.getNode(X86ISD::PTEST, DL, MVT::i32,
13562 VecIns.back(), VecIns.back());
13565 /// \brief return true if \c Op has a use that doesn't just read flags.
13566 static bool hasNonFlagsUse(SDValue Op) {
13567 for (SDNode::use_iterator UI = Op->use_begin(), UE = Op->use_end(); UI != UE;
13569 SDNode *User = *UI;
13570 unsigned UOpNo = UI.getOperandNo();
13571 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
13572 // Look pass truncate.
13573 UOpNo = User->use_begin().getOperandNo();
13574 User = *User->use_begin();
13577 if (User->getOpcode() != ISD::BRCOND && User->getOpcode() != ISD::SETCC &&
13578 !(User->getOpcode() == ISD::SELECT && UOpNo == 0))
13584 /// Emit nodes that will be selected as "test Op0,Op0", or something
13586 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC, SDLoc dl,
13587 SelectionDAG &DAG) const {
13588 if (Op.getValueType() == MVT::i1) {
13589 SDValue ExtOp = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i8, Op);
13590 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, ExtOp,
13591 DAG.getConstant(0, dl, MVT::i8));
13593 // CF and OF aren't always set the way we want. Determine which
13594 // of these we need.
13595 bool NeedCF = false;
13596 bool NeedOF = false;
13599 case X86::COND_A: case X86::COND_AE:
13600 case X86::COND_B: case X86::COND_BE:
13603 case X86::COND_G: case X86::COND_GE:
13604 case X86::COND_L: case X86::COND_LE:
13605 case X86::COND_O: case X86::COND_NO: {
13606 // Check if we really need to set the
13607 // Overflow flag. If NoSignedWrap is present
13608 // that is not actually needed.
13609 switch (Op->getOpcode()) {
13614 const auto *BinNode = cast<BinaryWithFlagsSDNode>(Op.getNode());
13615 if (BinNode->Flags.hasNoSignedWrap())
13625 // See if we can use the EFLAGS value from the operand instead of
13626 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
13627 // we prove that the arithmetic won't overflow, we can't use OF or CF.
13628 if (Op.getResNo() != 0 || NeedOF || NeedCF) {
13629 // Emit a CMP with 0, which is the TEST pattern.
13630 //if (Op.getValueType() == MVT::i1)
13631 // return DAG.getNode(X86ISD::CMP, dl, MVT::i1, Op,
13632 // DAG.getConstant(0, MVT::i1));
13633 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
13634 DAG.getConstant(0, dl, Op.getValueType()));
13636 unsigned Opcode = 0;
13637 unsigned NumOperands = 0;
13639 // Truncate operations may prevent the merge of the SETCC instruction
13640 // and the arithmetic instruction before it. Attempt to truncate the operands
13641 // of the arithmetic instruction and use a reduced bit-width instruction.
13642 bool NeedTruncation = false;
13643 SDValue ArithOp = Op;
13644 if (Op->getOpcode() == ISD::TRUNCATE && Op->hasOneUse()) {
13645 SDValue Arith = Op->getOperand(0);
13646 // Both the trunc and the arithmetic op need to have one user each.
13647 if (Arith->hasOneUse())
13648 switch (Arith.getOpcode()) {
13655 NeedTruncation = true;
13661 // NOTICE: In the code below we use ArithOp to hold the arithmetic operation
13662 // which may be the result of a CAST. We use the variable 'Op', which is the
13663 // non-casted variable when we check for possible users.
13664 switch (ArithOp.getOpcode()) {
13666 // Due to an isel shortcoming, be conservative if this add is likely to be
13667 // selected as part of a load-modify-store instruction. When the root node
13668 // in a match is a store, isel doesn't know how to remap non-chain non-flag
13669 // uses of other nodes in the match, such as the ADD in this case. This
13670 // leads to the ADD being left around and reselected, with the result being
13671 // two adds in the output. Alas, even if none our users are stores, that
13672 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
13673 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
13674 // climbing the DAG back to the root, and it doesn't seem to be worth the
13676 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13677 UE = Op.getNode()->use_end(); UI != UE; ++UI)
13678 if (UI->getOpcode() != ISD::CopyToReg &&
13679 UI->getOpcode() != ISD::SETCC &&
13680 UI->getOpcode() != ISD::STORE)
13683 if (ConstantSDNode *C =
13684 dyn_cast<ConstantSDNode>(ArithOp.getNode()->getOperand(1))) {
13685 // An add of one will be selected as an INC.
13686 if (C->getAPIntValue() == 1 && !Subtarget->slowIncDec()) {
13687 Opcode = X86ISD::INC;
13692 // An add of negative one (subtract of one) will be selected as a DEC.
13693 if (C->getAPIntValue().isAllOnesValue() && !Subtarget->slowIncDec()) {
13694 Opcode = X86ISD::DEC;
13700 // Otherwise use a regular EFLAGS-setting add.
13701 Opcode = X86ISD::ADD;
13706 // If we have a constant logical shift that's only used in a comparison
13707 // against zero turn it into an equivalent AND. This allows turning it into
13708 // a TEST instruction later.
13709 if ((X86CC == X86::COND_E || X86CC == X86::COND_NE) && Op->hasOneUse() &&
13710 isa<ConstantSDNode>(Op->getOperand(1)) && !hasNonFlagsUse(Op)) {
13711 EVT VT = Op.getValueType();
13712 unsigned BitWidth = VT.getSizeInBits();
13713 unsigned ShAmt = Op->getConstantOperandVal(1);
13714 if (ShAmt >= BitWidth) // Avoid undefined shifts.
13716 APInt Mask = ArithOp.getOpcode() == ISD::SRL
13717 ? APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt)
13718 : APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt);
13719 if (!Mask.isSignedIntN(32)) // Avoid large immediates.
13721 SDValue New = DAG.getNode(ISD::AND, dl, VT, Op->getOperand(0),
13722 DAG.getConstant(Mask, dl, VT));
13723 DAG.ReplaceAllUsesWith(Op, New);
13729 // If the primary and result isn't used, don't bother using X86ISD::AND,
13730 // because a TEST instruction will be better.
13731 if (!hasNonFlagsUse(Op))
13737 // Due to the ISEL shortcoming noted above, be conservative if this op is
13738 // likely to be selected as part of a load-modify-store instruction.
13739 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13740 UE = Op.getNode()->use_end(); UI != UE; ++UI)
13741 if (UI->getOpcode() == ISD::STORE)
13744 // Otherwise use a regular EFLAGS-setting instruction.
13745 switch (ArithOp.getOpcode()) {
13746 default: llvm_unreachable("unexpected operator!");
13747 case ISD::SUB: Opcode = X86ISD::SUB; break;
13748 case ISD::XOR: Opcode = X86ISD::XOR; break;
13749 case ISD::AND: Opcode = X86ISD::AND; break;
13751 if (!NeedTruncation && (X86CC == X86::COND_E || X86CC == X86::COND_NE)) {
13752 SDValue EFLAGS = LowerVectorAllZeroTest(Op, Subtarget, DAG);
13753 if (EFLAGS.getNode())
13756 Opcode = X86ISD::OR;
13770 return SDValue(Op.getNode(), 1);
13776 // If we found that truncation is beneficial, perform the truncation and
13778 if (NeedTruncation) {
13779 EVT VT = Op.getValueType();
13780 SDValue WideVal = Op->getOperand(0);
13781 EVT WideVT = WideVal.getValueType();
13782 unsigned ConvertedOp = 0;
13783 // Use a target machine opcode to prevent further DAGCombine
13784 // optimizations that may separate the arithmetic operations
13785 // from the setcc node.
13786 switch (WideVal.getOpcode()) {
13788 case ISD::ADD: ConvertedOp = X86ISD::ADD; break;
13789 case ISD::SUB: ConvertedOp = X86ISD::SUB; break;
13790 case ISD::AND: ConvertedOp = X86ISD::AND; break;
13791 case ISD::OR: ConvertedOp = X86ISD::OR; break;
13792 case ISD::XOR: ConvertedOp = X86ISD::XOR; break;
13796 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13797 if (TLI.isOperationLegal(WideVal.getOpcode(), WideVT)) {
13798 SDValue V0 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(0));
13799 SDValue V1 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(1));
13800 Op = DAG.getNode(ConvertedOp, dl, VT, V0, V1);
13806 // Emit a CMP with 0, which is the TEST pattern.
13807 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
13808 DAG.getConstant(0, dl, Op.getValueType()));
13810 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
13811 SmallVector<SDValue, 4> Ops(Op->op_begin(), Op->op_begin() + NumOperands);
13813 SDValue New = DAG.getNode(Opcode, dl, VTs, Ops);
13814 DAG.ReplaceAllUsesWith(Op, New);
13815 return SDValue(New.getNode(), 1);
13818 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
13820 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
13821 SDLoc dl, SelectionDAG &DAG) const {
13822 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1)) {
13823 if (C->getAPIntValue() == 0)
13824 return EmitTest(Op0, X86CC, dl, DAG);
13826 assert(Op0.getValueType() != MVT::i1 &&
13827 "Unexpected comparison operation for MVT::i1 operands");
13830 if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 ||
13831 Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) {
13832 // Do the comparison at i32 if it's smaller, besides the Atom case.
13833 // This avoids subregister aliasing issues. Keep the smaller reference
13834 // if we're optimizing for size, however, as that'll allow better folding
13835 // of memory operations.
13836 if (Op0.getValueType() != MVT::i32 && Op0.getValueType() != MVT::i64 &&
13837 !DAG.getMachineFunction().getFunction()->optForMinSize() &&
13838 !Subtarget->isAtom()) {
13839 unsigned ExtendOp =
13840 isX86CCUnsigned(X86CC) ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
13841 Op0 = DAG.getNode(ExtendOp, dl, MVT::i32, Op0);
13842 Op1 = DAG.getNode(ExtendOp, dl, MVT::i32, Op1);
13844 // Use SUB instead of CMP to enable CSE between SUB and CMP.
13845 SDVTList VTs = DAG.getVTList(Op0.getValueType(), MVT::i32);
13846 SDValue Sub = DAG.getNode(X86ISD::SUB, dl, VTs,
13848 return SDValue(Sub.getNode(), 1);
13850 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
13853 /// Convert a comparison if required by the subtarget.
13854 SDValue X86TargetLowering::ConvertCmpIfNecessary(SDValue Cmp,
13855 SelectionDAG &DAG) const {
13856 // If the subtarget does not support the FUCOMI instruction, floating-point
13857 // comparisons have to be converted.
13858 if (Subtarget->hasCMov() ||
13859 Cmp.getOpcode() != X86ISD::CMP ||
13860 !Cmp.getOperand(0).getValueType().isFloatingPoint() ||
13861 !Cmp.getOperand(1).getValueType().isFloatingPoint())
13864 // The instruction selector will select an FUCOM instruction instead of
13865 // FUCOMI, which writes the comparison result to FPSW instead of EFLAGS. Hence
13866 // build an SDNode sequence that transfers the result from FPSW into EFLAGS:
13867 // (X86sahf (trunc (srl (X86fp_stsw (trunc (X86cmp ...)), 8))))
13869 SDValue TruncFPSW = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, Cmp);
13870 SDValue FNStSW = DAG.getNode(X86ISD::FNSTSW16r, dl, MVT::i16, TruncFPSW);
13871 SDValue Srl = DAG.getNode(ISD::SRL, dl, MVT::i16, FNStSW,
13872 DAG.getConstant(8, dl, MVT::i8));
13873 SDValue TruncSrl = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Srl);
13874 return DAG.getNode(X86ISD::SAHF, dl, MVT::i32, TruncSrl);
13877 /// The minimum architected relative accuracy is 2^-12. We need one
13878 /// Newton-Raphson step to have a good float result (24 bits of precision).
13879 SDValue X86TargetLowering::getRsqrtEstimate(SDValue Op,
13880 DAGCombinerInfo &DCI,
13881 unsigned &RefinementSteps,
13882 bool &UseOneConstNR) const {
13883 EVT VT = Op.getValueType();
13884 const char *RecipOp;
13886 // SSE1 has rsqrtss and rsqrtps. AVX adds a 256-bit variant for rsqrtps.
13887 // TODO: Add support for AVX512 (v16f32).
13888 // It is likely not profitable to do this for f64 because a double-precision
13889 // rsqrt estimate with refinement on x86 prior to FMA requires at least 16
13890 // instructions: convert to single, rsqrtss, convert back to double, refine
13891 // (3 steps = at least 13 insts). If an 'rsqrtsd' variant was added to the ISA
13892 // along with FMA, this could be a throughput win.
13893 if (VT == MVT::f32 && Subtarget->hasSSE1())
13895 else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
13896 (VT == MVT::v8f32 && Subtarget->hasAVX()))
13897 RecipOp = "vec-sqrtf";
13901 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
13902 if (!Recips.isEnabled(RecipOp))
13905 RefinementSteps = Recips.getRefinementSteps(RecipOp);
13906 UseOneConstNR = false;
13907 return DCI.DAG.getNode(X86ISD::FRSQRT, SDLoc(Op), VT, Op);
13910 /// The minimum architected relative accuracy is 2^-12. We need one
13911 /// Newton-Raphson step to have a good float result (24 bits of precision).
13912 SDValue X86TargetLowering::getRecipEstimate(SDValue Op,
13913 DAGCombinerInfo &DCI,
13914 unsigned &RefinementSteps) const {
13915 EVT VT = Op.getValueType();
13916 const char *RecipOp;
13918 // SSE1 has rcpss and rcpps. AVX adds a 256-bit variant for rcpps.
13919 // TODO: Add support for AVX512 (v16f32).
13920 // It is likely not profitable to do this for f64 because a double-precision
13921 // reciprocal estimate with refinement on x86 prior to FMA requires
13922 // 15 instructions: convert to single, rcpss, convert back to double, refine
13923 // (3 steps = 12 insts). If an 'rcpsd' variant was added to the ISA
13924 // along with FMA, this could be a throughput win.
13925 if (VT == MVT::f32 && Subtarget->hasSSE1())
13927 else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
13928 (VT == MVT::v8f32 && Subtarget->hasAVX()))
13929 RecipOp = "vec-divf";
13933 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
13934 if (!Recips.isEnabled(RecipOp))
13937 RefinementSteps = Recips.getRefinementSteps(RecipOp);
13938 return DCI.DAG.getNode(X86ISD::FRCP, SDLoc(Op), VT, Op);
13941 /// If we have at least two divisions that use the same divisor, convert to
13942 /// multplication by a reciprocal. This may need to be adjusted for a given
13943 /// CPU if a division's cost is not at least twice the cost of a multiplication.
13944 /// This is because we still need one division to calculate the reciprocal and
13945 /// then we need two multiplies by that reciprocal as replacements for the
13946 /// original divisions.
13947 unsigned X86TargetLowering::combineRepeatedFPDivisors() const {
13951 static bool isAllOnes(SDValue V) {
13952 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
13953 return C && C->isAllOnesValue();
13956 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
13957 /// if it's possible.
13958 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
13959 SDLoc dl, SelectionDAG &DAG) const {
13960 SDValue Op0 = And.getOperand(0);
13961 SDValue Op1 = And.getOperand(1);
13962 if (Op0.getOpcode() == ISD::TRUNCATE)
13963 Op0 = Op0.getOperand(0);
13964 if (Op1.getOpcode() == ISD::TRUNCATE)
13965 Op1 = Op1.getOperand(0);
13968 if (Op1.getOpcode() == ISD::SHL)
13969 std::swap(Op0, Op1);
13970 if (Op0.getOpcode() == ISD::SHL) {
13971 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
13972 if (And00C->getZExtValue() == 1) {
13973 // If we looked past a truncate, check that it's only truncating away
13975 unsigned BitWidth = Op0.getValueSizeInBits();
13976 unsigned AndBitWidth = And.getValueSizeInBits();
13977 if (BitWidth > AndBitWidth) {
13979 DAG.computeKnownBits(Op0, Zeros, Ones);
13980 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
13984 RHS = Op0.getOperand(1);
13986 } else if (Op1.getOpcode() == ISD::Constant) {
13987 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
13988 uint64_t AndRHSVal = AndRHS->getZExtValue();
13989 SDValue AndLHS = Op0;
13991 if (AndRHSVal == 1 && AndLHS.getOpcode() == ISD::SRL) {
13992 LHS = AndLHS.getOperand(0);
13993 RHS = AndLHS.getOperand(1);
13996 // Use BT if the immediate can't be encoded in a TEST instruction.
13997 if (!isUInt<32>(AndRHSVal) && isPowerOf2_64(AndRHSVal)) {
13999 RHS = DAG.getConstant(Log2_64_Ceil(AndRHSVal), dl, LHS.getValueType());
14003 if (LHS.getNode()) {
14004 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
14005 // instruction. Since the shift amount is in-range-or-undefined, we know
14006 // that doing a bittest on the i32 value is ok. We extend to i32 because
14007 // the encoding for the i16 version is larger than the i32 version.
14008 // Also promote i16 to i32 for performance / code size reason.
14009 if (LHS.getValueType() == MVT::i8 ||
14010 LHS.getValueType() == MVT::i16)
14011 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
14013 // If the operand types disagree, extend the shift amount to match. Since
14014 // BT ignores high bits (like shifts) we can use anyextend.
14015 if (LHS.getValueType() != RHS.getValueType())
14016 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
14018 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
14019 X86::CondCode Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
14020 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14021 DAG.getConstant(Cond, dl, MVT::i8), BT);
14027 /// \brief - Turns an ISD::CondCode into a value suitable for SSE floating point
14029 static int translateX86FSETCC(ISD::CondCode SetCCOpcode, SDValue &Op0,
14034 // SSE Condition code mapping:
14043 switch (SetCCOpcode) {
14044 default: llvm_unreachable("Unexpected SETCC condition");
14046 case ISD::SETEQ: SSECC = 0; break;
14048 case ISD::SETGT: Swap = true; // Fallthrough
14050 case ISD::SETOLT: SSECC = 1; break;
14052 case ISD::SETGE: Swap = true; // Fallthrough
14054 case ISD::SETOLE: SSECC = 2; break;
14055 case ISD::SETUO: SSECC = 3; break;
14057 case ISD::SETNE: SSECC = 4; break;
14058 case ISD::SETULE: Swap = true; // Fallthrough
14059 case ISD::SETUGE: SSECC = 5; break;
14060 case ISD::SETULT: Swap = true; // Fallthrough
14061 case ISD::SETUGT: SSECC = 6; break;
14062 case ISD::SETO: SSECC = 7; break;
14064 case ISD::SETONE: SSECC = 8; break;
14067 std::swap(Op0, Op1);
14072 // Lower256IntVSETCC - Break a VSETCC 256-bit integer VSETCC into two new 128
14073 // ones, and then concatenate the result back.
14074 static SDValue Lower256IntVSETCC(SDValue Op, SelectionDAG &DAG) {
14075 MVT VT = Op.getSimpleValueType();
14077 assert(VT.is256BitVector() && Op.getOpcode() == ISD::SETCC &&
14078 "Unsupported value type for operation");
14080 unsigned NumElems = VT.getVectorNumElements();
14082 SDValue CC = Op.getOperand(2);
14084 // Extract the LHS vectors
14085 SDValue LHS = Op.getOperand(0);
14086 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
14087 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
14089 // Extract the RHS vectors
14090 SDValue RHS = Op.getOperand(1);
14091 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
14092 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
14094 // Issue the operation on the smaller types and concatenate the result back
14095 MVT EltVT = VT.getVectorElementType();
14096 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
14097 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
14098 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC),
14099 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
14102 static SDValue LowerBoolVSETCC_AVX512(SDValue Op, SelectionDAG &DAG) {
14103 SDValue Op0 = Op.getOperand(0);
14104 SDValue Op1 = Op.getOperand(1);
14105 SDValue CC = Op.getOperand(2);
14106 MVT VT = Op.getSimpleValueType();
14109 assert(Op0.getSimpleValueType().getVectorElementType() == MVT::i1 &&
14110 "Unexpected type for boolean compare operation");
14111 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14112 SDValue NotOp0 = DAG.getNode(ISD::XOR, dl, VT, Op0,
14113 DAG.getConstant(-1, dl, VT));
14114 SDValue NotOp1 = DAG.getNode(ISD::XOR, dl, VT, Op1,
14115 DAG.getConstant(-1, dl, VT));
14116 switch (SetCCOpcode) {
14117 default: llvm_unreachable("Unexpected SETCC condition");
14119 // (x == y) -> ~(x ^ y)
14120 return DAG.getNode(ISD::XOR, dl, VT,
14121 DAG.getNode(ISD::XOR, dl, VT, Op0, Op1),
14122 DAG.getConstant(-1, dl, VT));
14124 // (x != y) -> (x ^ y)
14125 return DAG.getNode(ISD::XOR, dl, VT, Op0, Op1);
14128 // (x > y) -> (x & ~y)
14129 return DAG.getNode(ISD::AND, dl, VT, Op0, NotOp1);
14132 // (x < y) -> (~x & y)
14133 return DAG.getNode(ISD::AND, dl, VT, NotOp0, Op1);
14136 // (x <= y) -> (~x | y)
14137 return DAG.getNode(ISD::OR, dl, VT, NotOp0, Op1);
14140 // (x >=y) -> (x | ~y)
14141 return DAG.getNode(ISD::OR, dl, VT, Op0, NotOp1);
14145 static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG,
14146 const X86Subtarget *Subtarget) {
14147 SDValue Op0 = Op.getOperand(0);
14148 SDValue Op1 = Op.getOperand(1);
14149 SDValue CC = Op.getOperand(2);
14150 MVT VT = Op.getSimpleValueType();
14153 assert(Op0.getSimpleValueType().getVectorElementType().getSizeInBits() >= 8 &&
14154 Op.getSimpleValueType().getVectorElementType() == MVT::i1 &&
14155 "Cannot set masked compare for this operation");
14157 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14159 bool Unsigned = false;
14162 switch (SetCCOpcode) {
14163 default: llvm_unreachable("Unexpected SETCC condition");
14164 case ISD::SETNE: SSECC = 4; break;
14165 case ISD::SETEQ: Opc = X86ISD::PCMPEQM; break;
14166 case ISD::SETUGT: SSECC = 6; Unsigned = true; break;
14167 case ISD::SETLT: Swap = true; //fall-through
14168 case ISD::SETGT: Opc = X86ISD::PCMPGTM; break;
14169 case ISD::SETULT: SSECC = 1; Unsigned = true; break;
14170 case ISD::SETUGE: SSECC = 5; Unsigned = true; break; //NLT
14171 case ISD::SETGE: Swap = true; SSECC = 2; break; // LE + swap
14172 case ISD::SETULE: Unsigned = true; //fall-through
14173 case ISD::SETLE: SSECC = 2; break;
14177 std::swap(Op0, Op1);
14179 return DAG.getNode(Opc, dl, VT, Op0, Op1);
14180 Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM;
14181 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14182 DAG.getConstant(SSECC, dl, MVT::i8));
14185 /// \brief Try to turn a VSETULT into a VSETULE by modifying its second
14186 /// operand \p Op1. If non-trivial (for example because it's not constant)
14187 /// return an empty value.
14188 static SDValue ChangeVSETULTtoVSETULE(SDLoc dl, SDValue Op1, SelectionDAG &DAG)
14190 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Op1.getNode());
14194 MVT VT = Op1.getSimpleValueType();
14195 MVT EVT = VT.getVectorElementType();
14196 unsigned n = VT.getVectorNumElements();
14197 SmallVector<SDValue, 8> ULTOp1;
14199 for (unsigned i = 0; i < n; ++i) {
14200 ConstantSDNode *Elt = dyn_cast<ConstantSDNode>(BV->getOperand(i));
14201 if (!Elt || Elt->isOpaque() || Elt->getSimpleValueType(0) != EVT)
14204 // Avoid underflow.
14205 APInt Val = Elt->getAPIntValue();
14209 ULTOp1.push_back(DAG.getConstant(Val - 1, dl, EVT));
14212 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, ULTOp1);
14215 static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
14216 SelectionDAG &DAG) {
14217 SDValue Op0 = Op.getOperand(0);
14218 SDValue Op1 = Op.getOperand(1);
14219 SDValue CC = Op.getOperand(2);
14220 MVT VT = Op.getSimpleValueType();
14221 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14222 bool isFP = Op.getOperand(1).getSimpleValueType().isFloatingPoint();
14227 MVT EltVT = Op0.getSimpleValueType().getVectorElementType();
14228 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
14231 unsigned SSECC = translateX86FSETCC(SetCCOpcode, Op0, Op1);
14232 unsigned Opc = X86ISD::CMPP;
14233 if (Subtarget->hasAVX512() && VT.getVectorElementType() == MVT::i1) {
14234 assert(VT.getVectorNumElements() <= 16);
14235 Opc = X86ISD::CMPM;
14237 // In the two special cases we can't handle, emit two comparisons.
14240 unsigned CombineOpc;
14241 if (SetCCOpcode == ISD::SETUEQ) {
14242 CC0 = 3; CC1 = 0; CombineOpc = ISD::OR;
14244 assert(SetCCOpcode == ISD::SETONE);
14245 CC0 = 7; CC1 = 4; CombineOpc = ISD::AND;
14248 SDValue Cmp0 = DAG.getNode(Opc, dl, VT, Op0, Op1,
14249 DAG.getConstant(CC0, dl, MVT::i8));
14250 SDValue Cmp1 = DAG.getNode(Opc, dl, VT, Op0, Op1,
14251 DAG.getConstant(CC1, dl, MVT::i8));
14252 return DAG.getNode(CombineOpc, dl, VT, Cmp0, Cmp1);
14254 // Handle all other FP comparisons here.
14255 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14256 DAG.getConstant(SSECC, dl, MVT::i8));
14259 MVT VTOp0 = Op0.getSimpleValueType();
14260 assert(VTOp0 == Op1.getSimpleValueType() &&
14261 "Expected operands with same type!");
14262 assert(VT.getVectorNumElements() == VTOp0.getVectorNumElements() &&
14263 "Invalid number of packed elements for source and destination!");
14265 if (VT.is128BitVector() && VTOp0.is256BitVector()) {
14266 // On non-AVX512 targets, a vector of MVT::i1 is promoted by the type
14267 // legalizer to a wider vector type. In the case of 'vsetcc' nodes, the
14268 // legalizer firstly checks if the first operand in input to the setcc has
14269 // a legal type. If so, then it promotes the return type to that same type.
14270 // Otherwise, the return type is promoted to the 'next legal type' which,
14271 // for a vector of MVT::i1 is always a 128-bit integer vector type.
14273 // We reach this code only if the following two conditions are met:
14274 // 1. Both return type and operand type have been promoted to wider types
14275 // by the type legalizer.
14276 // 2. The original operand type has been promoted to a 256-bit vector.
14278 // Note that condition 2. only applies for AVX targets.
14279 SDValue NewOp = DAG.getSetCC(dl, VTOp0, Op0, Op1, SetCCOpcode);
14280 return DAG.getZExtOrTrunc(NewOp, dl, VT);
14283 // The non-AVX512 code below works under the assumption that source and
14284 // destination types are the same.
14285 assert((Subtarget->hasAVX512() || (VT == VTOp0)) &&
14286 "Value types for source and destination must be the same!");
14288 // Break 256-bit integer vector compare into smaller ones.
14289 if (VT.is256BitVector() && !Subtarget->hasInt256())
14290 return Lower256IntVSETCC(Op, DAG);
14292 MVT OpVT = Op1.getSimpleValueType();
14293 if (OpVT.getVectorElementType() == MVT::i1)
14294 return LowerBoolVSETCC_AVX512(Op, DAG);
14296 bool MaskResult = (VT.getVectorElementType() == MVT::i1);
14297 if (Subtarget->hasAVX512()) {
14298 if (Op1.getSimpleValueType().is512BitVector() ||
14299 (Subtarget->hasBWI() && Subtarget->hasVLX()) ||
14300 (MaskResult && OpVT.getVectorElementType().getSizeInBits() >= 32))
14301 return LowerIntVSETCC_AVX512(Op, DAG, Subtarget);
14303 // In AVX-512 architecture setcc returns mask with i1 elements,
14304 // But there is no compare instruction for i8 and i16 elements in KNL.
14305 // We are not talking about 512-bit operands in this case, these
14306 // types are illegal.
14308 (OpVT.getVectorElementType().getSizeInBits() < 32 &&
14309 OpVT.getVectorElementType().getSizeInBits() >= 8))
14310 return DAG.getNode(ISD::TRUNCATE, dl, VT,
14311 DAG.getNode(ISD::SETCC, dl, OpVT, Op0, Op1, CC));
14314 // Lower using XOP integer comparisons.
14315 if ((VT == MVT::v16i8 || VT == MVT::v8i16 ||
14316 VT == MVT::v4i32 || VT == MVT::v2i64) && Subtarget->hasXOP()) {
14317 // Translate compare code to XOP PCOM compare mode.
14318 unsigned CmpMode = 0;
14319 switch (SetCCOpcode) {
14320 default: llvm_unreachable("Unexpected SETCC condition");
14322 case ISD::SETLT: CmpMode = 0x00; break;
14324 case ISD::SETLE: CmpMode = 0x01; break;
14326 case ISD::SETGT: CmpMode = 0x02; break;
14328 case ISD::SETGE: CmpMode = 0x03; break;
14329 case ISD::SETEQ: CmpMode = 0x04; break;
14330 case ISD::SETNE: CmpMode = 0x05; break;
14333 // Are we comparing unsigned or signed integers?
14334 unsigned Opc = ISD::isUnsignedIntSetCC(SetCCOpcode)
14335 ? X86ISD::VPCOMU : X86ISD::VPCOM;
14337 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14338 DAG.getConstant(CmpMode, dl, MVT::i8));
14341 // We are handling one of the integer comparisons here. Since SSE only has
14342 // GT and EQ comparisons for integer, swapping operands and multiple
14343 // operations may be required for some comparisons.
14345 bool Swap = false, Invert = false, FlipSigns = false, MinMax = false;
14346 bool Subus = false;
14348 switch (SetCCOpcode) {
14349 default: llvm_unreachable("Unexpected SETCC condition");
14350 case ISD::SETNE: Invert = true;
14351 case ISD::SETEQ: Opc = X86ISD::PCMPEQ; break;
14352 case ISD::SETLT: Swap = true;
14353 case ISD::SETGT: Opc = X86ISD::PCMPGT; break;
14354 case ISD::SETGE: Swap = true;
14355 case ISD::SETLE: Opc = X86ISD::PCMPGT;
14356 Invert = true; break;
14357 case ISD::SETULT: Swap = true;
14358 case ISD::SETUGT: Opc = X86ISD::PCMPGT;
14359 FlipSigns = true; break;
14360 case ISD::SETUGE: Swap = true;
14361 case ISD::SETULE: Opc = X86ISD::PCMPGT;
14362 FlipSigns = true; Invert = true; break;
14365 // Special case: Use min/max operations for SETULE/SETUGE
14366 MVT VET = VT.getVectorElementType();
14368 (Subtarget->hasSSE41() && (VET >= MVT::i8 && VET <= MVT::i32))
14369 || (Subtarget->hasSSE2() && (VET == MVT::i8));
14372 switch (SetCCOpcode) {
14374 case ISD::SETULE: Opc = ISD::UMIN; MinMax = true; break;
14375 case ISD::SETUGE: Opc = ISD::UMAX; MinMax = true; break;
14378 if (MinMax) { Swap = false; Invert = false; FlipSigns = false; }
14381 bool hasSubus = Subtarget->hasSSE2() && (VET == MVT::i8 || VET == MVT::i16);
14382 if (!MinMax && hasSubus) {
14383 // As another special case, use PSUBUS[BW] when it's profitable. E.g. for
14385 // t = psubus Op0, Op1
14386 // pcmpeq t, <0..0>
14387 switch (SetCCOpcode) {
14389 case ISD::SETULT: {
14390 // If the comparison is against a constant we can turn this into a
14391 // setule. With psubus, setule does not require a swap. This is
14392 // beneficial because the constant in the register is no longer
14393 // destructed as the destination so it can be hoisted out of a loop.
14394 // Only do this pre-AVX since vpcmp* is no longer destructive.
14395 if (Subtarget->hasAVX())
14397 SDValue ULEOp1 = ChangeVSETULTtoVSETULE(dl, Op1, DAG);
14398 if (ULEOp1.getNode()) {
14400 Subus = true; Invert = false; Swap = false;
14404 // Psubus is better than flip-sign because it requires no inversion.
14405 case ISD::SETUGE: Subus = true; Invert = false; Swap = true; break;
14406 case ISD::SETULE: Subus = true; Invert = false; Swap = false; break;
14410 Opc = X86ISD::SUBUS;
14416 std::swap(Op0, Op1);
14418 // Check that the operation in question is available (most are plain SSE2,
14419 // but PCMPGTQ and PCMPEQQ have different requirements).
14420 if (VT == MVT::v2i64) {
14421 if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42()) {
14422 assert(Subtarget->hasSSE2() && "Don't know how to lower!");
14424 // First cast everything to the right type.
14425 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
14426 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
14428 // Since SSE has no unsigned integer comparisons, we need to flip the sign
14429 // bits of the inputs before performing those operations. The lower
14430 // compare is always unsigned.
14433 SB = DAG.getConstant(0x80000000U, dl, MVT::v4i32);
14435 SDValue Sign = DAG.getConstant(0x80000000U, dl, MVT::i32);
14436 SDValue Zero = DAG.getConstant(0x00000000U, dl, MVT::i32);
14437 SB = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
14438 Sign, Zero, Sign, Zero);
14440 Op0 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op0, SB);
14441 Op1 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op1, SB);
14443 // Emulate PCMPGTQ with (hi1 > hi2) | ((hi1 == hi2) & (lo1 > lo2))
14444 SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1);
14445 SDValue EQ = DAG.getNode(X86ISD::PCMPEQ, dl, MVT::v4i32, Op0, Op1);
14447 // Create masks for only the low parts/high parts of the 64 bit integers.
14448 static const int MaskHi[] = { 1, 1, 3, 3 };
14449 static const int MaskLo[] = { 0, 0, 2, 2 };
14450 SDValue EQHi = DAG.getVectorShuffle(MVT::v4i32, dl, EQ, EQ, MaskHi);
14451 SDValue GTLo = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskLo);
14452 SDValue GTHi = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi);
14454 SDValue Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, EQHi, GTLo);
14455 Result = DAG.getNode(ISD::OR, dl, MVT::v4i32, Result, GTHi);
14458 Result = DAG.getNOT(dl, Result, MVT::v4i32);
14460 return DAG.getBitcast(VT, Result);
14463 if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) {
14464 // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with
14465 // pcmpeqd + pshufd + pand.
14466 assert(Subtarget->hasSSE2() && !FlipSigns && "Don't know how to lower!");
14468 // First cast everything to the right type.
14469 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
14470 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
14473 SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1);
14475 // Make sure the lower and upper halves are both all-ones.
14476 static const int Mask[] = { 1, 0, 3, 2 };
14477 SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask);
14478 Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf);
14481 Result = DAG.getNOT(dl, Result, MVT::v4i32);
14483 return DAG.getBitcast(VT, Result);
14487 // Since SSE has no unsigned integer comparisons, we need to flip the sign
14488 // bits of the inputs before performing those operations.
14490 MVT EltVT = VT.getVectorElementType();
14491 SDValue SB = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()), dl,
14493 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SB);
14494 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SB);
14497 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
14499 // If the logical-not of the result is required, perform that now.
14501 Result = DAG.getNOT(dl, Result, VT);
14504 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Op0, Result);
14507 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Result,
14508 getZeroVector(VT, Subtarget, DAG, dl));
14513 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
14515 MVT VT = Op.getSimpleValueType();
14517 if (VT.isVector()) return LowerVSETCC(Op, Subtarget, DAG);
14519 assert(((!Subtarget->hasAVX512() && VT == MVT::i8) || (VT == MVT::i1))
14520 && "SetCC type must be 8-bit or 1-bit integer");
14521 SDValue Op0 = Op.getOperand(0);
14522 SDValue Op1 = Op.getOperand(1);
14524 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
14526 // Optimize to BT if possible.
14527 // Lower (X & (1 << N)) == 0 to BT(X, N).
14528 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
14529 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
14530 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
14531 Op1.getOpcode() == ISD::Constant &&
14532 cast<ConstantSDNode>(Op1)->isNullValue() &&
14533 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14534 if (SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG)) {
14536 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewSetCC);
14541 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
14543 if (Op1.getOpcode() == ISD::Constant &&
14544 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
14545 cast<ConstantSDNode>(Op1)->isNullValue()) &&
14546 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14548 // If the input is a setcc, then reuse the input setcc or use a new one with
14549 // the inverted condition.
14550 if (Op0.getOpcode() == X86ISD::SETCC) {
14551 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
14552 bool Invert = (CC == ISD::SETNE) ^
14553 cast<ConstantSDNode>(Op1)->isNullValue();
14557 CCode = X86::GetOppositeBranchCondition(CCode);
14558 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14559 DAG.getConstant(CCode, dl, MVT::i8),
14560 Op0.getOperand(1));
14562 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
14566 if ((Op0.getValueType() == MVT::i1) && (Op1.getOpcode() == ISD::Constant) &&
14567 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1) &&
14568 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14570 ISD::CondCode NewCC = ISD::getSetCCInverse(CC, true);
14571 return DAG.getSetCC(dl, VT, Op0, DAG.getConstant(0, dl, MVT::i1), NewCC);
14574 bool isFP = Op1.getSimpleValueType().isFloatingPoint();
14575 unsigned X86CC = TranslateX86CC(CC, dl, isFP, Op0, Op1, DAG);
14576 if (X86CC == X86::COND_INVALID)
14579 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, dl, DAG);
14580 EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
14581 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14582 DAG.getConstant(X86CC, dl, MVT::i8), EFLAGS);
14584 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
14588 SDValue X86TargetLowering::LowerSETCCE(SDValue Op, SelectionDAG &DAG) const {
14589 SDValue LHS = Op.getOperand(0);
14590 SDValue RHS = Op.getOperand(1);
14591 SDValue Carry = Op.getOperand(2);
14592 SDValue Cond = Op.getOperand(3);
14595 assert(LHS.getSimpleValueType().isInteger() && "SETCCE is integer only.");
14596 X86::CondCode CC = TranslateIntegerX86CC(cast<CondCodeSDNode>(Cond)->get());
14598 assert(Carry.getOpcode() != ISD::CARRY_FALSE);
14599 SDVTList VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
14600 SDValue Cmp = DAG.getNode(X86ISD::SBB, DL, VTs, LHS, RHS, Carry);
14601 return DAG.getNode(X86ISD::SETCC, DL, Op.getValueType(),
14602 DAG.getConstant(CC, DL, MVT::i8), Cmp.getValue(1));
14605 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
14606 static bool isX86LogicalCmp(SDValue Op) {
14607 unsigned Opc = Op.getNode()->getOpcode();
14608 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI ||
14609 Opc == X86ISD::SAHF)
14611 if (Op.getResNo() == 1 &&
14612 (Opc == X86ISD::ADD ||
14613 Opc == X86ISD::SUB ||
14614 Opc == X86ISD::ADC ||
14615 Opc == X86ISD::SBB ||
14616 Opc == X86ISD::SMUL ||
14617 Opc == X86ISD::UMUL ||
14618 Opc == X86ISD::INC ||
14619 Opc == X86ISD::DEC ||
14620 Opc == X86ISD::OR ||
14621 Opc == X86ISD::XOR ||
14622 Opc == X86ISD::AND))
14625 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
14631 static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) {
14632 if (V.getOpcode() != ISD::TRUNCATE)
14635 SDValue VOp0 = V.getOperand(0);
14636 unsigned InBits = VOp0.getValueSizeInBits();
14637 unsigned Bits = V.getValueSizeInBits();
14638 return DAG.MaskedValueIsZero(VOp0, APInt::getHighBitsSet(InBits,InBits-Bits));
14641 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
14642 bool addTest = true;
14643 SDValue Cond = Op.getOperand(0);
14644 SDValue Op1 = Op.getOperand(1);
14645 SDValue Op2 = Op.getOperand(2);
14647 MVT VT = Op1.getSimpleValueType();
14650 // Lower FP selects into a CMP/AND/ANDN/OR sequence when the necessary SSE ops
14651 // are available or VBLENDV if AVX is available.
14652 // Otherwise FP cmovs get lowered into a less efficient branch sequence later.
14653 if (Cond.getOpcode() == ISD::SETCC &&
14654 ((Subtarget->hasSSE2() && (VT == MVT::f32 || VT == MVT::f64)) ||
14655 (Subtarget->hasSSE1() && VT == MVT::f32)) &&
14656 VT == Cond.getOperand(0).getSimpleValueType() && Cond->hasOneUse()) {
14657 SDValue CondOp0 = Cond.getOperand(0), CondOp1 = Cond.getOperand(1);
14658 int SSECC = translateX86FSETCC(
14659 cast<CondCodeSDNode>(Cond.getOperand(2))->get(), CondOp0, CondOp1);
14662 if (Subtarget->hasAVX512()) {
14663 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CondOp0, CondOp1,
14664 DAG.getConstant(SSECC, DL, MVT::i8));
14665 return DAG.getNode(X86ISD::SELECT, DL, VT, Cmp, Op1, Op2);
14668 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, VT, CondOp0, CondOp1,
14669 DAG.getConstant(SSECC, DL, MVT::i8));
14671 // If we have AVX, we can use a variable vector select (VBLENDV) instead
14672 // of 3 logic instructions for size savings and potentially speed.
14673 // Unfortunately, there is no scalar form of VBLENDV.
14675 // If either operand is a constant, don't try this. We can expect to
14676 // optimize away at least one of the logic instructions later in that
14677 // case, so that sequence would be faster than a variable blend.
14679 // BLENDV was introduced with SSE 4.1, but the 2 register form implicitly
14680 // uses XMM0 as the selection register. That may need just as many
14681 // instructions as the AND/ANDN/OR sequence due to register moves, so
14684 if (Subtarget->hasAVX() &&
14685 !isa<ConstantFPSDNode>(Op1) && !isa<ConstantFPSDNode>(Op2)) {
14687 // Convert to vectors, do a VSELECT, and convert back to scalar.
14688 // All of the conversions should be optimized away.
14690 MVT VecVT = VT == MVT::f32 ? MVT::v4f32 : MVT::v2f64;
14691 SDValue VOp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op1);
14692 SDValue VOp2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op2);
14693 SDValue VCmp = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Cmp);
14695 MVT VCmpVT = VT == MVT::f32 ? MVT::v4i32 : MVT::v2i64;
14696 VCmp = DAG.getBitcast(VCmpVT, VCmp);
14698 SDValue VSel = DAG.getNode(ISD::VSELECT, DL, VecVT, VCmp, VOp1, VOp2);
14700 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT,
14701 VSel, DAG.getIntPtrConstant(0, DL));
14703 SDValue AndN = DAG.getNode(X86ISD::FANDN, DL, VT, Cmp, Op2);
14704 SDValue And = DAG.getNode(X86ISD::FAND, DL, VT, Cmp, Op1);
14705 return DAG.getNode(X86ISD::FOR, DL, VT, AndN, And);
14709 if (VT.isVector() && VT.getVectorElementType() == MVT::i1) {
14711 if (ISD::isBuildVectorOfConstantSDNodes(Op1.getNode()))
14712 Op1Scalar = ConvertI1VectorToInteger(Op1, DAG);
14713 else if (Op1.getOpcode() == ISD::BITCAST && Op1.getOperand(0))
14714 Op1Scalar = Op1.getOperand(0);
14716 if (ISD::isBuildVectorOfConstantSDNodes(Op2.getNode()))
14717 Op2Scalar = ConvertI1VectorToInteger(Op2, DAG);
14718 else if (Op2.getOpcode() == ISD::BITCAST && Op2.getOperand(0))
14719 Op2Scalar = Op2.getOperand(0);
14720 if (Op1Scalar.getNode() && Op2Scalar.getNode()) {
14721 SDValue newSelect = DAG.getNode(ISD::SELECT, DL,
14722 Op1Scalar.getValueType(),
14723 Cond, Op1Scalar, Op2Scalar);
14724 if (newSelect.getValueSizeInBits() == VT.getSizeInBits())
14725 return DAG.getBitcast(VT, newSelect);
14726 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, newSelect);
14727 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, ExtVec,
14728 DAG.getIntPtrConstant(0, DL));
14732 if (VT == MVT::v4i1 || VT == MVT::v2i1) {
14733 SDValue zeroConst = DAG.getIntPtrConstant(0, DL);
14734 Op1 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
14735 DAG.getUNDEF(MVT::v8i1), Op1, zeroConst);
14736 Op2 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
14737 DAG.getUNDEF(MVT::v8i1), Op2, zeroConst);
14738 SDValue newSelect = DAG.getNode(ISD::SELECT, DL, MVT::v8i1,
14740 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, newSelect, zeroConst);
14743 if (Cond.getOpcode() == ISD::SETCC) {
14744 SDValue NewCond = LowerSETCC(Cond, DAG);
14745 if (NewCond.getNode())
14749 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
14750 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
14751 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
14752 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
14753 if (Cond.getOpcode() == X86ISD::SETCC &&
14754 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
14755 isZero(Cond.getOperand(1).getOperand(1))) {
14756 SDValue Cmp = Cond.getOperand(1);
14758 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
14760 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
14761 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
14762 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
14764 SDValue CmpOp0 = Cmp.getOperand(0);
14765 // Apply further optimizations for special cases
14766 // (select (x != 0), -1, 0) -> neg & sbb
14767 // (select (x == 0), 0, -1) -> neg & sbb
14768 if (ConstantSDNode *YC = dyn_cast<ConstantSDNode>(Y))
14769 if (YC->isNullValue() &&
14770 (isAllOnes(Op1) == (CondCode == X86::COND_NE))) {
14771 SDVTList VTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32);
14772 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, VTs,
14773 DAG.getConstant(0, DL,
14774 CmpOp0.getValueType()),
14776 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14777 DAG.getConstant(X86::COND_B, DL, MVT::i8),
14778 SDValue(Neg.getNode(), 1));
14782 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
14783 CmpOp0, DAG.getConstant(1, DL, CmpOp0.getValueType()));
14784 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
14786 SDValue Res = // Res = 0 or -1.
14787 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14788 DAG.getConstant(X86::COND_B, DL, MVT::i8), Cmp);
14790 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
14791 Res = DAG.getNOT(DL, Res, Res.getValueType());
14793 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
14794 if (!N2C || !N2C->isNullValue())
14795 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
14800 // Look past (and (setcc_carry (cmp ...)), 1).
14801 if (Cond.getOpcode() == ISD::AND &&
14802 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
14803 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
14804 if (C && C->getAPIntValue() == 1)
14805 Cond = Cond.getOperand(0);
14808 // If condition flag is set by a X86ISD::CMP, then use it as the condition
14809 // setting operand in place of the X86ISD::SETCC.
14810 unsigned CondOpcode = Cond.getOpcode();
14811 if (CondOpcode == X86ISD::SETCC ||
14812 CondOpcode == X86ISD::SETCC_CARRY) {
14813 CC = Cond.getOperand(0);
14815 SDValue Cmp = Cond.getOperand(1);
14816 unsigned Opc = Cmp.getOpcode();
14817 MVT VT = Op.getSimpleValueType();
14819 bool IllegalFPCMov = false;
14820 if (VT.isFloatingPoint() && !VT.isVector() &&
14821 !isScalarFPTypeInSSEReg(VT)) // FPStack?
14822 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
14824 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
14825 Opc == X86ISD::BT) { // FIXME
14829 } else if (CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
14830 CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
14831 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
14832 Cond.getOperand(0).getValueType() != MVT::i8)) {
14833 SDValue LHS = Cond.getOperand(0);
14834 SDValue RHS = Cond.getOperand(1);
14835 unsigned X86Opcode;
14838 switch (CondOpcode) {
14839 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
14840 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
14841 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
14842 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
14843 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
14844 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
14845 default: llvm_unreachable("unexpected overflowing operator");
14847 if (CondOpcode == ISD::UMULO)
14848 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
14851 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
14853 SDValue X86Op = DAG.getNode(X86Opcode, DL, VTs, LHS, RHS);
14855 if (CondOpcode == ISD::UMULO)
14856 Cond = X86Op.getValue(2);
14858 Cond = X86Op.getValue(1);
14860 CC = DAG.getConstant(X86Cond, DL, MVT::i8);
14865 // Look past the truncate if the high bits are known zero.
14866 if (isTruncWithZeroHighBitsInput(Cond, DAG))
14867 Cond = Cond.getOperand(0);
14869 // We know the result of AND is compared against zero. Try to match
14871 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
14872 if (SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG)) {
14873 CC = NewSetCC.getOperand(0);
14874 Cond = NewSetCC.getOperand(1);
14881 CC = DAG.getConstant(X86::COND_NE, DL, MVT::i8);
14882 Cond = EmitTest(Cond, X86::COND_NE, DL, DAG);
14885 // a < b ? -1 : 0 -> RES = ~setcc_carry
14886 // a < b ? 0 : -1 -> RES = setcc_carry
14887 // a >= b ? -1 : 0 -> RES = setcc_carry
14888 // a >= b ? 0 : -1 -> RES = ~setcc_carry
14889 if (Cond.getOpcode() == X86ISD::SUB) {
14890 Cond = ConvertCmpIfNecessary(Cond, DAG);
14891 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
14893 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
14894 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
14895 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14896 DAG.getConstant(X86::COND_B, DL, MVT::i8),
14898 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
14899 return DAG.getNOT(DL, Res, Res.getValueType());
14904 // X86 doesn't have an i8 cmov. If both operands are the result of a truncate
14905 // widen the cmov and push the truncate through. This avoids introducing a new
14906 // branch during isel and doesn't add any extensions.
14907 if (Op.getValueType() == MVT::i8 &&
14908 Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) {
14909 SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0);
14910 if (T1.getValueType() == T2.getValueType() &&
14911 // Blacklist CopyFromReg to avoid partial register stalls.
14912 T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){
14913 SDVTList VTs = DAG.getVTList(T1.getValueType(), MVT::Glue);
14914 SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, VTs, T2, T1, CC, Cond);
14915 return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov);
14919 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
14920 // condition is true.
14921 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
14922 SDValue Ops[] = { Op2, Op1, CC, Cond };
14923 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops);
14926 static SDValue LowerSIGN_EXTEND_AVX512(SDValue Op,
14927 const X86Subtarget *Subtarget,
14928 SelectionDAG &DAG) {
14929 MVT VT = Op->getSimpleValueType(0);
14930 SDValue In = Op->getOperand(0);
14931 MVT InVT = In.getSimpleValueType();
14932 MVT VTElt = VT.getVectorElementType();
14933 MVT InVTElt = InVT.getVectorElementType();
14937 if ((InVTElt == MVT::i1) &&
14938 (((Subtarget->hasBWI() && Subtarget->hasVLX() &&
14939 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() <= 16)) ||
14941 ((Subtarget->hasBWI() && VT.is512BitVector() &&
14942 VTElt.getSizeInBits() <= 16)) ||
14944 ((Subtarget->hasDQI() && Subtarget->hasVLX() &&
14945 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() >= 32)) ||
14947 ((Subtarget->hasDQI() && VT.is512BitVector() &&
14948 VTElt.getSizeInBits() >= 32))))
14949 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14951 unsigned int NumElts = VT.getVectorNumElements();
14953 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
14956 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1) {
14957 if (In.getOpcode() == X86ISD::VSEXT || In.getOpcode() == X86ISD::VZEXT)
14958 return DAG.getNode(In.getOpcode(), dl, VT, In.getOperand(0));
14959 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14962 assert (InVT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
14963 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
14965 DAG.getConstant(APInt::getAllOnesValue(ExtVT.getScalarSizeInBits()), dl,
14968 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), dl, ExtVT);
14970 SDValue V = DAG.getNode(ISD::VSELECT, dl, ExtVT, In, NegOne, Zero);
14971 if (VT.is512BitVector())
14973 return DAG.getNode(X86ISD::VTRUNC, dl, VT, V);
14976 static SDValue LowerSIGN_EXTEND_VECTOR_INREG(SDValue Op,
14977 const X86Subtarget *Subtarget,
14978 SelectionDAG &DAG) {
14979 SDValue In = Op->getOperand(0);
14980 MVT VT = Op->getSimpleValueType(0);
14981 MVT InVT = In.getSimpleValueType();
14982 assert(VT.getSizeInBits() == InVT.getSizeInBits());
14984 MVT InSVT = InVT.getVectorElementType();
14985 assert(VT.getVectorElementType().getSizeInBits() > InSVT.getSizeInBits());
14987 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
14989 if (InSVT != MVT::i32 && InSVT != MVT::i16 && InSVT != MVT::i8)
14994 // SSE41 targets can use the pmovsx* instructions directly.
14995 if (Subtarget->hasSSE41())
14996 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14998 // pre-SSE41 targets unpack lower lanes and then sign-extend using SRAI.
15002 // As SRAI is only available on i16/i32 types, we expand only up to i32
15003 // and handle i64 separately.
15004 while (CurrVT != VT && CurrVT.getVectorElementType() != MVT::i32) {
15005 Curr = DAG.getNode(X86ISD::UNPCKL, dl, CurrVT, DAG.getUNDEF(CurrVT), Curr);
15006 MVT CurrSVT = MVT::getIntegerVT(CurrVT.getScalarSizeInBits() * 2);
15007 CurrVT = MVT::getVectorVT(CurrSVT, CurrVT.getVectorNumElements() / 2);
15008 Curr = DAG.getBitcast(CurrVT, Curr);
15011 SDValue SignExt = Curr;
15012 if (CurrVT != InVT) {
15013 unsigned SignExtShift =
15014 CurrVT.getVectorElementType().getSizeInBits() - InSVT.getSizeInBits();
15015 SignExt = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
15016 DAG.getConstant(SignExtShift, dl, MVT::i8));
15022 if (VT == MVT::v2i64 && CurrVT == MVT::v4i32) {
15023 SDValue Sign = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
15024 DAG.getConstant(31, dl, MVT::i8));
15025 SDValue Ext = DAG.getVectorShuffle(CurrVT, dl, SignExt, Sign, {0, 4, 1, 5});
15026 return DAG.getBitcast(VT, Ext);
15032 static SDValue LowerSIGN_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
15033 SelectionDAG &DAG) {
15034 MVT VT = Op->getSimpleValueType(0);
15035 SDValue In = Op->getOperand(0);
15036 MVT InVT = In.getSimpleValueType();
15039 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
15040 return LowerSIGN_EXTEND_AVX512(Op, Subtarget, DAG);
15042 if ((VT != MVT::v4i64 || InVT != MVT::v4i32) &&
15043 (VT != MVT::v8i32 || InVT != MVT::v8i16) &&
15044 (VT != MVT::v16i16 || InVT != MVT::v16i8))
15047 if (Subtarget->hasInt256())
15048 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
15050 // Optimize vectors in AVX mode
15051 // Sign extend v8i16 to v8i32 and
15054 // Divide input vector into two parts
15055 // for v4i32 the shuffle mask will be { 0, 1, -1, -1} {2, 3, -1, -1}
15056 // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32
15057 // concat the vectors to original VT
15059 unsigned NumElems = InVT.getVectorNumElements();
15060 SDValue Undef = DAG.getUNDEF(InVT);
15062 SmallVector<int,8> ShufMask1(NumElems, -1);
15063 for (unsigned i = 0; i != NumElems/2; ++i)
15066 SDValue OpLo = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask1[0]);
15068 SmallVector<int,8> ShufMask2(NumElems, -1);
15069 for (unsigned i = 0; i != NumElems/2; ++i)
15070 ShufMask2[i] = i + NumElems/2;
15072 SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask2[0]);
15074 MVT HalfVT = MVT::getVectorVT(VT.getVectorElementType(),
15075 VT.getVectorNumElements()/2);
15077 OpLo = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpLo);
15078 OpHi = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpHi);
15080 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
15083 // Lower vector extended loads using a shuffle. If SSSE3 is not available we
15084 // may emit an illegal shuffle but the expansion is still better than scalar
15085 // code. We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise
15086 // we'll emit a shuffle and a arithmetic shift.
15087 // FIXME: Is the expansion actually better than scalar code? It doesn't seem so.
15088 // TODO: It is possible to support ZExt by zeroing the undef values during
15089 // the shuffle phase or after the shuffle.
15090 static SDValue LowerExtendedLoad(SDValue Op, const X86Subtarget *Subtarget,
15091 SelectionDAG &DAG) {
15092 MVT RegVT = Op.getSimpleValueType();
15093 assert(RegVT.isVector() && "We only custom lower vector sext loads.");
15094 assert(RegVT.isInteger() &&
15095 "We only custom lower integer vector sext loads.");
15097 // Nothing useful we can do without SSE2 shuffles.
15098 assert(Subtarget->hasSSE2() && "We only custom lower sext loads with SSE2.");
15100 LoadSDNode *Ld = cast<LoadSDNode>(Op.getNode());
15102 EVT MemVT = Ld->getMemoryVT();
15103 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15104 unsigned RegSz = RegVT.getSizeInBits();
15106 ISD::LoadExtType Ext = Ld->getExtensionType();
15108 assert((Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)
15109 && "Only anyext and sext are currently implemented.");
15110 assert(MemVT != RegVT && "Cannot extend to the same type");
15111 assert(MemVT.isVector() && "Must load a vector from memory");
15113 unsigned NumElems = RegVT.getVectorNumElements();
15114 unsigned MemSz = MemVT.getSizeInBits();
15115 assert(RegSz > MemSz && "Register size must be greater than the mem size");
15117 if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256()) {
15118 // The only way in which we have a legal 256-bit vector result but not the
15119 // integer 256-bit operations needed to directly lower a sextload is if we
15120 // have AVX1 but not AVX2. In that case, we can always emit a sextload to
15121 // a 128-bit vector and a normal sign_extend to 256-bits that should get
15122 // correctly legalized. We do this late to allow the canonical form of
15123 // sextload to persist throughout the rest of the DAG combiner -- it wants
15124 // to fold together any extensions it can, and so will fuse a sign_extend
15125 // of an sextload into a sextload targeting a wider value.
15127 if (MemSz == 128) {
15128 // Just switch this to a normal load.
15129 assert(TLI.isTypeLegal(MemVT) && "If the memory type is a 128-bit type, "
15130 "it must be a legal 128-bit vector "
15132 Load = DAG.getLoad(MemVT, dl, Ld->getChain(), Ld->getBasePtr(),
15133 Ld->getPointerInfo(), Ld->isVolatile(), Ld->isNonTemporal(),
15134 Ld->isInvariant(), Ld->getAlignment());
15136 assert(MemSz < 128 &&
15137 "Can't extend a type wider than 128 bits to a 256 bit vector!");
15138 // Do an sext load to a 128-bit vector type. We want to use the same
15139 // number of elements, but elements half as wide. This will end up being
15140 // recursively lowered by this routine, but will succeed as we definitely
15141 // have all the necessary features if we're using AVX1.
15143 EVT::getIntegerVT(*DAG.getContext(), RegVT.getScalarSizeInBits() / 2);
15144 EVT HalfVecVT = EVT::getVectorVT(*DAG.getContext(), HalfEltVT, NumElems);
15146 DAG.getExtLoad(Ext, dl, HalfVecVT, Ld->getChain(), Ld->getBasePtr(),
15147 Ld->getPointerInfo(), MemVT, Ld->isVolatile(),
15148 Ld->isNonTemporal(), Ld->isInvariant(),
15149 Ld->getAlignment());
15152 // Replace chain users with the new chain.
15153 assert(Load->getNumValues() == 2 && "Loads must carry a chain!");
15154 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Load.getValue(1));
15156 // Finally, do a normal sign-extend to the desired register.
15157 return DAG.getSExtOrTrunc(Load, dl, RegVT);
15160 // All sizes must be a power of two.
15161 assert(isPowerOf2_32(RegSz * MemSz * NumElems) &&
15162 "Non-power-of-two elements are not custom lowered!");
15164 // Attempt to load the original value using scalar loads.
15165 // Find the largest scalar type that divides the total loaded size.
15166 MVT SclrLoadTy = MVT::i8;
15167 for (MVT Tp : MVT::integer_valuetypes()) {
15168 if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
15173 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
15174 if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
15176 SclrLoadTy = MVT::f64;
15178 // Calculate the number of scalar loads that we need to perform
15179 // in order to load our vector from memory.
15180 unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
15182 assert((Ext != ISD::SEXTLOAD || NumLoads == 1) &&
15183 "Can only lower sext loads with a single scalar load!");
15185 unsigned loadRegZize = RegSz;
15186 if (Ext == ISD::SEXTLOAD && RegSz >= 256)
15189 // Represent our vector as a sequence of elements which are the
15190 // largest scalar that we can load.
15191 EVT LoadUnitVecVT = EVT::getVectorVT(
15192 *DAG.getContext(), SclrLoadTy, loadRegZize / SclrLoadTy.getSizeInBits());
15194 // Represent the data using the same element type that is stored in
15195 // memory. In practice, we ''widen'' MemVT.
15197 EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
15198 loadRegZize / MemVT.getScalarSizeInBits());
15200 assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
15201 "Invalid vector type");
15203 // We can't shuffle using an illegal type.
15204 assert(TLI.isTypeLegal(WideVecVT) &&
15205 "We only lower types that form legal widened vector types");
15207 SmallVector<SDValue, 8> Chains;
15208 SDValue Ptr = Ld->getBasePtr();
15209 SDValue Increment = DAG.getConstant(SclrLoadTy.getSizeInBits() / 8, dl,
15210 TLI.getPointerTy(DAG.getDataLayout()));
15211 SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
15213 for (unsigned i = 0; i < NumLoads; ++i) {
15214 // Perform a single load.
15215 SDValue ScalarLoad =
15216 DAG.getLoad(SclrLoadTy, dl, Ld->getChain(), Ptr, Ld->getPointerInfo(),
15217 Ld->isVolatile(), Ld->isNonTemporal(), Ld->isInvariant(),
15218 Ld->getAlignment());
15219 Chains.push_back(ScalarLoad.getValue(1));
15220 // Create the first element type using SCALAR_TO_VECTOR in order to avoid
15221 // another round of DAGCombining.
15223 Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
15225 Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
15226 ScalarLoad, DAG.getIntPtrConstant(i, dl));
15228 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
15231 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
15233 // Bitcast the loaded value to a vector of the original element type, in
15234 // the size of the target vector type.
15235 SDValue SlicedVec = DAG.getBitcast(WideVecVT, Res);
15236 unsigned SizeRatio = RegSz / MemSz;
15238 if (Ext == ISD::SEXTLOAD) {
15239 // If we have SSE4.1, we can directly emit a VSEXT node.
15240 if (Subtarget->hasSSE41()) {
15241 SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
15242 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
15246 // Otherwise we'll use SIGN_EXTEND_VECTOR_INREG to sign extend the lowest
15248 assert(TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND_VECTOR_INREG, RegVT) &&
15249 "We can't implement a sext load without SIGN_EXTEND_VECTOR_INREG!");
15251 SDValue Shuff = DAG.getSignExtendVectorInReg(SlicedVec, dl, RegVT);
15252 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
15256 // Redistribute the loaded elements into the different locations.
15257 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
15258 for (unsigned i = 0; i != NumElems; ++i)
15259 ShuffleVec[i * SizeRatio] = i;
15261 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
15262 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
15264 // Bitcast to the requested type.
15265 Shuff = DAG.getBitcast(RegVT, Shuff);
15266 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
15270 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
15271 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
15272 // from the AND / OR.
15273 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
15274 Opc = Op.getOpcode();
15275 if (Opc != ISD::OR && Opc != ISD::AND)
15277 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
15278 Op.getOperand(0).hasOneUse() &&
15279 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
15280 Op.getOperand(1).hasOneUse());
15283 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
15284 // 1 and that the SETCC node has a single use.
15285 static bool isXor1OfSetCC(SDValue Op) {
15286 if (Op.getOpcode() != ISD::XOR)
15288 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
15289 if (N1C && N1C->getAPIntValue() == 1) {
15290 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
15291 Op.getOperand(0).hasOneUse();
15296 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
15297 bool addTest = true;
15298 SDValue Chain = Op.getOperand(0);
15299 SDValue Cond = Op.getOperand(1);
15300 SDValue Dest = Op.getOperand(2);
15303 bool Inverted = false;
15305 if (Cond.getOpcode() == ISD::SETCC) {
15306 // Check for setcc([su]{add,sub,mul}o == 0).
15307 if (cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETEQ &&
15308 isa<ConstantSDNode>(Cond.getOperand(1)) &&
15309 cast<ConstantSDNode>(Cond.getOperand(1))->isNullValue() &&
15310 Cond.getOperand(0).getResNo() == 1 &&
15311 (Cond.getOperand(0).getOpcode() == ISD::SADDO ||
15312 Cond.getOperand(0).getOpcode() == ISD::UADDO ||
15313 Cond.getOperand(0).getOpcode() == ISD::SSUBO ||
15314 Cond.getOperand(0).getOpcode() == ISD::USUBO ||
15315 Cond.getOperand(0).getOpcode() == ISD::SMULO ||
15316 Cond.getOperand(0).getOpcode() == ISD::UMULO)) {
15318 Cond = Cond.getOperand(0);
15320 SDValue NewCond = LowerSETCC(Cond, DAG);
15321 if (NewCond.getNode())
15326 // FIXME: LowerXALUO doesn't handle these!!
15327 else if (Cond.getOpcode() == X86ISD::ADD ||
15328 Cond.getOpcode() == X86ISD::SUB ||
15329 Cond.getOpcode() == X86ISD::SMUL ||
15330 Cond.getOpcode() == X86ISD::UMUL)
15331 Cond = LowerXALUO(Cond, DAG);
15334 // Look pass (and (setcc_carry (cmp ...)), 1).
15335 if (Cond.getOpcode() == ISD::AND &&
15336 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
15337 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
15338 if (C && C->getAPIntValue() == 1)
15339 Cond = Cond.getOperand(0);
15342 // If condition flag is set by a X86ISD::CMP, then use it as the condition
15343 // setting operand in place of the X86ISD::SETCC.
15344 unsigned CondOpcode = Cond.getOpcode();
15345 if (CondOpcode == X86ISD::SETCC ||
15346 CondOpcode == X86ISD::SETCC_CARRY) {
15347 CC = Cond.getOperand(0);
15349 SDValue Cmp = Cond.getOperand(1);
15350 unsigned Opc = Cmp.getOpcode();
15351 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
15352 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
15356 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
15360 // These can only come from an arithmetic instruction with overflow,
15361 // e.g. SADDO, UADDO.
15362 Cond = Cond.getNode()->getOperand(1);
15368 CondOpcode = Cond.getOpcode();
15369 if (CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
15370 CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
15371 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
15372 Cond.getOperand(0).getValueType() != MVT::i8)) {
15373 SDValue LHS = Cond.getOperand(0);
15374 SDValue RHS = Cond.getOperand(1);
15375 unsigned X86Opcode;
15378 // Keep this in sync with LowerXALUO, otherwise we might create redundant
15379 // instructions that can't be removed afterwards (i.e. X86ISD::ADD and
15381 switch (CondOpcode) {
15382 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
15384 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
15386 X86Opcode = X86ISD::INC; X86Cond = X86::COND_O;
15389 X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
15390 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
15392 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
15394 X86Opcode = X86ISD::DEC; X86Cond = X86::COND_O;
15397 X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
15398 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
15399 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
15400 default: llvm_unreachable("unexpected overflowing operator");
15403 X86Cond = X86::GetOppositeBranchCondition((X86::CondCode)X86Cond);
15404 if (CondOpcode == ISD::UMULO)
15405 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
15408 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
15410 SDValue X86Op = DAG.getNode(X86Opcode, dl, VTs, LHS, RHS);
15412 if (CondOpcode == ISD::UMULO)
15413 Cond = X86Op.getValue(2);
15415 Cond = X86Op.getValue(1);
15417 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
15421 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
15422 SDValue Cmp = Cond.getOperand(0).getOperand(1);
15423 if (CondOpc == ISD::OR) {
15424 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
15425 // two branches instead of an explicit OR instruction with a
15427 if (Cmp == Cond.getOperand(1).getOperand(1) &&
15428 isX86LogicalCmp(Cmp)) {
15429 CC = Cond.getOperand(0).getOperand(0);
15430 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15431 Chain, Dest, CC, Cmp);
15432 CC = Cond.getOperand(1).getOperand(0);
15436 } else { // ISD::AND
15437 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
15438 // two branches instead of an explicit AND instruction with a
15439 // separate test. However, we only do this if this block doesn't
15440 // have a fall-through edge, because this requires an explicit
15441 // jmp when the condition is false.
15442 if (Cmp == Cond.getOperand(1).getOperand(1) &&
15443 isX86LogicalCmp(Cmp) &&
15444 Op.getNode()->hasOneUse()) {
15445 X86::CondCode CCode =
15446 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
15447 CCode = X86::GetOppositeBranchCondition(CCode);
15448 CC = DAG.getConstant(CCode, dl, MVT::i8);
15449 SDNode *User = *Op.getNode()->use_begin();
15450 // Look for an unconditional branch following this conditional branch.
15451 // We need this because we need to reverse the successors in order
15452 // to implement FCMP_OEQ.
15453 if (User->getOpcode() == ISD::BR) {
15454 SDValue FalseBB = User->getOperand(1);
15456 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15457 assert(NewBR == User);
15461 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15462 Chain, Dest, CC, Cmp);
15463 X86::CondCode CCode =
15464 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
15465 CCode = X86::GetOppositeBranchCondition(CCode);
15466 CC = DAG.getConstant(CCode, dl, MVT::i8);
15472 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
15473 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
15474 // It should be transformed during dag combiner except when the condition
15475 // is set by a arithmetics with overflow node.
15476 X86::CondCode CCode =
15477 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
15478 CCode = X86::GetOppositeBranchCondition(CCode);
15479 CC = DAG.getConstant(CCode, dl, MVT::i8);
15480 Cond = Cond.getOperand(0).getOperand(1);
15482 } else if (Cond.getOpcode() == ISD::SETCC &&
15483 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETOEQ) {
15484 // For FCMP_OEQ, we can emit
15485 // two branches instead of an explicit AND instruction with a
15486 // separate test. However, we only do this if this block doesn't
15487 // have a fall-through edge, because this requires an explicit
15488 // jmp when the condition is false.
15489 if (Op.getNode()->hasOneUse()) {
15490 SDNode *User = *Op.getNode()->use_begin();
15491 // Look for an unconditional branch following this conditional branch.
15492 // We need this because we need to reverse the successors in order
15493 // to implement FCMP_OEQ.
15494 if (User->getOpcode() == ISD::BR) {
15495 SDValue FalseBB = User->getOperand(1);
15497 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15498 assert(NewBR == User);
15502 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
15503 Cond.getOperand(0), Cond.getOperand(1));
15504 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
15505 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
15506 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15507 Chain, Dest, CC, Cmp);
15508 CC = DAG.getConstant(X86::COND_P, dl, MVT::i8);
15513 } else if (Cond.getOpcode() == ISD::SETCC &&
15514 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETUNE) {
15515 // For FCMP_UNE, we can emit
15516 // two branches instead of an explicit AND instruction with a
15517 // separate test. However, we only do this if this block doesn't
15518 // have a fall-through edge, because this requires an explicit
15519 // jmp when the condition is false.
15520 if (Op.getNode()->hasOneUse()) {
15521 SDNode *User = *Op.getNode()->use_begin();
15522 // Look for an unconditional branch following this conditional branch.
15523 // We need this because we need to reverse the successors in order
15524 // to implement FCMP_UNE.
15525 if (User->getOpcode() == ISD::BR) {
15526 SDValue FalseBB = User->getOperand(1);
15528 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15529 assert(NewBR == User);
15532 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
15533 Cond.getOperand(0), Cond.getOperand(1));
15534 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
15535 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
15536 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15537 Chain, Dest, CC, Cmp);
15538 CC = DAG.getConstant(X86::COND_NP, dl, MVT::i8);
15548 // Look pass the truncate if the high bits are known zero.
15549 if (isTruncWithZeroHighBitsInput(Cond, DAG))
15550 Cond = Cond.getOperand(0);
15552 // We know the result of AND is compared against zero. Try to match
15554 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
15555 if (SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG)) {
15556 CC = NewSetCC.getOperand(0);
15557 Cond = NewSetCC.getOperand(1);
15564 X86::CondCode X86Cond = Inverted ? X86::COND_E : X86::COND_NE;
15565 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
15566 Cond = EmitTest(Cond, X86Cond, dl, DAG);
15568 Cond = ConvertCmpIfNecessary(Cond, DAG);
15569 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15570 Chain, Dest, CC, Cond);
15573 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
15574 // Calls to _alloca are needed to probe the stack when allocating more than 4k
15575 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
15576 // that the guard pages used by the OS virtual memory manager are allocated in
15577 // correct sequence.
15579 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
15580 SelectionDAG &DAG) const {
15581 MachineFunction &MF = DAG.getMachineFunction();
15582 bool SplitStack = MF.shouldSplitStack();
15583 bool Lower = (Subtarget->isOSWindows() && !Subtarget->isTargetMachO()) ||
15588 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15589 SDNode* Node = Op.getNode();
15591 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
15592 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
15593 " not tell us which reg is the stack pointer!");
15594 EVT VT = Node->getValueType(0);
15595 SDValue Tmp1 = SDValue(Node, 0);
15596 SDValue Tmp2 = SDValue(Node, 1);
15597 SDValue Tmp3 = Node->getOperand(2);
15598 SDValue Chain = Tmp1.getOperand(0);
15600 // Chain the dynamic stack allocation so that it doesn't modify the stack
15601 // pointer when other instructions are using the stack.
15602 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, dl, true),
15605 SDValue Size = Tmp2.getOperand(1);
15606 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
15607 Chain = SP.getValue(1);
15608 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
15609 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
15610 unsigned StackAlign = TFI.getStackAlignment();
15611 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
15612 if (Align > StackAlign)
15613 Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1,
15614 DAG.getConstant(-(uint64_t)Align, dl, VT));
15615 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
15617 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, dl, true),
15618 DAG.getIntPtrConstant(0, dl, true), SDValue(),
15621 SDValue Ops[2] = { Tmp1, Tmp2 };
15622 return DAG.getMergeValues(Ops, dl);
15626 SDValue Chain = Op.getOperand(0);
15627 SDValue Size = Op.getOperand(1);
15628 unsigned Align = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
15629 EVT VT = Op.getNode()->getValueType(0);
15631 bool Is64Bit = Subtarget->is64Bit();
15632 MVT SPTy = getPointerTy(DAG.getDataLayout());
15635 MachineRegisterInfo &MRI = MF.getRegInfo();
15638 // The 64 bit implementation of segmented stacks needs to clobber both r10
15639 // r11. This makes it impossible to use it along with nested parameters.
15640 const Function *F = MF.getFunction();
15642 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
15644 if (I->hasNestAttr())
15645 report_fatal_error("Cannot use segmented stacks with functions that "
15646 "have nested arguments.");
15649 const TargetRegisterClass *AddrRegClass = getRegClassFor(SPTy);
15650 unsigned Vreg = MRI.createVirtualRegister(AddrRegClass);
15651 Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size);
15652 SDValue Value = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain,
15653 DAG.getRegister(Vreg, SPTy));
15654 SDValue Ops1[2] = { Value, Chain };
15655 return DAG.getMergeValues(Ops1, dl);
15658 const unsigned Reg = (Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX);
15660 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
15661 Flag = Chain.getValue(1);
15662 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
15664 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
15666 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15667 unsigned SPReg = RegInfo->getStackRegister();
15668 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, SPTy);
15669 Chain = SP.getValue(1);
15672 SP = DAG.getNode(ISD::AND, dl, VT, SP.getValue(0),
15673 DAG.getConstant(-(uint64_t)Align, dl, VT));
15674 Chain = DAG.getCopyToReg(Chain, dl, SPReg, SP);
15677 SDValue Ops1[2] = { SP, Chain };
15678 return DAG.getMergeValues(Ops1, dl);
15682 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
15683 MachineFunction &MF = DAG.getMachineFunction();
15684 auto PtrVT = getPointerTy(MF.getDataLayout());
15685 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
15687 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
15690 if (!Subtarget->is64Bit() ||
15691 Subtarget->isCallingConvWin64(MF.getFunction()->getCallingConv())) {
15692 // vastart just stores the address of the VarArgsFrameIndex slot into the
15693 // memory location argument.
15694 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
15695 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
15696 MachinePointerInfo(SV), false, false, 0);
15700 // gp_offset (0 - 6 * 8)
15701 // fp_offset (48 - 48 + 8 * 16)
15702 // overflow_arg_area (point to parameters coming in memory).
15704 SmallVector<SDValue, 8> MemOps;
15705 SDValue FIN = Op.getOperand(1);
15707 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
15708 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
15710 FIN, MachinePointerInfo(SV), false, false, 0);
15711 MemOps.push_back(Store);
15714 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL));
15715 Store = DAG.getStore(Op.getOperand(0), DL,
15716 DAG.getConstant(FuncInfo->getVarArgsFPOffset(), DL,
15718 FIN, MachinePointerInfo(SV, 4), false, false, 0);
15719 MemOps.push_back(Store);
15721 // Store ptr to overflow_arg_area
15722 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL));
15723 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
15724 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
15725 MachinePointerInfo(SV, 8),
15727 MemOps.push_back(Store);
15729 // Store ptr to reg_save_area.
15730 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(
15731 Subtarget->isTarget64BitLP64() ? 8 : 4, DL));
15732 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(), PtrVT);
15733 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN, MachinePointerInfo(
15734 SV, Subtarget->isTarget64BitLP64() ? 16 : 12), false, false, 0);
15735 MemOps.push_back(Store);
15736 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
15739 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
15740 assert(Subtarget->is64Bit() &&
15741 "LowerVAARG only handles 64-bit va_arg!");
15742 assert(Op.getNode()->getNumOperands() == 4);
15744 MachineFunction &MF = DAG.getMachineFunction();
15745 if (Subtarget->isCallingConvWin64(MF.getFunction()->getCallingConv()))
15746 // The Win64 ABI uses char* instead of a structure.
15747 return DAG.expandVAArg(Op.getNode());
15749 SDValue Chain = Op.getOperand(0);
15750 SDValue SrcPtr = Op.getOperand(1);
15751 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
15752 unsigned Align = Op.getConstantOperandVal(3);
15755 EVT ArgVT = Op.getNode()->getValueType(0);
15756 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
15757 uint32_t ArgSize = DAG.getDataLayout().getTypeAllocSize(ArgTy);
15760 // Decide which area this value should be read from.
15761 // TODO: Implement the AMD64 ABI in its entirety. This simple
15762 // selection mechanism works only for the basic types.
15763 if (ArgVT == MVT::f80) {
15764 llvm_unreachable("va_arg for f80 not yet implemented");
15765 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
15766 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
15767 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
15768 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
15770 llvm_unreachable("Unhandled argument type in LowerVAARG");
15773 if (ArgMode == 2) {
15774 // Sanity Check: Make sure using fp_offset makes sense.
15775 assert(!Subtarget->useSoftFloat() &&
15776 !(MF.getFunction()->hasFnAttribute(Attribute::NoImplicitFloat)) &&
15777 Subtarget->hasSSE1());
15780 // Insert VAARG_64 node into the DAG
15781 // VAARG_64 returns two values: Variable Argument Address, Chain
15782 SDValue InstOps[] = {Chain, SrcPtr, DAG.getConstant(ArgSize, dl, MVT::i32),
15783 DAG.getConstant(ArgMode, dl, MVT::i8),
15784 DAG.getConstant(Align, dl, MVT::i32)};
15785 SDVTList VTs = DAG.getVTList(getPointerTy(DAG.getDataLayout()), MVT::Other);
15786 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
15787 VTs, InstOps, MVT::i64,
15788 MachinePointerInfo(SV),
15790 /*Volatile=*/false,
15792 /*WriteMem=*/true);
15793 Chain = VAARG.getValue(1);
15795 // Load the next argument and return it
15796 return DAG.getLoad(ArgVT, dl,
15799 MachinePointerInfo(),
15800 false, false, false, 0);
15803 static SDValue LowerVACOPY(SDValue Op, const X86Subtarget *Subtarget,
15804 SelectionDAG &DAG) {
15805 // X86-64 va_list is a struct { i32, i32, i8*, i8* }, except on Windows,
15806 // where a va_list is still an i8*.
15807 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
15808 if (Subtarget->isCallingConvWin64(
15809 DAG.getMachineFunction().getFunction()->getCallingConv()))
15810 // Probably a Win64 va_copy.
15811 return DAG.expandVACopy(Op.getNode());
15813 SDValue Chain = Op.getOperand(0);
15814 SDValue DstPtr = Op.getOperand(1);
15815 SDValue SrcPtr = Op.getOperand(2);
15816 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
15817 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
15820 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
15821 DAG.getIntPtrConstant(24, DL), 8, /*isVolatile*/false,
15823 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
15826 // getTargetVShiftByConstNode - Handle vector element shifts where the shift
15827 // amount is a constant. Takes immediate version of shift as input.
15828 static SDValue getTargetVShiftByConstNode(unsigned Opc, SDLoc dl, MVT VT,
15829 SDValue SrcOp, uint64_t ShiftAmt,
15830 SelectionDAG &DAG) {
15831 MVT ElementType = VT.getVectorElementType();
15833 // Fold this packed shift into its first operand if ShiftAmt is 0.
15837 // Check for ShiftAmt >= element width
15838 if (ShiftAmt >= ElementType.getSizeInBits()) {
15839 if (Opc == X86ISD::VSRAI)
15840 ShiftAmt = ElementType.getSizeInBits() - 1;
15842 return DAG.getConstant(0, dl, VT);
15845 assert((Opc == X86ISD::VSHLI || Opc == X86ISD::VSRLI || Opc == X86ISD::VSRAI)
15846 && "Unknown target vector shift-by-constant node");
15848 // Fold this packed vector shift into a build vector if SrcOp is a
15849 // vector of Constants or UNDEFs, and SrcOp valuetype is the same as VT.
15850 if (VT == SrcOp.getSimpleValueType() &&
15851 ISD::isBuildVectorOfConstantSDNodes(SrcOp.getNode())) {
15852 SmallVector<SDValue, 8> Elts;
15853 unsigned NumElts = SrcOp->getNumOperands();
15854 ConstantSDNode *ND;
15857 default: llvm_unreachable(nullptr);
15858 case X86ISD::VSHLI:
15859 for (unsigned i=0; i!=NumElts; ++i) {
15860 SDValue CurrentOp = SrcOp->getOperand(i);
15861 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15862 Elts.push_back(CurrentOp);
15865 ND = cast<ConstantSDNode>(CurrentOp);
15866 const APInt &C = ND->getAPIntValue();
15867 Elts.push_back(DAG.getConstant(C.shl(ShiftAmt), dl, ElementType));
15870 case X86ISD::VSRLI:
15871 for (unsigned i=0; i!=NumElts; ++i) {
15872 SDValue CurrentOp = SrcOp->getOperand(i);
15873 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15874 Elts.push_back(CurrentOp);
15877 ND = cast<ConstantSDNode>(CurrentOp);
15878 const APInt &C = ND->getAPIntValue();
15879 Elts.push_back(DAG.getConstant(C.lshr(ShiftAmt), dl, ElementType));
15882 case X86ISD::VSRAI:
15883 for (unsigned i=0; i!=NumElts; ++i) {
15884 SDValue CurrentOp = SrcOp->getOperand(i);
15885 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15886 Elts.push_back(CurrentOp);
15889 ND = cast<ConstantSDNode>(CurrentOp);
15890 const APInt &C = ND->getAPIntValue();
15891 Elts.push_back(DAG.getConstant(C.ashr(ShiftAmt), dl, ElementType));
15896 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
15899 return DAG.getNode(Opc, dl, VT, SrcOp,
15900 DAG.getConstant(ShiftAmt, dl, MVT::i8));
15903 // getTargetVShiftNode - Handle vector element shifts where the shift amount
15904 // may or may not be a constant. Takes immediate version of shift as input.
15905 static SDValue getTargetVShiftNode(unsigned Opc, SDLoc dl, MVT VT,
15906 SDValue SrcOp, SDValue ShAmt,
15907 SelectionDAG &DAG) {
15908 MVT SVT = ShAmt.getSimpleValueType();
15909 assert((SVT == MVT::i32 || SVT == MVT::i64) && "Unexpected value type!");
15911 // Catch shift-by-constant.
15912 if (ConstantSDNode *CShAmt = dyn_cast<ConstantSDNode>(ShAmt))
15913 return getTargetVShiftByConstNode(Opc, dl, VT, SrcOp,
15914 CShAmt->getZExtValue(), DAG);
15916 // Change opcode to non-immediate version
15918 default: llvm_unreachable("Unknown target vector shift node");
15919 case X86ISD::VSHLI: Opc = X86ISD::VSHL; break;
15920 case X86ISD::VSRLI: Opc = X86ISD::VSRL; break;
15921 case X86ISD::VSRAI: Opc = X86ISD::VSRA; break;
15924 const X86Subtarget &Subtarget =
15925 static_cast<const X86Subtarget &>(DAG.getSubtarget());
15926 if (Subtarget.hasSSE41() && ShAmt.getOpcode() == ISD::ZERO_EXTEND &&
15927 ShAmt.getOperand(0).getSimpleValueType() == MVT::i16) {
15928 // Let the shuffle legalizer expand this shift amount node.
15929 SDValue Op0 = ShAmt.getOperand(0);
15930 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(Op0), MVT::v8i16, Op0);
15931 ShAmt = getShuffleVectorZeroOrUndef(Op0, 0, true, &Subtarget, DAG);
15933 // Need to build a vector containing shift amount.
15934 // SSE/AVX packed shifts only use the lower 64-bit of the shift count.
15935 SmallVector<SDValue, 4> ShOps;
15936 ShOps.push_back(ShAmt);
15937 if (SVT == MVT::i32) {
15938 ShOps.push_back(DAG.getConstant(0, dl, SVT));
15939 ShOps.push_back(DAG.getUNDEF(SVT));
15941 ShOps.push_back(DAG.getUNDEF(SVT));
15943 MVT BVT = SVT == MVT::i32 ? MVT::v4i32 : MVT::v2i64;
15944 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, BVT, ShOps);
15947 // The return type has to be a 128-bit type with the same element
15948 // type as the input type.
15949 MVT EltVT = VT.getVectorElementType();
15950 MVT ShVT = MVT::getVectorVT(EltVT, 128/EltVT.getSizeInBits());
15952 ShAmt = DAG.getBitcast(ShVT, ShAmt);
15953 return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt);
15956 /// \brief Return (and \p Op, \p Mask) for compare instructions or
15957 /// (vselect \p Mask, \p Op, \p PreservedSrc) for others along with the
15958 /// necessary casting or extending for \p Mask when lowering masking intrinsics
15959 static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask,
15960 SDValue PreservedSrc,
15961 const X86Subtarget *Subtarget,
15962 SelectionDAG &DAG) {
15963 MVT VT = Op.getSimpleValueType();
15964 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
15966 unsigned OpcodeSelect = ISD::VSELECT;
15969 if (isAllOnes(Mask))
15972 if (MaskVT.bitsGT(Mask.getSimpleValueType())) {
15973 MVT newMaskVT = MVT::getIntegerVT(MaskVT.getSizeInBits());
15974 VMask = DAG.getBitcast(MaskVT,
15975 DAG.getNode(ISD::ANY_EXTEND, dl, newMaskVT, Mask));
15977 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
15978 Mask.getSimpleValueType().getSizeInBits());
15979 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
15980 // are extracted by EXTRACT_SUBVECTOR.
15981 VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15982 DAG.getBitcast(BitcastVT, Mask),
15983 DAG.getIntPtrConstant(0, dl));
15986 switch (Op.getOpcode()) {
15988 case X86ISD::PCMPEQM:
15989 case X86ISD::PCMPGTM:
15991 case X86ISD::CMPMU:
15992 return DAG.getNode(ISD::AND, dl, VT, Op, VMask);
15993 case X86ISD::VFPCLASS:
15994 return DAG.getNode(ISD::OR, dl, VT, Op, VMask);
15995 case X86ISD::VTRUNC:
15996 case X86ISD::VTRUNCS:
15997 case X86ISD::VTRUNCUS:
15998 // We can't use ISD::VSELECT here because it is not always "Legal"
15999 // for the destination type. For example vpmovqb require only AVX512
16000 // and vselect that can operate on byte element type require BWI
16001 OpcodeSelect = X86ISD::SELECT;
16004 if (PreservedSrc.getOpcode() == ISD::UNDEF)
16005 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
16006 return DAG.getNode(OpcodeSelect, dl, VT, VMask, Op, PreservedSrc);
16009 /// \brief Creates an SDNode for a predicated scalar operation.
16010 /// \returns (X86vselect \p Mask, \p Op, \p PreservedSrc).
16011 /// The mask is coming as MVT::i8 and it should be truncated
16012 /// to MVT::i1 while lowering masking intrinsics.
16013 /// The main difference between ScalarMaskingNode and VectorMaskingNode is using
16014 /// "X86select" instead of "vselect". We just can't create the "vselect" node
16015 /// for a scalar instruction.
16016 static SDValue getScalarMaskingNode(SDValue Op, SDValue Mask,
16017 SDValue PreservedSrc,
16018 const X86Subtarget *Subtarget,
16019 SelectionDAG &DAG) {
16020 if (isAllOnes(Mask))
16023 MVT VT = Op.getSimpleValueType();
16025 // The mask should be of type MVT::i1
16026 SDValue IMask = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Mask);
16028 if (Op.getOpcode() == X86ISD::FSETCC)
16029 return DAG.getNode(ISD::AND, dl, VT, Op, IMask);
16030 if (Op.getOpcode() == X86ISD::VFPCLASS)
16031 return DAG.getNode(ISD::OR, dl, VT, Op, IMask);
16033 if (PreservedSrc.getOpcode() == ISD::UNDEF)
16034 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
16035 return DAG.getNode(X86ISD::SELECT, dl, VT, IMask, Op, PreservedSrc);
16038 static int getSEHRegistrationNodeSize(const Function *Fn) {
16039 if (!Fn->hasPersonalityFn())
16040 report_fatal_error(
16041 "querying registration node size for function without personality");
16042 // The RegNodeSize is 6 32-bit words for SEH and 4 for C++ EH. See
16043 // WinEHStatePass for the full struct definition.
16044 switch (classifyEHPersonality(Fn->getPersonalityFn())) {
16045 case EHPersonality::MSVC_X86SEH: return 24;
16046 case EHPersonality::MSVC_CXX: return 16;
16049 report_fatal_error("can only recover FP for MSVC EH personality functions");
16052 /// When the 32-bit MSVC runtime transfers control to us, either to an outlined
16053 /// function or when returning to a parent frame after catching an exception, we
16054 /// recover the parent frame pointer by doing arithmetic on the incoming EBP.
16055 /// Here's the math:
16056 /// RegNodeBase = EntryEBP - RegNodeSize
16057 /// ParentFP = RegNodeBase - RegNodeFrameOffset
16058 /// Subtracting RegNodeSize takes us to the offset of the registration node, and
16059 /// subtracting the offset (negative on x86) takes us back to the parent FP.
16060 static SDValue recoverFramePointer(SelectionDAG &DAG, const Function *Fn,
16061 SDValue EntryEBP) {
16062 MachineFunction &MF = DAG.getMachineFunction();
16065 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16066 MVT PtrVT = TLI.getPointerTy(DAG.getDataLayout());
16068 // It's possible that the parent function no longer has a personality function
16069 // if the exceptional code was optimized away, in which case we just return
16070 // the incoming EBP.
16071 if (!Fn->hasPersonalityFn())
16074 int RegNodeSize = getSEHRegistrationNodeSize(Fn);
16076 // Get an MCSymbol that will ultimately resolve to the frame offset of the EH
16078 MCSymbol *OffsetSym =
16079 MF.getMMI().getContext().getOrCreateParentFrameOffsetSymbol(
16080 GlobalValue::getRealLinkageName(Fn->getName()));
16081 SDValue OffsetSymVal = DAG.getMCSymbol(OffsetSym, PtrVT);
16082 SDValue RegNodeFrameOffset =
16083 DAG.getNode(ISD::LOCAL_RECOVER, dl, PtrVT, OffsetSymVal);
16085 // RegNodeBase = EntryEBP - RegNodeSize
16086 // ParentFP = RegNodeBase - RegNodeFrameOffset
16087 SDValue RegNodeBase = DAG.getNode(ISD::SUB, dl, PtrVT, EntryEBP,
16088 DAG.getConstant(RegNodeSize, dl, PtrVT));
16089 return DAG.getNode(ISD::SUB, dl, PtrVT, RegNodeBase, RegNodeFrameOffset);
16092 static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
16093 SelectionDAG &DAG) {
16095 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
16096 MVT VT = Op.getSimpleValueType();
16097 const IntrinsicData* IntrData = getIntrinsicWithoutChain(IntNo);
16099 switch(IntrData->Type) {
16100 case INTR_TYPE_1OP:
16101 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1));
16102 case INTR_TYPE_2OP:
16103 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16105 case INTR_TYPE_2OP_IMM8:
16106 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16107 DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(2)));
16108 case INTR_TYPE_3OP:
16109 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16110 Op.getOperand(2), Op.getOperand(3));
16111 case INTR_TYPE_4OP:
16112 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16113 Op.getOperand(2), Op.getOperand(3), Op.getOperand(4));
16114 case INTR_TYPE_1OP_MASK_RM: {
16115 SDValue Src = Op.getOperand(1);
16116 SDValue PassThru = Op.getOperand(2);
16117 SDValue Mask = Op.getOperand(3);
16118 SDValue RoundingMode;
16119 // We allways add rounding mode to the Node.
16120 // If the rounding mode is not specified, we add the
16121 // "current direction" mode.
16122 if (Op.getNumOperands() == 4)
16124 DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
16126 RoundingMode = Op.getOperand(4);
16127 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16128 if (IntrWithRoundingModeOpcode != 0)
16129 if (cast<ConstantSDNode>(RoundingMode)->getZExtValue() !=
16130 X86::STATIC_ROUNDING::CUR_DIRECTION)
16131 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16132 dl, Op.getValueType(), Src, RoundingMode),
16133 Mask, PassThru, Subtarget, DAG);
16134 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src,
16136 Mask, PassThru, Subtarget, DAG);
16138 case INTR_TYPE_1OP_MASK: {
16139 SDValue Src = Op.getOperand(1);
16140 SDValue PassThru = Op.getOperand(2);
16141 SDValue Mask = Op.getOperand(3);
16142 // We add rounding mode to the Node when
16143 // - RM Opcode is specified and
16144 // - RM is not "current direction".
16145 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16146 if (IntrWithRoundingModeOpcode != 0) {
16147 SDValue Rnd = Op.getOperand(4);
16148 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
16149 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
16150 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16151 dl, Op.getValueType(),
16153 Mask, PassThru, Subtarget, DAG);
16156 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src),
16157 Mask, PassThru, Subtarget, DAG);
16159 case INTR_TYPE_SCALAR_MASK: {
16160 SDValue Src1 = Op.getOperand(1);
16161 SDValue Src2 = Op.getOperand(2);
16162 SDValue passThru = Op.getOperand(3);
16163 SDValue Mask = Op.getOperand(4);
16164 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2),
16165 Mask, passThru, Subtarget, DAG);
16167 case INTR_TYPE_SCALAR_MASK_RM: {
16168 SDValue Src1 = Op.getOperand(1);
16169 SDValue Src2 = Op.getOperand(2);
16170 SDValue Src0 = Op.getOperand(3);
16171 SDValue Mask = Op.getOperand(4);
16172 // There are 2 kinds of intrinsics in this group:
16173 // (1) With suppress-all-exceptions (sae) or rounding mode- 6 operands
16174 // (2) With rounding mode and sae - 7 operands.
16175 if (Op.getNumOperands() == 6) {
16176 SDValue Sae = Op.getOperand(5);
16177 unsigned Opc = IntrData->Opc1 ? IntrData->Opc1 : IntrData->Opc0;
16178 return getScalarMaskingNode(DAG.getNode(Opc, dl, VT, Src1, Src2,
16180 Mask, Src0, Subtarget, DAG);
16182 assert(Op.getNumOperands() == 7 && "Unexpected intrinsic form");
16183 SDValue RoundingMode = Op.getOperand(5);
16184 SDValue Sae = Op.getOperand(6);
16185 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2,
16186 RoundingMode, Sae),
16187 Mask, Src0, Subtarget, DAG);
16189 case INTR_TYPE_2OP_MASK:
16190 case INTR_TYPE_2OP_IMM8_MASK: {
16191 SDValue Src1 = Op.getOperand(1);
16192 SDValue Src2 = Op.getOperand(2);
16193 SDValue PassThru = Op.getOperand(3);
16194 SDValue Mask = Op.getOperand(4);
16196 if (IntrData->Type == INTR_TYPE_2OP_IMM8_MASK)
16197 Src2 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Src2);
16199 // We specify 2 possible opcodes for intrinsics with rounding modes.
16200 // First, we check if the intrinsic may have non-default rounding mode,
16201 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16202 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16203 if (IntrWithRoundingModeOpcode != 0) {
16204 SDValue Rnd = Op.getOperand(5);
16205 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
16206 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
16207 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16208 dl, Op.getValueType(),
16210 Mask, PassThru, Subtarget, DAG);
16213 // TODO: Intrinsics should have fast-math-flags to propagate.
16214 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,Src1,Src2),
16215 Mask, PassThru, Subtarget, DAG);
16217 case INTR_TYPE_2OP_MASK_RM: {
16218 SDValue Src1 = Op.getOperand(1);
16219 SDValue Src2 = Op.getOperand(2);
16220 SDValue PassThru = Op.getOperand(3);
16221 SDValue Mask = Op.getOperand(4);
16222 // We specify 2 possible modes for intrinsics, with/without rounding
16224 // First, we check if the intrinsic have rounding mode (6 operands),
16225 // if not, we set rounding mode to "current".
16227 if (Op.getNumOperands() == 6)
16228 Rnd = Op.getOperand(5);
16230 Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
16231 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16233 Mask, PassThru, Subtarget, DAG);
16235 case INTR_TYPE_3OP_SCALAR_MASK_RM: {
16236 SDValue Src1 = Op.getOperand(1);
16237 SDValue Src2 = Op.getOperand(2);
16238 SDValue Src3 = Op.getOperand(3);
16239 SDValue PassThru = Op.getOperand(4);
16240 SDValue Mask = Op.getOperand(5);
16241 SDValue Sae = Op.getOperand(6);
16243 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1,
16245 Mask, PassThru, Subtarget, DAG);
16247 case INTR_TYPE_3OP_MASK_RM: {
16248 SDValue Src1 = Op.getOperand(1);
16249 SDValue Src2 = Op.getOperand(2);
16250 SDValue Imm = Op.getOperand(3);
16251 SDValue PassThru = Op.getOperand(4);
16252 SDValue Mask = Op.getOperand(5);
16253 // We specify 2 possible modes for intrinsics, with/without rounding
16255 // First, we check if the intrinsic have rounding mode (7 operands),
16256 // if not, we set rounding mode to "current".
16258 if (Op.getNumOperands() == 7)
16259 Rnd = Op.getOperand(6);
16261 Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
16262 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16263 Src1, Src2, Imm, Rnd),
16264 Mask, PassThru, Subtarget, DAG);
16266 case INTR_TYPE_3OP_IMM8_MASK:
16267 case INTR_TYPE_3OP_MASK:
16268 case INSERT_SUBVEC: {
16269 SDValue Src1 = Op.getOperand(1);
16270 SDValue Src2 = Op.getOperand(2);
16271 SDValue Src3 = Op.getOperand(3);
16272 SDValue PassThru = Op.getOperand(4);
16273 SDValue Mask = Op.getOperand(5);
16275 if (IntrData->Type == INTR_TYPE_3OP_IMM8_MASK)
16276 Src3 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Src3);
16277 else if (IntrData->Type == INSERT_SUBVEC) {
16278 // imm should be adapted to ISD::INSERT_SUBVECTOR behavior
16279 assert(isa<ConstantSDNode>(Src3) && "Expected a ConstantSDNode here!");
16280 unsigned Imm = cast<ConstantSDNode>(Src3)->getZExtValue();
16281 Imm *= Src2.getSimpleValueType().getVectorNumElements();
16282 Src3 = DAG.getTargetConstant(Imm, dl, MVT::i32);
16285 // We specify 2 possible opcodes for intrinsics with rounding modes.
16286 // First, we check if the intrinsic may have non-default rounding mode,
16287 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16288 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16289 if (IntrWithRoundingModeOpcode != 0) {
16290 SDValue Rnd = Op.getOperand(6);
16291 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
16292 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
16293 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16294 dl, Op.getValueType(),
16295 Src1, Src2, Src3, Rnd),
16296 Mask, PassThru, Subtarget, DAG);
16299 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16301 Mask, PassThru, Subtarget, DAG);
16303 case VPERM_3OP_MASKZ:
16304 case VPERM_3OP_MASK:
16307 case FMA_OP_MASK: {
16308 SDValue Src1 = Op.getOperand(1);
16309 SDValue Src2 = Op.getOperand(2);
16310 SDValue Src3 = Op.getOperand(3);
16311 SDValue Mask = Op.getOperand(4);
16312 MVT VT = Op.getSimpleValueType();
16313 SDValue PassThru = SDValue();
16315 // set PassThru element
16316 if (IntrData->Type == VPERM_3OP_MASKZ || IntrData->Type == FMA_OP_MASKZ)
16317 PassThru = getZeroVector(VT, Subtarget, DAG, dl);
16318 else if (IntrData->Type == FMA_OP_MASK3)
16323 // We specify 2 possible opcodes for intrinsics with rounding modes.
16324 // First, we check if the intrinsic may have non-default rounding mode,
16325 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16326 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16327 if (IntrWithRoundingModeOpcode != 0) {
16328 SDValue Rnd = Op.getOperand(5);
16329 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16330 X86::STATIC_ROUNDING::CUR_DIRECTION)
16331 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16332 dl, Op.getValueType(),
16333 Src1, Src2, Src3, Rnd),
16334 Mask, PassThru, Subtarget, DAG);
16336 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0,
16337 dl, Op.getValueType(),
16339 Mask, PassThru, Subtarget, DAG);
16341 case TERLOG_OP_MASK:
16342 case TERLOG_OP_MASKZ: {
16343 SDValue Src1 = Op.getOperand(1);
16344 SDValue Src2 = Op.getOperand(2);
16345 SDValue Src3 = Op.getOperand(3);
16346 SDValue Src4 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(4));
16347 SDValue Mask = Op.getOperand(5);
16348 MVT VT = Op.getSimpleValueType();
16349 SDValue PassThru = Src1;
16350 // Set PassThru element.
16351 if (IntrData->Type == TERLOG_OP_MASKZ)
16352 PassThru = getZeroVector(VT, Subtarget, DAG, dl);
16354 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16355 Src1, Src2, Src3, Src4),
16356 Mask, PassThru, Subtarget, DAG);
16359 // FPclass intrinsics with mask
16360 SDValue Src1 = Op.getOperand(1);
16361 MVT VT = Src1.getSimpleValueType();
16362 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
16363 SDValue Imm = Op.getOperand(2);
16364 SDValue Mask = Op.getOperand(3);
16365 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
16366 Mask.getSimpleValueType().getSizeInBits());
16367 SDValue FPclass = DAG.getNode(IntrData->Opc0, dl, MaskVT, Src1, Imm);
16368 SDValue FPclassMask = getVectorMaskingNode(FPclass, Mask,
16369 DAG.getTargetConstant(0, dl, MaskVT),
16371 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
16372 DAG.getUNDEF(BitcastVT), FPclassMask,
16373 DAG.getIntPtrConstant(0, dl));
16374 return DAG.getBitcast(Op.getValueType(), Res);
16377 SDValue Src1 = Op.getOperand(1);
16378 SDValue Imm = Op.getOperand(2);
16379 SDValue Mask = Op.getOperand(3);
16380 SDValue FPclass = DAG.getNode(IntrData->Opc0, dl, MVT::i1, Src1, Imm);
16381 SDValue FPclassMask = getScalarMaskingNode(FPclass, Mask,
16382 DAG.getTargetConstant(0, dl, MVT::i1), Subtarget, DAG);
16383 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i8, FPclassMask);
16386 case CMP_MASK_CC: {
16387 // Comparison intrinsics with masks.
16388 // Example of transformation:
16389 // (i8 (int_x86_avx512_mask_pcmpeq_q_128
16390 // (v2i64 %a), (v2i64 %b), (i8 %mask))) ->
16392 // (v8i1 (insert_subvector undef,
16393 // (v2i1 (and (PCMPEQM %a, %b),
16394 // (extract_subvector
16395 // (v8i1 (bitcast %mask)), 0))), 0))))
16396 MVT VT = Op.getOperand(1).getSimpleValueType();
16397 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
16398 SDValue Mask = Op.getOperand((IntrData->Type == CMP_MASK_CC) ? 4 : 3);
16399 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
16400 Mask.getSimpleValueType().getSizeInBits());
16402 if (IntrData->Type == CMP_MASK_CC) {
16403 SDValue CC = Op.getOperand(3);
16404 CC = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, CC);
16405 // We specify 2 possible opcodes for intrinsics with rounding modes.
16406 // First, we check if the intrinsic may have non-default rounding mode,
16407 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16408 if (IntrData->Opc1 != 0) {
16409 SDValue Rnd = Op.getOperand(5);
16410 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16411 X86::STATIC_ROUNDING::CUR_DIRECTION)
16412 Cmp = DAG.getNode(IntrData->Opc1, dl, MaskVT, Op.getOperand(1),
16413 Op.getOperand(2), CC, Rnd);
16415 //default rounding mode
16417 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
16418 Op.getOperand(2), CC);
16421 assert(IntrData->Type == CMP_MASK && "Unexpected intrinsic type!");
16422 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
16425 SDValue CmpMask = getVectorMaskingNode(Cmp, Mask,
16426 DAG.getTargetConstant(0, dl,
16429 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
16430 DAG.getUNDEF(BitcastVT), CmpMask,
16431 DAG.getIntPtrConstant(0, dl));
16432 return DAG.getBitcast(Op.getValueType(), Res);
16434 case CMP_MASK_SCALAR_CC: {
16435 SDValue Src1 = Op.getOperand(1);
16436 SDValue Src2 = Op.getOperand(2);
16437 SDValue CC = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(3));
16438 SDValue Mask = Op.getOperand(4);
16441 if (IntrData->Opc1 != 0) {
16442 SDValue Rnd = Op.getOperand(5);
16443 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16444 X86::STATIC_ROUNDING::CUR_DIRECTION)
16445 Cmp = DAG.getNode(IntrData->Opc1, dl, MVT::i1, Src1, Src2, CC, Rnd);
16447 //default rounding mode
16449 Cmp = DAG.getNode(IntrData->Opc0, dl, MVT::i1, Src1, Src2, CC);
16451 SDValue CmpMask = getScalarMaskingNode(Cmp, Mask,
16452 DAG.getTargetConstant(0, dl,
16456 return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::i8,
16457 DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i8, CmpMask),
16458 DAG.getValueType(MVT::i1));
16460 case COMI: { // Comparison intrinsics
16461 ISD::CondCode CC = (ISD::CondCode)IntrData->Opc1;
16462 SDValue LHS = Op.getOperand(1);
16463 SDValue RHS = Op.getOperand(2);
16464 unsigned X86CC = TranslateX86CC(CC, dl, true, LHS, RHS, DAG);
16465 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
16466 SDValue Cond = DAG.getNode(IntrData->Opc0, dl, MVT::i32, LHS, RHS);
16467 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16468 DAG.getConstant(X86CC, dl, MVT::i8), Cond);
16469 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16472 return getTargetVShiftNode(IntrData->Opc0, dl, Op.getSimpleValueType(),
16473 Op.getOperand(1), Op.getOperand(2), DAG);
16475 return getVectorMaskingNode(getTargetVShiftNode(IntrData->Opc0, dl,
16476 Op.getSimpleValueType(),
16478 Op.getOperand(2), DAG),
16479 Op.getOperand(4), Op.getOperand(3), Subtarget,
16481 case COMPRESS_EXPAND_IN_REG: {
16482 SDValue Mask = Op.getOperand(3);
16483 SDValue DataToCompress = Op.getOperand(1);
16484 SDValue PassThru = Op.getOperand(2);
16485 if (isAllOnes(Mask)) // return data as is
16486 return Op.getOperand(1);
16488 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16490 Mask, PassThru, Subtarget, DAG);
16493 SDValue Mask = Op.getOperand(1);
16494 MVT MaskVT = MVT::getVectorVT(MVT::i1, Mask.getSimpleValueType().getSizeInBits());
16495 Mask = DAG.getBitcast(MaskVT, Mask);
16496 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Mask);
16499 SDValue Mask = Op.getOperand(3);
16500 MVT VT = Op.getSimpleValueType();
16501 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
16502 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
16503 Mask.getSimpleValueType().getSizeInBits());
16505 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16506 DAG.getBitcast(BitcastVT, Mask),
16507 DAG.getIntPtrConstant(0, dl));
16508 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, Op.getOperand(1),
16517 default: return SDValue(); // Don't custom lower most intrinsics.
16519 case Intrinsic::x86_avx2_permd:
16520 case Intrinsic::x86_avx2_permps:
16521 // Operands intentionally swapped. Mask is last operand to intrinsic,
16522 // but second operand for node/instruction.
16523 return DAG.getNode(X86ISD::VPERMV, dl, Op.getValueType(),
16524 Op.getOperand(2), Op.getOperand(1));
16526 // ptest and testp intrinsics. The intrinsic these come from are designed to
16527 // return an integer value, not just an instruction so lower it to the ptest
16528 // or testp pattern and a setcc for the result.
16529 case Intrinsic::x86_sse41_ptestz:
16530 case Intrinsic::x86_sse41_ptestc:
16531 case Intrinsic::x86_sse41_ptestnzc:
16532 case Intrinsic::x86_avx_ptestz_256:
16533 case Intrinsic::x86_avx_ptestc_256:
16534 case Intrinsic::x86_avx_ptestnzc_256:
16535 case Intrinsic::x86_avx_vtestz_ps:
16536 case Intrinsic::x86_avx_vtestc_ps:
16537 case Intrinsic::x86_avx_vtestnzc_ps:
16538 case Intrinsic::x86_avx_vtestz_pd:
16539 case Intrinsic::x86_avx_vtestc_pd:
16540 case Intrinsic::x86_avx_vtestnzc_pd:
16541 case Intrinsic::x86_avx_vtestz_ps_256:
16542 case Intrinsic::x86_avx_vtestc_ps_256:
16543 case Intrinsic::x86_avx_vtestnzc_ps_256:
16544 case Intrinsic::x86_avx_vtestz_pd_256:
16545 case Intrinsic::x86_avx_vtestc_pd_256:
16546 case Intrinsic::x86_avx_vtestnzc_pd_256: {
16547 bool IsTestPacked = false;
16550 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
16551 case Intrinsic::x86_avx_vtestz_ps:
16552 case Intrinsic::x86_avx_vtestz_pd:
16553 case Intrinsic::x86_avx_vtestz_ps_256:
16554 case Intrinsic::x86_avx_vtestz_pd_256:
16555 IsTestPacked = true; // Fallthrough
16556 case Intrinsic::x86_sse41_ptestz:
16557 case Intrinsic::x86_avx_ptestz_256:
16559 X86CC = X86::COND_E;
16561 case Intrinsic::x86_avx_vtestc_ps:
16562 case Intrinsic::x86_avx_vtestc_pd:
16563 case Intrinsic::x86_avx_vtestc_ps_256:
16564 case Intrinsic::x86_avx_vtestc_pd_256:
16565 IsTestPacked = true; // Fallthrough
16566 case Intrinsic::x86_sse41_ptestc:
16567 case Intrinsic::x86_avx_ptestc_256:
16569 X86CC = X86::COND_B;
16571 case Intrinsic::x86_avx_vtestnzc_ps:
16572 case Intrinsic::x86_avx_vtestnzc_pd:
16573 case Intrinsic::x86_avx_vtestnzc_ps_256:
16574 case Intrinsic::x86_avx_vtestnzc_pd_256:
16575 IsTestPacked = true; // Fallthrough
16576 case Intrinsic::x86_sse41_ptestnzc:
16577 case Intrinsic::x86_avx_ptestnzc_256:
16579 X86CC = X86::COND_A;
16583 SDValue LHS = Op.getOperand(1);
16584 SDValue RHS = Op.getOperand(2);
16585 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
16586 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
16587 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
16588 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
16589 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16591 case Intrinsic::x86_avx512_kortestz_w:
16592 case Intrinsic::x86_avx512_kortestc_w: {
16593 unsigned X86CC = (IntNo == Intrinsic::x86_avx512_kortestz_w)? X86::COND_E: X86::COND_B;
16594 SDValue LHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(1));
16595 SDValue RHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(2));
16596 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
16597 SDValue Test = DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, LHS, RHS);
16598 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i1, CC, Test);
16599 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16602 case Intrinsic::x86_sse42_pcmpistria128:
16603 case Intrinsic::x86_sse42_pcmpestria128:
16604 case Intrinsic::x86_sse42_pcmpistric128:
16605 case Intrinsic::x86_sse42_pcmpestric128:
16606 case Intrinsic::x86_sse42_pcmpistrio128:
16607 case Intrinsic::x86_sse42_pcmpestrio128:
16608 case Intrinsic::x86_sse42_pcmpistris128:
16609 case Intrinsic::x86_sse42_pcmpestris128:
16610 case Intrinsic::x86_sse42_pcmpistriz128:
16611 case Intrinsic::x86_sse42_pcmpestriz128: {
16615 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
16616 case Intrinsic::x86_sse42_pcmpistria128:
16617 Opcode = X86ISD::PCMPISTRI;
16618 X86CC = X86::COND_A;
16620 case Intrinsic::x86_sse42_pcmpestria128:
16621 Opcode = X86ISD::PCMPESTRI;
16622 X86CC = X86::COND_A;
16624 case Intrinsic::x86_sse42_pcmpistric128:
16625 Opcode = X86ISD::PCMPISTRI;
16626 X86CC = X86::COND_B;
16628 case Intrinsic::x86_sse42_pcmpestric128:
16629 Opcode = X86ISD::PCMPESTRI;
16630 X86CC = X86::COND_B;
16632 case Intrinsic::x86_sse42_pcmpistrio128:
16633 Opcode = X86ISD::PCMPISTRI;
16634 X86CC = X86::COND_O;
16636 case Intrinsic::x86_sse42_pcmpestrio128:
16637 Opcode = X86ISD::PCMPESTRI;
16638 X86CC = X86::COND_O;
16640 case Intrinsic::x86_sse42_pcmpistris128:
16641 Opcode = X86ISD::PCMPISTRI;
16642 X86CC = X86::COND_S;
16644 case Intrinsic::x86_sse42_pcmpestris128:
16645 Opcode = X86ISD::PCMPESTRI;
16646 X86CC = X86::COND_S;
16648 case Intrinsic::x86_sse42_pcmpistriz128:
16649 Opcode = X86ISD::PCMPISTRI;
16650 X86CC = X86::COND_E;
16652 case Intrinsic::x86_sse42_pcmpestriz128:
16653 Opcode = X86ISD::PCMPESTRI;
16654 X86CC = X86::COND_E;
16657 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
16658 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
16659 SDValue PCMP = DAG.getNode(Opcode, dl, VTs, NewOps);
16660 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16661 DAG.getConstant(X86CC, dl, MVT::i8),
16662 SDValue(PCMP.getNode(), 1));
16663 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16666 case Intrinsic::x86_sse42_pcmpistri128:
16667 case Intrinsic::x86_sse42_pcmpestri128: {
16669 if (IntNo == Intrinsic::x86_sse42_pcmpistri128)
16670 Opcode = X86ISD::PCMPISTRI;
16672 Opcode = X86ISD::PCMPESTRI;
16674 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
16675 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
16676 return DAG.getNode(Opcode, dl, VTs, NewOps);
16679 case Intrinsic::x86_seh_lsda: {
16680 // Compute the symbol for the LSDA. We know it'll get emitted later.
16681 MachineFunction &MF = DAG.getMachineFunction();
16682 SDValue Op1 = Op.getOperand(1);
16683 auto *Fn = cast<Function>(cast<GlobalAddressSDNode>(Op1)->getGlobal());
16684 MCSymbol *LSDASym = MF.getMMI().getContext().getOrCreateLSDASymbol(
16685 GlobalValue::getRealLinkageName(Fn->getName()));
16687 // Generate a simple absolute symbol reference. This intrinsic is only
16688 // supported on 32-bit Windows, which isn't PIC.
16689 SDValue Result = DAG.getMCSymbol(LSDASym, VT);
16690 return DAG.getNode(X86ISD::Wrapper, dl, VT, Result);
16693 case Intrinsic::x86_seh_recoverfp: {
16694 SDValue FnOp = Op.getOperand(1);
16695 SDValue IncomingFPOp = Op.getOperand(2);
16696 GlobalAddressSDNode *GSD = dyn_cast<GlobalAddressSDNode>(FnOp);
16697 auto *Fn = dyn_cast_or_null<Function>(GSD ? GSD->getGlobal() : nullptr);
16699 report_fatal_error(
16700 "llvm.x86.seh.recoverfp must take a function as the first argument");
16701 return recoverFramePointer(DAG, Fn, IncomingFPOp);
16704 case Intrinsic::localaddress: {
16705 // Returns one of the stack, base, or frame pointer registers, depending on
16706 // which is used to reference local variables.
16707 MachineFunction &MF = DAG.getMachineFunction();
16708 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16710 if (RegInfo->hasBasePointer(MF))
16711 Reg = RegInfo->getBaseRegister();
16712 else // This function handles the SP or FP case.
16713 Reg = RegInfo->getPtrSizedFrameRegister(MF);
16714 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
16719 static SDValue getGatherNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16720 SDValue Src, SDValue Mask, SDValue Base,
16721 SDValue Index, SDValue ScaleOp, SDValue Chain,
16722 const X86Subtarget * Subtarget) {
16724 auto *C = cast<ConstantSDNode>(ScaleOp);
16725 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16726 MVT MaskVT = MVT::getVectorVT(MVT::i1,
16727 Index.getSimpleValueType().getVectorNumElements());
16729 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16731 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16733 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
16734 Mask.getSimpleValueType().getSizeInBits());
16736 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16737 // are extracted by EXTRACT_SUBVECTOR.
16738 MaskInReg = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16739 DAG.getBitcast(BitcastVT, Mask),
16740 DAG.getIntPtrConstant(0, dl));
16742 SDVTList VTs = DAG.getVTList(Op.getValueType(), MaskVT, MVT::Other);
16743 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16744 SDValue Segment = DAG.getRegister(0, MVT::i32);
16745 if (Src.getOpcode() == ISD::UNDEF)
16746 Src = getZeroVector(Op.getValueType(), Subtarget, DAG, dl);
16747 SDValue Ops[] = {Src, MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
16748 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
16749 SDValue RetOps[] = { SDValue(Res, 0), SDValue(Res, 2) };
16750 return DAG.getMergeValues(RetOps, dl);
16753 static SDValue getScatterNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16754 SDValue Src, SDValue Mask, SDValue Base,
16755 SDValue Index, SDValue ScaleOp, SDValue Chain) {
16757 auto *C = cast<ConstantSDNode>(ScaleOp);
16758 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16759 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16760 SDValue Segment = DAG.getRegister(0, MVT::i32);
16761 MVT MaskVT = MVT::getVectorVT(MVT::i1,
16762 Index.getSimpleValueType().getVectorNumElements());
16764 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16766 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16768 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
16769 Mask.getSimpleValueType().getSizeInBits());
16771 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16772 // are extracted by EXTRACT_SUBVECTOR.
16773 MaskInReg = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16774 DAG.getBitcast(BitcastVT, Mask),
16775 DAG.getIntPtrConstant(0, dl));
16777 SDVTList VTs = DAG.getVTList(MaskVT, MVT::Other);
16778 SDValue Ops[] = {Base, Scale, Index, Disp, Segment, MaskInReg, Src, Chain};
16779 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
16780 return SDValue(Res, 1);
16783 static SDValue getPrefetchNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16784 SDValue Mask, SDValue Base, SDValue Index,
16785 SDValue ScaleOp, SDValue Chain) {
16787 auto *C = cast<ConstantSDNode>(ScaleOp);
16788 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16789 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16790 SDValue Segment = DAG.getRegister(0, MVT::i32);
16792 MVT::getVectorVT(MVT::i1, Index.getSimpleValueType().getVectorNumElements());
16794 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16796 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16798 MaskInReg = DAG.getBitcast(MaskVT, Mask);
16799 //SDVTList VTs = DAG.getVTList(MVT::Other);
16800 SDValue Ops[] = {MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
16801 SDNode *Res = DAG.getMachineNode(Opc, dl, MVT::Other, Ops);
16802 return SDValue(Res, 0);
16805 // getReadPerformanceCounter - Handles the lowering of builtin intrinsics that
16806 // read performance monitor counters (x86_rdpmc).
16807 static void getReadPerformanceCounter(SDNode *N, SDLoc DL,
16808 SelectionDAG &DAG, const X86Subtarget *Subtarget,
16809 SmallVectorImpl<SDValue> &Results) {
16810 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
16811 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
16814 // The ECX register is used to select the index of the performance counter
16816 SDValue Chain = DAG.getCopyToReg(N->getOperand(0), DL, X86::ECX,
16818 SDValue rd = DAG.getNode(X86ISD::RDPMC_DAG, DL, Tys, Chain);
16820 // Reads the content of a 64-bit performance counter and returns it in the
16821 // registers EDX:EAX.
16822 if (Subtarget->is64Bit()) {
16823 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
16824 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
16827 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
16828 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
16831 Chain = HI.getValue(1);
16833 if (Subtarget->is64Bit()) {
16834 // The EAX register is loaded with the low-order 32 bits. The EDX register
16835 // is loaded with the supported high-order bits of the counter.
16836 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
16837 DAG.getConstant(32, DL, MVT::i8));
16838 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
16839 Results.push_back(Chain);
16843 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
16844 SDValue Ops[] = { LO, HI };
16845 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
16846 Results.push_back(Pair);
16847 Results.push_back(Chain);
16850 // getReadTimeStampCounter - Handles the lowering of builtin intrinsics that
16851 // read the time stamp counter (x86_rdtsc and x86_rdtscp). This function is
16852 // also used to custom lower READCYCLECOUNTER nodes.
16853 static void getReadTimeStampCounter(SDNode *N, SDLoc DL, unsigned Opcode,
16854 SelectionDAG &DAG, const X86Subtarget *Subtarget,
16855 SmallVectorImpl<SDValue> &Results) {
16856 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
16857 SDValue rd = DAG.getNode(Opcode, DL, Tys, N->getOperand(0));
16860 // The processor's time-stamp counter (a 64-bit MSR) is stored into the
16861 // EDX:EAX registers. EDX is loaded with the high-order 32 bits of the MSR
16862 // and the EAX register is loaded with the low-order 32 bits.
16863 if (Subtarget->is64Bit()) {
16864 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
16865 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
16868 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
16869 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
16872 SDValue Chain = HI.getValue(1);
16874 if (Opcode == X86ISD::RDTSCP_DAG) {
16875 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
16877 // Instruction RDTSCP loads the IA32:TSC_AUX_MSR (address C000_0103H) into
16878 // the ECX register. Add 'ecx' explicitly to the chain.
16879 SDValue ecx = DAG.getCopyFromReg(Chain, DL, X86::ECX, MVT::i32,
16881 // Explicitly store the content of ECX at the location passed in input
16882 // to the 'rdtscp' intrinsic.
16883 Chain = DAG.getStore(ecx.getValue(1), DL, ecx, N->getOperand(2),
16884 MachinePointerInfo(), false, false, 0);
16887 if (Subtarget->is64Bit()) {
16888 // The EDX register is loaded with the high-order 32 bits of the MSR, and
16889 // the EAX register is loaded with the low-order 32 bits.
16890 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
16891 DAG.getConstant(32, DL, MVT::i8));
16892 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
16893 Results.push_back(Chain);
16897 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
16898 SDValue Ops[] = { LO, HI };
16899 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
16900 Results.push_back(Pair);
16901 Results.push_back(Chain);
16904 static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget *Subtarget,
16905 SelectionDAG &DAG) {
16906 SmallVector<SDValue, 2> Results;
16908 getReadTimeStampCounter(Op.getNode(), DL, X86ISD::RDTSC_DAG, DAG, Subtarget,
16910 return DAG.getMergeValues(Results, DL);
16913 static SDValue LowerSEHRESTOREFRAME(SDValue Op, const X86Subtarget *Subtarget,
16914 SelectionDAG &DAG) {
16915 MachineFunction &MF = DAG.getMachineFunction();
16916 const Function *Fn = MF.getFunction();
16918 SDValue Chain = Op.getOperand(0);
16920 assert(Subtarget->getFrameLowering()->hasFP(MF) &&
16921 "using llvm.x86.seh.restoreframe requires a frame pointer");
16923 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16924 MVT VT = TLI.getPointerTy(DAG.getDataLayout());
16926 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16927 unsigned FrameReg =
16928 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
16929 unsigned SPReg = RegInfo->getStackRegister();
16930 unsigned SlotSize = RegInfo->getSlotSize();
16932 // Get incoming EBP.
16933 SDValue IncomingEBP =
16934 DAG.getCopyFromReg(Chain, dl, FrameReg, VT);
16936 // SP is saved in the first field of every registration node, so load
16937 // [EBP-RegNodeSize] into SP.
16938 int RegNodeSize = getSEHRegistrationNodeSize(Fn);
16939 SDValue SPAddr = DAG.getNode(ISD::ADD, dl, VT, IncomingEBP,
16940 DAG.getConstant(-RegNodeSize, dl, VT));
16942 DAG.getLoad(VT, dl, Chain, SPAddr, MachinePointerInfo(), false, false,
16943 false, VT.getScalarSizeInBits() / 8);
16944 Chain = DAG.getCopyToReg(Chain, dl, SPReg, NewSP);
16946 if (!RegInfo->needsStackRealignment(MF)) {
16947 // Adjust EBP to point back to the original frame position.
16948 SDValue NewFP = recoverFramePointer(DAG, Fn, IncomingEBP);
16949 Chain = DAG.getCopyToReg(Chain, dl, FrameReg, NewFP);
16951 assert(RegInfo->hasBasePointer(MF) &&
16952 "functions with Win32 EH must use frame or base pointer register");
16954 // Reload the base pointer (ESI) with the adjusted incoming EBP.
16955 SDValue NewBP = recoverFramePointer(DAG, Fn, IncomingEBP);
16956 Chain = DAG.getCopyToReg(Chain, dl, RegInfo->getBaseRegister(), NewBP);
16958 // Reload the spilled EBP value, now that the stack and base pointers are
16960 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
16961 X86FI->setHasSEHFramePtrSave(true);
16962 int FI = MF.getFrameInfo()->CreateSpillStackObject(SlotSize, SlotSize);
16963 X86FI->setSEHFramePtrSaveIndex(FI);
16964 SDValue NewFP = DAG.getLoad(VT, dl, Chain, DAG.getFrameIndex(FI, VT),
16965 MachinePointerInfo(), false, false, false,
16966 VT.getScalarSizeInBits() / 8);
16967 Chain = DAG.getCopyToReg(NewFP, dl, FrameReg, NewFP);
16973 static SDValue MarkEHRegistrationNode(SDValue Op, SelectionDAG &DAG) {
16974 MachineFunction &MF = DAG.getMachineFunction();
16975 SDValue Chain = Op.getOperand(0);
16976 SDValue RegNode = Op.getOperand(2);
16977 WinEHFuncInfo *EHInfo = MF.getWinEHFuncInfo();
16979 report_fatal_error("EH registrations only live in functions using WinEH");
16981 // Cast the operand to an alloca, and remember the frame index.
16982 auto *FINode = dyn_cast<FrameIndexSDNode>(RegNode);
16984 report_fatal_error("llvm.x86.seh.ehregnode expects a static alloca");
16985 EHInfo->EHRegNodeFrameIndex = FINode->getIndex();
16987 // Return the chain operand without making any DAG nodes.
16991 /// \brief Lower intrinsics for TRUNCATE_TO_MEM case
16992 /// return truncate Store/MaskedStore Node
16993 static SDValue LowerINTRINSIC_TRUNCATE_TO_MEM(const SDValue & Op,
16997 SDValue Mask = Op.getOperand(4);
16998 SDValue DataToTruncate = Op.getOperand(3);
16999 SDValue Addr = Op.getOperand(2);
17000 SDValue Chain = Op.getOperand(0);
17002 MVT VT = DataToTruncate.getSimpleValueType();
17003 MVT SVT = MVT::getVectorVT(ElementType, VT.getVectorNumElements());
17005 if (isAllOnes(Mask)) // return just a truncate store
17006 return DAG.getTruncStore(Chain, dl, DataToTruncate, Addr,
17007 MachinePointerInfo(), SVT, false, false,
17008 SVT.getScalarSizeInBits()/8);
17010 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
17011 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
17012 Mask.getSimpleValueType().getSizeInBits());
17013 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
17014 // are extracted by EXTRACT_SUBVECTOR.
17015 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
17016 DAG.getBitcast(BitcastVT, Mask),
17017 DAG.getIntPtrConstant(0, dl));
17019 MachineMemOperand *MMO = DAG.getMachineFunction().
17020 getMachineMemOperand(MachinePointerInfo(),
17021 MachineMemOperand::MOStore, SVT.getStoreSize(),
17022 SVT.getScalarSizeInBits()/8);
17024 return DAG.getMaskedStore(Chain, dl, DataToTruncate, Addr,
17025 VMask, SVT, MMO, true);
17028 static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
17029 SelectionDAG &DAG) {
17030 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
17032 const IntrinsicData* IntrData = getIntrinsicWithChain(IntNo);
17034 if (IntNo == llvm::Intrinsic::x86_seh_restoreframe)
17035 return LowerSEHRESTOREFRAME(Op, Subtarget, DAG);
17036 else if (IntNo == llvm::Intrinsic::x86_seh_ehregnode)
17037 return MarkEHRegistrationNode(Op, DAG);
17042 switch(IntrData->Type) {
17043 default: llvm_unreachable("Unknown Intrinsic Type");
17046 // Emit the node with the right value type.
17047 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other);
17048 SDValue Result = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
17050 // If the value returned by RDRAND/RDSEED was valid (CF=1), return 1.
17051 // Otherwise return the value from Rand, which is always 0, casted to i32.
17052 SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)),
17053 DAG.getConstant(1, dl, Op->getValueType(1)),
17054 DAG.getConstant(X86::COND_B, dl, MVT::i32),
17055 SDValue(Result.getNode(), 1) };
17056 SDValue isValid = DAG.getNode(X86ISD::CMOV, dl,
17057 DAG.getVTList(Op->getValueType(1), MVT::Glue),
17060 // Return { result, isValid, chain }.
17061 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid,
17062 SDValue(Result.getNode(), 2));
17065 //gather(v1, mask, index, base, scale);
17066 SDValue Chain = Op.getOperand(0);
17067 SDValue Src = Op.getOperand(2);
17068 SDValue Base = Op.getOperand(3);
17069 SDValue Index = Op.getOperand(4);
17070 SDValue Mask = Op.getOperand(5);
17071 SDValue Scale = Op.getOperand(6);
17072 return getGatherNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale,
17076 //scatter(base, mask, index, v1, scale);
17077 SDValue Chain = Op.getOperand(0);
17078 SDValue Base = Op.getOperand(2);
17079 SDValue Mask = Op.getOperand(3);
17080 SDValue Index = Op.getOperand(4);
17081 SDValue Src = Op.getOperand(5);
17082 SDValue Scale = Op.getOperand(6);
17083 return getScatterNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index,
17087 SDValue Hint = Op.getOperand(6);
17088 unsigned HintVal = cast<ConstantSDNode>(Hint)->getZExtValue();
17089 assert(HintVal < 2 && "Wrong prefetch hint in intrinsic: should be 0 or 1");
17090 unsigned Opcode = (HintVal ? IntrData->Opc1 : IntrData->Opc0);
17091 SDValue Chain = Op.getOperand(0);
17092 SDValue Mask = Op.getOperand(2);
17093 SDValue Index = Op.getOperand(3);
17094 SDValue Base = Op.getOperand(4);
17095 SDValue Scale = Op.getOperand(5);
17096 return getPrefetchNode(Opcode, Op, DAG, Mask, Base, Index, Scale, Chain);
17098 // Read Time Stamp Counter (RDTSC) and Processor ID (RDTSCP).
17100 SmallVector<SDValue, 2> Results;
17101 getReadTimeStampCounter(Op.getNode(), dl, IntrData->Opc0, DAG, Subtarget,
17103 return DAG.getMergeValues(Results, dl);
17105 // Read Performance Monitoring Counters.
17107 SmallVector<SDValue, 2> Results;
17108 getReadPerformanceCounter(Op.getNode(), dl, DAG, Subtarget, Results);
17109 return DAG.getMergeValues(Results, dl);
17111 // XTEST intrinsics.
17113 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
17114 SDValue InTrans = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
17115 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17116 DAG.getConstant(X86::COND_NE, dl, MVT::i8),
17118 SDValue Ret = DAG.getNode(ISD::ZERO_EXTEND, dl, Op->getValueType(0), SetCC);
17119 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(),
17120 Ret, SDValue(InTrans.getNode(), 1));
17124 SmallVector<SDValue, 2> Results;
17125 SDVTList CFVTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
17126 SDVTList VTs = DAG.getVTList(Op.getOperand(3)->getValueType(0), MVT::Other);
17127 SDValue GenCF = DAG.getNode(X86ISD::ADD, dl, CFVTs, Op.getOperand(2),
17128 DAG.getConstant(-1, dl, MVT::i8));
17129 SDValue Res = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(3),
17130 Op.getOperand(4), GenCF.getValue(1));
17131 SDValue Store = DAG.getStore(Op.getOperand(0), dl, Res.getValue(0),
17132 Op.getOperand(5), MachinePointerInfo(),
17134 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17135 DAG.getConstant(X86::COND_B, dl, MVT::i8),
17137 Results.push_back(SetCC);
17138 Results.push_back(Store);
17139 return DAG.getMergeValues(Results, dl);
17141 case COMPRESS_TO_MEM: {
17143 SDValue Mask = Op.getOperand(4);
17144 SDValue DataToCompress = Op.getOperand(3);
17145 SDValue Addr = Op.getOperand(2);
17146 SDValue Chain = Op.getOperand(0);
17148 MVT VT = DataToCompress.getSimpleValueType();
17149 if (isAllOnes(Mask)) // return just a store
17150 return DAG.getStore(Chain, dl, DataToCompress, Addr,
17151 MachinePointerInfo(), false, false,
17152 VT.getScalarSizeInBits()/8);
17154 SDValue Compressed =
17155 getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, DataToCompress),
17156 Mask, DAG.getUNDEF(VT), Subtarget, DAG);
17157 return DAG.getStore(Chain, dl, Compressed, Addr,
17158 MachinePointerInfo(), false, false,
17159 VT.getScalarSizeInBits()/8);
17161 case TRUNCATE_TO_MEM_VI8:
17162 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i8);
17163 case TRUNCATE_TO_MEM_VI16:
17164 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i16);
17165 case TRUNCATE_TO_MEM_VI32:
17166 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i32);
17167 case EXPAND_FROM_MEM: {
17169 SDValue Mask = Op.getOperand(4);
17170 SDValue PassThru = Op.getOperand(3);
17171 SDValue Addr = Op.getOperand(2);
17172 SDValue Chain = Op.getOperand(0);
17173 MVT VT = Op.getSimpleValueType();
17175 if (isAllOnes(Mask)) // return just a load
17176 return DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(), false, false,
17177 false, VT.getScalarSizeInBits()/8);
17179 SDValue DataToExpand = DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(),
17180 false, false, false,
17181 VT.getScalarSizeInBits()/8);
17183 SDValue Results[] = {
17184 getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, DataToExpand),
17185 Mask, PassThru, Subtarget, DAG), Chain};
17186 return DAG.getMergeValues(Results, dl);
17191 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
17192 SelectionDAG &DAG) const {
17193 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
17194 MFI->setReturnAddressIsTaken(true);
17196 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
17199 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
17201 EVT PtrVT = getPointerTy(DAG.getDataLayout());
17204 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
17205 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17206 SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), dl, PtrVT);
17207 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
17208 DAG.getNode(ISD::ADD, dl, PtrVT,
17209 FrameAddr, Offset),
17210 MachinePointerInfo(), false, false, false, 0);
17213 // Just load the return address.
17214 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
17215 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
17216 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
17219 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
17220 MachineFunction &MF = DAG.getMachineFunction();
17221 MachineFrameInfo *MFI = MF.getFrameInfo();
17222 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
17223 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17224 EVT VT = Op.getValueType();
17226 MFI->setFrameAddressIsTaken(true);
17228 if (MF.getTarget().getMCAsmInfo()->usesWindowsCFI()) {
17229 // Depth > 0 makes no sense on targets which use Windows unwind codes. It
17230 // is not possible to crawl up the stack without looking at the unwind codes
17232 int FrameAddrIndex = FuncInfo->getFAIndex();
17233 if (!FrameAddrIndex) {
17234 // Set up a frame object for the return address.
17235 unsigned SlotSize = RegInfo->getSlotSize();
17236 FrameAddrIndex = MF.getFrameInfo()->CreateFixedObject(
17237 SlotSize, /*Offset=*/0, /*IsImmutable=*/false);
17238 FuncInfo->setFAIndex(FrameAddrIndex);
17240 return DAG.getFrameIndex(FrameAddrIndex, VT);
17243 unsigned FrameReg =
17244 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
17245 SDLoc dl(Op); // FIXME probably not meaningful
17246 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
17247 assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
17248 (FrameReg == X86::EBP && VT == MVT::i32)) &&
17249 "Invalid Frame Register!");
17250 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
17252 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
17253 MachinePointerInfo(),
17254 false, false, false, 0);
17258 // FIXME? Maybe this could be a TableGen attribute on some registers and
17259 // this table could be generated automatically from RegInfo.
17260 unsigned X86TargetLowering::getRegisterByName(const char* RegName, EVT VT,
17261 SelectionDAG &DAG) const {
17262 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
17263 const MachineFunction &MF = DAG.getMachineFunction();
17265 unsigned Reg = StringSwitch<unsigned>(RegName)
17266 .Case("esp", X86::ESP)
17267 .Case("rsp", X86::RSP)
17268 .Case("ebp", X86::EBP)
17269 .Case("rbp", X86::RBP)
17272 if (Reg == X86::EBP || Reg == X86::RBP) {
17273 if (!TFI.hasFP(MF))
17274 report_fatal_error("register " + StringRef(RegName) +
17275 " is allocatable: function has no frame pointer");
17278 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17279 unsigned FrameReg =
17280 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
17281 assert((FrameReg == X86::EBP || FrameReg == X86::RBP) &&
17282 "Invalid Frame Register!");
17290 report_fatal_error("Invalid register name global variable");
17293 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
17294 SelectionDAG &DAG) const {
17295 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17296 return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize(), SDLoc(Op));
17299 unsigned X86TargetLowering::getExceptionPointerRegister(
17300 const Constant *PersonalityFn) const {
17301 if (classifyEHPersonality(PersonalityFn) == EHPersonality::CoreCLR)
17302 return Subtarget->isTarget64BitLP64() ? X86::RDX : X86::EDX;
17304 return Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX;
17307 unsigned X86TargetLowering::getExceptionSelectorRegister(
17308 const Constant *PersonalityFn) const {
17309 // Funclet personalities don't use selectors (the runtime does the selection).
17310 assert(!isFuncletEHPersonality(classifyEHPersonality(PersonalityFn)));
17311 return Subtarget->isTarget64BitLP64() ? X86::RDX : X86::EDX;
17314 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
17315 SDValue Chain = Op.getOperand(0);
17316 SDValue Offset = Op.getOperand(1);
17317 SDValue Handler = Op.getOperand(2);
17320 EVT PtrVT = getPointerTy(DAG.getDataLayout());
17321 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17322 unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
17323 assert(((FrameReg == X86::RBP && PtrVT == MVT::i64) ||
17324 (FrameReg == X86::EBP && PtrVT == MVT::i32)) &&
17325 "Invalid Frame Register!");
17326 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, PtrVT);
17327 unsigned StoreAddrReg = (PtrVT == MVT::i64) ? X86::RCX : X86::ECX;
17329 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Frame,
17330 DAG.getIntPtrConstant(RegInfo->getSlotSize(),
17332 StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, StoreAddr, Offset);
17333 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
17335 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
17337 return DAG.getNode(X86ISD::EH_RETURN, dl, MVT::Other, Chain,
17338 DAG.getRegister(StoreAddrReg, PtrVT));
17341 SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
17342 SelectionDAG &DAG) const {
17344 return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL,
17345 DAG.getVTList(MVT::i32, MVT::Other),
17346 Op.getOperand(0), Op.getOperand(1));
17349 SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
17350 SelectionDAG &DAG) const {
17352 return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
17353 Op.getOperand(0), Op.getOperand(1));
17356 static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
17357 return Op.getOperand(0);
17360 SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
17361 SelectionDAG &DAG) const {
17362 SDValue Root = Op.getOperand(0);
17363 SDValue Trmp = Op.getOperand(1); // trampoline
17364 SDValue FPtr = Op.getOperand(2); // nested function
17365 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
17368 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
17369 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
17371 if (Subtarget->is64Bit()) {
17372 SDValue OutChains[6];
17374 // Large code-model.
17375 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
17376 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
17378 const unsigned char N86R10 = TRI->getEncodingValue(X86::R10) & 0x7;
17379 const unsigned char N86R11 = TRI->getEncodingValue(X86::R11) & 0x7;
17381 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
17383 // Load the pointer to the nested function into R11.
17384 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
17385 SDValue Addr = Trmp;
17386 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17387 Addr, MachinePointerInfo(TrmpAddr),
17390 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17391 DAG.getConstant(2, dl, MVT::i64));
17392 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
17393 MachinePointerInfo(TrmpAddr, 2),
17396 // Load the 'nest' parameter value into R10.
17397 // R10 is specified in X86CallingConv.td
17398 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
17399 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17400 DAG.getConstant(10, dl, MVT::i64));
17401 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17402 Addr, MachinePointerInfo(TrmpAddr, 10),
17405 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17406 DAG.getConstant(12, dl, MVT::i64));
17407 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
17408 MachinePointerInfo(TrmpAddr, 12),
17411 // Jump to the nested function.
17412 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
17413 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17414 DAG.getConstant(20, dl, MVT::i64));
17415 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17416 Addr, MachinePointerInfo(TrmpAddr, 20),
17419 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
17420 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17421 DAG.getConstant(22, dl, MVT::i64));
17422 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, dl, MVT::i8),
17423 Addr, MachinePointerInfo(TrmpAddr, 22),
17426 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
17428 const Function *Func =
17429 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
17430 CallingConv::ID CC = Func->getCallingConv();
17435 llvm_unreachable("Unsupported calling convention");
17436 case CallingConv::C:
17437 case CallingConv::X86_StdCall: {
17438 // Pass 'nest' parameter in ECX.
17439 // Must be kept in sync with X86CallingConv.td
17440 NestReg = X86::ECX;
17442 // Check that ECX wasn't needed by an 'inreg' parameter.
17443 FunctionType *FTy = Func->getFunctionType();
17444 const AttributeSet &Attrs = Func->getAttributes();
17446 if (!Attrs.isEmpty() && !Func->isVarArg()) {
17447 unsigned InRegCount = 0;
17450 for (FunctionType::param_iterator I = FTy->param_begin(),
17451 E = FTy->param_end(); I != E; ++I, ++Idx)
17452 if (Attrs.hasAttribute(Idx, Attribute::InReg)) {
17453 auto &DL = DAG.getDataLayout();
17454 // FIXME: should only count parameters that are lowered to integers.
17455 InRegCount += (DL.getTypeSizeInBits(*I) + 31) / 32;
17458 if (InRegCount > 2) {
17459 report_fatal_error("Nest register in use - reduce number of inreg"
17465 case CallingConv::X86_FastCall:
17466 case CallingConv::X86_ThisCall:
17467 case CallingConv::Fast:
17468 // Pass 'nest' parameter in EAX.
17469 // Must be kept in sync with X86CallingConv.td
17470 NestReg = X86::EAX;
17474 SDValue OutChains[4];
17475 SDValue Addr, Disp;
17477 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17478 DAG.getConstant(10, dl, MVT::i32));
17479 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
17481 // This is storing the opcode for MOV32ri.
17482 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
17483 const unsigned char N86Reg = TRI->getEncodingValue(NestReg) & 0x7;
17484 OutChains[0] = DAG.getStore(Root, dl,
17485 DAG.getConstant(MOV32ri|N86Reg, dl, MVT::i8),
17486 Trmp, MachinePointerInfo(TrmpAddr),
17489 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17490 DAG.getConstant(1, dl, MVT::i32));
17491 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
17492 MachinePointerInfo(TrmpAddr, 1),
17495 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
17496 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17497 DAG.getConstant(5, dl, MVT::i32));
17498 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, dl, MVT::i8),
17499 Addr, MachinePointerInfo(TrmpAddr, 5),
17502 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17503 DAG.getConstant(6, dl, MVT::i32));
17504 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
17505 MachinePointerInfo(TrmpAddr, 6),
17508 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
17512 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
17513 SelectionDAG &DAG) const {
17515 The rounding mode is in bits 11:10 of FPSR, and has the following
17517 00 Round to nearest
17522 FLT_ROUNDS, on the other hand, expects the following:
17529 To perform the conversion, we do:
17530 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
17533 MachineFunction &MF = DAG.getMachineFunction();
17534 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
17535 unsigned StackAlignment = TFI.getStackAlignment();
17536 MVT VT = Op.getSimpleValueType();
17539 // Save FP Control Word to stack slot
17540 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
17541 SDValue StackSlot =
17542 DAG.getFrameIndex(SSFI, getPointerTy(DAG.getDataLayout()));
17544 MachineMemOperand *MMO =
17545 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
17546 MachineMemOperand::MOStore, 2, 2);
17548 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
17549 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
17550 DAG.getVTList(MVT::Other),
17551 Ops, MVT::i16, MMO);
17553 // Load FP Control Word from stack slot
17554 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
17555 MachinePointerInfo(), false, false, false, 0);
17557 // Transform as necessary
17559 DAG.getNode(ISD::SRL, DL, MVT::i16,
17560 DAG.getNode(ISD::AND, DL, MVT::i16,
17561 CWD, DAG.getConstant(0x800, DL, MVT::i16)),
17562 DAG.getConstant(11, DL, MVT::i8));
17564 DAG.getNode(ISD::SRL, DL, MVT::i16,
17565 DAG.getNode(ISD::AND, DL, MVT::i16,
17566 CWD, DAG.getConstant(0x400, DL, MVT::i16)),
17567 DAG.getConstant(9, DL, MVT::i8));
17570 DAG.getNode(ISD::AND, DL, MVT::i16,
17571 DAG.getNode(ISD::ADD, DL, MVT::i16,
17572 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
17573 DAG.getConstant(1, DL, MVT::i16)),
17574 DAG.getConstant(3, DL, MVT::i16));
17576 return DAG.getNode((VT.getSizeInBits() < 16 ?
17577 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
17580 /// \brief Lower a vector CTLZ using native supported vector CTLZ instruction.
17582 // 1. i32/i64 128/256-bit vector (native support require VLX) are expended
17583 // to 512-bit vector.
17584 // 2. i8/i16 vector implemented using dword LZCNT vector instruction
17585 // ( sub(trunc(lzcnt(zext32(x)))) ). In case zext32(x) is illegal,
17586 // split the vector, perform operation on it's Lo a Hi part and
17587 // concatenate the results.
17588 static SDValue LowerVectorCTLZ_AVX512(SDValue Op, SelectionDAG &DAG) {
17590 MVT VT = Op.getSimpleValueType();
17591 MVT EltVT = VT.getVectorElementType();
17592 unsigned NumElems = VT.getVectorNumElements();
17594 if (EltVT == MVT::i64 || EltVT == MVT::i32) {
17595 // Extend to 512 bit vector.
17596 assert((VT.is256BitVector() || VT.is128BitVector()) &&
17597 "Unsupported value type for operation");
17599 MVT NewVT = MVT::getVectorVT(EltVT, 512 / VT.getScalarSizeInBits());
17600 SDValue Vec512 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, NewVT,
17601 DAG.getUNDEF(NewVT),
17603 DAG.getIntPtrConstant(0, dl));
17604 SDValue CtlzNode = DAG.getNode(ISD::CTLZ, dl, NewVT, Vec512);
17606 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, CtlzNode,
17607 DAG.getIntPtrConstant(0, dl));
17610 assert((EltVT == MVT::i8 || EltVT == MVT::i16) &&
17611 "Unsupported element type");
17613 if (16 < NumElems) {
17614 // Split vector, it's Lo and Hi parts will be handled in next iteration.
17616 std::tie(Lo, Hi) = DAG.SplitVector(Op.getOperand(0), dl);
17617 MVT OutVT = MVT::getVectorVT(EltVT, NumElems/2);
17619 Lo = DAG.getNode(Op.getOpcode(), dl, OutVT, Lo);
17620 Hi = DAG.getNode(Op.getOpcode(), dl, OutVT, Hi);
17622 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Lo, Hi);
17625 MVT NewVT = MVT::getVectorVT(MVT::i32, NumElems);
17627 assert((NewVT.is256BitVector() || NewVT.is512BitVector()) &&
17628 "Unsupported value type for operation");
17630 // Use native supported vector instruction vplzcntd.
17631 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, NewVT, Op.getOperand(0));
17632 SDValue CtlzNode = DAG.getNode(ISD::CTLZ, dl, NewVT, Op);
17633 SDValue TruncNode = DAG.getNode(ISD::TRUNCATE, dl, VT, CtlzNode);
17634 SDValue Delta = DAG.getConstant(32 - EltVT.getSizeInBits(), dl, VT);
17636 return DAG.getNode(ISD::SUB, dl, VT, TruncNode, Delta);
17639 static SDValue LowerCTLZ(SDValue Op, const X86Subtarget *Subtarget,
17640 SelectionDAG &DAG) {
17641 MVT VT = Op.getSimpleValueType();
17643 unsigned NumBits = VT.getSizeInBits();
17646 if (VT.isVector() && Subtarget->hasAVX512())
17647 return LowerVectorCTLZ_AVX512(Op, DAG);
17649 Op = Op.getOperand(0);
17650 if (VT == MVT::i8) {
17651 // Zero extend to i32 since there is not an i8 bsr.
17653 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
17656 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
17657 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
17658 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
17660 // If src is zero (i.e. bsr sets ZF), returns NumBits.
17663 DAG.getConstant(NumBits + NumBits - 1, dl, OpVT),
17664 DAG.getConstant(X86::COND_E, dl, MVT::i8),
17667 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops);
17669 // Finally xor with NumBits-1.
17670 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
17671 DAG.getConstant(NumBits - 1, dl, OpVT));
17674 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
17678 static SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, const X86Subtarget *Subtarget,
17679 SelectionDAG &DAG) {
17680 MVT VT = Op.getSimpleValueType();
17682 unsigned NumBits = VT.getSizeInBits();
17685 if (VT.isVector() && Subtarget->hasAVX512())
17686 return LowerVectorCTLZ_AVX512(Op, DAG);
17688 Op = Op.getOperand(0);
17689 if (VT == MVT::i8) {
17690 // Zero extend to i32 since there is not an i8 bsr.
17692 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
17695 // Issue a bsr (scan bits in reverse).
17696 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
17697 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
17699 // And xor with NumBits-1.
17700 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
17701 DAG.getConstant(NumBits - 1, dl, OpVT));
17704 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
17708 static SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
17709 MVT VT = Op.getSimpleValueType();
17710 unsigned NumBits = VT.getScalarSizeInBits();
17713 if (VT.isVector()) {
17714 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
17716 SDValue N0 = Op.getOperand(0);
17717 SDValue Zero = DAG.getConstant(0, dl, VT);
17719 // lsb(x) = (x & -x)
17720 SDValue LSB = DAG.getNode(ISD::AND, dl, VT, N0,
17721 DAG.getNode(ISD::SUB, dl, VT, Zero, N0));
17723 // cttz_undef(x) = (width - 1) - ctlz(lsb)
17724 if (Op.getOpcode() == ISD::CTTZ_ZERO_UNDEF &&
17725 TLI.isOperationLegal(ISD::CTLZ, VT)) {
17726 SDValue WidthMinusOne = DAG.getConstant(NumBits - 1, dl, VT);
17727 return DAG.getNode(ISD::SUB, dl, VT, WidthMinusOne,
17728 DAG.getNode(ISD::CTLZ, dl, VT, LSB));
17731 // cttz(x) = ctpop(lsb - 1)
17732 SDValue One = DAG.getConstant(1, dl, VT);
17733 return DAG.getNode(ISD::CTPOP, dl, VT,
17734 DAG.getNode(ISD::SUB, dl, VT, LSB, One));
17737 assert(Op.getOpcode() == ISD::CTTZ &&
17738 "Only scalar CTTZ requires custom lowering");
17740 // Issue a bsf (scan bits forward) which also sets EFLAGS.
17741 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
17742 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op.getOperand(0));
17744 // If src is zero (i.e. bsf sets ZF), returns NumBits.
17747 DAG.getConstant(NumBits, dl, VT),
17748 DAG.getConstant(X86::COND_E, dl, MVT::i8),
17751 return DAG.getNode(X86ISD::CMOV, dl, VT, Ops);
17754 // Lower256IntArith - Break a 256-bit integer operation into two new 128-bit
17755 // ones, and then concatenate the result back.
17756 static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) {
17757 MVT VT = Op.getSimpleValueType();
17759 assert(VT.is256BitVector() && VT.isInteger() &&
17760 "Unsupported value type for operation");
17762 unsigned NumElems = VT.getVectorNumElements();
17765 // Extract the LHS vectors
17766 SDValue LHS = Op.getOperand(0);
17767 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
17768 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
17770 // Extract the RHS vectors
17771 SDValue RHS = Op.getOperand(1);
17772 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
17773 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
17775 MVT EltVT = VT.getVectorElementType();
17776 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
17778 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
17779 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1),
17780 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
17783 static SDValue LowerADD(SDValue Op, SelectionDAG &DAG) {
17784 if (Op.getValueType() == MVT::i1)
17785 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
17786 Op.getOperand(0), Op.getOperand(1));
17787 assert(Op.getSimpleValueType().is256BitVector() &&
17788 Op.getSimpleValueType().isInteger() &&
17789 "Only handle AVX 256-bit vector integer operation");
17790 return Lower256IntArith(Op, DAG);
17793 static SDValue LowerSUB(SDValue Op, SelectionDAG &DAG) {
17794 if (Op.getValueType() == MVT::i1)
17795 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
17796 Op.getOperand(0), Op.getOperand(1));
17797 assert(Op.getSimpleValueType().is256BitVector() &&
17798 Op.getSimpleValueType().isInteger() &&
17799 "Only handle AVX 256-bit vector integer operation");
17800 return Lower256IntArith(Op, DAG);
17803 static SDValue LowerMINMAX(SDValue Op, SelectionDAG &DAG) {
17804 assert(Op.getSimpleValueType().is256BitVector() &&
17805 Op.getSimpleValueType().isInteger() &&
17806 "Only handle AVX 256-bit vector integer operation");
17807 return Lower256IntArith(Op, DAG);
17810 static SDValue LowerMUL(SDValue Op, const X86Subtarget *Subtarget,
17811 SelectionDAG &DAG) {
17813 MVT VT = Op.getSimpleValueType();
17816 return DAG.getNode(ISD::AND, dl, VT, Op.getOperand(0), Op.getOperand(1));
17818 // Decompose 256-bit ops into smaller 128-bit ops.
17819 if (VT.is256BitVector() && !Subtarget->hasInt256())
17820 return Lower256IntArith(Op, DAG);
17822 SDValue A = Op.getOperand(0);
17823 SDValue B = Op.getOperand(1);
17825 // Lower v16i8/v32i8 mul as promotion to v8i16/v16i16 vector
17826 // pairs, multiply and truncate.
17827 if (VT == MVT::v16i8 || VT == MVT::v32i8) {
17828 if (Subtarget->hasInt256()) {
17829 if (VT == MVT::v32i8) {
17830 MVT SubVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() / 2);
17831 SDValue Lo = DAG.getIntPtrConstant(0, dl);
17832 SDValue Hi = DAG.getIntPtrConstant(VT.getVectorNumElements() / 2, dl);
17833 SDValue ALo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Lo);
17834 SDValue BLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Lo);
17835 SDValue AHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Hi);
17836 SDValue BHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Hi);
17837 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
17838 DAG.getNode(ISD::MUL, dl, SubVT, ALo, BLo),
17839 DAG.getNode(ISD::MUL, dl, SubVT, AHi, BHi));
17842 MVT ExVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements());
17843 return DAG.getNode(
17844 ISD::TRUNCATE, dl, VT,
17845 DAG.getNode(ISD::MUL, dl, ExVT,
17846 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, A),
17847 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, B)));
17850 assert(VT == MVT::v16i8 &&
17851 "Pre-AVX2 support only supports v16i8 multiplication");
17852 MVT ExVT = MVT::v8i16;
17854 // Extract the lo parts and sign extend to i16
17856 if (Subtarget->hasSSE41()) {
17857 ALo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, A);
17858 BLo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, B);
17860 const int ShufMask[] = {-1, 0, -1, 1, -1, 2, -1, 3,
17861 -1, 4, -1, 5, -1, 6, -1, 7};
17862 ALo = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
17863 BLo = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
17864 ALo = DAG.getBitcast(ExVT, ALo);
17865 BLo = DAG.getBitcast(ExVT, BLo);
17866 ALo = DAG.getNode(ISD::SRA, dl, ExVT, ALo, DAG.getConstant(8, dl, ExVT));
17867 BLo = DAG.getNode(ISD::SRA, dl, ExVT, BLo, DAG.getConstant(8, dl, ExVT));
17870 // Extract the hi parts and sign extend to i16
17872 if (Subtarget->hasSSE41()) {
17873 const int ShufMask[] = {8, 9, 10, 11, 12, 13, 14, 15,
17874 -1, -1, -1, -1, -1, -1, -1, -1};
17875 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
17876 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
17877 AHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, AHi);
17878 BHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, BHi);
17880 const int ShufMask[] = {-1, 8, -1, 9, -1, 10, -1, 11,
17881 -1, 12, -1, 13, -1, 14, -1, 15};
17882 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
17883 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
17884 AHi = DAG.getBitcast(ExVT, AHi);
17885 BHi = DAG.getBitcast(ExVT, BHi);
17886 AHi = DAG.getNode(ISD::SRA, dl, ExVT, AHi, DAG.getConstant(8, dl, ExVT));
17887 BHi = DAG.getNode(ISD::SRA, dl, ExVT, BHi, DAG.getConstant(8, dl, ExVT));
17890 // Multiply, mask the lower 8bits of the lo/hi results and pack
17891 SDValue RLo = DAG.getNode(ISD::MUL, dl, ExVT, ALo, BLo);
17892 SDValue RHi = DAG.getNode(ISD::MUL, dl, ExVT, AHi, BHi);
17893 RLo = DAG.getNode(ISD::AND, dl, ExVT, RLo, DAG.getConstant(255, dl, ExVT));
17894 RHi = DAG.getNode(ISD::AND, dl, ExVT, RHi, DAG.getConstant(255, dl, ExVT));
17895 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
17898 // Lower v4i32 mul as 2x shuffle, 2x pmuludq, 2x shuffle.
17899 if (VT == MVT::v4i32) {
17900 assert(Subtarget->hasSSE2() && !Subtarget->hasSSE41() &&
17901 "Should not custom lower when pmuldq is available!");
17903 // Extract the odd parts.
17904 static const int UnpackMask[] = { 1, -1, 3, -1 };
17905 SDValue Aodds = DAG.getVectorShuffle(VT, dl, A, A, UnpackMask);
17906 SDValue Bodds = DAG.getVectorShuffle(VT, dl, B, B, UnpackMask);
17908 // Multiply the even parts.
17909 SDValue Evens = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, A, B);
17910 // Now multiply odd parts.
17911 SDValue Odds = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, Aodds, Bodds);
17913 Evens = DAG.getBitcast(VT, Evens);
17914 Odds = DAG.getBitcast(VT, Odds);
17916 // Merge the two vectors back together with a shuffle. This expands into 2
17918 static const int ShufMask[] = { 0, 4, 2, 6 };
17919 return DAG.getVectorShuffle(VT, dl, Evens, Odds, ShufMask);
17922 assert((VT == MVT::v2i64 || VT == MVT::v4i64 || VT == MVT::v8i64) &&
17923 "Only know how to lower V2I64/V4I64/V8I64 multiply");
17925 // Ahi = psrlqi(a, 32);
17926 // Bhi = psrlqi(b, 32);
17928 // AloBlo = pmuludq(a, b);
17929 // AloBhi = pmuludq(a, Bhi);
17930 // AhiBlo = pmuludq(Ahi, b);
17932 // AloBhi = psllqi(AloBhi, 32);
17933 // AhiBlo = psllqi(AhiBlo, 32);
17934 // return AloBlo + AloBhi + AhiBlo;
17936 SDValue Ahi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, A, 32, DAG);
17937 SDValue Bhi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, B, 32, DAG);
17939 SDValue AhiBlo = Ahi;
17940 SDValue AloBhi = Bhi;
17941 // Bit cast to 32-bit vectors for MULUDQ
17942 MVT MulVT = (VT == MVT::v2i64) ? MVT::v4i32 :
17943 (VT == MVT::v4i64) ? MVT::v8i32 : MVT::v16i32;
17944 A = DAG.getBitcast(MulVT, A);
17945 B = DAG.getBitcast(MulVT, B);
17946 Ahi = DAG.getBitcast(MulVT, Ahi);
17947 Bhi = DAG.getBitcast(MulVT, Bhi);
17949 SDValue AloBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, B);
17950 // After shifting right const values the result may be all-zero.
17951 if (!ISD::isBuildVectorAllZeros(Ahi.getNode())) {
17952 AhiBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Ahi, B);
17953 AhiBlo = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AhiBlo, 32, DAG);
17955 if (!ISD::isBuildVectorAllZeros(Bhi.getNode())) {
17956 AloBhi = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, Bhi);
17957 AloBhi = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AloBhi, 32, DAG);
17960 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
17961 return DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
17964 SDValue X86TargetLowering::LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const {
17965 assert(Subtarget->isTargetWin64() && "Unexpected target");
17966 EVT VT = Op.getValueType();
17967 assert(VT.isInteger() && VT.getSizeInBits() == 128 &&
17968 "Unexpected return type for lowering");
17972 switch (Op->getOpcode()) {
17973 default: llvm_unreachable("Unexpected request for libcall!");
17974 case ISD::SDIV: isSigned = true; LC = RTLIB::SDIV_I128; break;
17975 case ISD::UDIV: isSigned = false; LC = RTLIB::UDIV_I128; break;
17976 case ISD::SREM: isSigned = true; LC = RTLIB::SREM_I128; break;
17977 case ISD::UREM: isSigned = false; LC = RTLIB::UREM_I128; break;
17978 case ISD::SDIVREM: isSigned = true; LC = RTLIB::SDIVREM_I128; break;
17979 case ISD::UDIVREM: isSigned = false; LC = RTLIB::UDIVREM_I128; break;
17983 SDValue InChain = DAG.getEntryNode();
17985 TargetLowering::ArgListTy Args;
17986 TargetLowering::ArgListEntry Entry;
17987 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
17988 EVT ArgVT = Op->getOperand(i).getValueType();
17989 assert(ArgVT.isInteger() && ArgVT.getSizeInBits() == 128 &&
17990 "Unexpected argument type for lowering");
17991 SDValue StackPtr = DAG.CreateStackTemporary(ArgVT, 16);
17992 Entry.Node = StackPtr;
17993 InChain = DAG.getStore(InChain, dl, Op->getOperand(i), StackPtr, MachinePointerInfo(),
17995 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
17996 Entry.Ty = PointerType::get(ArgTy,0);
17997 Entry.isSExt = false;
17998 Entry.isZExt = false;
17999 Args.push_back(Entry);
18002 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
18003 getPointerTy(DAG.getDataLayout()));
18005 TargetLowering::CallLoweringInfo CLI(DAG);
18006 CLI.setDebugLoc(dl).setChain(InChain)
18007 .setCallee(getLibcallCallingConv(LC),
18008 static_cast<EVT>(MVT::v2i64).getTypeForEVT(*DAG.getContext()),
18009 Callee, std::move(Args), 0)
18010 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
18012 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
18013 return DAG.getBitcast(VT, CallInfo.first);
18016 static SDValue LowerMUL_LOHI(SDValue Op, const X86Subtarget *Subtarget,
18017 SelectionDAG &DAG) {
18018 SDValue Op0 = Op.getOperand(0), Op1 = Op.getOperand(1);
18019 MVT VT = Op0.getSimpleValueType();
18022 assert((VT == MVT::v4i32 && Subtarget->hasSSE2()) ||
18023 (VT == MVT::v8i32 && Subtarget->hasInt256()));
18025 // PMULxD operations multiply each even value (starting at 0) of LHS with
18026 // the related value of RHS and produce a widen result.
18027 // E.g., PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
18028 // => <2 x i64> <ae|cg>
18030 // In other word, to have all the results, we need to perform two PMULxD:
18031 // 1. one with the even values.
18032 // 2. one with the odd values.
18033 // To achieve #2, with need to place the odd values at an even position.
18035 // Place the odd value at an even position (basically, shift all values 1
18036 // step to the left):
18037 const int Mask[] = {1, -1, 3, -1, 5, -1, 7, -1};
18038 // <a|b|c|d> => <b|undef|d|undef>
18039 SDValue Odd0 = DAG.getVectorShuffle(VT, dl, Op0, Op0, Mask);
18040 // <e|f|g|h> => <f|undef|h|undef>
18041 SDValue Odd1 = DAG.getVectorShuffle(VT, dl, Op1, Op1, Mask);
18043 // Emit two multiplies, one for the lower 2 ints and one for the higher 2
18045 MVT MulVT = VT == MVT::v4i32 ? MVT::v2i64 : MVT::v4i64;
18046 bool IsSigned = Op->getOpcode() == ISD::SMUL_LOHI;
18048 (!IsSigned || !Subtarget->hasSSE41()) ? X86ISD::PMULUDQ : X86ISD::PMULDQ;
18049 // PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
18050 // => <2 x i64> <ae|cg>
18051 SDValue Mul1 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Op0, Op1));
18052 // PMULUDQ <4 x i32> <b|undef|d|undef>, <4 x i32> <f|undef|h|undef>
18053 // => <2 x i64> <bf|dh>
18054 SDValue Mul2 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Odd0, Odd1));
18056 // Shuffle it back into the right order.
18057 SDValue Highs, Lows;
18058 if (VT == MVT::v8i32) {
18059 const int HighMask[] = {1, 9, 3, 11, 5, 13, 7, 15};
18060 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
18061 const int LowMask[] = {0, 8, 2, 10, 4, 12, 6, 14};
18062 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
18064 const int HighMask[] = {1, 5, 3, 7};
18065 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
18066 const int LowMask[] = {0, 4, 2, 6};
18067 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
18070 // If we have a signed multiply but no PMULDQ fix up the high parts of a
18071 // unsigned multiply.
18072 if (IsSigned && !Subtarget->hasSSE41()) {
18073 SDValue ShAmt = DAG.getConstant(
18075 DAG.getTargetLoweringInfo().getShiftAmountTy(VT, DAG.getDataLayout()));
18076 SDValue T1 = DAG.getNode(ISD::AND, dl, VT,
18077 DAG.getNode(ISD::SRA, dl, VT, Op0, ShAmt), Op1);
18078 SDValue T2 = DAG.getNode(ISD::AND, dl, VT,
18079 DAG.getNode(ISD::SRA, dl, VT, Op1, ShAmt), Op0);
18081 SDValue Fixup = DAG.getNode(ISD::ADD, dl, VT, T1, T2);
18082 Highs = DAG.getNode(ISD::SUB, dl, VT, Highs, Fixup);
18085 // The first result of MUL_LOHI is actually the low value, followed by the
18087 SDValue Ops[] = {Lows, Highs};
18088 return DAG.getMergeValues(Ops, dl);
18091 // Return true if the required (according to Opcode) shift-imm form is natively
18092 // supported by the Subtarget
18093 static bool SupportedVectorShiftWithImm(MVT VT, const X86Subtarget *Subtarget,
18095 if (VT.getScalarSizeInBits() < 16)
18098 if (VT.is512BitVector() &&
18099 (VT.getScalarSizeInBits() > 16 || Subtarget->hasBWI()))
18102 bool LShift = VT.is128BitVector() ||
18103 (VT.is256BitVector() && Subtarget->hasInt256());
18105 bool AShift = LShift && (Subtarget->hasVLX() ||
18106 (VT != MVT::v2i64 && VT != MVT::v4i64));
18107 return (Opcode == ISD::SRA) ? AShift : LShift;
18110 // The shift amount is a variable, but it is the same for all vector lanes.
18111 // These instructions are defined together with shift-immediate.
18113 bool SupportedVectorShiftWithBaseAmnt(MVT VT, const X86Subtarget *Subtarget,
18115 return SupportedVectorShiftWithImm(VT, Subtarget, Opcode);
18118 // Return true if the required (according to Opcode) variable-shift form is
18119 // natively supported by the Subtarget
18120 static bool SupportedVectorVarShift(MVT VT, const X86Subtarget *Subtarget,
18123 if (!Subtarget->hasInt256() || VT.getScalarSizeInBits() < 16)
18126 // vXi16 supported only on AVX-512, BWI
18127 if (VT.getScalarSizeInBits() == 16 && !Subtarget->hasBWI())
18130 if (VT.is512BitVector() || Subtarget->hasVLX())
18133 bool LShift = VT.is128BitVector() || VT.is256BitVector();
18134 bool AShift = LShift && VT != MVT::v2i64 && VT != MVT::v4i64;
18135 return (Opcode == ISD::SRA) ? AShift : LShift;
18138 static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
18139 const X86Subtarget *Subtarget) {
18140 MVT VT = Op.getSimpleValueType();
18142 SDValue R = Op.getOperand(0);
18143 SDValue Amt = Op.getOperand(1);
18145 unsigned X86Opc = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
18146 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
18148 auto ArithmeticShiftRight64 = [&](uint64_t ShiftAmt) {
18149 assert((VT == MVT::v2i64 || VT == MVT::v4i64) && "Unexpected SRA type");
18150 MVT ExVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() * 2);
18151 SDValue Ex = DAG.getBitcast(ExVT, R);
18153 if (ShiftAmt >= 32) {
18154 // Splat sign to upper i32 dst, and SRA upper i32 src to lower i32.
18156 getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex, 31, DAG);
18157 SDValue Lower = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex,
18158 ShiftAmt - 32, DAG);
18159 if (VT == MVT::v2i64)
18160 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {5, 1, 7, 3});
18161 if (VT == MVT::v4i64)
18162 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower,
18163 {9, 1, 11, 3, 13, 5, 15, 7});
18165 // SRA upper i32, SHL whole i64 and select lower i32.
18166 SDValue Upper = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex,
18169 getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt, DAG);
18170 Lower = DAG.getBitcast(ExVT, Lower);
18171 if (VT == MVT::v2i64)
18172 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {4, 1, 6, 3});
18173 if (VT == MVT::v4i64)
18174 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower,
18175 {8, 1, 10, 3, 12, 5, 14, 7});
18177 return DAG.getBitcast(VT, Ex);
18180 // Optimize shl/srl/sra with constant shift amount.
18181 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
18182 if (auto *ShiftConst = BVAmt->getConstantSplatNode()) {
18183 uint64_t ShiftAmt = ShiftConst->getZExtValue();
18185 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
18186 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
18188 // i64 SRA needs to be performed as partial shifts.
18189 if ((VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
18190 Op.getOpcode() == ISD::SRA && !Subtarget->hasXOP())
18191 return ArithmeticShiftRight64(ShiftAmt);
18193 if (VT == MVT::v16i8 || (Subtarget->hasInt256() && VT == MVT::v32i8)) {
18194 unsigned NumElts = VT.getVectorNumElements();
18195 MVT ShiftVT = MVT::getVectorVT(MVT::i16, NumElts / 2);
18197 // Simple i8 add case
18198 if (Op.getOpcode() == ISD::SHL && ShiftAmt == 1)
18199 return DAG.getNode(ISD::ADD, dl, VT, R, R);
18201 // ashr(R, 7) === cmp_slt(R, 0)
18202 if (Op.getOpcode() == ISD::SRA && ShiftAmt == 7) {
18203 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
18204 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
18207 // XOP can shift v16i8 directly instead of as shift v8i16 + mask.
18208 if (VT == MVT::v16i8 && Subtarget->hasXOP())
18211 if (Op.getOpcode() == ISD::SHL) {
18212 // Make a large shift.
18213 SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, ShiftVT,
18215 SHL = DAG.getBitcast(VT, SHL);
18216 // Zero out the rightmost bits.
18217 SmallVector<SDValue, 32> V(
18218 NumElts, DAG.getConstant(uint8_t(-1U << ShiftAmt), dl, MVT::i8));
18219 return DAG.getNode(ISD::AND, dl, VT, SHL,
18220 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
18222 if (Op.getOpcode() == ISD::SRL) {
18223 // Make a large shift.
18224 SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, ShiftVT,
18226 SRL = DAG.getBitcast(VT, SRL);
18227 // Zero out the leftmost bits.
18228 SmallVector<SDValue, 32> V(
18229 NumElts, DAG.getConstant(uint8_t(-1U) >> ShiftAmt, dl, MVT::i8));
18230 return DAG.getNode(ISD::AND, dl, VT, SRL,
18231 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
18233 if (Op.getOpcode() == ISD::SRA) {
18234 // ashr(R, Amt) === sub(xor(lshr(R, Amt), Mask), Mask)
18235 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
18236 SmallVector<SDValue, 32> V(NumElts,
18237 DAG.getConstant(128 >> ShiftAmt, dl,
18239 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V);
18240 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
18241 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
18244 llvm_unreachable("Unknown shift opcode.");
18249 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
18250 if (!Subtarget->is64Bit() && !Subtarget->hasXOP() &&
18251 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64))) {
18253 // Peek through any splat that was introduced for i64 shift vectorization.
18254 int SplatIndex = -1;
18255 if (ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt.getNode()))
18256 if (SVN->isSplat()) {
18257 SplatIndex = SVN->getSplatIndex();
18258 Amt = Amt.getOperand(0);
18259 assert(SplatIndex < (int)VT.getVectorNumElements() &&
18260 "Splat shuffle referencing second operand");
18263 if (Amt.getOpcode() != ISD::BITCAST ||
18264 Amt.getOperand(0).getOpcode() != ISD::BUILD_VECTOR)
18267 Amt = Amt.getOperand(0);
18268 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
18269 VT.getVectorNumElements();
18270 unsigned RatioInLog2 = Log2_32_Ceil(Ratio);
18271 uint64_t ShiftAmt = 0;
18272 unsigned BaseOp = (SplatIndex < 0 ? 0 : SplatIndex * Ratio);
18273 for (unsigned i = 0; i != Ratio; ++i) {
18274 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i + BaseOp));
18278 ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2)));
18281 // Check remaining shift amounts (if not a splat).
18282 if (SplatIndex < 0) {
18283 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
18284 uint64_t ShAmt = 0;
18285 for (unsigned j = 0; j != Ratio; ++j) {
18286 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i + j));
18290 ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2)));
18292 if (ShAmt != ShiftAmt)
18297 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
18298 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
18300 if (Op.getOpcode() == ISD::SRA)
18301 return ArithmeticShiftRight64(ShiftAmt);
18307 static SDValue LowerScalarVariableShift(SDValue Op, SelectionDAG &DAG,
18308 const X86Subtarget* Subtarget) {
18309 MVT VT = Op.getSimpleValueType();
18311 SDValue R = Op.getOperand(0);
18312 SDValue Amt = Op.getOperand(1);
18314 unsigned X86OpcI = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
18315 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
18317 unsigned X86OpcV = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHL :
18318 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRL : X86ISD::VSRA;
18320 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode())) {
18322 MVT EltVT = VT.getVectorElementType();
18324 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Amt)) {
18325 // Check if this build_vector node is doing a splat.
18326 // If so, then set BaseShAmt equal to the splat value.
18327 BaseShAmt = BV->getSplatValue();
18328 if (BaseShAmt && BaseShAmt.getOpcode() == ISD::UNDEF)
18329 BaseShAmt = SDValue();
18331 if (Amt.getOpcode() == ISD::EXTRACT_SUBVECTOR)
18332 Amt = Amt.getOperand(0);
18334 ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt);
18335 if (SVN && SVN->isSplat()) {
18336 unsigned SplatIdx = (unsigned)SVN->getSplatIndex();
18337 SDValue InVec = Amt.getOperand(0);
18338 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
18339 assert((SplatIdx < InVec.getSimpleValueType().getVectorNumElements()) &&
18340 "Unexpected shuffle index found!");
18341 BaseShAmt = InVec.getOperand(SplatIdx);
18342 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
18343 if (ConstantSDNode *C =
18344 dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
18345 if (C->getZExtValue() == SplatIdx)
18346 BaseShAmt = InVec.getOperand(1);
18351 // Avoid introducing an extract element from a shuffle.
18352 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InVec,
18353 DAG.getIntPtrConstant(SplatIdx, dl));
18357 if (BaseShAmt.getNode()) {
18358 assert(EltVT.bitsLE(MVT::i64) && "Unexpected element type!");
18359 if (EltVT != MVT::i64 && EltVT.bitsGT(MVT::i32))
18360 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, BaseShAmt);
18361 else if (EltVT.bitsLT(MVT::i32))
18362 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt);
18364 return getTargetVShiftNode(X86OpcI, dl, VT, R, BaseShAmt, DAG);
18368 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
18369 if (!Subtarget->is64Bit() && VT == MVT::v2i64 &&
18370 Amt.getOpcode() == ISD::BITCAST &&
18371 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
18372 Amt = Amt.getOperand(0);
18373 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
18374 VT.getVectorNumElements();
18375 std::vector<SDValue> Vals(Ratio);
18376 for (unsigned i = 0; i != Ratio; ++i)
18377 Vals[i] = Amt.getOperand(i);
18378 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
18379 for (unsigned j = 0; j != Ratio; ++j)
18380 if (Vals[j] != Amt.getOperand(i + j))
18384 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode()))
18385 return DAG.getNode(X86OpcV, dl, VT, R, Op.getOperand(1));
18390 static SDValue LowerShift(SDValue Op, const X86Subtarget* Subtarget,
18391 SelectionDAG &DAG) {
18392 MVT VT = Op.getSimpleValueType();
18394 SDValue R = Op.getOperand(0);
18395 SDValue Amt = Op.getOperand(1);
18397 assert(VT.isVector() && "Custom lowering only for vector shifts!");
18398 assert(Subtarget->hasSSE2() && "Only custom lower when we have SSE2!");
18400 if (SDValue V = LowerScalarImmediateShift(Op, DAG, Subtarget))
18403 if (SDValue V = LowerScalarVariableShift(Op, DAG, Subtarget))
18406 if (SupportedVectorVarShift(VT, Subtarget, Op.getOpcode()))
18409 // XOP has 128-bit variable logical/arithmetic shifts.
18410 // +ve/-ve Amt = shift left/right.
18411 if (Subtarget->hasXOP() &&
18412 (VT == MVT::v2i64 || VT == MVT::v4i32 ||
18413 VT == MVT::v8i16 || VT == MVT::v16i8)) {
18414 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SRA) {
18415 SDValue Zero = getZeroVector(VT, Subtarget, DAG, dl);
18416 Amt = DAG.getNode(ISD::SUB, dl, VT, Zero, Amt);
18418 if (Op.getOpcode() == ISD::SHL || Op.getOpcode() == ISD::SRL)
18419 return DAG.getNode(X86ISD::VPSHL, dl, VT, R, Amt);
18420 if (Op.getOpcode() == ISD::SRA)
18421 return DAG.getNode(X86ISD::VPSHA, dl, VT, R, Amt);
18424 // 2i64 vector logical shifts can efficiently avoid scalarization - do the
18425 // shifts per-lane and then shuffle the partial results back together.
18426 if (VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) {
18427 // Splat the shift amounts so the scalar shifts above will catch it.
18428 SDValue Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {0, 0});
18429 SDValue Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {1, 1});
18430 SDValue R0 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt0);
18431 SDValue R1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt1);
18432 return DAG.getVectorShuffle(VT, dl, R0, R1, {0, 3});
18435 // i64 vector arithmetic shift can be emulated with the transform:
18436 // M = lshr(SIGN_BIT, Amt)
18437 // ashr(R, Amt) === sub(xor(lshr(R, Amt), M), M)
18438 if ((VT == MVT::v2i64 || (VT == MVT::v4i64 && Subtarget->hasInt256())) &&
18439 Op.getOpcode() == ISD::SRA) {
18440 SDValue S = DAG.getConstant(APInt::getSignBit(64), dl, VT);
18441 SDValue M = DAG.getNode(ISD::SRL, dl, VT, S, Amt);
18442 R = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
18443 R = DAG.getNode(ISD::XOR, dl, VT, R, M);
18444 R = DAG.getNode(ISD::SUB, dl, VT, R, M);
18448 // If possible, lower this packed shift into a vector multiply instead of
18449 // expanding it into a sequence of scalar shifts.
18450 // Do this only if the vector shift count is a constant build_vector.
18451 if (Op.getOpcode() == ISD::SHL &&
18452 (VT == MVT::v8i16 || VT == MVT::v4i32 ||
18453 (Subtarget->hasInt256() && VT == MVT::v16i16)) &&
18454 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18455 SmallVector<SDValue, 8> Elts;
18456 MVT SVT = VT.getVectorElementType();
18457 unsigned SVTBits = SVT.getSizeInBits();
18458 APInt One(SVTBits, 1);
18459 unsigned NumElems = VT.getVectorNumElements();
18461 for (unsigned i=0; i !=NumElems; ++i) {
18462 SDValue Op = Amt->getOperand(i);
18463 if (Op->getOpcode() == ISD::UNDEF) {
18464 Elts.push_back(Op);
18468 ConstantSDNode *ND = cast<ConstantSDNode>(Op);
18469 APInt C(SVTBits, ND->getAPIntValue().getZExtValue());
18470 uint64_t ShAmt = C.getZExtValue();
18471 if (ShAmt >= SVTBits) {
18472 Elts.push_back(DAG.getUNDEF(SVT));
18475 Elts.push_back(DAG.getConstant(One.shl(ShAmt), dl, SVT));
18477 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
18478 return DAG.getNode(ISD::MUL, dl, VT, R, BV);
18481 // Lower SHL with variable shift amount.
18482 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
18483 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, dl, VT));
18485 Op = DAG.getNode(ISD::ADD, dl, VT, Op,
18486 DAG.getConstant(0x3f800000U, dl, VT));
18487 Op = DAG.getBitcast(MVT::v4f32, Op);
18488 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
18489 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
18492 // If possible, lower this shift as a sequence of two shifts by
18493 // constant plus a MOVSS/MOVSD instead of scalarizing it.
18495 // (v4i32 (srl A, (build_vector < X, Y, Y, Y>)))
18497 // Could be rewritten as:
18498 // (v4i32 (MOVSS (srl A, <Y,Y,Y,Y>), (srl A, <X,X,X,X>)))
18500 // The advantage is that the two shifts from the example would be
18501 // lowered as X86ISD::VSRLI nodes. This would be cheaper than scalarizing
18502 // the vector shift into four scalar shifts plus four pairs of vector
18504 if ((VT == MVT::v8i16 || VT == MVT::v4i32) &&
18505 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18506 unsigned TargetOpcode = X86ISD::MOVSS;
18507 bool CanBeSimplified;
18508 // The splat value for the first packed shift (the 'X' from the example).
18509 SDValue Amt1 = Amt->getOperand(0);
18510 // The splat value for the second packed shift (the 'Y' from the example).
18511 SDValue Amt2 = (VT == MVT::v4i32) ? Amt->getOperand(1) :
18512 Amt->getOperand(2);
18514 // See if it is possible to replace this node with a sequence of
18515 // two shifts followed by a MOVSS/MOVSD
18516 if (VT == MVT::v4i32) {
18517 // Check if it is legal to use a MOVSS.
18518 CanBeSimplified = Amt2 == Amt->getOperand(2) &&
18519 Amt2 == Amt->getOperand(3);
18520 if (!CanBeSimplified) {
18521 // Otherwise, check if we can still simplify this node using a MOVSD.
18522 CanBeSimplified = Amt1 == Amt->getOperand(1) &&
18523 Amt->getOperand(2) == Amt->getOperand(3);
18524 TargetOpcode = X86ISD::MOVSD;
18525 Amt2 = Amt->getOperand(2);
18528 // Do similar checks for the case where the machine value type
18530 CanBeSimplified = Amt1 == Amt->getOperand(1);
18531 for (unsigned i=3; i != 8 && CanBeSimplified; ++i)
18532 CanBeSimplified = Amt2 == Amt->getOperand(i);
18534 if (!CanBeSimplified) {
18535 TargetOpcode = X86ISD::MOVSD;
18536 CanBeSimplified = true;
18537 Amt2 = Amt->getOperand(4);
18538 for (unsigned i=0; i != 4 && CanBeSimplified; ++i)
18539 CanBeSimplified = Amt1 == Amt->getOperand(i);
18540 for (unsigned j=4; j != 8 && CanBeSimplified; ++j)
18541 CanBeSimplified = Amt2 == Amt->getOperand(j);
18545 if (CanBeSimplified && isa<ConstantSDNode>(Amt1) &&
18546 isa<ConstantSDNode>(Amt2)) {
18547 // Replace this node with two shifts followed by a MOVSS/MOVSD.
18548 MVT CastVT = MVT::v4i32;
18550 DAG.getConstant(cast<ConstantSDNode>(Amt1)->getAPIntValue(), dl, VT);
18551 SDValue Shift1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat1);
18553 DAG.getConstant(cast<ConstantSDNode>(Amt2)->getAPIntValue(), dl, VT);
18554 SDValue Shift2 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat2);
18555 if (TargetOpcode == X86ISD::MOVSD)
18556 CastVT = MVT::v2i64;
18557 SDValue BitCast1 = DAG.getBitcast(CastVT, Shift1);
18558 SDValue BitCast2 = DAG.getBitcast(CastVT, Shift2);
18559 SDValue Result = getTargetShuffleNode(TargetOpcode, dl, CastVT, BitCast2,
18561 return DAG.getBitcast(VT, Result);
18565 // v4i32 Non Uniform Shifts.
18566 // If the shift amount is constant we can shift each lane using the SSE2
18567 // immediate shifts, else we need to zero-extend each lane to the lower i64
18568 // and shift using the SSE2 variable shifts.
18569 // The separate results can then be blended together.
18570 if (VT == MVT::v4i32) {
18571 unsigned Opc = Op.getOpcode();
18572 SDValue Amt0, Amt1, Amt2, Amt3;
18573 if (ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18574 Amt0 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {0, 0, 0, 0});
18575 Amt1 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {1, 1, 1, 1});
18576 Amt2 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {2, 2, 2, 2});
18577 Amt3 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {3, 3, 3, 3});
18579 // ISD::SHL is handled above but we include it here for completeness.
18582 llvm_unreachable("Unknown target vector shift node");
18584 Opc = X86ISD::VSHL;
18587 Opc = X86ISD::VSRL;
18590 Opc = X86ISD::VSRA;
18593 // The SSE2 shifts use the lower i64 as the same shift amount for
18594 // all lanes and the upper i64 is ignored. These shuffle masks
18595 // optimally zero-extend each lanes on SSE2/SSE41/AVX targets.
18596 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
18597 Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Z, {0, 4, -1, -1});
18598 Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Z, {1, 5, -1, -1});
18599 Amt2 = DAG.getVectorShuffle(VT, dl, Amt, Z, {2, 6, -1, -1});
18600 Amt3 = DAG.getVectorShuffle(VT, dl, Amt, Z, {3, 7, -1, -1});
18603 SDValue R0 = DAG.getNode(Opc, dl, VT, R, Amt0);
18604 SDValue R1 = DAG.getNode(Opc, dl, VT, R, Amt1);
18605 SDValue R2 = DAG.getNode(Opc, dl, VT, R, Amt2);
18606 SDValue R3 = DAG.getNode(Opc, dl, VT, R, Amt3);
18607 SDValue R02 = DAG.getVectorShuffle(VT, dl, R0, R2, {0, -1, 6, -1});
18608 SDValue R13 = DAG.getVectorShuffle(VT, dl, R1, R3, {-1, 1, -1, 7});
18609 return DAG.getVectorShuffle(VT, dl, R02, R13, {0, 5, 2, 7});
18612 if (VT == MVT::v16i8 ||
18613 (VT == MVT::v32i8 && Subtarget->hasInt256() && !Subtarget->hasXOP())) {
18614 MVT ExtVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements() / 2);
18615 unsigned ShiftOpcode = Op->getOpcode();
18617 auto SignBitSelect = [&](MVT SelVT, SDValue Sel, SDValue V0, SDValue V1) {
18618 // On SSE41 targets we make use of the fact that VSELECT lowers
18619 // to PBLENDVB which selects bytes based just on the sign bit.
18620 if (Subtarget->hasSSE41()) {
18621 V0 = DAG.getBitcast(VT, V0);
18622 V1 = DAG.getBitcast(VT, V1);
18623 Sel = DAG.getBitcast(VT, Sel);
18624 return DAG.getBitcast(SelVT,
18625 DAG.getNode(ISD::VSELECT, dl, VT, Sel, V0, V1));
18627 // On pre-SSE41 targets we test for the sign bit by comparing to
18628 // zero - a negative value will set all bits of the lanes to true
18629 // and VSELECT uses that in its OR(AND(V0,C),AND(V1,~C)) lowering.
18630 SDValue Z = getZeroVector(SelVT, Subtarget, DAG, dl);
18631 SDValue C = DAG.getNode(X86ISD::PCMPGT, dl, SelVT, Z, Sel);
18632 return DAG.getNode(ISD::VSELECT, dl, SelVT, C, V0, V1);
18635 // Turn 'a' into a mask suitable for VSELECT: a = a << 5;
18636 // We can safely do this using i16 shifts as we're only interested in
18637 // the 3 lower bits of each byte.
18638 Amt = DAG.getBitcast(ExtVT, Amt);
18639 Amt = DAG.getNode(ISD::SHL, dl, ExtVT, Amt, DAG.getConstant(5, dl, ExtVT));
18640 Amt = DAG.getBitcast(VT, Amt);
18642 if (Op->getOpcode() == ISD::SHL || Op->getOpcode() == ISD::SRL) {
18643 // r = VSELECT(r, shift(r, 4), a);
18645 DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT));
18646 R = SignBitSelect(VT, Amt, M, R);
18649 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18651 // r = VSELECT(r, shift(r, 2), a);
18652 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT));
18653 R = SignBitSelect(VT, Amt, M, R);
18656 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18658 // return VSELECT(r, shift(r, 1), a);
18659 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT));
18660 R = SignBitSelect(VT, Amt, M, R);
18664 if (Op->getOpcode() == ISD::SRA) {
18665 // For SRA we need to unpack each byte to the higher byte of a i16 vector
18666 // so we can correctly sign extend. We don't care what happens to the
18668 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), Amt);
18669 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), Amt);
18670 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), R);
18671 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), R);
18672 ALo = DAG.getBitcast(ExtVT, ALo);
18673 AHi = DAG.getBitcast(ExtVT, AHi);
18674 RLo = DAG.getBitcast(ExtVT, RLo);
18675 RHi = DAG.getBitcast(ExtVT, RHi);
18677 // r = VSELECT(r, shift(r, 4), a);
18678 SDValue MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
18679 DAG.getConstant(4, dl, ExtVT));
18680 SDValue MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
18681 DAG.getConstant(4, dl, ExtVT));
18682 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
18683 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
18686 ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo);
18687 AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi);
18689 // r = VSELECT(r, shift(r, 2), a);
18690 MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
18691 DAG.getConstant(2, dl, ExtVT));
18692 MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
18693 DAG.getConstant(2, dl, ExtVT));
18694 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
18695 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
18698 ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo);
18699 AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi);
18701 // r = VSELECT(r, shift(r, 1), a);
18702 MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
18703 DAG.getConstant(1, dl, ExtVT));
18704 MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
18705 DAG.getConstant(1, dl, ExtVT));
18706 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
18707 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
18709 // Logical shift the result back to the lower byte, leaving a zero upper
18711 // meaning that we can safely pack with PACKUSWB.
18713 DAG.getNode(ISD::SRL, dl, ExtVT, RLo, DAG.getConstant(8, dl, ExtVT));
18715 DAG.getNode(ISD::SRL, dl, ExtVT, RHi, DAG.getConstant(8, dl, ExtVT));
18716 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
18720 // It's worth extending once and using the v8i32 shifts for 16-bit types, but
18721 // the extra overheads to get from v16i8 to v8i32 make the existing SSE
18722 // solution better.
18723 if (Subtarget->hasInt256() && VT == MVT::v8i16) {
18724 MVT ExtVT = MVT::v8i32;
18726 Op.getOpcode() == ISD::SRA ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
18727 R = DAG.getNode(ExtOpc, dl, ExtVT, R);
18728 Amt = DAG.getNode(ISD::ANY_EXTEND, dl, ExtVT, Amt);
18729 return DAG.getNode(ISD::TRUNCATE, dl, VT,
18730 DAG.getNode(Op.getOpcode(), dl, ExtVT, R, Amt));
18733 if (Subtarget->hasInt256() && !Subtarget->hasXOP() && VT == MVT::v16i16) {
18734 MVT ExtVT = MVT::v8i32;
18735 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
18736 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, Amt, Z);
18737 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, Amt, Z);
18738 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, R, R);
18739 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, R, R);
18740 ALo = DAG.getBitcast(ExtVT, ALo);
18741 AHi = DAG.getBitcast(ExtVT, AHi);
18742 RLo = DAG.getBitcast(ExtVT, RLo);
18743 RHi = DAG.getBitcast(ExtVT, RHi);
18744 SDValue Lo = DAG.getNode(Op.getOpcode(), dl, ExtVT, RLo, ALo);
18745 SDValue Hi = DAG.getNode(Op.getOpcode(), dl, ExtVT, RHi, AHi);
18746 Lo = DAG.getNode(ISD::SRL, dl, ExtVT, Lo, DAG.getConstant(16, dl, ExtVT));
18747 Hi = DAG.getNode(ISD::SRL, dl, ExtVT, Hi, DAG.getConstant(16, dl, ExtVT));
18748 return DAG.getNode(X86ISD::PACKUS, dl, VT, Lo, Hi);
18751 if (VT == MVT::v8i16) {
18752 unsigned ShiftOpcode = Op->getOpcode();
18754 auto SignBitSelect = [&](SDValue Sel, SDValue V0, SDValue V1) {
18755 // On SSE41 targets we make use of the fact that VSELECT lowers
18756 // to PBLENDVB which selects bytes based just on the sign bit.
18757 if (Subtarget->hasSSE41()) {
18758 MVT ExtVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() * 2);
18759 V0 = DAG.getBitcast(ExtVT, V0);
18760 V1 = DAG.getBitcast(ExtVT, V1);
18761 Sel = DAG.getBitcast(ExtVT, Sel);
18762 return DAG.getBitcast(
18763 VT, DAG.getNode(ISD::VSELECT, dl, ExtVT, Sel, V0, V1));
18765 // On pre-SSE41 targets we splat the sign bit - a negative value will
18766 // set all bits of the lanes to true and VSELECT uses that in
18767 // its OR(AND(V0,C),AND(V1,~C)) lowering.
18769 DAG.getNode(ISD::SRA, dl, VT, Sel, DAG.getConstant(15, dl, VT));
18770 return DAG.getNode(ISD::VSELECT, dl, VT, C, V0, V1);
18773 // Turn 'a' into a mask suitable for VSELECT: a = a << 12;
18774 if (Subtarget->hasSSE41()) {
18775 // On SSE41 targets we need to replicate the shift mask in both
18776 // bytes for PBLENDVB.
18779 DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(4, dl, VT)),
18780 DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT)));
18782 Amt = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT));
18785 // r = VSELECT(r, shift(r, 8), a);
18786 SDValue M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(8, dl, VT));
18787 R = SignBitSelect(Amt, M, R);
18790 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18792 // r = VSELECT(r, shift(r, 4), a);
18793 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT));
18794 R = SignBitSelect(Amt, M, R);
18797 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18799 // r = VSELECT(r, shift(r, 2), a);
18800 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT));
18801 R = SignBitSelect(Amt, M, R);
18804 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18806 // return VSELECT(r, shift(r, 1), a);
18807 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT));
18808 R = SignBitSelect(Amt, M, R);
18812 // Decompose 256-bit shifts into smaller 128-bit shifts.
18813 if (VT.is256BitVector()) {
18814 unsigned NumElems = VT.getVectorNumElements();
18815 MVT EltVT = VT.getVectorElementType();
18816 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
18818 // Extract the two vectors
18819 SDValue V1 = Extract128BitVector(R, 0, DAG, dl);
18820 SDValue V2 = Extract128BitVector(R, NumElems/2, DAG, dl);
18822 // Recreate the shift amount vectors
18823 SDValue Amt1, Amt2;
18824 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
18825 // Constant shift amount
18826 SmallVector<SDValue, 8> Ops(Amt->op_begin(), Amt->op_begin() + NumElems);
18827 ArrayRef<SDValue> Amt1Csts = makeArrayRef(Ops).slice(0, NumElems / 2);
18828 ArrayRef<SDValue> Amt2Csts = makeArrayRef(Ops).slice(NumElems / 2);
18830 Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt1Csts);
18831 Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt2Csts);
18833 // Variable shift amount
18834 Amt1 = Extract128BitVector(Amt, 0, DAG, dl);
18835 Amt2 = Extract128BitVector(Amt, NumElems/2, DAG, dl);
18838 // Issue new vector shifts for the smaller types
18839 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
18840 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
18842 // Concatenate the result back
18843 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
18849 static SDValue LowerRotate(SDValue Op, const X86Subtarget *Subtarget,
18850 SelectionDAG &DAG) {
18851 MVT VT = Op.getSimpleValueType();
18853 SDValue R = Op.getOperand(0);
18854 SDValue Amt = Op.getOperand(1);
18856 assert(VT.isVector() && "Custom lowering only for vector rotates!");
18857 assert(Subtarget->hasXOP() && "XOP support required for vector rotates!");
18858 assert((Op.getOpcode() == ISD::ROTL) && "Only ROTL supported");
18860 // XOP has 128-bit vector variable + immediate rotates.
18861 // +ve/-ve Amt = rotate left/right.
18863 // Split 256-bit integers.
18864 if (VT.is256BitVector())
18865 return Lower256IntArith(Op, DAG);
18867 assert(VT.is128BitVector() && "Only rotate 128-bit vectors!");
18869 // Attempt to rotate by immediate.
18870 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
18871 if (auto *RotateConst = BVAmt->getConstantSplatNode()) {
18872 uint64_t RotateAmt = RotateConst->getAPIntValue().getZExtValue();
18873 assert(RotateAmt < VT.getScalarSizeInBits() && "Rotation out of range");
18874 return DAG.getNode(X86ISD::VPROTI, DL, VT, R,
18875 DAG.getConstant(RotateAmt, DL, MVT::i8));
18879 // Use general rotate by variable (per-element).
18880 return DAG.getNode(X86ISD::VPROT, DL, VT, R, Amt);
18883 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
18884 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
18885 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
18886 // looks for this combo and may remove the "setcc" instruction if the "setcc"
18887 // has only one use.
18888 SDNode *N = Op.getNode();
18889 SDValue LHS = N->getOperand(0);
18890 SDValue RHS = N->getOperand(1);
18891 unsigned BaseOp = 0;
18894 switch (Op.getOpcode()) {
18895 default: llvm_unreachable("Unknown ovf instruction!");
18897 // A subtract of one will be selected as a INC. Note that INC doesn't
18898 // set CF, so we can't do this for UADDO.
18899 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
18901 BaseOp = X86ISD::INC;
18902 Cond = X86::COND_O;
18905 BaseOp = X86ISD::ADD;
18906 Cond = X86::COND_O;
18909 BaseOp = X86ISD::ADD;
18910 Cond = X86::COND_B;
18913 // A subtract of one will be selected as a DEC. Note that DEC doesn't
18914 // set CF, so we can't do this for USUBO.
18915 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
18917 BaseOp = X86ISD::DEC;
18918 Cond = X86::COND_O;
18921 BaseOp = X86ISD::SUB;
18922 Cond = X86::COND_O;
18925 BaseOp = X86ISD::SUB;
18926 Cond = X86::COND_B;
18929 BaseOp = N->getValueType(0) == MVT::i8 ? X86ISD::SMUL8 : X86ISD::SMUL;
18930 Cond = X86::COND_O;
18932 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
18933 if (N->getValueType(0) == MVT::i8) {
18934 BaseOp = X86ISD::UMUL8;
18935 Cond = X86::COND_O;
18938 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
18940 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
18943 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
18944 DAG.getConstant(X86::COND_O, DL, MVT::i32),
18945 SDValue(Sum.getNode(), 2));
18947 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
18951 // Also sets EFLAGS.
18952 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
18953 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
18956 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
18957 DAG.getConstant(Cond, DL, MVT::i32),
18958 SDValue(Sum.getNode(), 1));
18960 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
18963 /// Returns true if the operand type is exactly twice the native width, and
18964 /// the corresponding cmpxchg8b or cmpxchg16b instruction is available.
18965 /// Used to know whether to use cmpxchg8/16b when expanding atomic operations
18966 /// (otherwise we leave them alone to become __sync_fetch_and_... calls).
18967 bool X86TargetLowering::needsCmpXchgNb(Type *MemType) const {
18968 unsigned OpWidth = MemType->getPrimitiveSizeInBits();
18971 return !Subtarget->is64Bit(); // FIXME this should be Subtarget.hasCmpxchg8b
18972 else if (OpWidth == 128)
18973 return Subtarget->hasCmpxchg16b();
18978 bool X86TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
18979 return needsCmpXchgNb(SI->getValueOperand()->getType());
18982 // Note: this turns large loads into lock cmpxchg8b/16b.
18983 // FIXME: On 32 bits x86, fild/movq might be faster than lock cmpxchg8b.
18984 TargetLowering::AtomicExpansionKind
18985 X86TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
18986 auto PTy = cast<PointerType>(LI->getPointerOperand()->getType());
18987 return needsCmpXchgNb(PTy->getElementType()) ? AtomicExpansionKind::CmpXChg
18988 : AtomicExpansionKind::None;
18991 TargetLowering::AtomicExpansionKind
18992 X86TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
18993 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
18994 Type *MemType = AI->getType();
18996 // If the operand is too big, we must see if cmpxchg8/16b is available
18997 // and default to library calls otherwise.
18998 if (MemType->getPrimitiveSizeInBits() > NativeWidth) {
18999 return needsCmpXchgNb(MemType) ? AtomicExpansionKind::CmpXChg
19000 : AtomicExpansionKind::None;
19003 AtomicRMWInst::BinOp Op = AI->getOperation();
19006 llvm_unreachable("Unknown atomic operation");
19007 case AtomicRMWInst::Xchg:
19008 case AtomicRMWInst::Add:
19009 case AtomicRMWInst::Sub:
19010 // It's better to use xadd, xsub or xchg for these in all cases.
19011 return AtomicExpansionKind::None;
19012 case AtomicRMWInst::Or:
19013 case AtomicRMWInst::And:
19014 case AtomicRMWInst::Xor:
19015 // If the atomicrmw's result isn't actually used, we can just add a "lock"
19016 // prefix to a normal instruction for these operations.
19017 return !AI->use_empty() ? AtomicExpansionKind::CmpXChg
19018 : AtomicExpansionKind::None;
19019 case AtomicRMWInst::Nand:
19020 case AtomicRMWInst::Max:
19021 case AtomicRMWInst::Min:
19022 case AtomicRMWInst::UMax:
19023 case AtomicRMWInst::UMin:
19024 // These always require a non-trivial set of data operations on x86. We must
19025 // use a cmpxchg loop.
19026 return AtomicExpansionKind::CmpXChg;
19030 static bool hasMFENCE(const X86Subtarget& Subtarget) {
19031 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
19032 // no-sse2). There isn't any reason to disable it if the target processor
19034 return Subtarget.hasSSE2() || Subtarget.is64Bit();
19038 X86TargetLowering::lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const {
19039 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
19040 Type *MemType = AI->getType();
19041 // Accesses larger than the native width are turned into cmpxchg/libcalls, so
19042 // there is no benefit in turning such RMWs into loads, and it is actually
19043 // harmful as it introduces a mfence.
19044 if (MemType->getPrimitiveSizeInBits() > NativeWidth)
19047 auto Builder = IRBuilder<>(AI);
19048 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
19049 auto SynchScope = AI->getSynchScope();
19050 // We must restrict the ordering to avoid generating loads with Release or
19051 // ReleaseAcquire orderings.
19052 auto Order = AtomicCmpXchgInst::getStrongestFailureOrdering(AI->getOrdering());
19053 auto Ptr = AI->getPointerOperand();
19055 // Before the load we need a fence. Here is an example lifted from
19056 // http://www.hpl.hp.com/techreports/2012/HPL-2012-68.pdf showing why a fence
19059 // x.store(1, relaxed);
19060 // r1 = y.fetch_add(0, release);
19062 // y.fetch_add(42, acquire);
19063 // r2 = x.load(relaxed);
19064 // r1 = r2 = 0 is impossible, but becomes possible if the idempotent rmw is
19065 // lowered to just a load without a fence. A mfence flushes the store buffer,
19066 // making the optimization clearly correct.
19067 // FIXME: it is required if isAtLeastRelease(Order) but it is not clear
19068 // otherwise, we might be able to be more aggressive on relaxed idempotent
19069 // rmw. In practice, they do not look useful, so we don't try to be
19070 // especially clever.
19071 if (SynchScope == SingleThread)
19072 // FIXME: we could just insert an X86ISD::MEMBARRIER here, except we are at
19073 // the IR level, so we must wrap it in an intrinsic.
19076 if (!hasMFENCE(*Subtarget))
19077 // FIXME: it might make sense to use a locked operation here but on a
19078 // different cache-line to prevent cache-line bouncing. In practice it
19079 // is probably a small win, and x86 processors without mfence are rare
19080 // enough that we do not bother.
19084 llvm::Intrinsic::getDeclaration(M, Intrinsic::x86_sse2_mfence);
19085 Builder.CreateCall(MFence, {});
19087 // Finally we can emit the atomic load.
19088 LoadInst *Loaded = Builder.CreateAlignedLoad(Ptr,
19089 AI->getType()->getPrimitiveSizeInBits());
19090 Loaded->setAtomic(Order, SynchScope);
19091 AI->replaceAllUsesWith(Loaded);
19092 AI->eraseFromParent();
19096 static SDValue LowerATOMIC_FENCE(SDValue Op, const X86Subtarget *Subtarget,
19097 SelectionDAG &DAG) {
19099 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
19100 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
19101 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
19102 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
19104 // The only fence that needs an instruction is a sequentially-consistent
19105 // cross-thread fence.
19106 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
19107 if (hasMFENCE(*Subtarget))
19108 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
19110 SDValue Chain = Op.getOperand(0);
19111 SDValue Zero = DAG.getConstant(0, dl, MVT::i32);
19113 DAG.getRegister(X86::ESP, MVT::i32), // Base
19114 DAG.getTargetConstant(1, dl, MVT::i8), // Scale
19115 DAG.getRegister(0, MVT::i32), // Index
19116 DAG.getTargetConstant(0, dl, MVT::i32), // Disp
19117 DAG.getRegister(0, MVT::i32), // Segment.
19121 SDNode *Res = DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops);
19122 return SDValue(Res, 0);
19125 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
19126 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
19129 static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget *Subtarget,
19130 SelectionDAG &DAG) {
19131 MVT T = Op.getSimpleValueType();
19135 switch(T.SimpleTy) {
19136 default: llvm_unreachable("Invalid value type!");
19137 case MVT::i8: Reg = X86::AL; size = 1; break;
19138 case MVT::i16: Reg = X86::AX; size = 2; break;
19139 case MVT::i32: Reg = X86::EAX; size = 4; break;
19141 assert(Subtarget->is64Bit() && "Node not type legal!");
19142 Reg = X86::RAX; size = 8;
19145 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
19146 Op.getOperand(2), SDValue());
19147 SDValue Ops[] = { cpIn.getValue(0),
19150 DAG.getTargetConstant(size, DL, MVT::i8),
19151 cpIn.getValue(1) };
19152 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
19153 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
19154 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
19158 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
19159 SDValue EFLAGS = DAG.getCopyFromReg(cpOut.getValue(1), DL, X86::EFLAGS,
19160 MVT::i32, cpOut.getValue(2));
19161 SDValue Success = DAG.getNode(X86ISD::SETCC, DL, Op->getValueType(1),
19162 DAG.getConstant(X86::COND_E, DL, MVT::i8),
19165 DAG.ReplaceAllUsesOfValueWith(Op.getValue(0), cpOut);
19166 DAG.ReplaceAllUsesOfValueWith(Op.getValue(1), Success);
19167 DAG.ReplaceAllUsesOfValueWith(Op.getValue(2), EFLAGS.getValue(1));
19171 static SDValue LowerBITCAST(SDValue Op, const X86Subtarget *Subtarget,
19172 SelectionDAG &DAG) {
19173 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
19174 MVT DstVT = Op.getSimpleValueType();
19176 if (SrcVT == MVT::v2i32 || SrcVT == MVT::v4i16 || SrcVT == MVT::v8i8) {
19177 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19178 if (DstVT != MVT::f64)
19179 // This conversion needs to be expanded.
19182 SDValue InVec = Op->getOperand(0);
19184 unsigned NumElts = SrcVT.getVectorNumElements();
19185 MVT SVT = SrcVT.getVectorElementType();
19187 // Widen the vector in input in the case of MVT::v2i32.
19188 // Example: from MVT::v2i32 to MVT::v4i32.
19189 SmallVector<SDValue, 16> Elts;
19190 for (unsigned i = 0, e = NumElts; i != e; ++i)
19191 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT, InVec,
19192 DAG.getIntPtrConstant(i, dl)));
19194 // Explicitly mark the extra elements as Undef.
19195 Elts.append(NumElts, DAG.getUNDEF(SVT));
19197 EVT NewVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
19198 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Elts);
19199 SDValue ToV2F64 = DAG.getBitcast(MVT::v2f64, BV);
19200 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, ToV2F64,
19201 DAG.getIntPtrConstant(0, dl));
19204 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
19205 Subtarget->hasMMX() && "Unexpected custom BITCAST");
19206 assert((DstVT == MVT::i64 ||
19207 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
19208 "Unexpected custom BITCAST");
19209 // i64 <=> MMX conversions are Legal.
19210 if (SrcVT==MVT::i64 && DstVT.isVector())
19212 if (DstVT==MVT::i64 && SrcVT.isVector())
19214 // MMX <=> MMX conversions are Legal.
19215 if (SrcVT.isVector() && DstVT.isVector())
19217 // All other conversions need to be expanded.
19221 /// Compute the horizontal sum of bytes in V for the elements of VT.
19223 /// Requires V to be a byte vector and VT to be an integer vector type with
19224 /// wider elements than V's type. The width of the elements of VT determines
19225 /// how many bytes of V are summed horizontally to produce each element of the
19227 static SDValue LowerHorizontalByteSum(SDValue V, MVT VT,
19228 const X86Subtarget *Subtarget,
19229 SelectionDAG &DAG) {
19231 MVT ByteVecVT = V.getSimpleValueType();
19232 MVT EltVT = VT.getVectorElementType();
19233 int NumElts = VT.getVectorNumElements();
19234 assert(ByteVecVT.getVectorElementType() == MVT::i8 &&
19235 "Expected value to have byte element type.");
19236 assert(EltVT != MVT::i8 &&
19237 "Horizontal byte sum only makes sense for wider elements!");
19238 unsigned VecSize = VT.getSizeInBits();
19239 assert(ByteVecVT.getSizeInBits() == VecSize && "Cannot change vector size!");
19241 // PSADBW instruction horizontally add all bytes and leave the result in i64
19242 // chunks, thus directly computes the pop count for v2i64 and v4i64.
19243 if (EltVT == MVT::i64) {
19244 SDValue Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
19245 V = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT, V, Zeros);
19246 return DAG.getBitcast(VT, V);
19249 if (EltVT == MVT::i32) {
19250 // We unpack the low half and high half into i32s interleaved with zeros so
19251 // that we can use PSADBW to horizontally sum them. The most useful part of
19252 // this is that it lines up the results of two PSADBW instructions to be
19253 // two v2i64 vectors which concatenated are the 4 population counts. We can
19254 // then use PACKUSWB to shrink and concatenate them into a v4i32 again.
19255 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, DL);
19256 SDValue Low = DAG.getNode(X86ISD::UNPCKL, DL, VT, V, Zeros);
19257 SDValue High = DAG.getNode(X86ISD::UNPCKH, DL, VT, V, Zeros);
19259 // Do the horizontal sums into two v2i64s.
19260 Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
19261 Low = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT,
19262 DAG.getBitcast(ByteVecVT, Low), Zeros);
19263 High = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT,
19264 DAG.getBitcast(ByteVecVT, High), Zeros);
19266 // Merge them together.
19267 MVT ShortVecVT = MVT::getVectorVT(MVT::i16, VecSize / 16);
19268 V = DAG.getNode(X86ISD::PACKUS, DL, ByteVecVT,
19269 DAG.getBitcast(ShortVecVT, Low),
19270 DAG.getBitcast(ShortVecVT, High));
19272 return DAG.getBitcast(VT, V);
19275 // The only element type left is i16.
19276 assert(EltVT == MVT::i16 && "Unknown how to handle type");
19278 // To obtain pop count for each i16 element starting from the pop count for
19279 // i8 elements, shift the i16s left by 8, sum as i8s, and then shift as i16s
19280 // right by 8. It is important to shift as i16s as i8 vector shift isn't
19281 // directly supported.
19282 SmallVector<SDValue, 16> Shifters(NumElts, DAG.getConstant(8, DL, EltVT));
19283 SDValue Shifter = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters);
19284 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, DAG.getBitcast(VT, V), Shifter);
19285 V = DAG.getNode(ISD::ADD, DL, ByteVecVT, DAG.getBitcast(ByteVecVT, Shl),
19286 DAG.getBitcast(ByteVecVT, V));
19287 return DAG.getNode(ISD::SRL, DL, VT, DAG.getBitcast(VT, V), Shifter);
19290 static SDValue LowerVectorCTPOPInRegLUT(SDValue Op, SDLoc DL,
19291 const X86Subtarget *Subtarget,
19292 SelectionDAG &DAG) {
19293 MVT VT = Op.getSimpleValueType();
19294 MVT EltVT = VT.getVectorElementType();
19295 unsigned VecSize = VT.getSizeInBits();
19297 // Implement a lookup table in register by using an algorithm based on:
19298 // http://wm.ite.pl/articles/sse-popcount.html
19300 // The general idea is that every lower byte nibble in the input vector is an
19301 // index into a in-register pre-computed pop count table. We then split up the
19302 // input vector in two new ones: (1) a vector with only the shifted-right
19303 // higher nibbles for each byte and (2) a vector with the lower nibbles (and
19304 // masked out higher ones) for each byte. PSHUB is used separately with both
19305 // to index the in-register table. Next, both are added and the result is a
19306 // i8 vector where each element contains the pop count for input byte.
19308 // To obtain the pop count for elements != i8, we follow up with the same
19309 // approach and use additional tricks as described below.
19311 const int LUT[16] = {/* 0 */ 0, /* 1 */ 1, /* 2 */ 1, /* 3 */ 2,
19312 /* 4 */ 1, /* 5 */ 2, /* 6 */ 2, /* 7 */ 3,
19313 /* 8 */ 1, /* 9 */ 2, /* a */ 2, /* b */ 3,
19314 /* c */ 2, /* d */ 3, /* e */ 3, /* f */ 4};
19316 int NumByteElts = VecSize / 8;
19317 MVT ByteVecVT = MVT::getVectorVT(MVT::i8, NumByteElts);
19318 SDValue In = DAG.getBitcast(ByteVecVT, Op);
19319 SmallVector<SDValue, 16> LUTVec;
19320 for (int i = 0; i < NumByteElts; ++i)
19321 LUTVec.push_back(DAG.getConstant(LUT[i % 16], DL, MVT::i8));
19322 SDValue InRegLUT = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, LUTVec);
19323 SmallVector<SDValue, 16> Mask0F(NumByteElts,
19324 DAG.getConstant(0x0F, DL, MVT::i8));
19325 SDValue M0F = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Mask0F);
19328 SmallVector<SDValue, 16> Four(NumByteElts, DAG.getConstant(4, DL, MVT::i8));
19329 SDValue FourV = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Four);
19330 SDValue HighNibbles = DAG.getNode(ISD::SRL, DL, ByteVecVT, In, FourV);
19333 SDValue LowNibbles = DAG.getNode(ISD::AND, DL, ByteVecVT, In, M0F);
19335 // The input vector is used as the shuffle mask that index elements into the
19336 // LUT. After counting low and high nibbles, add the vector to obtain the
19337 // final pop count per i8 element.
19338 SDValue HighPopCnt =
19339 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, HighNibbles);
19340 SDValue LowPopCnt =
19341 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, LowNibbles);
19342 SDValue PopCnt = DAG.getNode(ISD::ADD, DL, ByteVecVT, HighPopCnt, LowPopCnt);
19344 if (EltVT == MVT::i8)
19347 return LowerHorizontalByteSum(PopCnt, VT, Subtarget, DAG);
19350 static SDValue LowerVectorCTPOPBitmath(SDValue Op, SDLoc DL,
19351 const X86Subtarget *Subtarget,
19352 SelectionDAG &DAG) {
19353 MVT VT = Op.getSimpleValueType();
19354 assert(VT.is128BitVector() &&
19355 "Only 128-bit vector bitmath lowering supported.");
19357 int VecSize = VT.getSizeInBits();
19358 MVT EltVT = VT.getVectorElementType();
19359 int Len = EltVT.getSizeInBits();
19361 // This is the vectorized version of the "best" algorithm from
19362 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
19363 // with a minor tweak to use a series of adds + shifts instead of vector
19364 // multiplications. Implemented for all integer vector types. We only use
19365 // this when we don't have SSSE3 which allows a LUT-based lowering that is
19366 // much faster, even faster than using native popcnt instructions.
19368 auto GetShift = [&](unsigned OpCode, SDValue V, int Shifter) {
19369 MVT VT = V.getSimpleValueType();
19370 SmallVector<SDValue, 32> Shifters(
19371 VT.getVectorNumElements(),
19372 DAG.getConstant(Shifter, DL, VT.getVectorElementType()));
19373 return DAG.getNode(OpCode, DL, VT, V,
19374 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters));
19376 auto GetMask = [&](SDValue V, APInt Mask) {
19377 MVT VT = V.getSimpleValueType();
19378 SmallVector<SDValue, 32> Masks(
19379 VT.getVectorNumElements(),
19380 DAG.getConstant(Mask, DL, VT.getVectorElementType()));
19381 return DAG.getNode(ISD::AND, DL, VT, V,
19382 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Masks));
19385 // We don't want to incur the implicit masks required to SRL vNi8 vectors on
19386 // x86, so set the SRL type to have elements at least i16 wide. This is
19387 // correct because all of our SRLs are followed immediately by a mask anyways
19388 // that handles any bits that sneak into the high bits of the byte elements.
19389 MVT SrlVT = Len > 8 ? VT : MVT::getVectorVT(MVT::i16, VecSize / 16);
19393 // v = v - ((v >> 1) & 0x55555555...)
19395 DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 1));
19396 SDValue And = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x55)));
19397 V = DAG.getNode(ISD::SUB, DL, VT, V, And);
19399 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
19400 SDValue AndLHS = GetMask(V, APInt::getSplat(Len, APInt(8, 0x33)));
19401 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 2));
19402 SDValue AndRHS = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x33)));
19403 V = DAG.getNode(ISD::ADD, DL, VT, AndLHS, AndRHS);
19405 // v = (v + (v >> 4)) & 0x0F0F0F0F...
19406 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 4));
19407 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, V, Srl);
19408 V = GetMask(Add, APInt::getSplat(Len, APInt(8, 0x0F)));
19410 // At this point, V contains the byte-wise population count, and we are
19411 // merely doing a horizontal sum if necessary to get the wider element
19413 if (EltVT == MVT::i8)
19416 return LowerHorizontalByteSum(
19417 DAG.getBitcast(MVT::getVectorVT(MVT::i8, VecSize / 8), V), VT, Subtarget,
19421 static SDValue LowerVectorCTPOP(SDValue Op, const X86Subtarget *Subtarget,
19422 SelectionDAG &DAG) {
19423 MVT VT = Op.getSimpleValueType();
19424 // FIXME: Need to add AVX-512 support here!
19425 assert((VT.is256BitVector() || VT.is128BitVector()) &&
19426 "Unknown CTPOP type to handle");
19427 SDLoc DL(Op.getNode());
19428 SDValue Op0 = Op.getOperand(0);
19430 if (!Subtarget->hasSSSE3()) {
19431 // We can't use the fast LUT approach, so fall back on vectorized bitmath.
19432 assert(VT.is128BitVector() && "Only 128-bit vectors supported in SSE!");
19433 return LowerVectorCTPOPBitmath(Op0, DL, Subtarget, DAG);
19436 if (VT.is256BitVector() && !Subtarget->hasInt256()) {
19437 unsigned NumElems = VT.getVectorNumElements();
19439 // Extract each 128-bit vector, compute pop count and concat the result.
19440 SDValue LHS = Extract128BitVector(Op0, 0, DAG, DL);
19441 SDValue RHS = Extract128BitVector(Op0, NumElems/2, DAG, DL);
19443 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT,
19444 LowerVectorCTPOPInRegLUT(LHS, DL, Subtarget, DAG),
19445 LowerVectorCTPOPInRegLUT(RHS, DL, Subtarget, DAG));
19448 return LowerVectorCTPOPInRegLUT(Op0, DL, Subtarget, DAG);
19451 static SDValue LowerCTPOP(SDValue Op, const X86Subtarget *Subtarget,
19452 SelectionDAG &DAG) {
19453 assert(Op.getSimpleValueType().isVector() &&
19454 "We only do custom lowering for vector population count.");
19455 return LowerVectorCTPOP(Op, Subtarget, DAG);
19458 static SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
19459 SDNode *Node = Op.getNode();
19461 EVT T = Node->getValueType(0);
19462 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
19463 DAG.getConstant(0, dl, T), Node->getOperand(2));
19464 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
19465 cast<AtomicSDNode>(Node)->getMemoryVT(),
19466 Node->getOperand(0),
19467 Node->getOperand(1), negOp,
19468 cast<AtomicSDNode>(Node)->getMemOperand(),
19469 cast<AtomicSDNode>(Node)->getOrdering(),
19470 cast<AtomicSDNode>(Node)->getSynchScope());
19473 static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) {
19474 SDNode *Node = Op.getNode();
19476 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
19478 // Convert seq_cst store -> xchg
19479 // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b)
19480 // FIXME: On 32-bit, store -> fist or movq would be more efficient
19481 // (The only way to get a 16-byte store is cmpxchg16b)
19482 // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment.
19483 if (cast<AtomicSDNode>(Node)->getOrdering() == SequentiallyConsistent ||
19484 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
19485 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
19486 cast<AtomicSDNode>(Node)->getMemoryVT(),
19487 Node->getOperand(0),
19488 Node->getOperand(1), Node->getOperand(2),
19489 cast<AtomicSDNode>(Node)->getMemOperand(),
19490 cast<AtomicSDNode>(Node)->getOrdering(),
19491 cast<AtomicSDNode>(Node)->getSynchScope());
19492 return Swap.getValue(1);
19494 // Other atomic stores have a simple pattern.
19498 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
19499 MVT VT = Op.getNode()->getSimpleValueType(0);
19501 // Let legalize expand this if it isn't a legal type yet.
19502 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
19505 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
19508 bool ExtraOp = false;
19509 switch (Op.getOpcode()) {
19510 default: llvm_unreachable("Invalid code");
19511 case ISD::ADDC: Opc = X86ISD::ADD; break;
19512 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
19513 case ISD::SUBC: Opc = X86ISD::SUB; break;
19514 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
19518 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
19520 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
19521 Op.getOperand(1), Op.getOperand(2));
19524 static SDValue LowerFSINCOS(SDValue Op, const X86Subtarget *Subtarget,
19525 SelectionDAG &DAG) {
19526 assert(Subtarget->isTargetDarwin() && Subtarget->is64Bit());
19528 // For MacOSX, we want to call an alternative entry point: __sincos_stret,
19529 // which returns the values as { float, float } (in XMM0) or
19530 // { double, double } (which is returned in XMM0, XMM1).
19532 SDValue Arg = Op.getOperand(0);
19533 EVT ArgVT = Arg.getValueType();
19534 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
19536 TargetLowering::ArgListTy Args;
19537 TargetLowering::ArgListEntry Entry;
19541 Entry.isSExt = false;
19542 Entry.isZExt = false;
19543 Args.push_back(Entry);
19545 bool isF64 = ArgVT == MVT::f64;
19546 // Only optimize x86_64 for now. i386 is a bit messy. For f32,
19547 // the small struct {f32, f32} is returned in (eax, edx). For f64,
19548 // the results are returned via SRet in memory.
19549 const char *LibcallName = isF64 ? "__sincos_stret" : "__sincosf_stret";
19550 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
19552 DAG.getExternalSymbol(LibcallName, TLI.getPointerTy(DAG.getDataLayout()));
19554 Type *RetTy = isF64
19555 ? (Type*)StructType::get(ArgTy, ArgTy, nullptr)
19556 : (Type*)VectorType::get(ArgTy, 4);
19558 TargetLowering::CallLoweringInfo CLI(DAG);
19559 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
19560 .setCallee(CallingConv::C, RetTy, Callee, std::move(Args), 0);
19562 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
19565 // Returned in xmm0 and xmm1.
19566 return CallResult.first;
19568 // Returned in bits 0:31 and 32:64 xmm0.
19569 SDValue SinVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
19570 CallResult.first, DAG.getIntPtrConstant(0, dl));
19571 SDValue CosVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
19572 CallResult.first, DAG.getIntPtrConstant(1, dl));
19573 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
19574 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys, SinVal, CosVal);
19577 static SDValue LowerMSCATTER(SDValue Op, const X86Subtarget *Subtarget,
19578 SelectionDAG &DAG) {
19579 assert(Subtarget->hasAVX512() &&
19580 "MGATHER/MSCATTER are supported on AVX-512 arch only");
19582 MaskedScatterSDNode *N = cast<MaskedScatterSDNode>(Op.getNode());
19583 MVT VT = N->getValue().getSimpleValueType();
19584 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported scatter op");
19587 // X86 scatter kills mask register, so its type should be added to
19588 // the list of return values
19589 if (N->getNumValues() == 1) {
19590 SDValue Index = N->getIndex();
19591 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
19592 !Index.getSimpleValueType().is512BitVector())
19593 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
19595 SDVTList VTs = DAG.getVTList(N->getMask().getValueType(), MVT::Other);
19596 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
19597 N->getOperand(3), Index };
19599 SDValue NewScatter = DAG.getMaskedScatter(VTs, VT, dl, Ops, N->getMemOperand());
19600 DAG.ReplaceAllUsesWith(Op, SDValue(NewScatter.getNode(), 1));
19601 return SDValue(NewScatter.getNode(), 0);
19606 static SDValue LowerMGATHER(SDValue Op, const X86Subtarget *Subtarget,
19607 SelectionDAG &DAG) {
19608 assert(Subtarget->hasAVX512() &&
19609 "MGATHER/MSCATTER are supported on AVX-512 arch only");
19611 MaskedGatherSDNode *N = cast<MaskedGatherSDNode>(Op.getNode());
19612 MVT VT = Op.getSimpleValueType();
19613 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported gather op");
19616 SDValue Index = N->getIndex();
19617 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
19618 !Index.getSimpleValueType().is512BitVector()) {
19619 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
19620 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
19621 N->getOperand(3), Index };
19622 DAG.UpdateNodeOperands(N, Ops);
19627 SDValue X86TargetLowering::LowerGC_TRANSITION_START(SDValue Op,
19628 SelectionDAG &DAG) const {
19629 // TODO: Eventually, the lowering of these nodes should be informed by or
19630 // deferred to the GC strategy for the function in which they appear. For
19631 // now, however, they must be lowered to something. Since they are logically
19632 // no-ops in the case of a null GC strategy (or a GC strategy which does not
19633 // require special handling for these nodes), lower them as literal NOOPs for
19635 SmallVector<SDValue, 2> Ops;
19637 Ops.push_back(Op.getOperand(0));
19638 if (Op->getGluedNode())
19639 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
19642 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
19643 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
19648 SDValue X86TargetLowering::LowerGC_TRANSITION_END(SDValue Op,
19649 SelectionDAG &DAG) const {
19650 // TODO: Eventually, the lowering of these nodes should be informed by or
19651 // deferred to the GC strategy for the function in which they appear. For
19652 // now, however, they must be lowered to something. Since they are logically
19653 // no-ops in the case of a null GC strategy (or a GC strategy which does not
19654 // require special handling for these nodes), lower them as literal NOOPs for
19656 SmallVector<SDValue, 2> Ops;
19658 Ops.push_back(Op.getOperand(0));
19659 if (Op->getGluedNode())
19660 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
19663 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
19664 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
19669 /// LowerOperation - Provide custom lowering hooks for some operations.
19671 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
19672 switch (Op.getOpcode()) {
19673 default: llvm_unreachable("Should not custom lower this!");
19674 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, Subtarget, DAG);
19675 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
19676 return LowerCMP_SWAP(Op, Subtarget, DAG);
19677 case ISD::CTPOP: return LowerCTPOP(Op, Subtarget, DAG);
19678 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
19679 case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG);
19680 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
19681 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, Subtarget, DAG);
19682 case ISD::VECTOR_SHUFFLE: return lowerVectorShuffle(Op, Subtarget, DAG);
19683 case ISD::VSELECT: return LowerVSELECT(Op, DAG);
19684 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
19685 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
19686 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op,Subtarget,DAG);
19687 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, Subtarget,DAG);
19688 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
19689 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
19690 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
19691 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
19692 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
19693 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
19694 case ISD::SHL_PARTS:
19695 case ISD::SRA_PARTS:
19696 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
19697 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
19698 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
19699 case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG);
19700 case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, Subtarget, DAG);
19701 case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, Subtarget, DAG);
19702 case ISD::ANY_EXTEND: return LowerANY_EXTEND(Op, Subtarget, DAG);
19703 case ISD::SIGN_EXTEND_VECTOR_INREG:
19704 return LowerSIGN_EXTEND_VECTOR_INREG(Op, Subtarget, DAG);
19705 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
19706 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
19707 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
19708 case ISD::LOAD: return LowerExtendedLoad(Op, Subtarget, DAG);
19710 case ISD::FNEG: return LowerFABSorFNEG(Op, DAG);
19711 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
19712 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
19713 case ISD::SETCC: return LowerSETCC(Op, DAG);
19714 case ISD::SETCCE: return LowerSETCCE(Op, DAG);
19715 case ISD::SELECT: return LowerSELECT(Op, DAG);
19716 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
19717 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
19718 case ISD::VASTART: return LowerVASTART(Op, DAG);
19719 case ISD::VAARG: return LowerVAARG(Op, DAG);
19720 case ISD::VACOPY: return LowerVACOPY(Op, Subtarget, DAG);
19721 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, Subtarget, DAG);
19722 case ISD::INTRINSIC_VOID:
19723 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, Subtarget, DAG);
19724 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
19725 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
19726 case ISD::FRAME_TO_ARGS_OFFSET:
19727 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
19728 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
19729 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
19730 case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
19731 case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
19732 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
19733 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
19734 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
19735 case ISD::CTLZ: return LowerCTLZ(Op, Subtarget, DAG);
19736 case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, Subtarget, DAG);
19738 case ISD::CTTZ_ZERO_UNDEF: return LowerCTTZ(Op, DAG);
19739 case ISD::MUL: return LowerMUL(Op, Subtarget, DAG);
19740 case ISD::UMUL_LOHI:
19741 case ISD::SMUL_LOHI: return LowerMUL_LOHI(Op, Subtarget, DAG);
19742 case ISD::ROTL: return LowerRotate(Op, Subtarget, DAG);
19745 case ISD::SHL: return LowerShift(Op, Subtarget, DAG);
19751 case ISD::UMULO: return LowerXALUO(Op, DAG);
19752 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, Subtarget,DAG);
19753 case ISD::BITCAST: return LowerBITCAST(Op, Subtarget, DAG);
19757 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
19758 case ISD::ADD: return LowerADD(Op, DAG);
19759 case ISD::SUB: return LowerSUB(Op, DAG);
19763 case ISD::UMIN: return LowerMINMAX(Op, DAG);
19764 case ISD::FSINCOS: return LowerFSINCOS(Op, Subtarget, DAG);
19765 case ISD::MGATHER: return LowerMGATHER(Op, Subtarget, DAG);
19766 case ISD::MSCATTER: return LowerMSCATTER(Op, Subtarget, DAG);
19767 case ISD::GC_TRANSITION_START:
19768 return LowerGC_TRANSITION_START(Op, DAG);
19769 case ISD::GC_TRANSITION_END: return LowerGC_TRANSITION_END(Op, DAG);
19773 /// ReplaceNodeResults - Replace a node with an illegal result type
19774 /// with a new node built out of custom code.
19775 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
19776 SmallVectorImpl<SDValue>&Results,
19777 SelectionDAG &DAG) const {
19779 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
19780 switch (N->getOpcode()) {
19782 llvm_unreachable("Do not know how to custom type legalize this operation!");
19783 // We might have generated v2f32 FMIN/FMAX operations. Widen them to v4f32.
19784 case X86ISD::FMINC:
19786 case X86ISD::FMAXC:
19787 case X86ISD::FMAX: {
19788 EVT VT = N->getValueType(0);
19789 assert(VT == MVT::v2f32 && "Unexpected type (!= v2f32) on FMIN/FMAX.");
19790 SDValue UNDEF = DAG.getUNDEF(VT);
19791 SDValue LHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
19792 N->getOperand(0), UNDEF);
19793 SDValue RHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
19794 N->getOperand(1), UNDEF);
19795 Results.push_back(DAG.getNode(N->getOpcode(), dl, MVT::v4f32, LHS, RHS));
19798 case ISD::SIGN_EXTEND_INREG:
19803 // We don't want to expand or promote these.
19810 case ISD::UDIVREM: {
19811 SDValue V = LowerWin64_i128OP(SDValue(N,0), DAG);
19812 Results.push_back(V);
19815 case ISD::FP_TO_SINT:
19816 case ISD::FP_TO_UINT: {
19817 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
19819 std::pair<SDValue,SDValue> Vals =
19820 FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true);
19821 SDValue FIST = Vals.first, StackSlot = Vals.second;
19822 if (FIST.getNode()) {
19823 EVT VT = N->getValueType(0);
19824 // Return a load from the stack slot.
19825 if (StackSlot.getNode())
19826 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
19827 MachinePointerInfo(),
19828 false, false, false, 0));
19830 Results.push_back(FIST);
19834 case ISD::UINT_TO_FP: {
19835 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19836 if (N->getOperand(0).getValueType() != MVT::v2i32 ||
19837 N->getValueType(0) != MVT::v2f32)
19839 SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64,
19841 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
19843 SDValue VBias = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2f64, Bias, Bias);
19844 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn,
19845 DAG.getBitcast(MVT::v2i64, VBias));
19846 Or = DAG.getBitcast(MVT::v2f64, Or);
19847 // TODO: Are there any fast-math-flags to propagate here?
19848 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias);
19849 Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub));
19852 case ISD::FP_ROUND: {
19853 if (!TLI.isTypeLegal(N->getOperand(0).getValueType()))
19855 SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0));
19856 Results.push_back(V);
19859 case ISD::FP_EXTEND: {
19860 // Right now, only MVT::v2f32 has OperationAction for FP_EXTEND.
19861 // No other ValueType for FP_EXTEND should reach this point.
19862 assert(N->getValueType(0) == MVT::v2f32 &&
19863 "Do not know how to legalize this Node");
19866 case ISD::INTRINSIC_W_CHAIN: {
19867 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
19869 default : llvm_unreachable("Do not know how to custom type "
19870 "legalize this intrinsic operation!");
19871 case Intrinsic::x86_rdtsc:
19872 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
19874 case Intrinsic::x86_rdtscp:
19875 return getReadTimeStampCounter(N, dl, X86ISD::RDTSCP_DAG, DAG, Subtarget,
19877 case Intrinsic::x86_rdpmc:
19878 return getReadPerformanceCounter(N, dl, DAG, Subtarget, Results);
19881 case ISD::READCYCLECOUNTER: {
19882 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
19885 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
19886 EVT T = N->getValueType(0);
19887 assert((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair");
19888 bool Regs64bit = T == MVT::i128;
19889 MVT HalfT = Regs64bit ? MVT::i64 : MVT::i32;
19890 SDValue cpInL, cpInH;
19891 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
19892 DAG.getConstant(0, dl, HalfT));
19893 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
19894 DAG.getConstant(1, dl, HalfT));
19895 cpInL = DAG.getCopyToReg(N->getOperand(0), dl,
19896 Regs64bit ? X86::RAX : X86::EAX,
19898 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl,
19899 Regs64bit ? X86::RDX : X86::EDX,
19900 cpInH, cpInL.getValue(1));
19901 SDValue swapInL, swapInH;
19902 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
19903 DAG.getConstant(0, dl, HalfT));
19904 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
19905 DAG.getConstant(1, dl, HalfT));
19906 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl,
19907 Regs64bit ? X86::RBX : X86::EBX,
19908 swapInL, cpInH.getValue(1));
19909 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl,
19910 Regs64bit ? X86::RCX : X86::ECX,
19911 swapInH, swapInL.getValue(1));
19912 SDValue Ops[] = { swapInH.getValue(0),
19914 swapInH.getValue(1) };
19915 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
19916 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
19917 unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_DAG :
19918 X86ISD::LCMPXCHG8_DAG;
19919 SDValue Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys, Ops, T, MMO);
19920 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl,
19921 Regs64bit ? X86::RAX : X86::EAX,
19922 HalfT, Result.getValue(1));
19923 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl,
19924 Regs64bit ? X86::RDX : X86::EDX,
19925 HalfT, cpOutL.getValue(2));
19926 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
19928 SDValue EFLAGS = DAG.getCopyFromReg(cpOutH.getValue(1), dl, X86::EFLAGS,
19929 MVT::i32, cpOutH.getValue(2));
19931 DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
19932 DAG.getConstant(X86::COND_E, dl, MVT::i8), EFLAGS);
19933 Success = DAG.getZExtOrTrunc(Success, dl, N->getValueType(1));
19935 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF));
19936 Results.push_back(Success);
19937 Results.push_back(EFLAGS.getValue(1));
19940 case ISD::ATOMIC_SWAP:
19941 case ISD::ATOMIC_LOAD_ADD:
19942 case ISD::ATOMIC_LOAD_SUB:
19943 case ISD::ATOMIC_LOAD_AND:
19944 case ISD::ATOMIC_LOAD_OR:
19945 case ISD::ATOMIC_LOAD_XOR:
19946 case ISD::ATOMIC_LOAD_NAND:
19947 case ISD::ATOMIC_LOAD_MIN:
19948 case ISD::ATOMIC_LOAD_MAX:
19949 case ISD::ATOMIC_LOAD_UMIN:
19950 case ISD::ATOMIC_LOAD_UMAX:
19951 case ISD::ATOMIC_LOAD: {
19952 // Delegate to generic TypeLegalization. Situations we can really handle
19953 // should have already been dealt with by AtomicExpandPass.cpp.
19956 case ISD::BITCAST: {
19957 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19958 EVT DstVT = N->getValueType(0);
19959 EVT SrcVT = N->getOperand(0)->getValueType(0);
19961 if (SrcVT != MVT::f64 ||
19962 (DstVT != MVT::v2i32 && DstVT != MVT::v4i16 && DstVT != MVT::v8i8))
19965 unsigned NumElts = DstVT.getVectorNumElements();
19966 EVT SVT = DstVT.getVectorElementType();
19967 EVT WiderVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
19968 SDValue Expanded = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
19969 MVT::v2f64, N->getOperand(0));
19970 SDValue ToVecInt = DAG.getBitcast(WiderVT, Expanded);
19972 if (ExperimentalVectorWideningLegalization) {
19973 // If we are legalizing vectors by widening, we already have the desired
19974 // legal vector type, just return it.
19975 Results.push_back(ToVecInt);
19979 SmallVector<SDValue, 8> Elts;
19980 for (unsigned i = 0, e = NumElts; i != e; ++i)
19981 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT,
19982 ToVecInt, DAG.getIntPtrConstant(i, dl)));
19984 Results.push_back(DAG.getNode(ISD::BUILD_VECTOR, dl, DstVT, Elts));
19989 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
19990 switch ((X86ISD::NodeType)Opcode) {
19991 case X86ISD::FIRST_NUMBER: break;
19992 case X86ISD::BSF: return "X86ISD::BSF";
19993 case X86ISD::BSR: return "X86ISD::BSR";
19994 case X86ISD::SHLD: return "X86ISD::SHLD";
19995 case X86ISD::SHRD: return "X86ISD::SHRD";
19996 case X86ISD::FAND: return "X86ISD::FAND";
19997 case X86ISD::FANDN: return "X86ISD::FANDN";
19998 case X86ISD::FOR: return "X86ISD::FOR";
19999 case X86ISD::FXOR: return "X86ISD::FXOR";
20000 case X86ISD::FILD: return "X86ISD::FILD";
20001 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
20002 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
20003 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
20004 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
20005 case X86ISD::FLD: return "X86ISD::FLD";
20006 case X86ISD::FST: return "X86ISD::FST";
20007 case X86ISD::CALL: return "X86ISD::CALL";
20008 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
20009 case X86ISD::RDTSCP_DAG: return "X86ISD::RDTSCP_DAG";
20010 case X86ISD::RDPMC_DAG: return "X86ISD::RDPMC_DAG";
20011 case X86ISD::BT: return "X86ISD::BT";
20012 case X86ISD::CMP: return "X86ISD::CMP";
20013 case X86ISD::COMI: return "X86ISD::COMI";
20014 case X86ISD::UCOMI: return "X86ISD::UCOMI";
20015 case X86ISD::CMPM: return "X86ISD::CMPM";
20016 case X86ISD::CMPMU: return "X86ISD::CMPMU";
20017 case X86ISD::CMPM_RND: return "X86ISD::CMPM_RND";
20018 case X86ISD::SETCC: return "X86ISD::SETCC";
20019 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
20020 case X86ISD::FSETCC: return "X86ISD::FSETCC";
20021 case X86ISD::FGETSIGNx86: return "X86ISD::FGETSIGNx86";
20022 case X86ISD::CMOV: return "X86ISD::CMOV";
20023 case X86ISD::BRCOND: return "X86ISD::BRCOND";
20024 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
20025 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
20026 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
20027 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
20028 case X86ISD::Wrapper: return "X86ISD::Wrapper";
20029 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
20030 case X86ISD::MOVDQ2Q: return "X86ISD::MOVDQ2Q";
20031 case X86ISD::MMX_MOVD2W: return "X86ISD::MMX_MOVD2W";
20032 case X86ISD::MMX_MOVW2D: return "X86ISD::MMX_MOVW2D";
20033 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
20034 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
20035 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
20036 case X86ISD::PINSRB: return "X86ISD::PINSRB";
20037 case X86ISD::PINSRW: return "X86ISD::PINSRW";
20038 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
20039 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
20040 case X86ISD::ANDNP: return "X86ISD::ANDNP";
20041 case X86ISD::PSIGN: return "X86ISD::PSIGN";
20042 case X86ISD::BLENDI: return "X86ISD::BLENDI";
20043 case X86ISD::SHRUNKBLEND: return "X86ISD::SHRUNKBLEND";
20044 case X86ISD::ADDUS: return "X86ISD::ADDUS";
20045 case X86ISD::SUBUS: return "X86ISD::SUBUS";
20046 case X86ISD::HADD: return "X86ISD::HADD";
20047 case X86ISD::HSUB: return "X86ISD::HSUB";
20048 case X86ISD::FHADD: return "X86ISD::FHADD";
20049 case X86ISD::FHSUB: return "X86ISD::FHSUB";
20050 case X86ISD::ABS: return "X86ISD::ABS";
20051 case X86ISD::CONFLICT: return "X86ISD::CONFLICT";
20052 case X86ISD::FMAX: return "X86ISD::FMAX";
20053 case X86ISD::FMAX_RND: return "X86ISD::FMAX_RND";
20054 case X86ISD::FMIN: return "X86ISD::FMIN";
20055 case X86ISD::FMIN_RND: return "X86ISD::FMIN_RND";
20056 case X86ISD::FMAXC: return "X86ISD::FMAXC";
20057 case X86ISD::FMINC: return "X86ISD::FMINC";
20058 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
20059 case X86ISD::FRCP: return "X86ISD::FRCP";
20060 case X86ISD::EXTRQI: return "X86ISD::EXTRQI";
20061 case X86ISD::INSERTQI: return "X86ISD::INSERTQI";
20062 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
20063 case X86ISD::TLSBASEADDR: return "X86ISD::TLSBASEADDR";
20064 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
20065 case X86ISD::EH_SJLJ_SETJMP: return "X86ISD::EH_SJLJ_SETJMP";
20066 case X86ISD::EH_SJLJ_LONGJMP: return "X86ISD::EH_SJLJ_LONGJMP";
20067 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
20068 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
20069 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
20070 case X86ISD::FNSTSW16r: return "X86ISD::FNSTSW16r";
20071 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
20072 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
20073 case X86ISD::LCMPXCHG16_DAG: return "X86ISD::LCMPXCHG16_DAG";
20074 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
20075 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
20076 case X86ISD::VZEXT: return "X86ISD::VZEXT";
20077 case X86ISD::VSEXT: return "X86ISD::VSEXT";
20078 case X86ISD::VTRUNC: return "X86ISD::VTRUNC";
20079 case X86ISD::VTRUNCS: return "X86ISD::VTRUNCS";
20080 case X86ISD::VTRUNCUS: return "X86ISD::VTRUNCUS";
20081 case X86ISD::VINSERT: return "X86ISD::VINSERT";
20082 case X86ISD::VFPEXT: return "X86ISD::VFPEXT";
20083 case X86ISD::VFPROUND: return "X86ISD::VFPROUND";
20084 case X86ISD::CVTDQ2PD: return "X86ISD::CVTDQ2PD";
20085 case X86ISD::CVTUDQ2PD: return "X86ISD::CVTUDQ2PD";
20086 case X86ISD::VSHLDQ: return "X86ISD::VSHLDQ";
20087 case X86ISD::VSRLDQ: return "X86ISD::VSRLDQ";
20088 case X86ISD::VSHL: return "X86ISD::VSHL";
20089 case X86ISD::VSRL: return "X86ISD::VSRL";
20090 case X86ISD::VSRA: return "X86ISD::VSRA";
20091 case X86ISD::VSHLI: return "X86ISD::VSHLI";
20092 case X86ISD::VSRLI: return "X86ISD::VSRLI";
20093 case X86ISD::VSRAI: return "X86ISD::VSRAI";
20094 case X86ISD::CMPP: return "X86ISD::CMPP";
20095 case X86ISD::PCMPEQ: return "X86ISD::PCMPEQ";
20096 case X86ISD::PCMPGT: return "X86ISD::PCMPGT";
20097 case X86ISD::PCMPEQM: return "X86ISD::PCMPEQM";
20098 case X86ISD::PCMPGTM: return "X86ISD::PCMPGTM";
20099 case X86ISD::ADD: return "X86ISD::ADD";
20100 case X86ISD::SUB: return "X86ISD::SUB";
20101 case X86ISD::ADC: return "X86ISD::ADC";
20102 case X86ISD::SBB: return "X86ISD::SBB";
20103 case X86ISD::SMUL: return "X86ISD::SMUL";
20104 case X86ISD::UMUL: return "X86ISD::UMUL";
20105 case X86ISD::SMUL8: return "X86ISD::SMUL8";
20106 case X86ISD::UMUL8: return "X86ISD::UMUL8";
20107 case X86ISD::SDIVREM8_SEXT_HREG: return "X86ISD::SDIVREM8_SEXT_HREG";
20108 case X86ISD::UDIVREM8_ZEXT_HREG: return "X86ISD::UDIVREM8_ZEXT_HREG";
20109 case X86ISD::INC: return "X86ISD::INC";
20110 case X86ISD::DEC: return "X86ISD::DEC";
20111 case X86ISD::OR: return "X86ISD::OR";
20112 case X86ISD::XOR: return "X86ISD::XOR";
20113 case X86ISD::AND: return "X86ISD::AND";
20114 case X86ISD::BEXTR: return "X86ISD::BEXTR";
20115 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
20116 case X86ISD::PTEST: return "X86ISD::PTEST";
20117 case X86ISD::TESTP: return "X86ISD::TESTP";
20118 case X86ISD::TESTM: return "X86ISD::TESTM";
20119 case X86ISD::TESTNM: return "X86ISD::TESTNM";
20120 case X86ISD::KORTEST: return "X86ISD::KORTEST";
20121 case X86ISD::KTEST: return "X86ISD::KTEST";
20122 case X86ISD::PACKSS: return "X86ISD::PACKSS";
20123 case X86ISD::PACKUS: return "X86ISD::PACKUS";
20124 case X86ISD::PALIGNR: return "X86ISD::PALIGNR";
20125 case X86ISD::VALIGN: return "X86ISD::VALIGN";
20126 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
20127 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
20128 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
20129 case X86ISD::SHUFP: return "X86ISD::SHUFP";
20130 case X86ISD::SHUF128: return "X86ISD::SHUF128";
20131 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
20132 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
20133 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
20134 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
20135 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
20136 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
20137 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
20138 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
20139 case X86ISD::MOVSD: return "X86ISD::MOVSD";
20140 case X86ISD::MOVSS: return "X86ISD::MOVSS";
20141 case X86ISD::UNPCKL: return "X86ISD::UNPCKL";
20142 case X86ISD::UNPCKH: return "X86ISD::UNPCKH";
20143 case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
20144 case X86ISD::VBROADCASTM: return "X86ISD::VBROADCASTM";
20145 case X86ISD::SUBV_BROADCAST: return "X86ISD::SUBV_BROADCAST";
20146 case X86ISD::VEXTRACT: return "X86ISD::VEXTRACT";
20147 case X86ISD::VPERMILPV: return "X86ISD::VPERMILPV";
20148 case X86ISD::VPERMILPI: return "X86ISD::VPERMILPI";
20149 case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128";
20150 case X86ISD::VPERMV: return "X86ISD::VPERMV";
20151 case X86ISD::VPERMV3: return "X86ISD::VPERMV3";
20152 case X86ISD::VPERMIV3: return "X86ISD::VPERMIV3";
20153 case X86ISD::VPERMI: return "X86ISD::VPERMI";
20154 case X86ISD::VPTERNLOG: return "X86ISD::VPTERNLOG";
20155 case X86ISD::VFIXUPIMM: return "X86ISD::VFIXUPIMM";
20156 case X86ISD::VRANGE: return "X86ISD::VRANGE";
20157 case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
20158 case X86ISD::PMULDQ: return "X86ISD::PMULDQ";
20159 case X86ISD::PSADBW: return "X86ISD::PSADBW";
20160 case X86ISD::DBPSADBW: return "X86ISD::DBPSADBW";
20161 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
20162 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
20163 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
20164 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
20165 case X86ISD::MFENCE: return "X86ISD::MFENCE";
20166 case X86ISD::SFENCE: return "X86ISD::SFENCE";
20167 case X86ISD::LFENCE: return "X86ISD::LFENCE";
20168 case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA";
20169 case X86ISD::SAHF: return "X86ISD::SAHF";
20170 case X86ISD::RDRAND: return "X86ISD::RDRAND";
20171 case X86ISD::RDSEED: return "X86ISD::RDSEED";
20172 case X86ISD::VPMADDUBSW: return "X86ISD::VPMADDUBSW";
20173 case X86ISD::VPMADDWD: return "X86ISD::VPMADDWD";
20174 case X86ISD::VPROT: return "X86ISD::VPROT";
20175 case X86ISD::VPROTI: return "X86ISD::VPROTI";
20176 case X86ISD::VPSHA: return "X86ISD::VPSHA";
20177 case X86ISD::VPSHL: return "X86ISD::VPSHL";
20178 case X86ISD::VPCOM: return "X86ISD::VPCOM";
20179 case X86ISD::VPCOMU: return "X86ISD::VPCOMU";
20180 case X86ISD::FMADD: return "X86ISD::FMADD";
20181 case X86ISD::FMSUB: return "X86ISD::FMSUB";
20182 case X86ISD::FNMADD: return "X86ISD::FNMADD";
20183 case X86ISD::FNMSUB: return "X86ISD::FNMSUB";
20184 case X86ISD::FMADDSUB: return "X86ISD::FMADDSUB";
20185 case X86ISD::FMSUBADD: return "X86ISD::FMSUBADD";
20186 case X86ISD::FMADD_RND: return "X86ISD::FMADD_RND";
20187 case X86ISD::FNMADD_RND: return "X86ISD::FNMADD_RND";
20188 case X86ISD::FMSUB_RND: return "X86ISD::FMSUB_RND";
20189 case X86ISD::FNMSUB_RND: return "X86ISD::FNMSUB_RND";
20190 case X86ISD::FMADDSUB_RND: return "X86ISD::FMADDSUB_RND";
20191 case X86ISD::FMSUBADD_RND: return "X86ISD::FMSUBADD_RND";
20192 case X86ISD::VRNDSCALE: return "X86ISD::VRNDSCALE";
20193 case X86ISD::VREDUCE: return "X86ISD::VREDUCE";
20194 case X86ISD::VGETMANT: return "X86ISD::VGETMANT";
20195 case X86ISD::PCMPESTRI: return "X86ISD::PCMPESTRI";
20196 case X86ISD::PCMPISTRI: return "X86ISD::PCMPISTRI";
20197 case X86ISD::XTEST: return "X86ISD::XTEST";
20198 case X86ISD::COMPRESS: return "X86ISD::COMPRESS";
20199 case X86ISD::EXPAND: return "X86ISD::EXPAND";
20200 case X86ISD::SELECT: return "X86ISD::SELECT";
20201 case X86ISD::ADDSUB: return "X86ISD::ADDSUB";
20202 case X86ISD::RCP28: return "X86ISD::RCP28";
20203 case X86ISD::EXP2: return "X86ISD::EXP2";
20204 case X86ISD::RSQRT28: return "X86ISD::RSQRT28";
20205 case X86ISD::FADD_RND: return "X86ISD::FADD_RND";
20206 case X86ISD::FSUB_RND: return "X86ISD::FSUB_RND";
20207 case X86ISD::FMUL_RND: return "X86ISD::FMUL_RND";
20208 case X86ISD::FDIV_RND: return "X86ISD::FDIV_RND";
20209 case X86ISD::FSQRT_RND: return "X86ISD::FSQRT_RND";
20210 case X86ISD::FGETEXP_RND: return "X86ISD::FGETEXP_RND";
20211 case X86ISD::SCALEF: return "X86ISD::SCALEF";
20212 case X86ISD::ADDS: return "X86ISD::ADDS";
20213 case X86ISD::SUBS: return "X86ISD::SUBS";
20214 case X86ISD::AVG: return "X86ISD::AVG";
20215 case X86ISD::MULHRS: return "X86ISD::MULHRS";
20216 case X86ISD::SINT_TO_FP_RND: return "X86ISD::SINT_TO_FP_RND";
20217 case X86ISD::UINT_TO_FP_RND: return "X86ISD::UINT_TO_FP_RND";
20218 case X86ISD::FP_TO_SINT_RND: return "X86ISD::FP_TO_SINT_RND";
20219 case X86ISD::FP_TO_UINT_RND: return "X86ISD::FP_TO_UINT_RND";
20220 case X86ISD::VFPCLASS: return "X86ISD::VFPCLASS";
20225 // isLegalAddressingMode - Return true if the addressing mode represented
20226 // by AM is legal for this target, for a load/store of the specified type.
20227 bool X86TargetLowering::isLegalAddressingMode(const DataLayout &DL,
20228 const AddrMode &AM, Type *Ty,
20229 unsigned AS) const {
20230 // X86 supports extremely general addressing modes.
20231 CodeModel::Model M = getTargetMachine().getCodeModel();
20232 Reloc::Model R = getTargetMachine().getRelocationModel();
20234 // X86 allows a sign-extended 32-bit immediate field as a displacement.
20235 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != nullptr))
20240 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
20242 // If a reference to this global requires an extra load, we can't fold it.
20243 if (isGlobalStubReference(GVFlags))
20246 // If BaseGV requires a register for the PIC base, we cannot also have a
20247 // BaseReg specified.
20248 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
20251 // If lower 4G is not available, then we must use rip-relative addressing.
20252 if ((M != CodeModel::Small || R != Reloc::Static) &&
20253 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
20257 switch (AM.Scale) {
20263 // These scales always work.
20268 // These scales are formed with basereg+scalereg. Only accept if there is
20273 default: // Other stuff never works.
20280 bool X86TargetLowering::isVectorShiftByScalarCheap(Type *Ty) const {
20281 unsigned Bits = Ty->getScalarSizeInBits();
20283 // 8-bit shifts are always expensive, but versions with a scalar amount aren't
20284 // particularly cheaper than those without.
20288 // On AVX2 there are new vpsllv[dq] instructions (and other shifts), that make
20289 // variable shifts just as cheap as scalar ones.
20290 if (Subtarget->hasInt256() && (Bits == 32 || Bits == 64))
20293 // Otherwise, it's significantly cheaper to shift by a scalar amount than by a
20294 // fully general vector.
20298 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
20299 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
20301 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
20302 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
20303 return NumBits1 > NumBits2;
20306 bool X86TargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
20307 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
20310 if (!isTypeLegal(EVT::getEVT(Ty1)))
20313 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
20315 // Assuming the caller doesn't have a zeroext or signext return parameter,
20316 // truncation all the way down to i1 is valid.
20320 bool X86TargetLowering::isLegalICmpImmediate(int64_t Imm) const {
20321 return isInt<32>(Imm);
20324 bool X86TargetLowering::isLegalAddImmediate(int64_t Imm) const {
20325 // Can also use sub to handle negated immediates.
20326 return isInt<32>(Imm);
20329 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
20330 if (!VT1.isInteger() || !VT2.isInteger())
20332 unsigned NumBits1 = VT1.getSizeInBits();
20333 unsigned NumBits2 = VT2.getSizeInBits();
20334 return NumBits1 > NumBits2;
20337 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
20338 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
20339 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
20342 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
20343 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
20344 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
20347 bool X86TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
20348 EVT VT1 = Val.getValueType();
20349 if (isZExtFree(VT1, VT2))
20352 if (Val.getOpcode() != ISD::LOAD)
20355 if (!VT1.isSimple() || !VT1.isInteger() ||
20356 !VT2.isSimple() || !VT2.isInteger())
20359 switch (VT1.getSimpleVT().SimpleTy) {
20364 // X86 has 8, 16, and 32-bit zero-extending loads.
20371 bool X86TargetLowering::isVectorLoadExtDesirable(SDValue) const { return true; }
20374 X86TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
20375 if (!(Subtarget->hasFMA() || Subtarget->hasFMA4() || Subtarget->hasAVX512()))
20378 VT = VT.getScalarType();
20380 if (!VT.isSimple())
20383 switch (VT.getSimpleVT().SimpleTy) {
20394 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
20395 // i16 instructions are longer (0x66 prefix) and potentially slower.
20396 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
20399 /// isShuffleMaskLegal - Targets can use this to indicate that they only
20400 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
20401 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
20402 /// are assumed to be legal.
20404 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
20406 if (!VT.isSimple())
20409 // Not for i1 vectors
20410 if (VT.getSimpleVT().getScalarType() == MVT::i1)
20413 // Very little shuffling can be done for 64-bit vectors right now.
20414 if (VT.getSimpleVT().getSizeInBits() == 64)
20417 // We only care that the types being shuffled are legal. The lowering can
20418 // handle any possible shuffle mask that results.
20419 return isTypeLegal(VT.getSimpleVT());
20423 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
20425 // Just delegate to the generic legality, clear masks aren't special.
20426 return isShuffleMaskLegal(Mask, VT);
20429 //===----------------------------------------------------------------------===//
20430 // X86 Scheduler Hooks
20431 //===----------------------------------------------------------------------===//
20433 /// Utility function to emit xbegin specifying the start of an RTM region.
20434 static MachineBasicBlock *EmitXBegin(MachineInstr *MI, MachineBasicBlock *MBB,
20435 const TargetInstrInfo *TII) {
20436 DebugLoc DL = MI->getDebugLoc();
20438 const BasicBlock *BB = MBB->getBasicBlock();
20439 MachineFunction::iterator I = ++MBB->getIterator();
20441 // For the v = xbegin(), we generate
20452 MachineBasicBlock *thisMBB = MBB;
20453 MachineFunction *MF = MBB->getParent();
20454 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
20455 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
20456 MF->insert(I, mainMBB);
20457 MF->insert(I, sinkMBB);
20459 // Transfer the remainder of BB and its successor edges to sinkMBB.
20460 sinkMBB->splice(sinkMBB->begin(), MBB,
20461 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
20462 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
20466 // # fallthrough to mainMBB
20467 // # abortion to sinkMBB
20468 BuildMI(thisMBB, DL, TII->get(X86::XBEGIN_4)).addMBB(sinkMBB);
20469 thisMBB->addSuccessor(mainMBB);
20470 thisMBB->addSuccessor(sinkMBB);
20474 BuildMI(mainMBB, DL, TII->get(X86::MOV32ri), X86::EAX).addImm(-1);
20475 mainMBB->addSuccessor(sinkMBB);
20478 // EAX is live into the sinkMBB
20479 sinkMBB->addLiveIn(X86::EAX);
20480 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
20481 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20484 MI->eraseFromParent();
20488 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
20489 // or XMM0_V32I8 in AVX all of this code can be replaced with that
20490 // in the .td file.
20491 static MachineBasicBlock *EmitPCMPSTRM(MachineInstr *MI, MachineBasicBlock *BB,
20492 const TargetInstrInfo *TII) {
20494 switch (MI->getOpcode()) {
20495 default: llvm_unreachable("illegal opcode!");
20496 case X86::PCMPISTRM128REG: Opc = X86::PCMPISTRM128rr; break;
20497 case X86::VPCMPISTRM128REG: Opc = X86::VPCMPISTRM128rr; break;
20498 case X86::PCMPISTRM128MEM: Opc = X86::PCMPISTRM128rm; break;
20499 case X86::VPCMPISTRM128MEM: Opc = X86::VPCMPISTRM128rm; break;
20500 case X86::PCMPESTRM128REG: Opc = X86::PCMPESTRM128rr; break;
20501 case X86::VPCMPESTRM128REG: Opc = X86::VPCMPESTRM128rr; break;
20502 case X86::PCMPESTRM128MEM: Opc = X86::PCMPESTRM128rm; break;
20503 case X86::VPCMPESTRM128MEM: Opc = X86::VPCMPESTRM128rm; break;
20506 DebugLoc dl = MI->getDebugLoc();
20507 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
20509 unsigned NumArgs = MI->getNumOperands();
20510 for (unsigned i = 1; i < NumArgs; ++i) {
20511 MachineOperand &Op = MI->getOperand(i);
20512 if (!(Op.isReg() && Op.isImplicit()))
20513 MIB.addOperand(Op);
20515 if (MI->hasOneMemOperand())
20516 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
20518 BuildMI(*BB, MI, dl,
20519 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20520 .addReg(X86::XMM0);
20522 MI->eraseFromParent();
20526 // FIXME: Custom handling because TableGen doesn't support multiple implicit
20527 // defs in an instruction pattern
20528 static MachineBasicBlock *EmitPCMPSTRI(MachineInstr *MI, MachineBasicBlock *BB,
20529 const TargetInstrInfo *TII) {
20531 switch (MI->getOpcode()) {
20532 default: llvm_unreachable("illegal opcode!");
20533 case X86::PCMPISTRIREG: Opc = X86::PCMPISTRIrr; break;
20534 case X86::VPCMPISTRIREG: Opc = X86::VPCMPISTRIrr; break;
20535 case X86::PCMPISTRIMEM: Opc = X86::PCMPISTRIrm; break;
20536 case X86::VPCMPISTRIMEM: Opc = X86::VPCMPISTRIrm; break;
20537 case X86::PCMPESTRIREG: Opc = X86::PCMPESTRIrr; break;
20538 case X86::VPCMPESTRIREG: Opc = X86::VPCMPESTRIrr; break;
20539 case X86::PCMPESTRIMEM: Opc = X86::PCMPESTRIrm; break;
20540 case X86::VPCMPESTRIMEM: Opc = X86::VPCMPESTRIrm; break;
20543 DebugLoc dl = MI->getDebugLoc();
20544 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
20546 unsigned NumArgs = MI->getNumOperands(); // remove the results
20547 for (unsigned i = 1; i < NumArgs; ++i) {
20548 MachineOperand &Op = MI->getOperand(i);
20549 if (!(Op.isReg() && Op.isImplicit()))
20550 MIB.addOperand(Op);
20552 if (MI->hasOneMemOperand())
20553 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
20555 BuildMI(*BB, MI, dl,
20556 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20559 MI->eraseFromParent();
20563 static MachineBasicBlock *EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB,
20564 const X86Subtarget *Subtarget) {
20565 DebugLoc dl = MI->getDebugLoc();
20566 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20567 // Address into RAX/EAX, other two args into ECX, EDX.
20568 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
20569 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
20570 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
20571 for (int i = 0; i < X86::AddrNumOperands; ++i)
20572 MIB.addOperand(MI->getOperand(i));
20574 unsigned ValOps = X86::AddrNumOperands;
20575 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
20576 .addReg(MI->getOperand(ValOps).getReg());
20577 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
20578 .addReg(MI->getOperand(ValOps+1).getReg());
20580 // The instruction doesn't actually take any operands though.
20581 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
20583 MI->eraseFromParent(); // The pseudo is gone now.
20587 MachineBasicBlock *
20588 X86TargetLowering::EmitVAARG64WithCustomInserter(MachineInstr *MI,
20589 MachineBasicBlock *MBB) const {
20590 // Emit va_arg instruction on X86-64.
20592 // Operands to this pseudo-instruction:
20593 // 0 ) Output : destination address (reg)
20594 // 1-5) Input : va_list address (addr, i64mem)
20595 // 6 ) ArgSize : Size (in bytes) of vararg type
20596 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
20597 // 8 ) Align : Alignment of type
20598 // 9 ) EFLAGS (implicit-def)
20600 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
20601 static_assert(X86::AddrNumOperands == 5,
20602 "VAARG_64 assumes 5 address operands");
20604 unsigned DestReg = MI->getOperand(0).getReg();
20605 MachineOperand &Base = MI->getOperand(1);
20606 MachineOperand &Scale = MI->getOperand(2);
20607 MachineOperand &Index = MI->getOperand(3);
20608 MachineOperand &Disp = MI->getOperand(4);
20609 MachineOperand &Segment = MI->getOperand(5);
20610 unsigned ArgSize = MI->getOperand(6).getImm();
20611 unsigned ArgMode = MI->getOperand(7).getImm();
20612 unsigned Align = MI->getOperand(8).getImm();
20614 // Memory Reference
20615 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
20616 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
20617 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
20619 // Machine Information
20620 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20621 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
20622 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
20623 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
20624 DebugLoc DL = MI->getDebugLoc();
20626 // struct va_list {
20629 // i64 overflow_area (address)
20630 // i64 reg_save_area (address)
20632 // sizeof(va_list) = 24
20633 // alignment(va_list) = 8
20635 unsigned TotalNumIntRegs = 6;
20636 unsigned TotalNumXMMRegs = 8;
20637 bool UseGPOffset = (ArgMode == 1);
20638 bool UseFPOffset = (ArgMode == 2);
20639 unsigned MaxOffset = TotalNumIntRegs * 8 +
20640 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
20642 /* Align ArgSize to a multiple of 8 */
20643 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
20644 bool NeedsAlign = (Align > 8);
20646 MachineBasicBlock *thisMBB = MBB;
20647 MachineBasicBlock *overflowMBB;
20648 MachineBasicBlock *offsetMBB;
20649 MachineBasicBlock *endMBB;
20651 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
20652 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
20653 unsigned OffsetReg = 0;
20655 if (!UseGPOffset && !UseFPOffset) {
20656 // If we only pull from the overflow region, we don't create a branch.
20657 // We don't need to alter control flow.
20658 OffsetDestReg = 0; // unused
20659 OverflowDestReg = DestReg;
20661 offsetMBB = nullptr;
20662 overflowMBB = thisMBB;
20665 // First emit code to check if gp_offset (or fp_offset) is below the bound.
20666 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
20667 // If not, pull from overflow_area. (branch to overflowMBB)
20672 // offsetMBB overflowMBB
20677 // Registers for the PHI in endMBB
20678 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
20679 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
20681 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
20682 MachineFunction *MF = MBB->getParent();
20683 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20684 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20685 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20687 MachineFunction::iterator MBBIter = ++MBB->getIterator();
20689 // Insert the new basic blocks
20690 MF->insert(MBBIter, offsetMBB);
20691 MF->insert(MBBIter, overflowMBB);
20692 MF->insert(MBBIter, endMBB);
20694 // Transfer the remainder of MBB and its successor edges to endMBB.
20695 endMBB->splice(endMBB->begin(), thisMBB,
20696 std::next(MachineBasicBlock::iterator(MI)), thisMBB->end());
20697 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
20699 // Make offsetMBB and overflowMBB successors of thisMBB
20700 thisMBB->addSuccessor(offsetMBB);
20701 thisMBB->addSuccessor(overflowMBB);
20703 // endMBB is a successor of both offsetMBB and overflowMBB
20704 offsetMBB->addSuccessor(endMBB);
20705 overflowMBB->addSuccessor(endMBB);
20707 // Load the offset value into a register
20708 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
20709 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
20713 .addDisp(Disp, UseFPOffset ? 4 : 0)
20714 .addOperand(Segment)
20715 .setMemRefs(MMOBegin, MMOEnd);
20717 // Check if there is enough room left to pull this argument.
20718 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
20720 .addImm(MaxOffset + 8 - ArgSizeA8);
20722 // Branch to "overflowMBB" if offset >= max
20723 // Fall through to "offsetMBB" otherwise
20724 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
20725 .addMBB(overflowMBB);
20728 // In offsetMBB, emit code to use the reg_save_area.
20730 assert(OffsetReg != 0);
20732 // Read the reg_save_area address.
20733 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
20734 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
20739 .addOperand(Segment)
20740 .setMemRefs(MMOBegin, MMOEnd);
20742 // Zero-extend the offset
20743 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
20744 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
20747 .addImm(X86::sub_32bit);
20749 // Add the offset to the reg_save_area to get the final address.
20750 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
20751 .addReg(OffsetReg64)
20752 .addReg(RegSaveReg);
20754 // Compute the offset for the next argument
20755 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
20756 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
20758 .addImm(UseFPOffset ? 16 : 8);
20760 // Store it back into the va_list.
20761 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
20765 .addDisp(Disp, UseFPOffset ? 4 : 0)
20766 .addOperand(Segment)
20767 .addReg(NextOffsetReg)
20768 .setMemRefs(MMOBegin, MMOEnd);
20771 BuildMI(offsetMBB, DL, TII->get(X86::JMP_1))
20776 // Emit code to use overflow area
20779 // Load the overflow_area address into a register.
20780 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
20781 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
20786 .addOperand(Segment)
20787 .setMemRefs(MMOBegin, MMOEnd);
20789 // If we need to align it, do so. Otherwise, just copy the address
20790 // to OverflowDestReg.
20792 // Align the overflow address
20793 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
20794 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
20796 // aligned_addr = (addr + (align-1)) & ~(align-1)
20797 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
20798 .addReg(OverflowAddrReg)
20801 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
20803 .addImm(~(uint64_t)(Align-1));
20805 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
20806 .addReg(OverflowAddrReg);
20809 // Compute the next overflow address after this argument.
20810 // (the overflow address should be kept 8-byte aligned)
20811 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
20812 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
20813 .addReg(OverflowDestReg)
20814 .addImm(ArgSizeA8);
20816 // Store the new overflow address.
20817 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
20822 .addOperand(Segment)
20823 .addReg(NextAddrReg)
20824 .setMemRefs(MMOBegin, MMOEnd);
20826 // If we branched, emit the PHI to the front of endMBB.
20828 BuildMI(*endMBB, endMBB->begin(), DL,
20829 TII->get(X86::PHI), DestReg)
20830 .addReg(OffsetDestReg).addMBB(offsetMBB)
20831 .addReg(OverflowDestReg).addMBB(overflowMBB);
20834 // Erase the pseudo instruction
20835 MI->eraseFromParent();
20840 MachineBasicBlock *
20841 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
20843 MachineBasicBlock *MBB) const {
20844 // Emit code to save XMM registers to the stack. The ABI says that the
20845 // number of registers to save is given in %al, so it's theoretically
20846 // possible to do an indirect jump trick to avoid saving all of them,
20847 // however this code takes a simpler approach and just executes all
20848 // of the stores if %al is non-zero. It's less code, and it's probably
20849 // easier on the hardware branch predictor, and stores aren't all that
20850 // expensive anyway.
20852 // Create the new basic blocks. One block contains all the XMM stores,
20853 // and one block is the final destination regardless of whether any
20854 // stores were performed.
20855 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
20856 MachineFunction *F = MBB->getParent();
20857 MachineFunction::iterator MBBIter = ++MBB->getIterator();
20858 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
20859 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
20860 F->insert(MBBIter, XMMSaveMBB);
20861 F->insert(MBBIter, EndMBB);
20863 // Transfer the remainder of MBB and its successor edges to EndMBB.
20864 EndMBB->splice(EndMBB->begin(), MBB,
20865 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
20866 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
20868 // The original block will now fall through to the XMM save block.
20869 MBB->addSuccessor(XMMSaveMBB);
20870 // The XMMSaveMBB will fall through to the end block.
20871 XMMSaveMBB->addSuccessor(EndMBB);
20873 // Now add the instructions.
20874 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20875 DebugLoc DL = MI->getDebugLoc();
20877 unsigned CountReg = MI->getOperand(0).getReg();
20878 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
20879 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
20881 if (!Subtarget->isCallingConvWin64(F->getFunction()->getCallingConv())) {
20882 // If %al is 0, branch around the XMM save block.
20883 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
20884 BuildMI(MBB, DL, TII->get(X86::JE_1)).addMBB(EndMBB);
20885 MBB->addSuccessor(EndMBB);
20888 // Make sure the last operand is EFLAGS, which gets clobbered by the branch
20889 // that was just emitted, but clearly shouldn't be "saved".
20890 assert((MI->getNumOperands() <= 3 ||
20891 !MI->getOperand(MI->getNumOperands() - 1).isReg() ||
20892 MI->getOperand(MI->getNumOperands() - 1).getReg() == X86::EFLAGS)
20893 && "Expected last argument to be EFLAGS");
20894 unsigned MOVOpc = Subtarget->hasFp256() ? X86::VMOVAPSmr : X86::MOVAPSmr;
20895 // In the XMM save block, save all the XMM argument registers.
20896 for (int i = 3, e = MI->getNumOperands() - 1; i != e; ++i) {
20897 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
20898 MachineMemOperand *MMO = F->getMachineMemOperand(
20899 MachinePointerInfo::getFixedStack(*F, RegSaveFrameIndex, Offset),
20900 MachineMemOperand::MOStore,
20901 /*Size=*/16, /*Align=*/16);
20902 BuildMI(XMMSaveMBB, DL, TII->get(MOVOpc))
20903 .addFrameIndex(RegSaveFrameIndex)
20904 .addImm(/*Scale=*/1)
20905 .addReg(/*IndexReg=*/0)
20906 .addImm(/*Disp=*/Offset)
20907 .addReg(/*Segment=*/0)
20908 .addReg(MI->getOperand(i).getReg())
20909 .addMemOperand(MMO);
20912 MI->eraseFromParent(); // The pseudo instruction is gone now.
20917 // The EFLAGS operand of SelectItr might be missing a kill marker
20918 // because there were multiple uses of EFLAGS, and ISel didn't know
20919 // which to mark. Figure out whether SelectItr should have had a
20920 // kill marker, and set it if it should. Returns the correct kill
20922 static bool checkAndUpdateEFLAGSKill(MachineBasicBlock::iterator SelectItr,
20923 MachineBasicBlock* BB,
20924 const TargetRegisterInfo* TRI) {
20925 // Scan forward through BB for a use/def of EFLAGS.
20926 MachineBasicBlock::iterator miI(std::next(SelectItr));
20927 for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) {
20928 const MachineInstr& mi = *miI;
20929 if (mi.readsRegister(X86::EFLAGS))
20931 if (mi.definesRegister(X86::EFLAGS))
20932 break; // Should have kill-flag - update below.
20935 // If we hit the end of the block, check whether EFLAGS is live into a
20937 if (miI == BB->end()) {
20938 for (MachineBasicBlock::succ_iterator sItr = BB->succ_begin(),
20939 sEnd = BB->succ_end();
20940 sItr != sEnd; ++sItr) {
20941 MachineBasicBlock* succ = *sItr;
20942 if (succ->isLiveIn(X86::EFLAGS))
20947 // We found a def, or hit the end of the basic block and EFLAGS wasn't live
20948 // out. SelectMI should have a kill flag on EFLAGS.
20949 SelectItr->addRegisterKilled(X86::EFLAGS, TRI);
20953 // Return true if it is OK for this CMOV pseudo-opcode to be cascaded
20954 // together with other CMOV pseudo-opcodes into a single basic-block with
20955 // conditional jump around it.
20956 static bool isCMOVPseudo(MachineInstr *MI) {
20957 switch (MI->getOpcode()) {
20958 case X86::CMOV_FR32:
20959 case X86::CMOV_FR64:
20960 case X86::CMOV_GR8:
20961 case X86::CMOV_GR16:
20962 case X86::CMOV_GR32:
20963 case X86::CMOV_RFP32:
20964 case X86::CMOV_RFP64:
20965 case X86::CMOV_RFP80:
20966 case X86::CMOV_V2F64:
20967 case X86::CMOV_V2I64:
20968 case X86::CMOV_V4F32:
20969 case X86::CMOV_V4F64:
20970 case X86::CMOV_V4I64:
20971 case X86::CMOV_V16F32:
20972 case X86::CMOV_V8F32:
20973 case X86::CMOV_V8F64:
20974 case X86::CMOV_V8I64:
20975 case X86::CMOV_V8I1:
20976 case X86::CMOV_V16I1:
20977 case X86::CMOV_V32I1:
20978 case X86::CMOV_V64I1:
20986 MachineBasicBlock *
20987 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
20988 MachineBasicBlock *BB) const {
20989 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20990 DebugLoc DL = MI->getDebugLoc();
20992 // To "insert" a SELECT_CC instruction, we actually have to insert the
20993 // diamond control-flow pattern. The incoming instruction knows the
20994 // destination vreg to set, the condition code register to branch on, the
20995 // true/false values to select between, and a branch opcode to use.
20996 const BasicBlock *LLVM_BB = BB->getBasicBlock();
20997 MachineFunction::iterator It = ++BB->getIterator();
21002 // cmpTY ccX, r1, r2
21004 // fallthrough --> copy0MBB
21005 MachineBasicBlock *thisMBB = BB;
21006 MachineFunction *F = BB->getParent();
21008 // This code lowers all pseudo-CMOV instructions. Generally it lowers these
21009 // as described above, by inserting a BB, and then making a PHI at the join
21010 // point to select the true and false operands of the CMOV in the PHI.
21012 // The code also handles two different cases of multiple CMOV opcodes
21016 // In this case, there are multiple CMOVs in a row, all which are based on
21017 // the same condition setting (or the exact opposite condition setting).
21018 // In this case we can lower all the CMOVs using a single inserted BB, and
21019 // then make a number of PHIs at the join point to model the CMOVs. The only
21020 // trickiness here, is that in a case like:
21022 // t2 = CMOV cond1 t1, f1
21023 // t3 = CMOV cond1 t2, f2
21025 // when rewriting this into PHIs, we have to perform some renaming on the
21026 // temps since you cannot have a PHI operand refer to a PHI result earlier
21027 // in the same block. The "simple" but wrong lowering would be:
21029 // t2 = PHI t1(BB1), f1(BB2)
21030 // t3 = PHI t2(BB1), f2(BB2)
21032 // but clearly t2 is not defined in BB1, so that is incorrect. The proper
21033 // renaming is to note that on the path through BB1, t2 is really just a
21034 // copy of t1, and do that renaming, properly generating:
21036 // t2 = PHI t1(BB1), f1(BB2)
21037 // t3 = PHI t1(BB1), f2(BB2)
21039 // Case 2, we lower cascaded CMOVs such as
21041 // (CMOV (CMOV F, T, cc1), T, cc2)
21043 // to two successives branches. For that, we look for another CMOV as the
21044 // following instruction.
21046 // Without this, we would add a PHI between the two jumps, which ends up
21047 // creating a few copies all around. For instance, for
21049 // (sitofp (zext (fcmp une)))
21051 // we would generate:
21053 // ucomiss %xmm1, %xmm0
21054 // movss <1.0f>, %xmm0
21055 // movaps %xmm0, %xmm1
21057 // xorps %xmm1, %xmm1
21060 // movaps %xmm1, %xmm0
21064 // because this custom-inserter would have generated:
21076 // A: X = ...; Y = ...
21078 // C: Z = PHI [X, A], [Y, B]
21080 // E: PHI [X, C], [Z, D]
21082 // If we lower both CMOVs in a single step, we can instead generate:
21094 // A: X = ...; Y = ...
21096 // E: PHI [X, A], [X, C], [Y, D]
21098 // Which, in our sitofp/fcmp example, gives us something like:
21100 // ucomiss %xmm1, %xmm0
21101 // movss <1.0f>, %xmm0
21104 // xorps %xmm0, %xmm0
21108 MachineInstr *CascadedCMOV = nullptr;
21109 MachineInstr *LastCMOV = MI;
21110 X86::CondCode CC = X86::CondCode(MI->getOperand(3).getImm());
21111 X86::CondCode OppCC = X86::GetOppositeBranchCondition(CC);
21112 MachineBasicBlock::iterator NextMIIt =
21113 std::next(MachineBasicBlock::iterator(MI));
21115 // Check for case 1, where there are multiple CMOVs with the same condition
21116 // first. Of the two cases of multiple CMOV lowerings, case 1 reduces the
21117 // number of jumps the most.
21119 if (isCMOVPseudo(MI)) {
21120 // See if we have a string of CMOVS with the same condition.
21121 while (NextMIIt != BB->end() &&
21122 isCMOVPseudo(NextMIIt) &&
21123 (NextMIIt->getOperand(3).getImm() == CC ||
21124 NextMIIt->getOperand(3).getImm() == OppCC)) {
21125 LastCMOV = &*NextMIIt;
21130 // This checks for case 2, but only do this if we didn't already find
21131 // case 1, as indicated by LastCMOV == MI.
21132 if (LastCMOV == MI &&
21133 NextMIIt != BB->end() && NextMIIt->getOpcode() == MI->getOpcode() &&
21134 NextMIIt->getOperand(2).getReg() == MI->getOperand(2).getReg() &&
21135 NextMIIt->getOperand(1).getReg() == MI->getOperand(0).getReg()) {
21136 CascadedCMOV = &*NextMIIt;
21139 MachineBasicBlock *jcc1MBB = nullptr;
21141 // If we have a cascaded CMOV, we lower it to two successive branches to
21142 // the same block. EFLAGS is used by both, so mark it as live in the second.
21143 if (CascadedCMOV) {
21144 jcc1MBB = F->CreateMachineBasicBlock(LLVM_BB);
21145 F->insert(It, jcc1MBB);
21146 jcc1MBB->addLiveIn(X86::EFLAGS);
21149 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
21150 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
21151 F->insert(It, copy0MBB);
21152 F->insert(It, sinkMBB);
21154 // If the EFLAGS register isn't dead in the terminator, then claim that it's
21155 // live into the sink and copy blocks.
21156 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
21158 MachineInstr *LastEFLAGSUser = CascadedCMOV ? CascadedCMOV : LastCMOV;
21159 if (!LastEFLAGSUser->killsRegister(X86::EFLAGS) &&
21160 !checkAndUpdateEFLAGSKill(LastEFLAGSUser, BB, TRI)) {
21161 copy0MBB->addLiveIn(X86::EFLAGS);
21162 sinkMBB->addLiveIn(X86::EFLAGS);
21165 // Transfer the remainder of BB and its successor edges to sinkMBB.
21166 sinkMBB->splice(sinkMBB->begin(), BB,
21167 std::next(MachineBasicBlock::iterator(LastCMOV)), BB->end());
21168 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
21170 // Add the true and fallthrough blocks as its successors.
21171 if (CascadedCMOV) {
21172 // The fallthrough block may be jcc1MBB, if we have a cascaded CMOV.
21173 BB->addSuccessor(jcc1MBB);
21175 // In that case, jcc1MBB will itself fallthrough the copy0MBB, and
21176 // jump to the sinkMBB.
21177 jcc1MBB->addSuccessor(copy0MBB);
21178 jcc1MBB->addSuccessor(sinkMBB);
21180 BB->addSuccessor(copy0MBB);
21183 // The true block target of the first (or only) branch is always sinkMBB.
21184 BB->addSuccessor(sinkMBB);
21186 // Create the conditional branch instruction.
21187 unsigned Opc = X86::GetCondBranchFromCond(CC);
21188 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
21190 if (CascadedCMOV) {
21191 unsigned Opc2 = X86::GetCondBranchFromCond(
21192 (X86::CondCode)CascadedCMOV->getOperand(3).getImm());
21193 BuildMI(jcc1MBB, DL, TII->get(Opc2)).addMBB(sinkMBB);
21197 // %FalseValue = ...
21198 // # fallthrough to sinkMBB
21199 copy0MBB->addSuccessor(sinkMBB);
21202 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
21204 MachineBasicBlock::iterator MIItBegin = MachineBasicBlock::iterator(MI);
21205 MachineBasicBlock::iterator MIItEnd =
21206 std::next(MachineBasicBlock::iterator(LastCMOV));
21207 MachineBasicBlock::iterator SinkInsertionPoint = sinkMBB->begin();
21208 DenseMap<unsigned, std::pair<unsigned, unsigned>> RegRewriteTable;
21209 MachineInstrBuilder MIB;
21211 // As we are creating the PHIs, we have to be careful if there is more than
21212 // one. Later CMOVs may reference the results of earlier CMOVs, but later
21213 // PHIs have to reference the individual true/false inputs from earlier PHIs.
21214 // That also means that PHI construction must work forward from earlier to
21215 // later, and that the code must maintain a mapping from earlier PHI's
21216 // destination registers, and the registers that went into the PHI.
21218 for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd; ++MIIt) {
21219 unsigned DestReg = MIIt->getOperand(0).getReg();
21220 unsigned Op1Reg = MIIt->getOperand(1).getReg();
21221 unsigned Op2Reg = MIIt->getOperand(2).getReg();
21223 // If this CMOV we are generating is the opposite condition from
21224 // the jump we generated, then we have to swap the operands for the
21225 // PHI that is going to be generated.
21226 if (MIIt->getOperand(3).getImm() == OppCC)
21227 std::swap(Op1Reg, Op2Reg);
21229 if (RegRewriteTable.find(Op1Reg) != RegRewriteTable.end())
21230 Op1Reg = RegRewriteTable[Op1Reg].first;
21232 if (RegRewriteTable.find(Op2Reg) != RegRewriteTable.end())
21233 Op2Reg = RegRewriteTable[Op2Reg].second;
21235 MIB = BuildMI(*sinkMBB, SinkInsertionPoint, DL,
21236 TII->get(X86::PHI), DestReg)
21237 .addReg(Op1Reg).addMBB(copy0MBB)
21238 .addReg(Op2Reg).addMBB(thisMBB);
21240 // Add this PHI to the rewrite table.
21241 RegRewriteTable[DestReg] = std::make_pair(Op1Reg, Op2Reg);
21244 // If we have a cascaded CMOV, the second Jcc provides the same incoming
21245 // value as the first Jcc (the True operand of the SELECT_CC/CMOV nodes).
21246 if (CascadedCMOV) {
21247 MIB.addReg(MI->getOperand(2).getReg()).addMBB(jcc1MBB);
21248 // Copy the PHI result to the register defined by the second CMOV.
21249 BuildMI(*sinkMBB, std::next(MachineBasicBlock::iterator(MIB.getInstr())),
21250 DL, TII->get(TargetOpcode::COPY),
21251 CascadedCMOV->getOperand(0).getReg())
21252 .addReg(MI->getOperand(0).getReg());
21253 CascadedCMOV->eraseFromParent();
21256 // Now remove the CMOV(s).
21257 for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd; )
21258 (MIIt++)->eraseFromParent();
21263 MachineBasicBlock *
21264 X86TargetLowering::EmitLoweredAtomicFP(MachineInstr *MI,
21265 MachineBasicBlock *BB) const {
21266 // Combine the following atomic floating-point modification pattern:
21267 // a.store(reg OP a.load(acquire), release)
21268 // Transform them into:
21269 // OPss (%gpr), %xmm
21270 // movss %xmm, (%gpr)
21271 // Or sd equivalent for 64-bit operations.
21273 switch (MI->getOpcode()) {
21274 default: llvm_unreachable("unexpected instr type for EmitLoweredAtomicFP");
21275 case X86::RELEASE_FADD32mr: MOp = X86::MOVSSmr; FOp = X86::ADDSSrm; break;
21276 case X86::RELEASE_FADD64mr: MOp = X86::MOVSDmr; FOp = X86::ADDSDrm; break;
21278 const X86InstrInfo *TII = Subtarget->getInstrInfo();
21279 DebugLoc DL = MI->getDebugLoc();
21280 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
21281 MachineOperand MSrc = MI->getOperand(0);
21282 unsigned VSrc = MI->getOperand(5).getReg();
21283 const MachineOperand &Disp = MI->getOperand(3);
21284 MachineOperand ZeroDisp = MachineOperand::CreateImm(0);
21285 bool hasDisp = Disp.isGlobal() || Disp.isImm();
21286 if (hasDisp && MSrc.isReg())
21287 MSrc.setIsKill(false);
21288 MachineInstrBuilder MIM = BuildMI(*BB, MI, DL, TII->get(MOp))
21289 .addOperand(/*Base=*/MSrc)
21290 .addImm(/*Scale=*/1)
21291 .addReg(/*Index=*/0)
21292 .addDisp(hasDisp ? Disp : ZeroDisp, /*off=*/0)
21294 MachineInstr *MIO = BuildMI(*BB, (MachineInstr *)MIM, DL, TII->get(FOp),
21295 MRI.createVirtualRegister(MRI.getRegClass(VSrc)))
21297 .addOperand(/*Base=*/MSrc)
21298 .addImm(/*Scale=*/1)
21299 .addReg(/*Index=*/0)
21300 .addDisp(hasDisp ? Disp : ZeroDisp, /*off=*/0)
21301 .addReg(/*Segment=*/0);
21302 MIM.addReg(MIO->getOperand(0).getReg(), RegState::Kill);
21303 MI->eraseFromParent(); // The pseudo instruction is gone now.
21307 MachineBasicBlock *
21308 X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI,
21309 MachineBasicBlock *BB) const {
21310 MachineFunction *MF = BB->getParent();
21311 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21312 DebugLoc DL = MI->getDebugLoc();
21313 const BasicBlock *LLVM_BB = BB->getBasicBlock();
21315 assert(MF->shouldSplitStack());
21317 const bool Is64Bit = Subtarget->is64Bit();
21318 const bool IsLP64 = Subtarget->isTarget64BitLP64();
21320 const unsigned TlsReg = Is64Bit ? X86::FS : X86::GS;
21321 const unsigned TlsOffset = IsLP64 ? 0x70 : Is64Bit ? 0x40 : 0x30;
21324 // ... [Till the alloca]
21325 // If stacklet is not large enough, jump to mallocMBB
21328 // Allocate by subtracting from RSP
21329 // Jump to continueMBB
21332 // Allocate by call to runtime
21336 // [rest of original BB]
21339 MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21340 MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21341 MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21343 MachineRegisterInfo &MRI = MF->getRegInfo();
21344 const TargetRegisterClass *AddrRegClass =
21345 getRegClassFor(getPointerTy(MF->getDataLayout()));
21347 unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass),
21348 bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass),
21349 tmpSPVReg = MRI.createVirtualRegister(AddrRegClass),
21350 SPLimitVReg = MRI.createVirtualRegister(AddrRegClass),
21351 sizeVReg = MI->getOperand(1).getReg(),
21352 physSPReg = IsLP64 || Subtarget->isTargetNaCl64() ? X86::RSP : X86::ESP;
21354 MachineFunction::iterator MBBIter = ++BB->getIterator();
21356 MF->insert(MBBIter, bumpMBB);
21357 MF->insert(MBBIter, mallocMBB);
21358 MF->insert(MBBIter, continueMBB);
21360 continueMBB->splice(continueMBB->begin(), BB,
21361 std::next(MachineBasicBlock::iterator(MI)), BB->end());
21362 continueMBB->transferSuccessorsAndUpdatePHIs(BB);
21364 // Add code to the main basic block to check if the stack limit has been hit,
21365 // and if so, jump to mallocMBB otherwise to bumpMBB.
21366 BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg);
21367 BuildMI(BB, DL, TII->get(IsLP64 ? X86::SUB64rr:X86::SUB32rr), SPLimitVReg)
21368 .addReg(tmpSPVReg).addReg(sizeVReg);
21369 BuildMI(BB, DL, TII->get(IsLP64 ? X86::CMP64mr:X86::CMP32mr))
21370 .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg)
21371 .addReg(SPLimitVReg);
21372 BuildMI(BB, DL, TII->get(X86::JG_1)).addMBB(mallocMBB);
21374 // bumpMBB simply decreases the stack pointer, since we know the current
21375 // stacklet has enough space.
21376 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg)
21377 .addReg(SPLimitVReg);
21378 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg)
21379 .addReg(SPLimitVReg);
21380 BuildMI(bumpMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
21382 // Calls into a routine in libgcc to allocate more space from the heap.
21383 const uint32_t *RegMask =
21384 Subtarget->getRegisterInfo()->getCallPreservedMask(*MF, CallingConv::C);
21386 BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
21388 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
21389 .addExternalSymbol("__morestack_allocate_stack_space")
21390 .addRegMask(RegMask)
21391 .addReg(X86::RDI, RegState::Implicit)
21392 .addReg(X86::RAX, RegState::ImplicitDefine);
21393 } else if (Is64Bit) {
21394 BuildMI(mallocMBB, DL, TII->get(X86::MOV32rr), X86::EDI)
21396 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
21397 .addExternalSymbol("__morestack_allocate_stack_space")
21398 .addRegMask(RegMask)
21399 .addReg(X86::EDI, RegState::Implicit)
21400 .addReg(X86::EAX, RegState::ImplicitDefine);
21402 BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg)
21404 BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg);
21405 BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32))
21406 .addExternalSymbol("__morestack_allocate_stack_space")
21407 .addRegMask(RegMask)
21408 .addReg(X86::EAX, RegState::ImplicitDefine);
21412 BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg)
21415 BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg)
21416 .addReg(IsLP64 ? X86::RAX : X86::EAX);
21417 BuildMI(mallocMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
21419 // Set up the CFG correctly.
21420 BB->addSuccessor(bumpMBB);
21421 BB->addSuccessor(mallocMBB);
21422 mallocMBB->addSuccessor(continueMBB);
21423 bumpMBB->addSuccessor(continueMBB);
21425 // Take care of the PHI nodes.
21426 BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI),
21427 MI->getOperand(0).getReg())
21428 .addReg(mallocPtrVReg).addMBB(mallocMBB)
21429 .addReg(bumpSPPtrVReg).addMBB(bumpMBB);
21431 // Delete the original pseudo instruction.
21432 MI->eraseFromParent();
21435 return continueMBB;
21438 MachineBasicBlock *
21439 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
21440 MachineBasicBlock *BB) const {
21441 assert(!Subtarget->isTargetMachO());
21442 DebugLoc DL = MI->getDebugLoc();
21443 MachineInstr *ResumeMI = Subtarget->getFrameLowering()->emitStackProbe(
21444 *BB->getParent(), *BB, MI, DL, false);
21445 MachineBasicBlock *ResumeBB = ResumeMI->getParent();
21446 MI->eraseFromParent(); // The pseudo instruction is gone now.
21450 MachineBasicBlock *
21451 X86TargetLowering::EmitLoweredCatchRet(MachineInstr *MI,
21452 MachineBasicBlock *BB) const {
21453 MachineFunction *MF = BB->getParent();
21454 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
21455 MachineBasicBlock *TargetMBB = MI->getOperand(0).getMBB();
21456 DebugLoc DL = MI->getDebugLoc();
21458 assert(!isAsynchronousEHPersonality(
21459 classifyEHPersonality(MF->getFunction()->getPersonalityFn())) &&
21460 "SEH does not use catchret!");
21462 // Only 32-bit EH needs to worry about manually restoring stack pointers.
21463 if (!Subtarget->is32Bit())
21466 // C++ EH creates a new target block to hold the restore code, and wires up
21467 // the new block to the return destination with a normal JMP_4.
21468 MachineBasicBlock *RestoreMBB =
21469 MF->CreateMachineBasicBlock(BB->getBasicBlock());
21470 assert(BB->succ_size() == 1);
21471 MF->insert(std::next(BB->getIterator()), RestoreMBB);
21472 RestoreMBB->transferSuccessorsAndUpdatePHIs(BB);
21473 BB->addSuccessor(RestoreMBB);
21474 MI->getOperand(0).setMBB(RestoreMBB);
21476 auto RestoreMBBI = RestoreMBB->begin();
21477 BuildMI(*RestoreMBB, RestoreMBBI, DL, TII.get(X86::EH_RESTORE));
21478 BuildMI(*RestoreMBB, RestoreMBBI, DL, TII.get(X86::JMP_4)).addMBB(TargetMBB);
21482 MachineBasicBlock *
21483 X86TargetLowering::EmitLoweredCatchPad(MachineInstr *MI,
21484 MachineBasicBlock *BB) const {
21485 MachineFunction *MF = BB->getParent();
21486 const Constant *PerFn = MF->getFunction()->getPersonalityFn();
21487 bool IsSEH = isAsynchronousEHPersonality(classifyEHPersonality(PerFn));
21488 // Only 32-bit SEH requires special handling for catchpad.
21489 if (IsSEH && Subtarget->is32Bit()) {
21490 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
21491 DebugLoc DL = MI->getDebugLoc();
21492 BuildMI(*BB, MI, DL, TII.get(X86::EH_RESTORE));
21494 MI->eraseFromParent();
21498 MachineBasicBlock *
21499 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
21500 MachineBasicBlock *BB) const {
21501 // This is pretty easy. We're taking the value that we received from
21502 // our load from the relocation, sticking it in either RDI (x86-64)
21503 // or EAX and doing an indirect call. The return value will then
21504 // be in the normal return register.
21505 MachineFunction *F = BB->getParent();
21506 const X86InstrInfo *TII = Subtarget->getInstrInfo();
21507 DebugLoc DL = MI->getDebugLoc();
21509 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
21510 assert(MI->getOperand(3).isGlobal() && "This should be a global");
21512 // Get a register mask for the lowered call.
21513 // FIXME: The 32-bit calls have non-standard calling conventions. Use a
21514 // proper register mask.
21515 const uint32_t *RegMask =
21516 Subtarget->is64Bit() ?
21517 Subtarget->getRegisterInfo()->getDarwinTLSCallPreservedMask() :
21518 Subtarget->getRegisterInfo()->getCallPreservedMask(*F, CallingConv::C);
21519 if (Subtarget->is64Bit()) {
21520 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
21521 TII->get(X86::MOV64rm), X86::RDI)
21523 .addImm(0).addReg(0)
21524 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
21525 MI->getOperand(3).getTargetFlags())
21527 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
21528 addDirectMem(MIB, X86::RDI);
21529 MIB.addReg(X86::RAX, RegState::ImplicitDefine).addRegMask(RegMask);
21530 } else if (F->getTarget().getRelocationModel() != Reloc::PIC_) {
21531 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
21532 TII->get(X86::MOV32rm), X86::EAX)
21534 .addImm(0).addReg(0)
21535 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
21536 MI->getOperand(3).getTargetFlags())
21538 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
21539 addDirectMem(MIB, X86::EAX);
21540 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
21542 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
21543 TII->get(X86::MOV32rm), X86::EAX)
21544 .addReg(TII->getGlobalBaseReg(F))
21545 .addImm(0).addReg(0)
21546 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
21547 MI->getOperand(3).getTargetFlags())
21549 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
21550 addDirectMem(MIB, X86::EAX);
21551 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
21554 MI->eraseFromParent(); // The pseudo instruction is gone now.
21558 MachineBasicBlock *
21559 X86TargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
21560 MachineBasicBlock *MBB) const {
21561 DebugLoc DL = MI->getDebugLoc();
21562 MachineFunction *MF = MBB->getParent();
21563 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21564 MachineRegisterInfo &MRI = MF->getRegInfo();
21566 const BasicBlock *BB = MBB->getBasicBlock();
21567 MachineFunction::iterator I = ++MBB->getIterator();
21569 // Memory Reference
21570 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
21571 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
21574 unsigned MemOpndSlot = 0;
21576 unsigned CurOp = 0;
21578 DstReg = MI->getOperand(CurOp++).getReg();
21579 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
21580 assert(RC->hasType(MVT::i32) && "Invalid destination!");
21581 unsigned mainDstReg = MRI.createVirtualRegister(RC);
21582 unsigned restoreDstReg = MRI.createVirtualRegister(RC);
21584 MemOpndSlot = CurOp;
21586 MVT PVT = getPointerTy(MF->getDataLayout());
21587 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
21588 "Invalid Pointer Size!");
21590 // For v = setjmp(buf), we generate
21593 // buf[LabelOffset] = restoreMBB <-- takes address of restoreMBB
21594 // SjLjSetup restoreMBB
21600 // v = phi(main, restore)
21603 // if base pointer being used, load it from frame
21606 MachineBasicBlock *thisMBB = MBB;
21607 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
21608 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
21609 MachineBasicBlock *restoreMBB = MF->CreateMachineBasicBlock(BB);
21610 MF->insert(I, mainMBB);
21611 MF->insert(I, sinkMBB);
21612 MF->push_back(restoreMBB);
21613 restoreMBB->setHasAddressTaken();
21615 MachineInstrBuilder MIB;
21617 // Transfer the remainder of BB and its successor edges to sinkMBB.
21618 sinkMBB->splice(sinkMBB->begin(), MBB,
21619 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
21620 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
21623 unsigned PtrStoreOpc = 0;
21624 unsigned LabelReg = 0;
21625 const int64_t LabelOffset = 1 * PVT.getStoreSize();
21626 Reloc::Model RM = MF->getTarget().getRelocationModel();
21627 bool UseImmLabel = (MF->getTarget().getCodeModel() == CodeModel::Small) &&
21628 (RM == Reloc::Static || RM == Reloc::DynamicNoPIC);
21630 // Prepare IP either in reg or imm.
21631 if (!UseImmLabel) {
21632 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr;
21633 const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
21634 LabelReg = MRI.createVirtualRegister(PtrRC);
21635 if (Subtarget->is64Bit()) {
21636 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA64r), LabelReg)
21640 .addMBB(restoreMBB)
21643 const X86InstrInfo *XII = static_cast<const X86InstrInfo*>(TII);
21644 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA32r), LabelReg)
21645 .addReg(XII->getGlobalBaseReg(MF))
21648 .addMBB(restoreMBB, Subtarget->ClassifyBlockAddressReference())
21652 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi;
21654 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PtrStoreOpc));
21655 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
21656 if (i == X86::AddrDisp)
21657 MIB.addDisp(MI->getOperand(MemOpndSlot + i), LabelOffset);
21659 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
21662 MIB.addReg(LabelReg);
21664 MIB.addMBB(restoreMBB);
21665 MIB.setMemRefs(MMOBegin, MMOEnd);
21667 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
21668 .addMBB(restoreMBB);
21670 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
21671 MIB.addRegMask(RegInfo->getNoPreservedMask());
21672 thisMBB->addSuccessor(mainMBB);
21673 thisMBB->addSuccessor(restoreMBB);
21677 BuildMI(mainMBB, DL, TII->get(X86::MOV32r0), mainDstReg);
21678 mainMBB->addSuccessor(sinkMBB);
21681 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
21682 TII->get(X86::PHI), DstReg)
21683 .addReg(mainDstReg).addMBB(mainMBB)
21684 .addReg(restoreDstReg).addMBB(restoreMBB);
21687 if (RegInfo->hasBasePointer(*MF)) {
21688 const bool Uses64BitFramePtr =
21689 Subtarget->isTarget64BitLP64() || Subtarget->isTargetNaCl64();
21690 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
21691 X86FI->setRestoreBasePointer(MF);
21692 unsigned FramePtr = RegInfo->getFrameRegister(*MF);
21693 unsigned BasePtr = RegInfo->getBaseRegister();
21694 unsigned Opm = Uses64BitFramePtr ? X86::MOV64rm : X86::MOV32rm;
21695 addRegOffset(BuildMI(restoreMBB, DL, TII->get(Opm), BasePtr),
21696 FramePtr, true, X86FI->getRestoreBasePointerOffset())
21697 .setMIFlag(MachineInstr::FrameSetup);
21699 BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1);
21700 BuildMI(restoreMBB, DL, TII->get(X86::JMP_1)).addMBB(sinkMBB);
21701 restoreMBB->addSuccessor(sinkMBB);
21703 MI->eraseFromParent();
21707 MachineBasicBlock *
21708 X86TargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
21709 MachineBasicBlock *MBB) const {
21710 DebugLoc DL = MI->getDebugLoc();
21711 MachineFunction *MF = MBB->getParent();
21712 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21713 MachineRegisterInfo &MRI = MF->getRegInfo();
21715 // Memory Reference
21716 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
21717 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
21719 MVT PVT = getPointerTy(MF->getDataLayout());
21720 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
21721 "Invalid Pointer Size!");
21723 const TargetRegisterClass *RC =
21724 (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass;
21725 unsigned Tmp = MRI.createVirtualRegister(RC);
21726 // Since FP is only updated here but NOT referenced, it's treated as GPR.
21727 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
21728 unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
21729 unsigned SP = RegInfo->getStackRegister();
21731 MachineInstrBuilder MIB;
21733 const int64_t LabelOffset = 1 * PVT.getStoreSize();
21734 const int64_t SPOffset = 2 * PVT.getStoreSize();
21736 unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm;
21737 unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r;
21740 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP);
21741 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
21742 MIB.addOperand(MI->getOperand(i));
21743 MIB.setMemRefs(MMOBegin, MMOEnd);
21745 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp);
21746 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
21747 if (i == X86::AddrDisp)
21748 MIB.addDisp(MI->getOperand(i), LabelOffset);
21750 MIB.addOperand(MI->getOperand(i));
21752 MIB.setMemRefs(MMOBegin, MMOEnd);
21754 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP);
21755 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
21756 if (i == X86::AddrDisp)
21757 MIB.addDisp(MI->getOperand(i), SPOffset);
21759 MIB.addOperand(MI->getOperand(i));
21761 MIB.setMemRefs(MMOBegin, MMOEnd);
21763 BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp);
21765 MI->eraseFromParent();
21769 // Replace 213-type (isel default) FMA3 instructions with 231-type for
21770 // accumulator loops. Writing back to the accumulator allows the coalescer
21771 // to remove extra copies in the loop.
21772 // FIXME: Do this on AVX512. We don't support 231 variants yet (PR23937).
21773 MachineBasicBlock *
21774 X86TargetLowering::emitFMA3Instr(MachineInstr *MI,
21775 MachineBasicBlock *MBB) const {
21776 MachineOperand &AddendOp = MI->getOperand(3);
21778 // Bail out early if the addend isn't a register - we can't switch these.
21779 if (!AddendOp.isReg())
21782 MachineFunction &MF = *MBB->getParent();
21783 MachineRegisterInfo &MRI = MF.getRegInfo();
21785 // Check whether the addend is defined by a PHI:
21786 assert(MRI.hasOneDef(AddendOp.getReg()) && "Multiple defs in SSA?");
21787 MachineInstr &AddendDef = *MRI.def_instr_begin(AddendOp.getReg());
21788 if (!AddendDef.isPHI())
21791 // Look for the following pattern:
21793 // %addend = phi [%entry, 0], [%loop, %result]
21795 // %result<tied1> = FMA213 %m2<tied0>, %m1, %addend
21799 // %addend = phi [%entry, 0], [%loop, %result]
21801 // %result<tied1> = FMA231 %addend<tied0>, %m1, %m2
21803 for (unsigned i = 1, e = AddendDef.getNumOperands(); i < e; i += 2) {
21804 assert(AddendDef.getOperand(i).isReg());
21805 MachineOperand PHISrcOp = AddendDef.getOperand(i);
21806 MachineInstr &PHISrcInst = *MRI.def_instr_begin(PHISrcOp.getReg());
21807 if (&PHISrcInst == MI) {
21808 // Found a matching instruction.
21809 unsigned NewFMAOpc = 0;
21810 switch (MI->getOpcode()) {
21811 case X86::VFMADDPDr213r: NewFMAOpc = X86::VFMADDPDr231r; break;
21812 case X86::VFMADDPSr213r: NewFMAOpc = X86::VFMADDPSr231r; break;
21813 case X86::VFMADDSDr213r: NewFMAOpc = X86::VFMADDSDr231r; break;
21814 case X86::VFMADDSSr213r: NewFMAOpc = X86::VFMADDSSr231r; break;
21815 case X86::VFMSUBPDr213r: NewFMAOpc = X86::VFMSUBPDr231r; break;
21816 case X86::VFMSUBPSr213r: NewFMAOpc = X86::VFMSUBPSr231r; break;
21817 case X86::VFMSUBSDr213r: NewFMAOpc = X86::VFMSUBSDr231r; break;
21818 case X86::VFMSUBSSr213r: NewFMAOpc = X86::VFMSUBSSr231r; break;
21819 case X86::VFNMADDPDr213r: NewFMAOpc = X86::VFNMADDPDr231r; break;
21820 case X86::VFNMADDPSr213r: NewFMAOpc = X86::VFNMADDPSr231r; break;
21821 case X86::VFNMADDSDr213r: NewFMAOpc = X86::VFNMADDSDr231r; break;
21822 case X86::VFNMADDSSr213r: NewFMAOpc = X86::VFNMADDSSr231r; break;
21823 case X86::VFNMSUBPDr213r: NewFMAOpc = X86::VFNMSUBPDr231r; break;
21824 case X86::VFNMSUBPSr213r: NewFMAOpc = X86::VFNMSUBPSr231r; break;
21825 case X86::VFNMSUBSDr213r: NewFMAOpc = X86::VFNMSUBSDr231r; break;
21826 case X86::VFNMSUBSSr213r: NewFMAOpc = X86::VFNMSUBSSr231r; break;
21827 case X86::VFMADDSUBPDr213r: NewFMAOpc = X86::VFMADDSUBPDr231r; break;
21828 case X86::VFMADDSUBPSr213r: NewFMAOpc = X86::VFMADDSUBPSr231r; break;
21829 case X86::VFMSUBADDPDr213r: NewFMAOpc = X86::VFMSUBADDPDr231r; break;
21830 case X86::VFMSUBADDPSr213r: NewFMAOpc = X86::VFMSUBADDPSr231r; break;
21832 case X86::VFMADDPDr213rY: NewFMAOpc = X86::VFMADDPDr231rY; break;
21833 case X86::VFMADDPSr213rY: NewFMAOpc = X86::VFMADDPSr231rY; break;
21834 case X86::VFMSUBPDr213rY: NewFMAOpc = X86::VFMSUBPDr231rY; break;
21835 case X86::VFMSUBPSr213rY: NewFMAOpc = X86::VFMSUBPSr231rY; break;
21836 case X86::VFNMADDPDr213rY: NewFMAOpc = X86::VFNMADDPDr231rY; break;
21837 case X86::VFNMADDPSr213rY: NewFMAOpc = X86::VFNMADDPSr231rY; break;
21838 case X86::VFNMSUBPDr213rY: NewFMAOpc = X86::VFNMSUBPDr231rY; break;
21839 case X86::VFNMSUBPSr213rY: NewFMAOpc = X86::VFNMSUBPSr231rY; break;
21840 case X86::VFMADDSUBPDr213rY: NewFMAOpc = X86::VFMADDSUBPDr231rY; break;
21841 case X86::VFMADDSUBPSr213rY: NewFMAOpc = X86::VFMADDSUBPSr231rY; break;
21842 case X86::VFMSUBADDPDr213rY: NewFMAOpc = X86::VFMSUBADDPDr231rY; break;
21843 case X86::VFMSUBADDPSr213rY: NewFMAOpc = X86::VFMSUBADDPSr231rY; break;
21844 default: llvm_unreachable("Unrecognized FMA variant.");
21847 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
21848 MachineInstrBuilder MIB =
21849 BuildMI(MF, MI->getDebugLoc(), TII.get(NewFMAOpc))
21850 .addOperand(MI->getOperand(0))
21851 .addOperand(MI->getOperand(3))
21852 .addOperand(MI->getOperand(2))
21853 .addOperand(MI->getOperand(1));
21854 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
21855 MI->eraseFromParent();
21862 MachineBasicBlock *
21863 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
21864 MachineBasicBlock *BB) const {
21865 switch (MI->getOpcode()) {
21866 default: llvm_unreachable("Unexpected instr type to insert");
21867 case X86::TAILJMPd64:
21868 case X86::TAILJMPr64:
21869 case X86::TAILJMPm64:
21870 case X86::TAILJMPd64_REX:
21871 case X86::TAILJMPr64_REX:
21872 case X86::TAILJMPm64_REX:
21873 llvm_unreachable("TAILJMP64 would not be touched here.");
21874 case X86::TCRETURNdi64:
21875 case X86::TCRETURNri64:
21876 case X86::TCRETURNmi64:
21878 case X86::WIN_ALLOCA:
21879 return EmitLoweredWinAlloca(MI, BB);
21880 case X86::CATCHRET:
21881 return EmitLoweredCatchRet(MI, BB);
21882 case X86::CATCHPAD:
21883 return EmitLoweredCatchPad(MI, BB);
21884 case X86::SEG_ALLOCA_32:
21885 case X86::SEG_ALLOCA_64:
21886 return EmitLoweredSegAlloca(MI, BB);
21887 case X86::TLSCall_32:
21888 case X86::TLSCall_64:
21889 return EmitLoweredTLSCall(MI, BB);
21890 case X86::CMOV_FR32:
21891 case X86::CMOV_FR64:
21892 case X86::CMOV_GR8:
21893 case X86::CMOV_GR16:
21894 case X86::CMOV_GR32:
21895 case X86::CMOV_RFP32:
21896 case X86::CMOV_RFP64:
21897 case X86::CMOV_RFP80:
21898 case X86::CMOV_V2F64:
21899 case X86::CMOV_V2I64:
21900 case X86::CMOV_V4F32:
21901 case X86::CMOV_V4F64:
21902 case X86::CMOV_V4I64:
21903 case X86::CMOV_V16F32:
21904 case X86::CMOV_V8F32:
21905 case X86::CMOV_V8F64:
21906 case X86::CMOV_V8I64:
21907 case X86::CMOV_V8I1:
21908 case X86::CMOV_V16I1:
21909 case X86::CMOV_V32I1:
21910 case X86::CMOV_V64I1:
21911 return EmitLoweredSelect(MI, BB);
21913 case X86::RELEASE_FADD32mr:
21914 case X86::RELEASE_FADD64mr:
21915 return EmitLoweredAtomicFP(MI, BB);
21917 case X86::FP32_TO_INT16_IN_MEM:
21918 case X86::FP32_TO_INT32_IN_MEM:
21919 case X86::FP32_TO_INT64_IN_MEM:
21920 case X86::FP64_TO_INT16_IN_MEM:
21921 case X86::FP64_TO_INT32_IN_MEM:
21922 case X86::FP64_TO_INT64_IN_MEM:
21923 case X86::FP80_TO_INT16_IN_MEM:
21924 case X86::FP80_TO_INT32_IN_MEM:
21925 case X86::FP80_TO_INT64_IN_MEM: {
21926 MachineFunction *F = BB->getParent();
21927 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21928 DebugLoc DL = MI->getDebugLoc();
21930 // Change the floating point control register to use "round towards zero"
21931 // mode when truncating to an integer value.
21932 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
21933 addFrameReference(BuildMI(*BB, MI, DL,
21934 TII->get(X86::FNSTCW16m)), CWFrameIdx);
21936 // Load the old value of the high byte of the control word...
21938 F->getRegInfo().createVirtualRegister(&X86::GR16RegClass);
21939 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
21942 // Set the high part to be round to zero...
21943 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
21946 // Reload the modified control word now...
21947 addFrameReference(BuildMI(*BB, MI, DL,
21948 TII->get(X86::FLDCW16m)), CWFrameIdx);
21950 // Restore the memory image of control word to original value
21951 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
21954 // Get the X86 opcode to use.
21956 switch (MI->getOpcode()) {
21957 default: llvm_unreachable("illegal opcode!");
21958 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
21959 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
21960 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
21961 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
21962 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
21963 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
21964 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
21965 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
21966 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
21970 MachineOperand &Op = MI->getOperand(0);
21972 AM.BaseType = X86AddressMode::RegBase;
21973 AM.Base.Reg = Op.getReg();
21975 AM.BaseType = X86AddressMode::FrameIndexBase;
21976 AM.Base.FrameIndex = Op.getIndex();
21978 Op = MI->getOperand(1);
21980 AM.Scale = Op.getImm();
21981 Op = MI->getOperand(2);
21983 AM.IndexReg = Op.getImm();
21984 Op = MI->getOperand(3);
21985 if (Op.isGlobal()) {
21986 AM.GV = Op.getGlobal();
21988 AM.Disp = Op.getImm();
21990 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
21991 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
21993 // Reload the original control word now.
21994 addFrameReference(BuildMI(*BB, MI, DL,
21995 TII->get(X86::FLDCW16m)), CWFrameIdx);
21997 MI->eraseFromParent(); // The pseudo instruction is gone now.
22000 // String/text processing lowering.
22001 case X86::PCMPISTRM128REG:
22002 case X86::VPCMPISTRM128REG:
22003 case X86::PCMPISTRM128MEM:
22004 case X86::VPCMPISTRM128MEM:
22005 case X86::PCMPESTRM128REG:
22006 case X86::VPCMPESTRM128REG:
22007 case X86::PCMPESTRM128MEM:
22008 case X86::VPCMPESTRM128MEM:
22009 assert(Subtarget->hasSSE42() &&
22010 "Target must have SSE4.2 or AVX features enabled");
22011 return EmitPCMPSTRM(MI, BB, Subtarget->getInstrInfo());
22013 // String/text processing lowering.
22014 case X86::PCMPISTRIREG:
22015 case X86::VPCMPISTRIREG:
22016 case X86::PCMPISTRIMEM:
22017 case X86::VPCMPISTRIMEM:
22018 case X86::PCMPESTRIREG:
22019 case X86::VPCMPESTRIREG:
22020 case X86::PCMPESTRIMEM:
22021 case X86::VPCMPESTRIMEM:
22022 assert(Subtarget->hasSSE42() &&
22023 "Target must have SSE4.2 or AVX features enabled");
22024 return EmitPCMPSTRI(MI, BB, Subtarget->getInstrInfo());
22026 // Thread synchronization.
22028 return EmitMonitor(MI, BB, Subtarget);
22032 return EmitXBegin(MI, BB, Subtarget->getInstrInfo());
22034 case X86::VASTART_SAVE_XMM_REGS:
22035 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
22037 case X86::VAARG_64:
22038 return EmitVAARG64WithCustomInserter(MI, BB);
22040 case X86::EH_SjLj_SetJmp32:
22041 case X86::EH_SjLj_SetJmp64:
22042 return emitEHSjLjSetJmp(MI, BB);
22044 case X86::EH_SjLj_LongJmp32:
22045 case X86::EH_SjLj_LongJmp64:
22046 return emitEHSjLjLongJmp(MI, BB);
22048 case TargetOpcode::STATEPOINT:
22049 // As an implementation detail, STATEPOINT shares the STACKMAP format at
22050 // this point in the process. We diverge later.
22051 return emitPatchPoint(MI, BB);
22053 case TargetOpcode::STACKMAP:
22054 case TargetOpcode::PATCHPOINT:
22055 return emitPatchPoint(MI, BB);
22057 case X86::VFMADDPDr213r:
22058 case X86::VFMADDPSr213r:
22059 case X86::VFMADDSDr213r:
22060 case X86::VFMADDSSr213r:
22061 case X86::VFMSUBPDr213r:
22062 case X86::VFMSUBPSr213r:
22063 case X86::VFMSUBSDr213r:
22064 case X86::VFMSUBSSr213r:
22065 case X86::VFNMADDPDr213r:
22066 case X86::VFNMADDPSr213r:
22067 case X86::VFNMADDSDr213r:
22068 case X86::VFNMADDSSr213r:
22069 case X86::VFNMSUBPDr213r:
22070 case X86::VFNMSUBPSr213r:
22071 case X86::VFNMSUBSDr213r:
22072 case X86::VFNMSUBSSr213r:
22073 case X86::VFMADDSUBPDr213r:
22074 case X86::VFMADDSUBPSr213r:
22075 case X86::VFMSUBADDPDr213r:
22076 case X86::VFMSUBADDPSr213r:
22077 case X86::VFMADDPDr213rY:
22078 case X86::VFMADDPSr213rY:
22079 case X86::VFMSUBPDr213rY:
22080 case X86::VFMSUBPSr213rY:
22081 case X86::VFNMADDPDr213rY:
22082 case X86::VFNMADDPSr213rY:
22083 case X86::VFNMSUBPDr213rY:
22084 case X86::VFNMSUBPSr213rY:
22085 case X86::VFMADDSUBPDr213rY:
22086 case X86::VFMADDSUBPSr213rY:
22087 case X86::VFMSUBADDPDr213rY:
22088 case X86::VFMSUBADDPSr213rY:
22089 return emitFMA3Instr(MI, BB);
22093 //===----------------------------------------------------------------------===//
22094 // X86 Optimization Hooks
22095 //===----------------------------------------------------------------------===//
22097 void X86TargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
22100 const SelectionDAG &DAG,
22101 unsigned Depth) const {
22102 unsigned BitWidth = KnownZero.getBitWidth();
22103 unsigned Opc = Op.getOpcode();
22104 assert((Opc >= ISD::BUILTIN_OP_END ||
22105 Opc == ISD::INTRINSIC_WO_CHAIN ||
22106 Opc == ISD::INTRINSIC_W_CHAIN ||
22107 Opc == ISD::INTRINSIC_VOID) &&
22108 "Should use MaskedValueIsZero if you don't know whether Op"
22109 " is a target node!");
22111 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
22125 // These nodes' second result is a boolean.
22126 if (Op.getResNo() == 0)
22129 case X86ISD::SETCC:
22130 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
22132 case ISD::INTRINSIC_WO_CHAIN: {
22133 unsigned IntId = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
22134 unsigned NumLoBits = 0;
22137 case Intrinsic::x86_sse_movmsk_ps:
22138 case Intrinsic::x86_avx_movmsk_ps_256:
22139 case Intrinsic::x86_sse2_movmsk_pd:
22140 case Intrinsic::x86_avx_movmsk_pd_256:
22141 case Intrinsic::x86_mmx_pmovmskb:
22142 case Intrinsic::x86_sse2_pmovmskb_128:
22143 case Intrinsic::x86_avx2_pmovmskb: {
22144 // High bits of movmskp{s|d}, pmovmskb are known zero.
22146 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
22147 case Intrinsic::x86_sse_movmsk_ps: NumLoBits = 4; break;
22148 case Intrinsic::x86_avx_movmsk_ps_256: NumLoBits = 8; break;
22149 case Intrinsic::x86_sse2_movmsk_pd: NumLoBits = 2; break;
22150 case Intrinsic::x86_avx_movmsk_pd_256: NumLoBits = 4; break;
22151 case Intrinsic::x86_mmx_pmovmskb: NumLoBits = 8; break;
22152 case Intrinsic::x86_sse2_pmovmskb_128: NumLoBits = 16; break;
22153 case Intrinsic::x86_avx2_pmovmskb: NumLoBits = 32; break;
22155 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - NumLoBits);
22164 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(
22166 const SelectionDAG &,
22167 unsigned Depth) const {
22168 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
22169 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
22170 return Op.getValueType().getScalarSizeInBits();
22176 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
22177 /// node is a GlobalAddress + offset.
22178 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
22179 const GlobalValue* &GA,
22180 int64_t &Offset) const {
22181 if (N->getOpcode() == X86ISD::Wrapper) {
22182 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
22183 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
22184 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
22188 return TargetLowering::isGAPlusOffset(N, GA, Offset);
22191 /// isShuffleHigh128VectorInsertLow - Checks whether the shuffle node is the
22192 /// same as extracting the high 128-bit part of 256-bit vector and then
22193 /// inserting the result into the low part of a new 256-bit vector
22194 static bool isShuffleHigh128VectorInsertLow(ShuffleVectorSDNode *SVOp) {
22195 EVT VT = SVOp->getValueType(0);
22196 unsigned NumElems = VT.getVectorNumElements();
22198 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
22199 for (unsigned i = 0, j = NumElems/2; i != NumElems/2; ++i, ++j)
22200 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
22201 SVOp->getMaskElt(j) >= 0)
22207 /// isShuffleLow128VectorInsertHigh - Checks whether the shuffle node is the
22208 /// same as extracting the low 128-bit part of 256-bit vector and then
22209 /// inserting the result into the high part of a new 256-bit vector
22210 static bool isShuffleLow128VectorInsertHigh(ShuffleVectorSDNode *SVOp) {
22211 EVT VT = SVOp->getValueType(0);
22212 unsigned NumElems = VT.getVectorNumElements();
22214 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
22215 for (unsigned i = NumElems/2, j = 0; i != NumElems; ++i, ++j)
22216 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
22217 SVOp->getMaskElt(j) >= 0)
22223 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
22224 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
22225 TargetLowering::DAGCombinerInfo &DCI,
22226 const X86Subtarget* Subtarget) {
22228 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
22229 SDValue V1 = SVOp->getOperand(0);
22230 SDValue V2 = SVOp->getOperand(1);
22231 EVT VT = SVOp->getValueType(0);
22232 unsigned NumElems = VT.getVectorNumElements();
22234 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
22235 V2.getOpcode() == ISD::CONCAT_VECTORS) {
22239 // V UNDEF BUILD_VECTOR UNDEF
22241 // CONCAT_VECTOR CONCAT_VECTOR
22244 // RESULT: V + zero extended
22246 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
22247 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
22248 V1.getOperand(1).getOpcode() != ISD::UNDEF)
22251 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
22254 // To match the shuffle mask, the first half of the mask should
22255 // be exactly the first vector, and all the rest a splat with the
22256 // first element of the second one.
22257 for (unsigned i = 0; i != NumElems/2; ++i)
22258 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
22259 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
22262 // If V1 is coming from a vector load then just fold to a VZEXT_LOAD.
22263 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(V1.getOperand(0))) {
22264 if (Ld->hasNUsesOfValue(1, 0)) {
22265 SDVTList Tys = DAG.getVTList(MVT::v4i64, MVT::Other);
22266 SDValue Ops[] = { Ld->getChain(), Ld->getBasePtr() };
22268 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops,
22270 Ld->getPointerInfo(),
22271 Ld->getAlignment(),
22272 false/*isVolatile*/, true/*ReadMem*/,
22273 false/*WriteMem*/);
22275 // Make sure the newly-created LOAD is in the same position as Ld in
22276 // terms of dependency. We create a TokenFactor for Ld and ResNode,
22277 // and update uses of Ld's output chain to use the TokenFactor.
22278 if (Ld->hasAnyUseOfValue(1)) {
22279 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
22280 SDValue(Ld, 1), SDValue(ResNode.getNode(), 1));
22281 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), NewChain);
22282 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(Ld, 1),
22283 SDValue(ResNode.getNode(), 1));
22286 return DAG.getBitcast(VT, ResNode);
22290 // Emit a zeroed vector and insert the desired subvector on its
22292 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
22293 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0), 0, DAG, dl);
22294 return DCI.CombineTo(N, InsV);
22297 //===--------------------------------------------------------------------===//
22298 // Combine some shuffles into subvector extracts and inserts:
22301 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
22302 if (isShuffleHigh128VectorInsertLow(SVOp)) {
22303 SDValue V = Extract128BitVector(V1, NumElems/2, DAG, dl);
22304 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, 0, DAG, dl);
22305 return DCI.CombineTo(N, InsV);
22308 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
22309 if (isShuffleLow128VectorInsertHigh(SVOp)) {
22310 SDValue V = Extract128BitVector(V1, 0, DAG, dl);
22311 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, NumElems/2, DAG, dl);
22312 return DCI.CombineTo(N, InsV);
22318 /// \brief Combine an arbitrary chain of shuffles into a single instruction if
22321 /// This is the leaf of the recursive combinine below. When we have found some
22322 /// chain of single-use x86 shuffle instructions and accumulated the combined
22323 /// shuffle mask represented by them, this will try to pattern match that mask
22324 /// into either a single instruction if there is a special purpose instruction
22325 /// for this operation, or into a PSHUFB instruction which is a fully general
22326 /// instruction but should only be used to replace chains over a certain depth.
22327 static bool combineX86ShuffleChain(SDValue Op, SDValue Root, ArrayRef<int> Mask,
22328 int Depth, bool HasPSHUFB, SelectionDAG &DAG,
22329 TargetLowering::DAGCombinerInfo &DCI,
22330 const X86Subtarget *Subtarget) {
22331 assert(!Mask.empty() && "Cannot combine an empty shuffle mask!");
22333 // Find the operand that enters the chain. Note that multiple uses are OK
22334 // here, we're not going to remove the operand we find.
22335 SDValue Input = Op.getOperand(0);
22336 while (Input.getOpcode() == ISD::BITCAST)
22337 Input = Input.getOperand(0);
22339 MVT VT = Input.getSimpleValueType();
22340 MVT RootVT = Root.getSimpleValueType();
22343 if (Mask.size() == 1) {
22344 int Index = Mask[0];
22345 assert((Index >= 0 || Index == SM_SentinelUndef ||
22346 Index == SM_SentinelZero) &&
22347 "Invalid shuffle index found!");
22349 // We may end up with an accumulated mask of size 1 as a result of
22350 // widening of shuffle operands (see function canWidenShuffleElements).
22351 // If the only shuffle index is equal to SM_SentinelZero then propagate
22352 // a zero vector. Otherwise, the combine shuffle mask is a no-op shuffle
22353 // mask, and therefore the entire chain of shuffles can be folded away.
22354 if (Index == SM_SentinelZero)
22355 DCI.CombineTo(Root.getNode(), getZeroVector(RootVT, Subtarget, DAG, DL));
22357 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Input),
22362 // Use the float domain if the operand type is a floating point type.
22363 bool FloatDomain = VT.isFloatingPoint();
22365 // For floating point shuffles, we don't have free copies in the shuffle
22366 // instructions or the ability to load as part of the instruction, so
22367 // canonicalize their shuffles to UNPCK or MOV variants.
22369 // Note that even with AVX we prefer the PSHUFD form of shuffle for integer
22370 // vectors because it can have a load folded into it that UNPCK cannot. This
22371 // doesn't preclude something switching to the shorter encoding post-RA.
22373 // FIXME: Should teach these routines about AVX vector widths.
22374 if (FloatDomain && VT.is128BitVector()) {
22375 if (Mask.equals({0, 0}) || Mask.equals({1, 1})) {
22376 bool Lo = Mask.equals({0, 0});
22379 // Check if we have SSE3 which will let us use MOVDDUP. That instruction
22380 // is no slower than UNPCKLPD but has the option to fold the input operand
22381 // into even an unaligned memory load.
22382 if (Lo && Subtarget->hasSSE3()) {
22383 Shuffle = X86ISD::MOVDDUP;
22384 ShuffleVT = MVT::v2f64;
22386 // We have MOVLHPS and MOVHLPS throughout SSE and they encode smaller
22387 // than the UNPCK variants.
22388 Shuffle = Lo ? X86ISD::MOVLHPS : X86ISD::MOVHLPS;
22389 ShuffleVT = MVT::v4f32;
22391 if (Depth == 1 && Root->getOpcode() == Shuffle)
22392 return false; // Nothing to do!
22393 Op = DAG.getBitcast(ShuffleVT, Input);
22394 DCI.AddToWorklist(Op.getNode());
22395 if (Shuffle == X86ISD::MOVDDUP)
22396 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
22398 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
22399 DCI.AddToWorklist(Op.getNode());
22400 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22404 if (Subtarget->hasSSE3() &&
22405 (Mask.equals({0, 0, 2, 2}) || Mask.equals({1, 1, 3, 3}))) {
22406 bool Lo = Mask.equals({0, 0, 2, 2});
22407 unsigned Shuffle = Lo ? X86ISD::MOVSLDUP : X86ISD::MOVSHDUP;
22408 MVT ShuffleVT = MVT::v4f32;
22409 if (Depth == 1 && Root->getOpcode() == Shuffle)
22410 return false; // Nothing to do!
22411 Op = DAG.getBitcast(ShuffleVT, Input);
22412 DCI.AddToWorklist(Op.getNode());
22413 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
22414 DCI.AddToWorklist(Op.getNode());
22415 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22419 if (Mask.equals({0, 0, 1, 1}) || Mask.equals({2, 2, 3, 3})) {
22420 bool Lo = Mask.equals({0, 0, 1, 1});
22421 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
22422 MVT ShuffleVT = MVT::v4f32;
22423 if (Depth == 1 && Root->getOpcode() == Shuffle)
22424 return false; // Nothing to do!
22425 Op = DAG.getBitcast(ShuffleVT, Input);
22426 DCI.AddToWorklist(Op.getNode());
22427 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
22428 DCI.AddToWorklist(Op.getNode());
22429 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22435 // We always canonicalize the 8 x i16 and 16 x i8 shuffles into their UNPCK
22436 // variants as none of these have single-instruction variants that are
22437 // superior to the UNPCK formulation.
22438 if (!FloatDomain && VT.is128BitVector() &&
22439 (Mask.equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
22440 Mask.equals({4, 4, 5, 5, 6, 6, 7, 7}) ||
22441 Mask.equals({0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7}) ||
22443 {8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14, 14, 15, 15}))) {
22444 bool Lo = Mask[0] == 0;
22445 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
22446 if (Depth == 1 && Root->getOpcode() == Shuffle)
22447 return false; // Nothing to do!
22449 switch (Mask.size()) {
22451 ShuffleVT = MVT::v8i16;
22454 ShuffleVT = MVT::v16i8;
22457 llvm_unreachable("Impossible mask size!");
22459 Op = DAG.getBitcast(ShuffleVT, Input);
22460 DCI.AddToWorklist(Op.getNode());
22461 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
22462 DCI.AddToWorklist(Op.getNode());
22463 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22468 // Don't try to re-form single instruction chains under any circumstances now
22469 // that we've done encoding canonicalization for them.
22473 // If we have 3 or more shuffle instructions or a chain involving PSHUFB, we
22474 // can replace them with a single PSHUFB instruction profitably. Intel's
22475 // manuals suggest only using PSHUFB if doing so replacing 5 instructions, but
22476 // in practice PSHUFB tends to be *very* fast so we're more aggressive.
22477 if ((Depth >= 3 || HasPSHUFB) && Subtarget->hasSSSE3()) {
22478 SmallVector<SDValue, 16> PSHUFBMask;
22479 int NumBytes = VT.getSizeInBits() / 8;
22480 int Ratio = NumBytes / Mask.size();
22481 for (int i = 0; i < NumBytes; ++i) {
22482 if (Mask[i / Ratio] == SM_SentinelUndef) {
22483 PSHUFBMask.push_back(DAG.getUNDEF(MVT::i8));
22486 int M = Mask[i / Ratio] != SM_SentinelZero
22487 ? Ratio * Mask[i / Ratio] + i % Ratio
22489 PSHUFBMask.push_back(DAG.getConstant(M, DL, MVT::i8));
22491 MVT ByteVT = MVT::getVectorVT(MVT::i8, NumBytes);
22492 Op = DAG.getBitcast(ByteVT, Input);
22493 DCI.AddToWorklist(Op.getNode());
22494 SDValue PSHUFBMaskOp =
22495 DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVT, PSHUFBMask);
22496 DCI.AddToWorklist(PSHUFBMaskOp.getNode());
22497 Op = DAG.getNode(X86ISD::PSHUFB, DL, ByteVT, Op, PSHUFBMaskOp);
22498 DCI.AddToWorklist(Op.getNode());
22499 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22504 // Failed to find any combines.
22508 /// \brief Fully generic combining of x86 shuffle instructions.
22510 /// This should be the last combine run over the x86 shuffle instructions. Once
22511 /// they have been fully optimized, this will recursively consider all chains
22512 /// of single-use shuffle instructions, build a generic model of the cumulative
22513 /// shuffle operation, and check for simpler instructions which implement this
22514 /// operation. We use this primarily for two purposes:
22516 /// 1) Collapse generic shuffles to specialized single instructions when
22517 /// equivalent. In most cases, this is just an encoding size win, but
22518 /// sometimes we will collapse multiple generic shuffles into a single
22519 /// special-purpose shuffle.
22520 /// 2) Look for sequences of shuffle instructions with 3 or more total
22521 /// instructions, and replace them with the slightly more expensive SSSE3
22522 /// PSHUFB instruction if available. We do this as the last combining step
22523 /// to ensure we avoid using PSHUFB if we can implement the shuffle with
22524 /// a suitable short sequence of other instructions. The PHUFB will either
22525 /// use a register or have to read from memory and so is slightly (but only
22526 /// slightly) more expensive than the other shuffle instructions.
22528 /// Because this is inherently a quadratic operation (for each shuffle in
22529 /// a chain, we recurse up the chain), the depth is limited to 8 instructions.
22530 /// This should never be an issue in practice as the shuffle lowering doesn't
22531 /// produce sequences of more than 8 instructions.
22533 /// FIXME: We will currently miss some cases where the redundant shuffling
22534 /// would simplify under the threshold for PSHUFB formation because of
22535 /// combine-ordering. To fix this, we should do the redundant instruction
22536 /// combining in this recursive walk.
22537 static bool combineX86ShufflesRecursively(SDValue Op, SDValue Root,
22538 ArrayRef<int> RootMask,
22539 int Depth, bool HasPSHUFB,
22541 TargetLowering::DAGCombinerInfo &DCI,
22542 const X86Subtarget *Subtarget) {
22543 // Bound the depth of our recursive combine because this is ultimately
22544 // quadratic in nature.
22548 // Directly rip through bitcasts to find the underlying operand.
22549 while (Op.getOpcode() == ISD::BITCAST && Op.getOperand(0).hasOneUse())
22550 Op = Op.getOperand(0);
22552 MVT VT = Op.getSimpleValueType();
22553 if (!VT.isVector())
22554 return false; // Bail if we hit a non-vector.
22556 assert(Root.getSimpleValueType().isVector() &&
22557 "Shuffles operate on vector types!");
22558 assert(VT.getSizeInBits() == Root.getSimpleValueType().getSizeInBits() &&
22559 "Can only combine shuffles of the same vector register size.");
22561 if (!isTargetShuffle(Op.getOpcode()))
22563 SmallVector<int, 16> OpMask;
22565 bool HaveMask = getTargetShuffleMask(Op.getNode(), VT, OpMask, IsUnary);
22566 // We only can combine unary shuffles which we can decode the mask for.
22567 if (!HaveMask || !IsUnary)
22570 assert(VT.getVectorNumElements() == OpMask.size() &&
22571 "Different mask size from vector size!");
22572 assert(((RootMask.size() > OpMask.size() &&
22573 RootMask.size() % OpMask.size() == 0) ||
22574 (OpMask.size() > RootMask.size() &&
22575 OpMask.size() % RootMask.size() == 0) ||
22576 OpMask.size() == RootMask.size()) &&
22577 "The smaller number of elements must divide the larger.");
22578 int RootRatio = std::max<int>(1, OpMask.size() / RootMask.size());
22579 int OpRatio = std::max<int>(1, RootMask.size() / OpMask.size());
22580 assert(((RootRatio == 1 && OpRatio == 1) ||
22581 (RootRatio == 1) != (OpRatio == 1)) &&
22582 "Must not have a ratio for both incoming and op masks!");
22584 SmallVector<int, 16> Mask;
22585 Mask.reserve(std::max(OpMask.size(), RootMask.size()));
22587 // Merge this shuffle operation's mask into our accumulated mask. Note that
22588 // this shuffle's mask will be the first applied to the input, followed by the
22589 // root mask to get us all the way to the root value arrangement. The reason
22590 // for this order is that we are recursing up the operation chain.
22591 for (int i = 0, e = std::max(OpMask.size(), RootMask.size()); i < e; ++i) {
22592 int RootIdx = i / RootRatio;
22593 if (RootMask[RootIdx] < 0) {
22594 // This is a zero or undef lane, we're done.
22595 Mask.push_back(RootMask[RootIdx]);
22599 int RootMaskedIdx = RootMask[RootIdx] * RootRatio + i % RootRatio;
22600 int OpIdx = RootMaskedIdx / OpRatio;
22601 if (OpMask[OpIdx] < 0) {
22602 // The incoming lanes are zero or undef, it doesn't matter which ones we
22604 Mask.push_back(OpMask[OpIdx]);
22608 // Ok, we have non-zero lanes, map them through.
22609 Mask.push_back(OpMask[OpIdx] * OpRatio +
22610 RootMaskedIdx % OpRatio);
22613 // See if we can recurse into the operand to combine more things.
22614 switch (Op.getOpcode()) {
22615 case X86ISD::PSHUFB:
22617 case X86ISD::PSHUFD:
22618 case X86ISD::PSHUFHW:
22619 case X86ISD::PSHUFLW:
22620 if (Op.getOperand(0).hasOneUse() &&
22621 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
22622 HasPSHUFB, DAG, DCI, Subtarget))
22626 case X86ISD::UNPCKL:
22627 case X86ISD::UNPCKH:
22628 assert(Op.getOperand(0) == Op.getOperand(1) &&
22629 "We only combine unary shuffles!");
22630 // We can't check for single use, we have to check that this shuffle is the
22632 if (Op->isOnlyUserOf(Op.getOperand(0).getNode()) &&
22633 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
22634 HasPSHUFB, DAG, DCI, Subtarget))
22639 // Minor canonicalization of the accumulated shuffle mask to make it easier
22640 // to match below. All this does is detect masks with squential pairs of
22641 // elements, and shrink them to the half-width mask. It does this in a loop
22642 // so it will reduce the size of the mask to the minimal width mask which
22643 // performs an equivalent shuffle.
22644 SmallVector<int, 16> WidenedMask;
22645 while (Mask.size() > 1 && canWidenShuffleElements(Mask, WidenedMask)) {
22646 Mask = std::move(WidenedMask);
22647 WidenedMask.clear();
22650 return combineX86ShuffleChain(Op, Root, Mask, Depth, HasPSHUFB, DAG, DCI,
22654 /// \brief Get the PSHUF-style mask from PSHUF node.
22656 /// This is a very minor wrapper around getTargetShuffleMask to easy forming v4
22657 /// PSHUF-style masks that can be reused with such instructions.
22658 static SmallVector<int, 4> getPSHUFShuffleMask(SDValue N) {
22659 MVT VT = N.getSimpleValueType();
22660 SmallVector<int, 4> Mask;
22662 bool HaveMask = getTargetShuffleMask(N.getNode(), VT, Mask, IsUnary);
22666 // If we have more than 128-bits, only the low 128-bits of shuffle mask
22667 // matter. Check that the upper masks are repeats and remove them.
22668 if (VT.getSizeInBits() > 128) {
22669 int LaneElts = 128 / VT.getScalarSizeInBits();
22671 for (int i = 1, NumLanes = VT.getSizeInBits() / 128; i < NumLanes; ++i)
22672 for (int j = 0; j < LaneElts; ++j)
22673 assert(Mask[j] == Mask[i * LaneElts + j] - (LaneElts * i) &&
22674 "Mask doesn't repeat in high 128-bit lanes!");
22676 Mask.resize(LaneElts);
22679 switch (N.getOpcode()) {
22680 case X86ISD::PSHUFD:
22682 case X86ISD::PSHUFLW:
22685 case X86ISD::PSHUFHW:
22686 Mask.erase(Mask.begin(), Mask.begin() + 4);
22687 for (int &M : Mask)
22691 llvm_unreachable("No valid shuffle instruction found!");
22695 /// \brief Search for a combinable shuffle across a chain ending in pshufd.
22697 /// We walk up the chain and look for a combinable shuffle, skipping over
22698 /// shuffles that we could hoist this shuffle's transformation past without
22699 /// altering anything.
22701 combineRedundantDWordShuffle(SDValue N, MutableArrayRef<int> Mask,
22703 TargetLowering::DAGCombinerInfo &DCI) {
22704 assert(N.getOpcode() == X86ISD::PSHUFD &&
22705 "Called with something other than an x86 128-bit half shuffle!");
22708 // Walk up a single-use chain looking for a combinable shuffle. Keep a stack
22709 // of the shuffles in the chain so that we can form a fresh chain to replace
22711 SmallVector<SDValue, 8> Chain;
22712 SDValue V = N.getOperand(0);
22713 for (; V.hasOneUse(); V = V.getOperand(0)) {
22714 switch (V.getOpcode()) {
22716 return SDValue(); // Nothing combined!
22719 // Skip bitcasts as we always know the type for the target specific
22723 case X86ISD::PSHUFD:
22724 // Found another dword shuffle.
22727 case X86ISD::PSHUFLW:
22728 // Check that the low words (being shuffled) are the identity in the
22729 // dword shuffle, and the high words are self-contained.
22730 if (Mask[0] != 0 || Mask[1] != 1 ||
22731 !(Mask[2] >= 2 && Mask[2] < 4 && Mask[3] >= 2 && Mask[3] < 4))
22734 Chain.push_back(V);
22737 case X86ISD::PSHUFHW:
22738 // Check that the high words (being shuffled) are the identity in the
22739 // dword shuffle, and the low words are self-contained.
22740 if (Mask[2] != 2 || Mask[3] != 3 ||
22741 !(Mask[0] >= 0 && Mask[0] < 2 && Mask[1] >= 0 && Mask[1] < 2))
22744 Chain.push_back(V);
22747 case X86ISD::UNPCKL:
22748 case X86ISD::UNPCKH:
22749 // For either i8 -> i16 or i16 -> i32 unpacks, we can combine a dword
22750 // shuffle into a preceding word shuffle.
22751 if (V.getSimpleValueType().getVectorElementType() != MVT::i8 &&
22752 V.getSimpleValueType().getVectorElementType() != MVT::i16)
22755 // Search for a half-shuffle which we can combine with.
22756 unsigned CombineOp =
22757 V.getOpcode() == X86ISD::UNPCKL ? X86ISD::PSHUFLW : X86ISD::PSHUFHW;
22758 if (V.getOperand(0) != V.getOperand(1) ||
22759 !V->isOnlyUserOf(V.getOperand(0).getNode()))
22761 Chain.push_back(V);
22762 V = V.getOperand(0);
22764 switch (V.getOpcode()) {
22766 return SDValue(); // Nothing to combine.
22768 case X86ISD::PSHUFLW:
22769 case X86ISD::PSHUFHW:
22770 if (V.getOpcode() == CombineOp)
22773 Chain.push_back(V);
22777 V = V.getOperand(0);
22781 } while (V.hasOneUse());
22784 // Break out of the loop if we break out of the switch.
22788 if (!V.hasOneUse())
22789 // We fell out of the loop without finding a viable combining instruction.
22792 // Merge this node's mask and our incoming mask.
22793 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22794 for (int &M : Mask)
22796 V = DAG.getNode(V.getOpcode(), DL, V.getValueType(), V.getOperand(0),
22797 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
22799 // Rebuild the chain around this new shuffle.
22800 while (!Chain.empty()) {
22801 SDValue W = Chain.pop_back_val();
22803 if (V.getValueType() != W.getOperand(0).getValueType())
22804 V = DAG.getBitcast(W.getOperand(0).getValueType(), V);
22806 switch (W.getOpcode()) {
22808 llvm_unreachable("Only PSHUF and UNPCK instructions get here!");
22810 case X86ISD::UNPCKL:
22811 case X86ISD::UNPCKH:
22812 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, V);
22815 case X86ISD::PSHUFD:
22816 case X86ISD::PSHUFLW:
22817 case X86ISD::PSHUFHW:
22818 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, W.getOperand(1));
22822 if (V.getValueType() != N.getValueType())
22823 V = DAG.getBitcast(N.getValueType(), V);
22825 // Return the new chain to replace N.
22829 /// \brief Search for a combinable shuffle across a chain ending in pshuflw or
22832 /// We walk up the chain, skipping shuffles of the other half and looking
22833 /// through shuffles which switch halves trying to find a shuffle of the same
22834 /// pair of dwords.
22835 static bool combineRedundantHalfShuffle(SDValue N, MutableArrayRef<int> Mask,
22837 TargetLowering::DAGCombinerInfo &DCI) {
22839 (N.getOpcode() == X86ISD::PSHUFLW || N.getOpcode() == X86ISD::PSHUFHW) &&
22840 "Called with something other than an x86 128-bit half shuffle!");
22842 unsigned CombineOpcode = N.getOpcode();
22844 // Walk up a single-use chain looking for a combinable shuffle.
22845 SDValue V = N.getOperand(0);
22846 for (; V.hasOneUse(); V = V.getOperand(0)) {
22847 switch (V.getOpcode()) {
22849 return false; // Nothing combined!
22852 // Skip bitcasts as we always know the type for the target specific
22856 case X86ISD::PSHUFLW:
22857 case X86ISD::PSHUFHW:
22858 if (V.getOpcode() == CombineOpcode)
22861 // Other-half shuffles are no-ops.
22864 // Break out of the loop if we break out of the switch.
22868 if (!V.hasOneUse())
22869 // We fell out of the loop without finding a viable combining instruction.
22872 // Combine away the bottom node as its shuffle will be accumulated into
22873 // a preceding shuffle.
22874 DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
22876 // Record the old value.
22879 // Merge this node's mask and our incoming mask (adjusted to account for all
22880 // the pshufd instructions encountered).
22881 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22882 for (int &M : Mask)
22884 V = DAG.getNode(V.getOpcode(), DL, MVT::v8i16, V.getOperand(0),
22885 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
22887 // Check that the shuffles didn't cancel each other out. If not, we need to
22888 // combine to the new one.
22890 // Replace the combinable shuffle with the combined one, updating all users
22891 // so that we re-evaluate the chain here.
22892 DCI.CombineTo(Old.getNode(), V, /*AddTo*/ true);
22897 /// \brief Try to combine x86 target specific shuffles.
22898 static SDValue PerformTargetShuffleCombine(SDValue N, SelectionDAG &DAG,
22899 TargetLowering::DAGCombinerInfo &DCI,
22900 const X86Subtarget *Subtarget) {
22902 MVT VT = N.getSimpleValueType();
22903 SmallVector<int, 4> Mask;
22905 switch (N.getOpcode()) {
22906 case X86ISD::PSHUFD:
22907 case X86ISD::PSHUFLW:
22908 case X86ISD::PSHUFHW:
22909 Mask = getPSHUFShuffleMask(N);
22910 assert(Mask.size() == 4);
22912 case X86ISD::UNPCKL: {
22913 // Combine X86ISD::UNPCKL and ISD::VECTOR_SHUFFLE into X86ISD::UNPCKH, in
22914 // which X86ISD::UNPCKL has a ISD::UNDEF operand, and ISD::VECTOR_SHUFFLE
22915 // moves upper half elements into the lower half part. For example:
22917 // t2: v16i8 = vector_shuffle<8,9,10,11,12,13,14,15,u,u,u,u,u,u,u,u> t1,
22919 // t3: v16i8 = X86ISD::UNPCKL undef:v16i8, t2
22921 // will be combined to:
22923 // t3: v16i8 = X86ISD::UNPCKH undef:v16i8, t1
22925 // This is only for 128-bit vectors. From SSE4.1 onward this combine may not
22926 // happen due to advanced instructions.
22927 if (!VT.is128BitVector())
22930 auto Op0 = N.getOperand(0);
22931 auto Op1 = N.getOperand(1);
22932 if (Op0.getOpcode() == ISD::UNDEF &&
22933 Op1.getNode()->getOpcode() == ISD::VECTOR_SHUFFLE) {
22934 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Op1.getNode())->getMask();
22936 unsigned NumElts = VT.getVectorNumElements();
22937 SmallVector<int, 8> ExpectedMask(NumElts, -1);
22938 std::iota(ExpectedMask.begin(), ExpectedMask.begin() + NumElts / 2,
22941 auto ShufOp = Op1.getOperand(0);
22942 if (isShuffleEquivalent(Op1, ShufOp, Mask, ExpectedMask))
22943 return DAG.getNode(X86ISD::UNPCKH, DL, VT, N.getOperand(0), ShufOp);
22951 // Nuke no-op shuffles that show up after combining.
22952 if (isNoopShuffleMask(Mask))
22953 return DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
22955 // Look for simplifications involving one or two shuffle instructions.
22956 SDValue V = N.getOperand(0);
22957 switch (N.getOpcode()) {
22960 case X86ISD::PSHUFLW:
22961 case X86ISD::PSHUFHW:
22962 assert(VT.getVectorElementType() == MVT::i16 && "Bad word shuffle type!");
22964 if (combineRedundantHalfShuffle(N, Mask, DAG, DCI))
22965 return SDValue(); // We combined away this shuffle, so we're done.
22967 // See if this reduces to a PSHUFD which is no more expensive and can
22968 // combine with more operations. Note that it has to at least flip the
22969 // dwords as otherwise it would have been removed as a no-op.
22970 if (makeArrayRef(Mask).equals({2, 3, 0, 1})) {
22971 int DMask[] = {0, 1, 2, 3};
22972 int DOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 2;
22973 DMask[DOffset + 0] = DOffset + 1;
22974 DMask[DOffset + 1] = DOffset + 0;
22975 MVT DVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
22976 V = DAG.getBitcast(DVT, V);
22977 DCI.AddToWorklist(V.getNode());
22978 V = DAG.getNode(X86ISD::PSHUFD, DL, DVT, V,
22979 getV4X86ShuffleImm8ForMask(DMask, DL, DAG));
22980 DCI.AddToWorklist(V.getNode());
22981 return DAG.getBitcast(VT, V);
22984 // Look for shuffle patterns which can be implemented as a single unpack.
22985 // FIXME: This doesn't handle the location of the PSHUFD generically, and
22986 // only works when we have a PSHUFD followed by two half-shuffles.
22987 if (Mask[0] == Mask[1] && Mask[2] == Mask[3] &&
22988 (V.getOpcode() == X86ISD::PSHUFLW ||
22989 V.getOpcode() == X86ISD::PSHUFHW) &&
22990 V.getOpcode() != N.getOpcode() &&
22992 SDValue D = V.getOperand(0);
22993 while (D.getOpcode() == ISD::BITCAST && D.hasOneUse())
22994 D = D.getOperand(0);
22995 if (D.getOpcode() == X86ISD::PSHUFD && D.hasOneUse()) {
22996 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22997 SmallVector<int, 4> DMask = getPSHUFShuffleMask(D);
22998 int NOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
22999 int VOffset = V.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
23001 for (int i = 0; i < 4; ++i) {
23002 WordMask[i + NOffset] = Mask[i] + NOffset;
23003 WordMask[i + VOffset] = VMask[i] + VOffset;
23005 // Map the word mask through the DWord mask.
23007 for (int i = 0; i < 8; ++i)
23008 MappedMask[i] = 2 * DMask[WordMask[i] / 2] + WordMask[i] % 2;
23009 if (makeArrayRef(MappedMask).equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
23010 makeArrayRef(MappedMask).equals({4, 4, 5, 5, 6, 6, 7, 7})) {
23011 // We can replace all three shuffles with an unpack.
23012 V = DAG.getBitcast(VT, D.getOperand(0));
23013 DCI.AddToWorklist(V.getNode());
23014 return DAG.getNode(MappedMask[0] == 0 ? X86ISD::UNPCKL
23023 case X86ISD::PSHUFD:
23024 if (SDValue NewN = combineRedundantDWordShuffle(N, Mask, DAG, DCI))
23033 /// \brief Try to combine a shuffle into a target-specific add-sub node.
23035 /// We combine this directly on the abstract vector shuffle nodes so it is
23036 /// easier to generically match. We also insert dummy vector shuffle nodes for
23037 /// the operands which explicitly discard the lanes which are unused by this
23038 /// operation to try to flow through the rest of the combiner the fact that
23039 /// they're unused.
23040 static SDValue combineShuffleToAddSub(SDNode *N, SelectionDAG &DAG) {
23042 EVT VT = N->getValueType(0);
23044 // We only handle target-independent shuffles.
23045 // FIXME: It would be easy and harmless to use the target shuffle mask
23046 // extraction tool to support more.
23047 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
23050 auto *SVN = cast<ShuffleVectorSDNode>(N);
23051 ArrayRef<int> Mask = SVN->getMask();
23052 SDValue V1 = N->getOperand(0);
23053 SDValue V2 = N->getOperand(1);
23055 // We require the first shuffle operand to be the SUB node, and the second to
23056 // be the ADD node.
23057 // FIXME: We should support the commuted patterns.
23058 if (V1->getOpcode() != ISD::FSUB || V2->getOpcode() != ISD::FADD)
23061 // If there are other uses of these operations we can't fold them.
23062 if (!V1->hasOneUse() || !V2->hasOneUse())
23065 // Ensure that both operations have the same operands. Note that we can
23066 // commute the FADD operands.
23067 SDValue LHS = V1->getOperand(0), RHS = V1->getOperand(1);
23068 if ((V2->getOperand(0) != LHS || V2->getOperand(1) != RHS) &&
23069 (V2->getOperand(0) != RHS || V2->getOperand(1) != LHS))
23072 // We're looking for blends between FADD and FSUB nodes. We insist on these
23073 // nodes being lined up in a specific expected pattern.
23074 if (!(isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
23075 isShuffleEquivalent(V1, V2, Mask, {0, 5, 2, 7}) ||
23076 isShuffleEquivalent(V1, V2, Mask, {0, 9, 2, 11, 4, 13, 6, 15})))
23079 // Only specific types are legal at this point, assert so we notice if and
23080 // when these change.
23081 assert((VT == MVT::v4f32 || VT == MVT::v2f64 || VT == MVT::v8f32 ||
23082 VT == MVT::v4f64) &&
23083 "Unknown vector type encountered!");
23085 return DAG.getNode(X86ISD::ADDSUB, DL, VT, LHS, RHS);
23088 /// PerformShuffleCombine - Performs several different shuffle combines.
23089 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
23090 TargetLowering::DAGCombinerInfo &DCI,
23091 const X86Subtarget *Subtarget) {
23093 SDValue N0 = N->getOperand(0);
23094 SDValue N1 = N->getOperand(1);
23095 EVT VT = N->getValueType(0);
23097 // Don't create instructions with illegal types after legalize types has run.
23098 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23099 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
23102 // If we have legalized the vector types, look for blends of FADD and FSUB
23103 // nodes that we can fuse into an ADDSUB node.
23104 if (TLI.isTypeLegal(VT) && Subtarget->hasSSE3())
23105 if (SDValue AddSub = combineShuffleToAddSub(N, DAG))
23108 // Combine 256-bit vector shuffles. This is only profitable when in AVX mode
23109 if (Subtarget->hasFp256() && VT.is256BitVector() &&
23110 N->getOpcode() == ISD::VECTOR_SHUFFLE)
23111 return PerformShuffleCombine256(N, DAG, DCI, Subtarget);
23113 // During Type Legalization, when promoting illegal vector types,
23114 // the backend might introduce new shuffle dag nodes and bitcasts.
23116 // This code performs the following transformation:
23117 // fold: (shuffle (bitcast (BINOP A, B)), Undef, <Mask>) ->
23118 // (shuffle (BINOP (bitcast A), (bitcast B)), Undef, <Mask>)
23120 // We do this only if both the bitcast and the BINOP dag nodes have
23121 // one use. Also, perform this transformation only if the new binary
23122 // operation is legal. This is to avoid introducing dag nodes that
23123 // potentially need to be further expanded (or custom lowered) into a
23124 // less optimal sequence of dag nodes.
23125 if (!DCI.isBeforeLegalize() && DCI.isBeforeLegalizeOps() &&
23126 N1.getOpcode() == ISD::UNDEF && N0.hasOneUse() &&
23127 N0.getOpcode() == ISD::BITCAST) {
23128 SDValue BC0 = N0.getOperand(0);
23129 EVT SVT = BC0.getValueType();
23130 unsigned Opcode = BC0.getOpcode();
23131 unsigned NumElts = VT.getVectorNumElements();
23133 if (BC0.hasOneUse() && SVT.isVector() &&
23134 SVT.getVectorNumElements() * 2 == NumElts &&
23135 TLI.isOperationLegal(Opcode, VT)) {
23136 bool CanFold = false;
23148 unsigned SVTNumElts = SVT.getVectorNumElements();
23149 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
23150 for (unsigned i = 0, e = SVTNumElts; i != e && CanFold; ++i)
23151 CanFold = SVOp->getMaskElt(i) == (int)(i * 2);
23152 for (unsigned i = SVTNumElts, e = NumElts; i != e && CanFold; ++i)
23153 CanFold = SVOp->getMaskElt(i) < 0;
23156 SDValue BC00 = DAG.getBitcast(VT, BC0.getOperand(0));
23157 SDValue BC01 = DAG.getBitcast(VT, BC0.getOperand(1));
23158 SDValue NewBinOp = DAG.getNode(BC0.getOpcode(), dl, VT, BC00, BC01);
23159 return DAG.getVectorShuffle(VT, dl, NewBinOp, N1, &SVOp->getMask()[0]);
23164 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
23165 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
23166 // consecutive, non-overlapping, and in the right order.
23167 SmallVector<SDValue, 16> Elts;
23168 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
23169 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
23171 if (SDValue LD = EltsFromConsecutiveLoads(VT, Elts, dl, DAG, true))
23174 if (isTargetShuffle(N->getOpcode())) {
23176 PerformTargetShuffleCombine(SDValue(N, 0), DAG, DCI, Subtarget);
23177 if (Shuffle.getNode())
23180 // Try recursively combining arbitrary sequences of x86 shuffle
23181 // instructions into higher-order shuffles. We do this after combining
23182 // specific PSHUF instruction sequences into their minimal form so that we
23183 // can evaluate how many specialized shuffle instructions are involved in
23184 // a particular chain.
23185 SmallVector<int, 1> NonceMask; // Just a placeholder.
23186 NonceMask.push_back(0);
23187 if (combineX86ShufflesRecursively(SDValue(N, 0), SDValue(N, 0), NonceMask,
23188 /*Depth*/ 1, /*HasPSHUFB*/ false, DAG,
23190 return SDValue(); // This routine will use CombineTo to replace N.
23196 /// XFormVExtractWithShuffleIntoLoad - Check if a vector extract from a target
23197 /// specific shuffle of a load can be folded into a single element load.
23198 /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
23199 /// shuffles have been custom lowered so we need to handle those here.
23200 static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
23201 TargetLowering::DAGCombinerInfo &DCI) {
23202 if (DCI.isBeforeLegalizeOps())
23205 SDValue InVec = N->getOperand(0);
23206 SDValue EltNo = N->getOperand(1);
23208 if (!isa<ConstantSDNode>(EltNo))
23211 EVT OriginalVT = InVec.getValueType();
23213 if (InVec.getOpcode() == ISD::BITCAST) {
23214 // Don't duplicate a load with other uses.
23215 if (!InVec.hasOneUse())
23217 EVT BCVT = InVec.getOperand(0).getValueType();
23218 if (!BCVT.isVector() ||
23219 BCVT.getVectorNumElements() != OriginalVT.getVectorNumElements())
23221 InVec = InVec.getOperand(0);
23224 EVT CurrentVT = InVec.getValueType();
23226 if (!isTargetShuffle(InVec.getOpcode()))
23229 // Don't duplicate a load with other uses.
23230 if (!InVec.hasOneUse())
23233 SmallVector<int, 16> ShuffleMask;
23235 if (!getTargetShuffleMask(InVec.getNode(), CurrentVT.getSimpleVT(),
23236 ShuffleMask, UnaryShuffle))
23239 // Select the input vector, guarding against out of range extract vector.
23240 unsigned NumElems = CurrentVT.getVectorNumElements();
23241 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
23242 int Idx = (Elt > (int)NumElems) ? -1 : ShuffleMask[Elt];
23243 SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0)
23244 : InVec.getOperand(1);
23246 // If inputs to shuffle are the same for both ops, then allow 2 uses
23247 unsigned AllowedUses = InVec.getNumOperands() > 1 &&
23248 InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
23250 if (LdNode.getOpcode() == ISD::BITCAST) {
23251 // Don't duplicate a load with other uses.
23252 if (!LdNode.getNode()->hasNUsesOfValue(AllowedUses, 0))
23255 AllowedUses = 1; // only allow 1 load use if we have a bitcast
23256 LdNode = LdNode.getOperand(0);
23259 if (!ISD::isNormalLoad(LdNode.getNode()))
23262 LoadSDNode *LN0 = cast<LoadSDNode>(LdNode);
23264 if (!LN0 ||!LN0->hasNUsesOfValue(AllowedUses, 0) || LN0->isVolatile())
23267 EVT EltVT = N->getValueType(0);
23268 // If there's a bitcast before the shuffle, check if the load type and
23269 // alignment is valid.
23270 unsigned Align = LN0->getAlignment();
23271 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23272 unsigned NewAlign = DAG.getDataLayout().getABITypeAlignment(
23273 EltVT.getTypeForEVT(*DAG.getContext()));
23275 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, EltVT))
23278 // All checks match so transform back to vector_shuffle so that DAG combiner
23279 // can finish the job
23282 // Create shuffle node taking into account the case that its a unary shuffle
23283 SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(CurrentVT)
23284 : InVec.getOperand(1);
23285 Shuffle = DAG.getVectorShuffle(CurrentVT, dl,
23286 InVec.getOperand(0), Shuffle,
23288 Shuffle = DAG.getBitcast(OriginalVT, Shuffle);
23289 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle,
23293 static SDValue PerformBITCASTCombine(SDNode *N, SelectionDAG &DAG,
23294 const X86Subtarget *Subtarget) {
23295 SDValue N0 = N->getOperand(0);
23296 EVT VT = N->getValueType(0);
23298 // Detect bitcasts between i32 to x86mmx low word. Since MMX types are
23299 // special and don't usually play with other vector types, it's better to
23300 // handle them early to be sure we emit efficient code by avoiding
23301 // store-load conversions.
23302 if (VT == MVT::x86mmx && N0.getOpcode() == ISD::BUILD_VECTOR &&
23303 N0.getValueType() == MVT::v2i32 &&
23304 isa<ConstantSDNode>(N0.getOperand(1))) {
23305 SDValue N00 = N0->getOperand(0);
23306 if (N0.getConstantOperandVal(1) == 0 && N00.getValueType() == MVT::i32)
23307 return DAG.getNode(X86ISD::MMX_MOVW2D, SDLoc(N00), VT, N00);
23310 // Convert a bitcasted integer logic operation that has one bitcasted
23311 // floating-point operand and one constant operand into a floating-point
23312 // logic operation. This may create a load of the constant, but that is
23313 // cheaper than materializing the constant in an integer register and
23314 // transferring it to an SSE register or transferring the SSE operand to
23315 // integer register and back.
23317 switch (N0.getOpcode()) {
23318 case ISD::AND: FPOpcode = X86ISD::FAND; break;
23319 case ISD::OR: FPOpcode = X86ISD::FOR; break;
23320 case ISD::XOR: FPOpcode = X86ISD::FXOR; break;
23321 default: return SDValue();
23323 if (((Subtarget->hasSSE1() && VT == MVT::f32) ||
23324 (Subtarget->hasSSE2() && VT == MVT::f64)) &&
23325 isa<ConstantSDNode>(N0.getOperand(1)) &&
23326 N0.getOperand(0).getOpcode() == ISD::BITCAST &&
23327 N0.getOperand(0).getOperand(0).getValueType() == VT) {
23328 SDValue N000 = N0.getOperand(0).getOperand(0);
23329 SDValue FPConst = DAG.getBitcast(VT, N0.getOperand(1));
23330 return DAG.getNode(FPOpcode, SDLoc(N0), VT, N000, FPConst);
23336 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
23337 /// generation and convert it from being a bunch of shuffles and extracts
23338 /// into a somewhat faster sequence. For i686, the best sequence is apparently
23339 /// storing the value and loading scalars back, while for x64 we should
23340 /// use 64-bit extracts and shifts.
23341 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
23342 TargetLowering::DAGCombinerInfo &DCI) {
23343 if (SDValue NewOp = XFormVExtractWithShuffleIntoLoad(N, DAG, DCI))
23346 SDValue InputVector = N->getOperand(0);
23347 SDLoc dl(InputVector);
23348 // Detect mmx to i32 conversion through a v2i32 elt extract.
23349 if (InputVector.getOpcode() == ISD::BITCAST && InputVector.hasOneUse() &&
23350 N->getValueType(0) == MVT::i32 &&
23351 InputVector.getValueType() == MVT::v2i32) {
23353 // The bitcast source is a direct mmx result.
23354 SDValue MMXSrc = InputVector.getNode()->getOperand(0);
23355 if (MMXSrc.getValueType() == MVT::x86mmx)
23356 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
23357 N->getValueType(0),
23358 InputVector.getNode()->getOperand(0));
23360 // The mmx is indirect: (i64 extract_elt (v1i64 bitcast (x86mmx ...))).
23361 if (MMXSrc.getOpcode() == ISD::EXTRACT_VECTOR_ELT && MMXSrc.hasOneUse() &&
23362 MMXSrc.getValueType() == MVT::i64) {
23363 SDValue MMXSrcOp = MMXSrc.getOperand(0);
23364 if (MMXSrcOp.hasOneUse() && MMXSrcOp.getOpcode() == ISD::BITCAST &&
23365 MMXSrcOp.getValueType() == MVT::v1i64 &&
23366 MMXSrcOp.getOperand(0).getValueType() == MVT::x86mmx)
23367 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
23368 N->getValueType(0), MMXSrcOp.getOperand(0));
23372 EVT VT = N->getValueType(0);
23374 if (VT == MVT::i1 && isa<ConstantSDNode>(N->getOperand(1)) &&
23375 InputVector.getOpcode() == ISD::BITCAST &&
23376 isa<ConstantSDNode>(InputVector.getOperand(0))) {
23377 uint64_t ExtractedElt =
23378 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
23379 uint64_t InputValue =
23380 cast<ConstantSDNode>(InputVector.getOperand(0))->getZExtValue();
23381 uint64_t Res = (InputValue >> ExtractedElt) & 1;
23382 return DAG.getConstant(Res, dl, MVT::i1);
23384 // Only operate on vectors of 4 elements, where the alternative shuffling
23385 // gets to be more expensive.
23386 if (InputVector.getValueType() != MVT::v4i32)
23389 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
23390 // single use which is a sign-extend or zero-extend, and all elements are
23392 SmallVector<SDNode *, 4> Uses;
23393 unsigned ExtractedElements = 0;
23394 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
23395 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
23396 if (UI.getUse().getResNo() != InputVector.getResNo())
23399 SDNode *Extract = *UI;
23400 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
23403 if (Extract->getValueType(0) != MVT::i32)
23405 if (!Extract->hasOneUse())
23407 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
23408 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
23410 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
23413 // Record which element was extracted.
23414 ExtractedElements |=
23415 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
23417 Uses.push_back(Extract);
23420 // If not all the elements were used, this may not be worthwhile.
23421 if (ExtractedElements != 15)
23424 // Ok, we've now decided to do the transformation.
23425 // If 64-bit shifts are legal, use the extract-shift sequence,
23426 // otherwise bounce the vector off the cache.
23427 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23430 if (TLI.isOperationLegal(ISD::SRA, MVT::i64)) {
23431 SDValue Cst = DAG.getBitcast(MVT::v2i64, InputVector);
23432 auto &DL = DAG.getDataLayout();
23433 EVT VecIdxTy = DAG.getTargetLoweringInfo().getVectorIdxTy(DL);
23434 SDValue BottomHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
23435 DAG.getConstant(0, dl, VecIdxTy));
23436 SDValue TopHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
23437 DAG.getConstant(1, dl, VecIdxTy));
23439 SDValue ShAmt = DAG.getConstant(
23440 32, dl, DAG.getTargetLoweringInfo().getShiftAmountTy(MVT::i64, DL));
23441 Vals[0] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BottomHalf);
23442 Vals[1] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
23443 DAG.getNode(ISD::SRA, dl, MVT::i64, BottomHalf, ShAmt));
23444 Vals[2] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, TopHalf);
23445 Vals[3] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
23446 DAG.getNode(ISD::SRA, dl, MVT::i64, TopHalf, ShAmt));
23448 // Store the value to a temporary stack slot.
23449 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
23450 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
23451 MachinePointerInfo(), false, false, 0);
23453 EVT ElementType = InputVector.getValueType().getVectorElementType();
23454 unsigned EltSize = ElementType.getSizeInBits() / 8;
23456 // Replace each use (extract) with a load of the appropriate element.
23457 for (unsigned i = 0; i < 4; ++i) {
23458 uint64_t Offset = EltSize * i;
23459 auto PtrVT = TLI.getPointerTy(DAG.getDataLayout());
23460 SDValue OffsetVal = DAG.getConstant(Offset, dl, PtrVT);
23462 SDValue ScalarAddr =
23463 DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, OffsetVal);
23465 // Load the scalar.
23466 Vals[i] = DAG.getLoad(ElementType, dl, Ch,
23467 ScalarAddr, MachinePointerInfo(),
23468 false, false, false, 0);
23473 // Replace the extracts
23474 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
23475 UE = Uses.end(); UI != UE; ++UI) {
23476 SDNode *Extract = *UI;
23478 SDValue Idx = Extract->getOperand(1);
23479 uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
23480 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), Vals[IdxVal]);
23483 // The replacement was made in place; don't return anything.
23488 transformVSELECTtoBlendVECTOR_SHUFFLE(SDNode *N, SelectionDAG &DAG,
23489 const X86Subtarget *Subtarget) {
23491 SDValue Cond = N->getOperand(0);
23492 SDValue LHS = N->getOperand(1);
23493 SDValue RHS = N->getOperand(2);
23495 if (Cond.getOpcode() == ISD::SIGN_EXTEND) {
23496 SDValue CondSrc = Cond->getOperand(0);
23497 if (CondSrc->getOpcode() == ISD::SIGN_EXTEND_INREG)
23498 Cond = CondSrc->getOperand(0);
23501 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
23504 // A vselect where all conditions and data are constants can be optimized into
23505 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
23506 if (ISD::isBuildVectorOfConstantSDNodes(LHS.getNode()) &&
23507 ISD::isBuildVectorOfConstantSDNodes(RHS.getNode()))
23510 unsigned MaskValue = 0;
23511 if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue))
23514 MVT VT = N->getSimpleValueType(0);
23515 unsigned NumElems = VT.getVectorNumElements();
23516 SmallVector<int, 8> ShuffleMask(NumElems, -1);
23517 for (unsigned i = 0; i < NumElems; ++i) {
23518 // Be sure we emit undef where we can.
23519 if (Cond.getOperand(i)->getOpcode() == ISD::UNDEF)
23520 ShuffleMask[i] = -1;
23522 ShuffleMask[i] = i + NumElems * ((MaskValue >> i) & 1);
23525 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23526 if (!TLI.isShuffleMaskLegal(ShuffleMask, VT))
23528 return DAG.getVectorShuffle(VT, dl, LHS, RHS, &ShuffleMask[0]);
23531 /// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT
23533 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
23534 TargetLowering::DAGCombinerInfo &DCI,
23535 const X86Subtarget *Subtarget) {
23537 SDValue Cond = N->getOperand(0);
23538 // Get the LHS/RHS of the select.
23539 SDValue LHS = N->getOperand(1);
23540 SDValue RHS = N->getOperand(2);
23541 EVT VT = LHS.getValueType();
23542 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23544 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
23545 // instructions match the semantics of the common C idiom x<y?x:y but not
23546 // x<=y?x:y, because of how they handle negative zero (which can be
23547 // ignored in unsafe-math mode).
23548 // We also try to create v2f32 min/max nodes, which we later widen to v4f32.
23549 if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() &&
23550 VT != MVT::f80 && (TLI.isTypeLegal(VT) || VT == MVT::v2f32) &&
23551 (Subtarget->hasSSE2() ||
23552 (Subtarget->hasSSE1() && VT.getScalarType() == MVT::f32))) {
23553 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23555 unsigned Opcode = 0;
23556 // Check for x CC y ? x : y.
23557 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
23558 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
23562 // Converting this to a min would handle NaNs incorrectly, and swapping
23563 // the operands would cause it to handle comparisons between positive
23564 // and negative zero incorrectly.
23565 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
23566 if (!DAG.getTarget().Options.UnsafeFPMath &&
23567 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
23569 std::swap(LHS, RHS);
23571 Opcode = X86ISD::FMIN;
23574 // Converting this to a min would handle comparisons between positive
23575 // and negative zero incorrectly.
23576 if (!DAG.getTarget().Options.UnsafeFPMath &&
23577 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
23579 Opcode = X86ISD::FMIN;
23582 // Converting this to a min would handle both negative zeros and NaNs
23583 // incorrectly, but we can swap the operands to fix both.
23584 std::swap(LHS, RHS);
23588 Opcode = X86ISD::FMIN;
23592 // Converting this to a max would handle comparisons between positive
23593 // and negative zero incorrectly.
23594 if (!DAG.getTarget().Options.UnsafeFPMath &&
23595 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
23597 Opcode = X86ISD::FMAX;
23600 // Converting this to a max would handle NaNs incorrectly, and swapping
23601 // the operands would cause it to handle comparisons between positive
23602 // and negative zero incorrectly.
23603 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
23604 if (!DAG.getTarget().Options.UnsafeFPMath &&
23605 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
23607 std::swap(LHS, RHS);
23609 Opcode = X86ISD::FMAX;
23612 // Converting this to a max would handle both negative zeros and NaNs
23613 // incorrectly, but we can swap the operands to fix both.
23614 std::swap(LHS, RHS);
23618 Opcode = X86ISD::FMAX;
23621 // Check for x CC y ? y : x -- a min/max with reversed arms.
23622 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
23623 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
23627 // Converting this to a min would handle comparisons between positive
23628 // and negative zero incorrectly, and swapping the operands would
23629 // cause it to handle NaNs incorrectly.
23630 if (!DAG.getTarget().Options.UnsafeFPMath &&
23631 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
23632 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
23634 std::swap(LHS, RHS);
23636 Opcode = X86ISD::FMIN;
23639 // Converting this to a min would handle NaNs incorrectly.
23640 if (!DAG.getTarget().Options.UnsafeFPMath &&
23641 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
23643 Opcode = X86ISD::FMIN;
23646 // Converting this to a min would handle both negative zeros and NaNs
23647 // incorrectly, but we can swap the operands to fix both.
23648 std::swap(LHS, RHS);
23652 Opcode = X86ISD::FMIN;
23656 // Converting this to a max would handle NaNs incorrectly.
23657 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
23659 Opcode = X86ISD::FMAX;
23662 // Converting this to a max would handle comparisons between positive
23663 // and negative zero incorrectly, and swapping the operands would
23664 // cause it to handle NaNs incorrectly.
23665 if (!DAG.getTarget().Options.UnsafeFPMath &&
23666 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
23667 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
23669 std::swap(LHS, RHS);
23671 Opcode = X86ISD::FMAX;
23674 // Converting this to a max would handle both negative zeros and NaNs
23675 // incorrectly, but we can swap the operands to fix both.
23676 std::swap(LHS, RHS);
23680 Opcode = X86ISD::FMAX;
23686 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
23689 EVT CondVT = Cond.getValueType();
23690 if (Subtarget->hasAVX512() && VT.isVector() && CondVT.isVector() &&
23691 CondVT.getVectorElementType() == MVT::i1) {
23692 // v16i8 (select v16i1, v16i8, v16i8) does not have a proper
23693 // lowering on KNL. In this case we convert it to
23694 // v16i8 (select v16i8, v16i8, v16i8) and use AVX instruction.
23695 // The same situation for all 128 and 256-bit vectors of i8 and i16.
23696 // Since SKX these selects have a proper lowering.
23697 EVT OpVT = LHS.getValueType();
23698 if ((OpVT.is128BitVector() || OpVT.is256BitVector()) &&
23699 (OpVT.getVectorElementType() == MVT::i8 ||
23700 OpVT.getVectorElementType() == MVT::i16) &&
23701 !(Subtarget->hasBWI() && Subtarget->hasVLX())) {
23702 Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, OpVT, Cond);
23703 DCI.AddToWorklist(Cond.getNode());
23704 return DAG.getNode(N->getOpcode(), DL, OpVT, Cond, LHS, RHS);
23707 // If this is a select between two integer constants, try to do some
23709 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
23710 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
23711 // Don't do this for crazy integer types.
23712 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
23713 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
23714 // so that TrueC (the true value) is larger than FalseC.
23715 bool NeedsCondInvert = false;
23717 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
23718 // Efficiently invertible.
23719 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
23720 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
23721 isa<ConstantSDNode>(Cond.getOperand(1))))) {
23722 NeedsCondInvert = true;
23723 std::swap(TrueC, FalseC);
23726 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
23727 if (FalseC->getAPIntValue() == 0 &&
23728 TrueC->getAPIntValue().isPowerOf2()) {
23729 if (NeedsCondInvert) // Invert the condition if needed.
23730 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
23731 DAG.getConstant(1, DL, Cond.getValueType()));
23733 // Zero extend the condition if needed.
23734 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
23736 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
23737 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
23738 DAG.getConstant(ShAmt, DL, MVT::i8));
23741 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
23742 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
23743 if (NeedsCondInvert) // Invert the condition if needed.
23744 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
23745 DAG.getConstant(1, DL, Cond.getValueType()));
23747 // Zero extend the condition if needed.
23748 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
23749 FalseC->getValueType(0), Cond);
23750 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23751 SDValue(FalseC, 0));
23754 // Optimize cases that will turn into an LEA instruction. This requires
23755 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
23756 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
23757 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
23758 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
23760 bool isFastMultiplier = false;
23762 switch ((unsigned char)Diff) {
23764 case 1: // result = add base, cond
23765 case 2: // result = lea base( , cond*2)
23766 case 3: // result = lea base(cond, cond*2)
23767 case 4: // result = lea base( , cond*4)
23768 case 5: // result = lea base(cond, cond*4)
23769 case 8: // result = lea base( , cond*8)
23770 case 9: // result = lea base(cond, cond*8)
23771 isFastMultiplier = true;
23776 if (isFastMultiplier) {
23777 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
23778 if (NeedsCondInvert) // Invert the condition if needed.
23779 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
23780 DAG.getConstant(1, DL, Cond.getValueType()));
23782 // Zero extend the condition if needed.
23783 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
23785 // Scale the condition by the difference.
23787 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
23788 DAG.getConstant(Diff, DL,
23789 Cond.getValueType()));
23791 // Add the base if non-zero.
23792 if (FalseC->getAPIntValue() != 0)
23793 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23794 SDValue(FalseC, 0));
23801 // Canonicalize max and min:
23802 // (x > y) ? x : y -> (x >= y) ? x : y
23803 // (x < y) ? x : y -> (x <= y) ? x : y
23804 // This allows use of COND_S / COND_NS (see TranslateX86CC) which eliminates
23805 // the need for an extra compare
23806 // against zero. e.g.
23807 // (x - y) > 0 : (x - y) ? 0 -> (x - y) >= 0 : (x - y) ? 0
23809 // testl %edi, %edi
23811 // cmovgl %edi, %eax
23815 // cmovsl %eax, %edi
23816 if (N->getOpcode() == ISD::SELECT && Cond.getOpcode() == ISD::SETCC &&
23817 DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
23818 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
23819 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23824 ISD::CondCode NewCC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGE;
23825 Cond = DAG.getSetCC(SDLoc(Cond), Cond.getValueType(),
23826 Cond.getOperand(0), Cond.getOperand(1), NewCC);
23827 return DAG.getNode(ISD::SELECT, DL, VT, Cond, LHS, RHS);
23832 // Early exit check
23833 if (!TLI.isTypeLegal(VT))
23836 // Match VSELECTs into subs with unsigned saturation.
23837 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
23838 // psubus is available in SSE2 and AVX2 for i8 and i16 vectors.
23839 ((Subtarget->hasSSE2() && (VT == MVT::v16i8 || VT == MVT::v8i16)) ||
23840 (Subtarget->hasAVX2() && (VT == MVT::v32i8 || VT == MVT::v16i16)))) {
23841 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23843 // Check if one of the arms of the VSELECT is a zero vector. If it's on the
23844 // left side invert the predicate to simplify logic below.
23846 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
23848 CC = ISD::getSetCCInverse(CC, true);
23849 } else if (ISD::isBuildVectorAllZeros(RHS.getNode())) {
23853 if (Other.getNode() && Other->getNumOperands() == 2 &&
23854 DAG.isEqualTo(Other->getOperand(0), Cond.getOperand(0))) {
23855 SDValue OpLHS = Other->getOperand(0), OpRHS = Other->getOperand(1);
23856 SDValue CondRHS = Cond->getOperand(1);
23858 // Look for a general sub with unsigned saturation first.
23859 // x >= y ? x-y : 0 --> subus x, y
23860 // x > y ? x-y : 0 --> subus x, y
23861 if ((CC == ISD::SETUGE || CC == ISD::SETUGT) &&
23862 Other->getOpcode() == ISD::SUB && DAG.isEqualTo(OpRHS, CondRHS))
23863 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
23865 if (auto *OpRHSBV = dyn_cast<BuildVectorSDNode>(OpRHS))
23866 if (auto *OpRHSConst = OpRHSBV->getConstantSplatNode()) {
23867 if (auto *CondRHSBV = dyn_cast<BuildVectorSDNode>(CondRHS))
23868 if (auto *CondRHSConst = CondRHSBV->getConstantSplatNode())
23869 // If the RHS is a constant we have to reverse the const
23870 // canonicalization.
23871 // x > C-1 ? x+-C : 0 --> subus x, C
23872 if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD &&
23873 CondRHSConst->getAPIntValue() ==
23874 (-OpRHSConst->getAPIntValue() - 1))
23875 return DAG.getNode(
23876 X86ISD::SUBUS, DL, VT, OpLHS,
23877 DAG.getConstant(-OpRHSConst->getAPIntValue(), DL, VT));
23879 // Another special case: If C was a sign bit, the sub has been
23880 // canonicalized into a xor.
23881 // FIXME: Would it be better to use computeKnownBits to determine
23882 // whether it's safe to decanonicalize the xor?
23883 // x s< 0 ? x^C : 0 --> subus x, C
23884 if (CC == ISD::SETLT && Other->getOpcode() == ISD::XOR &&
23885 ISD::isBuildVectorAllZeros(CondRHS.getNode()) &&
23886 OpRHSConst->getAPIntValue().isSignBit())
23887 // Note that we have to rebuild the RHS constant here to ensure we
23888 // don't rely on particular values of undef lanes.
23889 return DAG.getNode(
23890 X86ISD::SUBUS, DL, VT, OpLHS,
23891 DAG.getConstant(OpRHSConst->getAPIntValue(), DL, VT));
23896 // Simplify vector selection if condition value type matches vselect
23898 if (N->getOpcode() == ISD::VSELECT && CondVT == VT) {
23899 assert(Cond.getValueType().isVector() &&
23900 "vector select expects a vector selector!");
23902 bool TValIsAllOnes = ISD::isBuildVectorAllOnes(LHS.getNode());
23903 bool FValIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
23905 // Try invert the condition if true value is not all 1s and false value
23907 if (!TValIsAllOnes && !FValIsAllZeros &&
23908 // Check if the selector will be produced by CMPP*/PCMP*
23909 Cond.getOpcode() == ISD::SETCC &&
23910 // Check if SETCC has already been promoted
23911 TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT) ==
23913 bool TValIsAllZeros = ISD::isBuildVectorAllZeros(LHS.getNode());
23914 bool FValIsAllOnes = ISD::isBuildVectorAllOnes(RHS.getNode());
23916 if (TValIsAllZeros || FValIsAllOnes) {
23917 SDValue CC = Cond.getOperand(2);
23918 ISD::CondCode NewCC =
23919 ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
23920 Cond.getOperand(0).getValueType().isInteger());
23921 Cond = DAG.getSetCC(DL, CondVT, Cond.getOperand(0), Cond.getOperand(1), NewCC);
23922 std::swap(LHS, RHS);
23923 TValIsAllOnes = FValIsAllOnes;
23924 FValIsAllZeros = TValIsAllZeros;
23928 if (TValIsAllOnes || FValIsAllZeros) {
23931 if (TValIsAllOnes && FValIsAllZeros)
23933 else if (TValIsAllOnes)
23935 DAG.getNode(ISD::OR, DL, CondVT, Cond, DAG.getBitcast(CondVT, RHS));
23936 else if (FValIsAllZeros)
23937 Ret = DAG.getNode(ISD::AND, DL, CondVT, Cond,
23938 DAG.getBitcast(CondVT, LHS));
23940 return DAG.getBitcast(VT, Ret);
23944 // We should generate an X86ISD::BLENDI from a vselect if its argument
23945 // is a sign_extend_inreg of an any_extend of a BUILD_VECTOR of
23946 // constants. This specific pattern gets generated when we split a
23947 // selector for a 512 bit vector in a machine without AVX512 (but with
23948 // 256-bit vectors), during legalization:
23950 // (vselect (sign_extend (any_extend (BUILD_VECTOR)) i1) LHS RHS)
23952 // Iff we find this pattern and the build_vectors are built from
23953 // constants, we translate the vselect into a shuffle_vector that we
23954 // know will be matched by LowerVECTOR_SHUFFLEtoBlend.
23955 if ((N->getOpcode() == ISD::VSELECT ||
23956 N->getOpcode() == X86ISD::SHRUNKBLEND) &&
23957 !DCI.isBeforeLegalize() && !VT.is512BitVector()) {
23958 SDValue Shuffle = transformVSELECTtoBlendVECTOR_SHUFFLE(N, DAG, Subtarget);
23959 if (Shuffle.getNode())
23963 // If this is a *dynamic* select (non-constant condition) and we can match
23964 // this node with one of the variable blend instructions, restructure the
23965 // condition so that the blends can use the high bit of each element and use
23966 // SimplifyDemandedBits to simplify the condition operand.
23967 if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() &&
23968 !DCI.isBeforeLegalize() &&
23969 !ISD::isBuildVectorOfConstantSDNodes(Cond.getNode())) {
23970 unsigned BitWidth = Cond.getValueType().getScalarSizeInBits();
23972 // Don't optimize vector selects that map to mask-registers.
23976 // We can only handle the cases where VSELECT is directly legal on the
23977 // subtarget. We custom lower VSELECT nodes with constant conditions and
23978 // this makes it hard to see whether a dynamic VSELECT will correctly
23979 // lower, so we both check the operation's status and explicitly handle the
23980 // cases where a *dynamic* blend will fail even though a constant-condition
23981 // blend could be custom lowered.
23982 // FIXME: We should find a better way to handle this class of problems.
23983 // Potentially, we should combine constant-condition vselect nodes
23984 // pre-legalization into shuffles and not mark as many types as custom
23986 if (!TLI.isOperationLegalOrCustom(ISD::VSELECT, VT))
23988 // FIXME: We don't support i16-element blends currently. We could and
23989 // should support them by making *all* the bits in the condition be set
23990 // rather than just the high bit and using an i8-element blend.
23991 if (VT.getVectorElementType() == MVT::i16)
23993 // Dynamic blending was only available from SSE4.1 onward.
23994 if (VT.is128BitVector() && !Subtarget->hasSSE41())
23996 // Byte blends are only available in AVX2
23997 if (VT == MVT::v32i8 && !Subtarget->hasAVX2())
24000 assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size");
24001 APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);
24003 APInt KnownZero, KnownOne;
24004 TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(),
24005 DCI.isBeforeLegalizeOps());
24006 if (TLO.ShrinkDemandedConstant(Cond, DemandedMask) ||
24007 TLI.SimplifyDemandedBits(Cond, DemandedMask, KnownZero, KnownOne,
24009 // If we changed the computation somewhere in the DAG, this change
24010 // will affect all users of Cond.
24011 // Make sure it is fine and update all the nodes so that we do not
24012 // use the generic VSELECT anymore. Otherwise, we may perform
24013 // wrong optimizations as we messed up with the actual expectation
24014 // for the vector boolean values.
24015 if (Cond != TLO.Old) {
24016 // Check all uses of that condition operand to check whether it will be
24017 // consumed by non-BLEND instructions, which may depend on all bits are
24019 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
24021 if (I->getOpcode() != ISD::VSELECT)
24022 // TODO: Add other opcodes eventually lowered into BLEND.
24025 // Update all the users of the condition, before committing the change,
24026 // so that the VSELECT optimizations that expect the correct vector
24027 // boolean value will not be triggered.
24028 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
24030 DAG.ReplaceAllUsesOfValueWith(
24032 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(*I), I->getValueType(0),
24033 Cond, I->getOperand(1), I->getOperand(2)));
24034 DCI.CommitTargetLoweringOpt(TLO);
24037 // At this point, only Cond is changed. Change the condition
24038 // just for N to keep the opportunity to optimize all other
24039 // users their own way.
24040 DAG.ReplaceAllUsesOfValueWith(
24042 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(N), N->getValueType(0),
24043 TLO.New, N->getOperand(1), N->getOperand(2)));
24051 // Check whether a boolean test is testing a boolean value generated by
24052 // X86ISD::SETCC. If so, return the operand of that SETCC and proper condition
24055 // Simplify the following patterns:
24056 // (Op (CMP (SETCC Cond EFLAGS) 1) EQ) or
24057 // (Op (CMP (SETCC Cond EFLAGS) 0) NEQ)
24058 // to (Op EFLAGS Cond)
24060 // (Op (CMP (SETCC Cond EFLAGS) 0) EQ) or
24061 // (Op (CMP (SETCC Cond EFLAGS) 1) NEQ)
24062 // to (Op EFLAGS !Cond)
24064 // where Op could be BRCOND or CMOV.
24066 static SDValue checkBoolTestSetCCCombine(SDValue Cmp, X86::CondCode &CC) {
24067 // Quit if not CMP and SUB with its value result used.
24068 if (Cmp.getOpcode() != X86ISD::CMP &&
24069 (Cmp.getOpcode() != X86ISD::SUB || Cmp.getNode()->hasAnyUseOfValue(0)))
24072 // Quit if not used as a boolean value.
24073 if (CC != X86::COND_E && CC != X86::COND_NE)
24076 // Check CMP operands. One of them should be 0 or 1 and the other should be
24077 // an SetCC or extended from it.
24078 SDValue Op1 = Cmp.getOperand(0);
24079 SDValue Op2 = Cmp.getOperand(1);
24082 const ConstantSDNode* C = nullptr;
24083 bool needOppositeCond = (CC == X86::COND_E);
24084 bool checkAgainstTrue = false; // Is it a comparison against 1?
24086 if ((C = dyn_cast<ConstantSDNode>(Op1)))
24088 else if ((C = dyn_cast<ConstantSDNode>(Op2)))
24090 else // Quit if all operands are not constants.
24093 if (C->getZExtValue() == 1) {
24094 needOppositeCond = !needOppositeCond;
24095 checkAgainstTrue = true;
24096 } else if (C->getZExtValue() != 0)
24097 // Quit if the constant is neither 0 or 1.
24100 bool truncatedToBoolWithAnd = false;
24101 // Skip (zext $x), (trunc $x), or (and $x, 1) node.
24102 while (SetCC.getOpcode() == ISD::ZERO_EXTEND ||
24103 SetCC.getOpcode() == ISD::TRUNCATE ||
24104 SetCC.getOpcode() == ISD::AND) {
24105 if (SetCC.getOpcode() == ISD::AND) {
24107 ConstantSDNode *CS;
24108 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(0))) &&
24109 CS->getZExtValue() == 1)
24111 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(1))) &&
24112 CS->getZExtValue() == 1)
24116 SetCC = SetCC.getOperand(OpIdx);
24117 truncatedToBoolWithAnd = true;
24119 SetCC = SetCC.getOperand(0);
24122 switch (SetCC.getOpcode()) {
24123 case X86ISD::SETCC_CARRY:
24124 // Since SETCC_CARRY gives output based on R = CF ? ~0 : 0, it's unsafe to
24125 // simplify it if the result of SETCC_CARRY is not canonicalized to 0 or 1,
24126 // i.e. it's a comparison against true but the result of SETCC_CARRY is not
24127 // truncated to i1 using 'and'.
24128 if (checkAgainstTrue && !truncatedToBoolWithAnd)
24130 assert(X86::CondCode(SetCC.getConstantOperandVal(0)) == X86::COND_B &&
24131 "Invalid use of SETCC_CARRY!");
24133 case X86ISD::SETCC:
24134 // Set the condition code or opposite one if necessary.
24135 CC = X86::CondCode(SetCC.getConstantOperandVal(0));
24136 if (needOppositeCond)
24137 CC = X86::GetOppositeBranchCondition(CC);
24138 return SetCC.getOperand(1);
24139 case X86ISD::CMOV: {
24140 // Check whether false/true value has canonical one, i.e. 0 or 1.
24141 ConstantSDNode *FVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(0));
24142 ConstantSDNode *TVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(1));
24143 // Quit if true value is not a constant.
24146 // Quit if false value is not a constant.
24148 SDValue Op = SetCC.getOperand(0);
24149 // Skip 'zext' or 'trunc' node.
24150 if (Op.getOpcode() == ISD::ZERO_EXTEND ||
24151 Op.getOpcode() == ISD::TRUNCATE)
24152 Op = Op.getOperand(0);
24153 // A special case for rdrand/rdseed, where 0 is set if false cond is
24155 if ((Op.getOpcode() != X86ISD::RDRAND &&
24156 Op.getOpcode() != X86ISD::RDSEED) || Op.getResNo() != 0)
24159 // Quit if false value is not the constant 0 or 1.
24160 bool FValIsFalse = true;
24161 if (FVal && FVal->getZExtValue() != 0) {
24162 if (FVal->getZExtValue() != 1)
24164 // If FVal is 1, opposite cond is needed.
24165 needOppositeCond = !needOppositeCond;
24166 FValIsFalse = false;
24168 // Quit if TVal is not the constant opposite of FVal.
24169 if (FValIsFalse && TVal->getZExtValue() != 1)
24171 if (!FValIsFalse && TVal->getZExtValue() != 0)
24173 CC = X86::CondCode(SetCC.getConstantOperandVal(2));
24174 if (needOppositeCond)
24175 CC = X86::GetOppositeBranchCondition(CC);
24176 return SetCC.getOperand(3);
24183 /// Check whether Cond is an AND/OR of SETCCs off of the same EFLAGS.
24185 /// (X86or (X86setcc) (X86setcc))
24186 /// (X86cmp (and (X86setcc) (X86setcc)), 0)
24187 static bool checkBoolTestAndOrSetCCCombine(SDValue Cond, X86::CondCode &CC0,
24188 X86::CondCode &CC1, SDValue &Flags,
24190 if (Cond->getOpcode() == X86ISD::CMP) {
24191 ConstantSDNode *CondOp1C = dyn_cast<ConstantSDNode>(Cond->getOperand(1));
24192 if (!CondOp1C || !CondOp1C->isNullValue())
24195 Cond = Cond->getOperand(0);
24200 SDValue SetCC0, SetCC1;
24201 switch (Cond->getOpcode()) {
24202 default: return false;
24209 SetCC0 = Cond->getOperand(0);
24210 SetCC1 = Cond->getOperand(1);
24214 // Make sure we have SETCC nodes, using the same flags value.
24215 if (SetCC0.getOpcode() != X86ISD::SETCC ||
24216 SetCC1.getOpcode() != X86ISD::SETCC ||
24217 SetCC0->getOperand(1) != SetCC1->getOperand(1))
24220 CC0 = (X86::CondCode)SetCC0->getConstantOperandVal(0);
24221 CC1 = (X86::CondCode)SetCC1->getConstantOperandVal(0);
24222 Flags = SetCC0->getOperand(1);
24226 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
24227 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
24228 TargetLowering::DAGCombinerInfo &DCI,
24229 const X86Subtarget *Subtarget) {
24232 // If the flag operand isn't dead, don't touch this CMOV.
24233 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
24236 SDValue FalseOp = N->getOperand(0);
24237 SDValue TrueOp = N->getOperand(1);
24238 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
24239 SDValue Cond = N->getOperand(3);
24241 if (CC == X86::COND_E || CC == X86::COND_NE) {
24242 switch (Cond.getOpcode()) {
24246 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
24247 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
24248 return (CC == X86::COND_E) ? FalseOp : TrueOp;
24254 Flags = checkBoolTestSetCCCombine(Cond, CC);
24255 if (Flags.getNode() &&
24256 // Extra check as FCMOV only supports a subset of X86 cond.
24257 (FalseOp.getValueType() != MVT::f80 || hasFPCMov(CC))) {
24258 SDValue Ops[] = { FalseOp, TrueOp,
24259 DAG.getConstant(CC, DL, MVT::i8), Flags };
24260 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
24263 // If this is a select between two integer constants, try to do some
24264 // optimizations. Note that the operands are ordered the opposite of SELECT
24266 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
24267 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
24268 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
24269 // larger than FalseC (the false value).
24270 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
24271 CC = X86::GetOppositeBranchCondition(CC);
24272 std::swap(TrueC, FalseC);
24273 std::swap(TrueOp, FalseOp);
24276 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
24277 // This is efficient for any integer data type (including i8/i16) and
24279 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
24280 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
24281 DAG.getConstant(CC, DL, MVT::i8), Cond);
24283 // Zero extend the condition if needed.
24284 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
24286 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
24287 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
24288 DAG.getConstant(ShAmt, DL, MVT::i8));
24289 if (N->getNumValues() == 2) // Dead flag value?
24290 return DCI.CombineTo(N, Cond, SDValue());
24294 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
24295 // for any integer data type, including i8/i16.
24296 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
24297 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
24298 DAG.getConstant(CC, DL, MVT::i8), Cond);
24300 // Zero extend the condition if needed.
24301 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
24302 FalseC->getValueType(0), Cond);
24303 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
24304 SDValue(FalseC, 0));
24306 if (N->getNumValues() == 2) // Dead flag value?
24307 return DCI.CombineTo(N, Cond, SDValue());
24311 // Optimize cases that will turn into an LEA instruction. This requires
24312 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
24313 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
24314 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
24315 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
24317 bool isFastMultiplier = false;
24319 switch ((unsigned char)Diff) {
24321 case 1: // result = add base, cond
24322 case 2: // result = lea base( , cond*2)
24323 case 3: // result = lea base(cond, cond*2)
24324 case 4: // result = lea base( , cond*4)
24325 case 5: // result = lea base(cond, cond*4)
24326 case 8: // result = lea base( , cond*8)
24327 case 9: // result = lea base(cond, cond*8)
24328 isFastMultiplier = true;
24333 if (isFastMultiplier) {
24334 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
24335 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
24336 DAG.getConstant(CC, DL, MVT::i8), Cond);
24337 // Zero extend the condition if needed.
24338 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
24340 // Scale the condition by the difference.
24342 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
24343 DAG.getConstant(Diff, DL, Cond.getValueType()));
24345 // Add the base if non-zero.
24346 if (FalseC->getAPIntValue() != 0)
24347 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
24348 SDValue(FalseC, 0));
24349 if (N->getNumValues() == 2) // Dead flag value?
24350 return DCI.CombineTo(N, Cond, SDValue());
24357 // Handle these cases:
24358 // (select (x != c), e, c) -> select (x != c), e, x),
24359 // (select (x == c), c, e) -> select (x == c), x, e)
24360 // where the c is an integer constant, and the "select" is the combination
24361 // of CMOV and CMP.
24363 // The rationale for this change is that the conditional-move from a constant
24364 // needs two instructions, however, conditional-move from a register needs
24365 // only one instruction.
24367 // CAVEAT: By replacing a constant with a symbolic value, it may obscure
24368 // some instruction-combining opportunities. This opt needs to be
24369 // postponed as late as possible.
24371 if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) {
24372 // the DCI.xxxx conditions are provided to postpone the optimization as
24373 // late as possible.
24375 ConstantSDNode *CmpAgainst = nullptr;
24376 if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) &&
24377 (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) &&
24378 !isa<ConstantSDNode>(Cond.getOperand(0))) {
24380 if (CC == X86::COND_NE &&
24381 CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) {
24382 CC = X86::GetOppositeBranchCondition(CC);
24383 std::swap(TrueOp, FalseOp);
24386 if (CC == X86::COND_E &&
24387 CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) {
24388 SDValue Ops[] = { FalseOp, Cond.getOperand(0),
24389 DAG.getConstant(CC, DL, MVT::i8), Cond };
24390 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops);
24395 // Fold and/or of setcc's to double CMOV:
24396 // (CMOV F, T, ((cc1 | cc2) != 0)) -> (CMOV (CMOV F, T, cc1), T, cc2)
24397 // (CMOV F, T, ((cc1 & cc2) != 0)) -> (CMOV (CMOV T, F, !cc1), F, !cc2)
24399 // This combine lets us generate:
24400 // cmovcc1 (jcc1 if we don't have CMOV)
24406 // cmovne (jne if we don't have CMOV)
24407 // When we can't use the CMOV instruction, it might increase branch
24409 // When we can use CMOV, or when there is no mispredict, this improves
24410 // throughput and reduces register pressure.
24412 if (CC == X86::COND_NE) {
24414 X86::CondCode CC0, CC1;
24416 if (checkBoolTestAndOrSetCCCombine(Cond, CC0, CC1, Flags, isAndSetCC)) {
24418 std::swap(FalseOp, TrueOp);
24419 CC0 = X86::GetOppositeBranchCondition(CC0);
24420 CC1 = X86::GetOppositeBranchCondition(CC1);
24423 SDValue LOps[] = {FalseOp, TrueOp, DAG.getConstant(CC0, DL, MVT::i8),
24425 SDValue LCMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), LOps);
24426 SDValue Ops[] = {LCMOV, TrueOp, DAG.getConstant(CC1, DL, MVT::i8), Flags};
24427 SDValue CMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
24428 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SDValue(CMOV.getNode(), 1));
24436 /// PerformMulCombine - Optimize a single multiply with constant into two
24437 /// in order to implement it with two cheaper instructions, e.g.
24438 /// LEA + SHL, LEA + LEA.
24439 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
24440 TargetLowering::DAGCombinerInfo &DCI) {
24441 // An imul is usually smaller than the alternative sequence.
24442 if (DAG.getMachineFunction().getFunction()->optForMinSize())
24445 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
24448 EVT VT = N->getValueType(0);
24449 if (VT != MVT::i64 && VT != MVT::i32)
24452 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
24455 uint64_t MulAmt = C->getZExtValue();
24456 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
24459 uint64_t MulAmt1 = 0;
24460 uint64_t MulAmt2 = 0;
24461 if ((MulAmt % 9) == 0) {
24463 MulAmt2 = MulAmt / 9;
24464 } else if ((MulAmt % 5) == 0) {
24466 MulAmt2 = MulAmt / 5;
24467 } else if ((MulAmt % 3) == 0) {
24469 MulAmt2 = MulAmt / 3;
24472 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
24475 if (isPowerOf2_64(MulAmt2) &&
24476 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
24477 // If second multiplifer is pow2, issue it first. We want the multiply by
24478 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
24480 std::swap(MulAmt1, MulAmt2);
24483 if (isPowerOf2_64(MulAmt1))
24484 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
24485 DAG.getConstant(Log2_64(MulAmt1), DL, MVT::i8));
24487 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
24488 DAG.getConstant(MulAmt1, DL, VT));
24490 if (isPowerOf2_64(MulAmt2))
24491 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
24492 DAG.getConstant(Log2_64(MulAmt2), DL, MVT::i8));
24494 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
24495 DAG.getConstant(MulAmt2, DL, VT));
24497 // Do not add new nodes to DAG combiner worklist.
24498 DCI.CombineTo(N, NewMul, false);
24503 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
24504 SDValue N0 = N->getOperand(0);
24505 SDValue N1 = N->getOperand(1);
24506 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
24507 EVT VT = N0.getValueType();
24509 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
24510 // since the result of setcc_c is all zero's or all ones.
24511 if (VT.isInteger() && !VT.isVector() &&
24512 N1C && N0.getOpcode() == ISD::AND &&
24513 N0.getOperand(1).getOpcode() == ISD::Constant) {
24514 SDValue N00 = N0.getOperand(0);
24515 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
24516 APInt ShAmt = N1C->getAPIntValue();
24517 Mask = Mask.shl(ShAmt);
24518 bool MaskOK = false;
24519 // We can handle cases concerning bit-widening nodes containing setcc_c if
24520 // we carefully interrogate the mask to make sure we are semantics
24522 // The transform is not safe if the result of C1 << C2 exceeds the bitwidth
24523 // of the underlying setcc_c operation if the setcc_c was zero extended.
24524 // Consider the following example:
24525 // zext(setcc_c) -> i32 0x0000FFFF
24526 // c1 -> i32 0x0000FFFF
24527 // c2 -> i32 0x00000001
24528 // (shl (and (setcc_c), c1), c2) -> i32 0x0001FFFE
24529 // (and setcc_c, (c1 << c2)) -> i32 0x0000FFFE
24530 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
24532 } else if (N00.getOpcode() == ISD::SIGN_EXTEND &&
24533 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
24535 } else if ((N00.getOpcode() == ISD::ZERO_EXTEND ||
24536 N00.getOpcode() == ISD::ANY_EXTEND) &&
24537 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
24538 MaskOK = Mask.isIntN(N00.getOperand(0).getValueSizeInBits());
24540 if (MaskOK && Mask != 0) {
24542 return DAG.getNode(ISD::AND, DL, VT, N00, DAG.getConstant(Mask, DL, VT));
24546 // Hardware support for vector shifts is sparse which makes us scalarize the
24547 // vector operations in many cases. Also, on sandybridge ADD is faster than
24549 // (shl V, 1) -> add V,V
24550 if (auto *N1BV = dyn_cast<BuildVectorSDNode>(N1))
24551 if (auto *N1SplatC = N1BV->getConstantSplatNode()) {
24552 assert(N0.getValueType().isVector() && "Invalid vector shift type");
24553 // We shift all of the values by one. In many cases we do not have
24554 // hardware support for this operation. This is better expressed as an ADD
24556 if (N1SplatC->getAPIntValue() == 1)
24557 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N0);
24563 /// \brief Returns a vector of 0s if the node in input is a vector logical
24564 /// shift by a constant amount which is known to be bigger than or equal
24565 /// to the vector element size in bits.
24566 static SDValue performShiftToAllZeros(SDNode *N, SelectionDAG &DAG,
24567 const X86Subtarget *Subtarget) {
24568 EVT VT = N->getValueType(0);
24570 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16 &&
24571 (!Subtarget->hasInt256() ||
24572 (VT != MVT::v4i64 && VT != MVT::v8i32 && VT != MVT::v16i16)))
24575 SDValue Amt = N->getOperand(1);
24577 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Amt))
24578 if (auto *AmtSplat = AmtBV->getConstantSplatNode()) {
24579 APInt ShiftAmt = AmtSplat->getAPIntValue();
24580 unsigned MaxAmount =
24581 VT.getSimpleVT().getVectorElementType().getSizeInBits();
24583 // SSE2/AVX2 logical shifts always return a vector of 0s
24584 // if the shift amount is bigger than or equal to
24585 // the element size. The constant shift amount will be
24586 // encoded as a 8-bit immediate.
24587 if (ShiftAmt.trunc(8).uge(MaxAmount))
24588 return getZeroVector(VT, Subtarget, DAG, DL);
24594 /// PerformShiftCombine - Combine shifts.
24595 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
24596 TargetLowering::DAGCombinerInfo &DCI,
24597 const X86Subtarget *Subtarget) {
24598 if (N->getOpcode() == ISD::SHL)
24599 if (SDValue V = PerformSHLCombine(N, DAG))
24602 // Try to fold this logical shift into a zero vector.
24603 if (N->getOpcode() != ISD::SRA)
24604 if (SDValue V = performShiftToAllZeros(N, DAG, Subtarget))
24610 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
24611 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
24612 // and friends. Likewise for OR -> CMPNEQSS.
24613 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
24614 TargetLowering::DAGCombinerInfo &DCI,
24615 const X86Subtarget *Subtarget) {
24618 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
24619 // we're requiring SSE2 for both.
24620 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
24621 SDValue N0 = N->getOperand(0);
24622 SDValue N1 = N->getOperand(1);
24623 SDValue CMP0 = N0->getOperand(1);
24624 SDValue CMP1 = N1->getOperand(1);
24627 // The SETCCs should both refer to the same CMP.
24628 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
24631 SDValue CMP00 = CMP0->getOperand(0);
24632 SDValue CMP01 = CMP0->getOperand(1);
24633 EVT VT = CMP00.getValueType();
24635 if (VT == MVT::f32 || VT == MVT::f64) {
24636 bool ExpectingFlags = false;
24637 // Check for any users that want flags:
24638 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
24639 !ExpectingFlags && UI != UE; ++UI)
24640 switch (UI->getOpcode()) {
24645 ExpectingFlags = true;
24647 case ISD::CopyToReg:
24648 case ISD::SIGN_EXTEND:
24649 case ISD::ZERO_EXTEND:
24650 case ISD::ANY_EXTEND:
24654 if (!ExpectingFlags) {
24655 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
24656 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
24658 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
24659 X86::CondCode tmp = cc0;
24664 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
24665 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
24666 // FIXME: need symbolic constants for these magic numbers.
24667 // See X86ATTInstPrinter.cpp:printSSECC().
24668 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
24669 if (Subtarget->hasAVX512()) {
24670 SDValue FSetCC = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CMP00,
24672 DAG.getConstant(x86cc, DL, MVT::i8));
24673 if (N->getValueType(0) != MVT::i1)
24674 return DAG.getNode(ISD::ZERO_EXTEND, DL, N->getValueType(0),
24678 SDValue OnesOrZeroesF = DAG.getNode(X86ISD::FSETCC, DL,
24679 CMP00.getValueType(), CMP00, CMP01,
24680 DAG.getConstant(x86cc, DL,
24683 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
24684 MVT IntVT = is64BitFP ? MVT::i64 : MVT::i32;
24686 if (is64BitFP && !Subtarget->is64Bit()) {
24687 // On a 32-bit target, we cannot bitcast the 64-bit float to a
24688 // 64-bit integer, since that's not a legal type. Since
24689 // OnesOrZeroesF is all ones of all zeroes, we don't need all the
24690 // bits, but can do this little dance to extract the lowest 32 bits
24691 // and work with those going forward.
24692 SDValue Vector64 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64,
24694 SDValue Vector32 = DAG.getBitcast(MVT::v4f32, Vector64);
24695 OnesOrZeroesF = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32,
24696 Vector32, DAG.getIntPtrConstant(0, DL));
24700 SDValue OnesOrZeroesI = DAG.getBitcast(IntVT, OnesOrZeroesF);
24701 SDValue ANDed = DAG.getNode(ISD::AND, DL, IntVT, OnesOrZeroesI,
24702 DAG.getConstant(1, DL, IntVT));
24703 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8,
24705 return OneBitOfTruth;
24713 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
24714 /// so it can be folded inside ANDNP.
24715 static bool CanFoldXORWithAllOnes(const SDNode *N) {
24716 EVT VT = N->getValueType(0);
24718 // Match direct AllOnes for 128 and 256-bit vectors
24719 if (ISD::isBuildVectorAllOnes(N))
24722 // Look through a bit convert.
24723 if (N->getOpcode() == ISD::BITCAST)
24724 N = N->getOperand(0).getNode();
24726 // Sometimes the operand may come from a insert_subvector building a 256-bit
24728 if (VT.is256BitVector() &&
24729 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
24730 SDValue V1 = N->getOperand(0);
24731 SDValue V2 = N->getOperand(1);
24733 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
24734 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
24735 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
24736 ISD::isBuildVectorAllOnes(V2.getNode()))
24743 // On AVX/AVX2 the type v8i1 is legalized to v8i16, which is an XMM sized
24744 // register. In most cases we actually compare or select YMM-sized registers
24745 // and mixing the two types creates horrible code. This method optimizes
24746 // some of the transition sequences.
24747 static SDValue WidenMaskArithmetic(SDNode *N, SelectionDAG &DAG,
24748 TargetLowering::DAGCombinerInfo &DCI,
24749 const X86Subtarget *Subtarget) {
24750 EVT VT = N->getValueType(0);
24751 if (!VT.is256BitVector())
24754 assert((N->getOpcode() == ISD::ANY_EXTEND ||
24755 N->getOpcode() == ISD::ZERO_EXTEND ||
24756 N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node");
24758 SDValue Narrow = N->getOperand(0);
24759 EVT NarrowVT = Narrow->getValueType(0);
24760 if (!NarrowVT.is128BitVector())
24763 if (Narrow->getOpcode() != ISD::XOR &&
24764 Narrow->getOpcode() != ISD::AND &&
24765 Narrow->getOpcode() != ISD::OR)
24768 SDValue N0 = Narrow->getOperand(0);
24769 SDValue N1 = Narrow->getOperand(1);
24772 // The Left side has to be a trunc.
24773 if (N0.getOpcode() != ISD::TRUNCATE)
24776 // The type of the truncated inputs.
24777 EVT WideVT = N0->getOperand(0)->getValueType(0);
24781 // The right side has to be a 'trunc' or a constant vector.
24782 bool RHSTrunc = N1.getOpcode() == ISD::TRUNCATE;
24783 ConstantSDNode *RHSConstSplat = nullptr;
24784 if (auto *RHSBV = dyn_cast<BuildVectorSDNode>(N1))
24785 RHSConstSplat = RHSBV->getConstantSplatNode();
24786 if (!RHSTrunc && !RHSConstSplat)
24789 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24791 if (!TLI.isOperationLegalOrPromote(Narrow->getOpcode(), WideVT))
24794 // Set N0 and N1 to hold the inputs to the new wide operation.
24795 N0 = N0->getOperand(0);
24796 if (RHSConstSplat) {
24797 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT.getVectorElementType(),
24798 SDValue(RHSConstSplat, 0));
24799 SmallVector<SDValue, 8> C(WideVT.getVectorNumElements(), N1);
24800 N1 = DAG.getNode(ISD::BUILD_VECTOR, DL, WideVT, C);
24801 } else if (RHSTrunc) {
24802 N1 = N1->getOperand(0);
24805 // Generate the wide operation.
24806 SDValue Op = DAG.getNode(Narrow->getOpcode(), DL, WideVT, N0, N1);
24807 unsigned Opcode = N->getOpcode();
24809 case ISD::ANY_EXTEND:
24811 case ISD::ZERO_EXTEND: {
24812 unsigned InBits = NarrowVT.getScalarSizeInBits();
24813 APInt Mask = APInt::getAllOnesValue(InBits);
24814 Mask = Mask.zext(VT.getScalarSizeInBits());
24815 return DAG.getNode(ISD::AND, DL, VT,
24816 Op, DAG.getConstant(Mask, DL, VT));
24818 case ISD::SIGN_EXTEND:
24819 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT,
24820 Op, DAG.getValueType(NarrowVT));
24822 llvm_unreachable("Unexpected opcode");
24826 static SDValue VectorZextCombine(SDNode *N, SelectionDAG &DAG,
24827 TargetLowering::DAGCombinerInfo &DCI,
24828 const X86Subtarget *Subtarget) {
24829 SDValue N0 = N->getOperand(0);
24830 SDValue N1 = N->getOperand(1);
24833 // A vector zext_in_reg may be represented as a shuffle,
24834 // feeding into a bitcast (this represents anyext) feeding into
24835 // an and with a mask.
24836 // We'd like to try to combine that into a shuffle with zero
24837 // plus a bitcast, removing the and.
24838 if (N0.getOpcode() != ISD::BITCAST ||
24839 N0.getOperand(0).getOpcode() != ISD::VECTOR_SHUFFLE)
24842 // The other side of the AND should be a splat of 2^C, where C
24843 // is the number of bits in the source type.
24844 if (N1.getOpcode() == ISD::BITCAST)
24845 N1 = N1.getOperand(0);
24846 if (N1.getOpcode() != ISD::BUILD_VECTOR)
24848 BuildVectorSDNode *Vector = cast<BuildVectorSDNode>(N1);
24850 ShuffleVectorSDNode *Shuffle = cast<ShuffleVectorSDNode>(N0.getOperand(0));
24851 EVT SrcType = Shuffle->getValueType(0);
24853 // We expect a single-source shuffle
24854 if (Shuffle->getOperand(1)->getOpcode() != ISD::UNDEF)
24857 unsigned SrcSize = SrcType.getScalarSizeInBits();
24859 APInt SplatValue, SplatUndef;
24860 unsigned SplatBitSize;
24862 if (!Vector->isConstantSplat(SplatValue, SplatUndef,
24863 SplatBitSize, HasAnyUndefs))
24866 unsigned ResSize = N1.getValueType().getScalarSizeInBits();
24867 // Make sure the splat matches the mask we expect
24868 if (SplatBitSize > ResSize ||
24869 (SplatValue + 1).exactLogBase2() != (int)SrcSize)
24872 // Make sure the input and output size make sense
24873 if (SrcSize >= ResSize || ResSize % SrcSize)
24876 // We expect a shuffle of the form <0, u, u, u, 1, u, u, u...>
24877 // The number of u's between each two values depends on the ratio between
24878 // the source and dest type.
24879 unsigned ZextRatio = ResSize / SrcSize;
24880 bool IsZext = true;
24881 for (unsigned i = 0; i < SrcType.getVectorNumElements(); ++i) {
24882 if (i % ZextRatio) {
24883 if (Shuffle->getMaskElt(i) > 0) {
24889 if (Shuffle->getMaskElt(i) != (int)(i / ZextRatio)) {
24890 // Expected element number
24900 // Ok, perform the transformation - replace the shuffle with
24901 // a shuffle of the form <0, k, k, k, 1, k, k, k> with zero
24902 // (instead of undef) where the k elements come from the zero vector.
24903 SmallVector<int, 8> Mask;
24904 unsigned NumElems = SrcType.getVectorNumElements();
24905 for (unsigned i = 0; i < NumElems; ++i)
24907 Mask.push_back(NumElems);
24909 Mask.push_back(i / ZextRatio);
24911 SDValue NewShuffle = DAG.getVectorShuffle(Shuffle->getValueType(0), DL,
24912 Shuffle->getOperand(0), DAG.getConstant(0, DL, SrcType), Mask);
24913 return DAG.getBitcast(N0.getValueType(), NewShuffle);
24916 /// If both input operands of a logic op are being cast from floating point
24917 /// types, try to convert this into a floating point logic node to avoid
24918 /// unnecessary moves from SSE to integer registers.
24919 static SDValue convertIntLogicToFPLogic(SDNode *N, SelectionDAG &DAG,
24920 const X86Subtarget *Subtarget) {
24921 unsigned FPOpcode = ISD::DELETED_NODE;
24922 if (N->getOpcode() == ISD::AND)
24923 FPOpcode = X86ISD::FAND;
24924 else if (N->getOpcode() == ISD::OR)
24925 FPOpcode = X86ISD::FOR;
24926 else if (N->getOpcode() == ISD::XOR)
24927 FPOpcode = X86ISD::FXOR;
24929 assert(FPOpcode != ISD::DELETED_NODE &&
24930 "Unexpected input node for FP logic conversion");
24932 EVT VT = N->getValueType(0);
24933 SDValue N0 = N->getOperand(0);
24934 SDValue N1 = N->getOperand(1);
24936 if (N0.getOpcode() == ISD::BITCAST && N1.getOpcode() == ISD::BITCAST &&
24937 ((Subtarget->hasSSE1() && VT == MVT::i32) ||
24938 (Subtarget->hasSSE2() && VT == MVT::i64))) {
24939 SDValue N00 = N0.getOperand(0);
24940 SDValue N10 = N1.getOperand(0);
24941 EVT N00Type = N00.getValueType();
24942 EVT N10Type = N10.getValueType();
24943 if (N00Type.isFloatingPoint() && N10Type.isFloatingPoint()) {
24944 SDValue FPLogic = DAG.getNode(FPOpcode, DL, N00Type, N00, N10);
24945 return DAG.getBitcast(VT, FPLogic);
24951 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
24952 TargetLowering::DAGCombinerInfo &DCI,
24953 const X86Subtarget *Subtarget) {
24954 if (DCI.isBeforeLegalizeOps())
24957 if (SDValue Zext = VectorZextCombine(N, DAG, DCI, Subtarget))
24960 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
24963 if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
24966 EVT VT = N->getValueType(0);
24967 SDValue N0 = N->getOperand(0);
24968 SDValue N1 = N->getOperand(1);
24971 // Create BEXTR instructions
24972 // BEXTR is ((X >> imm) & (2**size-1))
24973 if (VT == MVT::i32 || VT == MVT::i64) {
24974 // Check for BEXTR.
24975 if ((Subtarget->hasBMI() || Subtarget->hasTBM()) &&
24976 (N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::SRL)) {
24977 ConstantSDNode *MaskNode = dyn_cast<ConstantSDNode>(N1);
24978 ConstantSDNode *ShiftNode = dyn_cast<ConstantSDNode>(N0.getOperand(1));
24979 if (MaskNode && ShiftNode) {
24980 uint64_t Mask = MaskNode->getZExtValue();
24981 uint64_t Shift = ShiftNode->getZExtValue();
24982 if (isMask_64(Mask)) {
24983 uint64_t MaskSize = countPopulation(Mask);
24984 if (Shift + MaskSize <= VT.getSizeInBits())
24985 return DAG.getNode(X86ISD::BEXTR, DL, VT, N0.getOperand(0),
24986 DAG.getConstant(Shift | (MaskSize << 8), DL,
24995 // Want to form ANDNP nodes:
24996 // 1) In the hopes of then easily combining them with OR and AND nodes
24997 // to form PBLEND/PSIGN.
24998 // 2) To match ANDN packed intrinsics
24999 if (VT != MVT::v2i64 && VT != MVT::v4i64)
25002 // Check LHS for vnot
25003 if (N0.getOpcode() == ISD::XOR &&
25004 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
25005 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
25006 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
25008 // Check RHS for vnot
25009 if (N1.getOpcode() == ISD::XOR &&
25010 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
25011 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
25012 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
25017 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
25018 TargetLowering::DAGCombinerInfo &DCI,
25019 const X86Subtarget *Subtarget) {
25020 if (DCI.isBeforeLegalizeOps())
25023 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
25026 if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
25029 SDValue N0 = N->getOperand(0);
25030 SDValue N1 = N->getOperand(1);
25031 EVT VT = N->getValueType(0);
25033 // look for psign/blend
25034 if (VT == MVT::v2i64 || VT == MVT::v4i64) {
25035 if (!Subtarget->hasSSSE3() ||
25036 (VT == MVT::v4i64 && !Subtarget->hasInt256()))
25039 // Canonicalize pandn to RHS
25040 if (N0.getOpcode() == X86ISD::ANDNP)
25042 // or (and (m, y), (pandn m, x))
25043 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
25044 SDValue Mask = N1.getOperand(0);
25045 SDValue X = N1.getOperand(1);
25047 if (N0.getOperand(0) == Mask)
25048 Y = N0.getOperand(1);
25049 if (N0.getOperand(1) == Mask)
25050 Y = N0.getOperand(0);
25052 // Check to see if the mask appeared in both the AND and ANDNP and
25056 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
25057 // Look through mask bitcast.
25058 if (Mask.getOpcode() == ISD::BITCAST)
25059 Mask = Mask.getOperand(0);
25060 if (X.getOpcode() == ISD::BITCAST)
25061 X = X.getOperand(0);
25062 if (Y.getOpcode() == ISD::BITCAST)
25063 Y = Y.getOperand(0);
25065 EVT MaskVT = Mask.getValueType();
25067 // Validate that the Mask operand is a vector sra node.
25068 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
25069 // there is no psrai.b
25070 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
25071 unsigned SraAmt = ~0;
25072 if (Mask.getOpcode() == ISD::SRA) {
25073 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Mask.getOperand(1)))
25074 if (auto *AmtConst = AmtBV->getConstantSplatNode())
25075 SraAmt = AmtConst->getZExtValue();
25076 } else if (Mask.getOpcode() == X86ISD::VSRAI) {
25077 SDValue SraC = Mask.getOperand(1);
25078 SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
25080 if ((SraAmt + 1) != EltBits)
25085 // Now we know we at least have a plendvb with the mask val. See if
25086 // we can form a psignb/w/d.
25087 // psign = x.type == y.type == mask.type && y = sub(0, x);
25088 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
25089 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
25090 X.getValueType() == MaskVT && Y.getValueType() == MaskVT) {
25091 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
25092 "Unsupported VT for PSIGN");
25093 Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask.getOperand(0));
25094 return DAG.getBitcast(VT, Mask);
25096 // PBLENDVB only available on SSE 4.1
25097 if (!Subtarget->hasSSE41())
25100 MVT BlendVT = (VT == MVT::v4i64) ? MVT::v32i8 : MVT::v16i8;
25102 X = DAG.getBitcast(BlendVT, X);
25103 Y = DAG.getBitcast(BlendVT, Y);
25104 Mask = DAG.getBitcast(BlendVT, Mask);
25105 Mask = DAG.getNode(ISD::VSELECT, DL, BlendVT, Mask, Y, X);
25106 return DAG.getBitcast(VT, Mask);
25110 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
25113 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
25114 bool OptForSize = DAG.getMachineFunction().getFunction()->optForSize();
25116 // SHLD/SHRD instructions have lower register pressure, but on some
25117 // platforms they have higher latency than the equivalent
25118 // series of shifts/or that would otherwise be generated.
25119 // Don't fold (or (x << c) | (y >> (64 - c))) if SHLD/SHRD instructions
25120 // have higher latencies and we are not optimizing for size.
25121 if (!OptForSize && Subtarget->isSHLDSlow())
25124 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
25126 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
25128 if (!N0.hasOneUse() || !N1.hasOneUse())
25131 SDValue ShAmt0 = N0.getOperand(1);
25132 if (ShAmt0.getValueType() != MVT::i8)
25134 SDValue ShAmt1 = N1.getOperand(1);
25135 if (ShAmt1.getValueType() != MVT::i8)
25137 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
25138 ShAmt0 = ShAmt0.getOperand(0);
25139 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
25140 ShAmt1 = ShAmt1.getOperand(0);
25143 unsigned Opc = X86ISD::SHLD;
25144 SDValue Op0 = N0.getOperand(0);
25145 SDValue Op1 = N1.getOperand(0);
25146 if (ShAmt0.getOpcode() == ISD::SUB) {
25147 Opc = X86ISD::SHRD;
25148 std::swap(Op0, Op1);
25149 std::swap(ShAmt0, ShAmt1);
25152 unsigned Bits = VT.getSizeInBits();
25153 if (ShAmt1.getOpcode() == ISD::SUB) {
25154 SDValue Sum = ShAmt1.getOperand(0);
25155 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
25156 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
25157 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
25158 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
25159 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
25160 return DAG.getNode(Opc, DL, VT,
25162 DAG.getNode(ISD::TRUNCATE, DL,
25165 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
25166 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
25168 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
25169 return DAG.getNode(Opc, DL, VT,
25170 N0.getOperand(0), N1.getOperand(0),
25171 DAG.getNode(ISD::TRUNCATE, DL,
25178 // Generate NEG and CMOV for integer abs.
25179 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
25180 EVT VT = N->getValueType(0);
25182 // Since X86 does not have CMOV for 8-bit integer, we don't convert
25183 // 8-bit integer abs to NEG and CMOV.
25184 if (VT.isInteger() && VT.getSizeInBits() == 8)
25187 SDValue N0 = N->getOperand(0);
25188 SDValue N1 = N->getOperand(1);
25191 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
25192 // and change it to SUB and CMOV.
25193 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
25194 N0.getOpcode() == ISD::ADD &&
25195 N0.getOperand(1) == N1 &&
25196 N1.getOpcode() == ISD::SRA &&
25197 N1.getOperand(0) == N0.getOperand(0))
25198 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
25199 if (Y1C->getAPIntValue() == VT.getSizeInBits()-1) {
25200 // Generate SUB & CMOV.
25201 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32),
25202 DAG.getConstant(0, DL, VT), N0.getOperand(0));
25204 SDValue Ops[] = { N0.getOperand(0), Neg,
25205 DAG.getConstant(X86::COND_GE, DL, MVT::i8),
25206 SDValue(Neg.getNode(), 1) };
25207 return DAG.getNode(X86ISD::CMOV, DL, DAG.getVTList(VT, MVT::Glue), Ops);
25212 // Try to turn tests against the signbit in the form of:
25213 // XOR(TRUNCATE(SRL(X, size(X)-1)), 1)
25216 static SDValue foldXorTruncShiftIntoCmp(SDNode *N, SelectionDAG &DAG) {
25217 // This is only worth doing if the output type is i8.
25218 if (N->getValueType(0) != MVT::i8)
25221 SDValue N0 = N->getOperand(0);
25222 SDValue N1 = N->getOperand(1);
25224 // We should be performing an xor against a truncated shift.
25225 if (N0.getOpcode() != ISD::TRUNCATE || !N0.hasOneUse())
25228 // Make sure we are performing an xor against one.
25229 if (!isa<ConstantSDNode>(N1) || !cast<ConstantSDNode>(N1)->isOne())
25232 // SetCC on x86 zero extends so only act on this if it's a logical shift.
25233 SDValue Shift = N0.getOperand(0);
25234 if (Shift.getOpcode() != ISD::SRL || !Shift.hasOneUse())
25237 // Make sure we are truncating from one of i16, i32 or i64.
25238 EVT ShiftTy = Shift.getValueType();
25239 if (ShiftTy != MVT::i16 && ShiftTy != MVT::i32 && ShiftTy != MVT::i64)
25242 // Make sure the shift amount extracts the sign bit.
25243 if (!isa<ConstantSDNode>(Shift.getOperand(1)) ||
25244 Shift.getConstantOperandVal(1) != ShiftTy.getSizeInBits() - 1)
25247 // Create a greater-than comparison against -1.
25248 // N.B. Using SETGE against 0 works but we want a canonical looking
25249 // comparison, using SETGT matches up with what TranslateX86CC.
25251 SDValue ShiftOp = Shift.getOperand(0);
25252 EVT ShiftOpTy = ShiftOp.getValueType();
25253 SDValue Cond = DAG.getSetCC(DL, MVT::i8, ShiftOp,
25254 DAG.getConstant(-1, DL, ShiftOpTy), ISD::SETGT);
25258 static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
25259 TargetLowering::DAGCombinerInfo &DCI,
25260 const X86Subtarget *Subtarget) {
25261 if (DCI.isBeforeLegalizeOps())
25264 if (SDValue RV = foldXorTruncShiftIntoCmp(N, DAG))
25267 if (Subtarget->hasCMov())
25268 if (SDValue RV = performIntegerAbsCombine(N, DAG))
25271 if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
25277 /// PerformLOADCombine - Do target-specific dag combines on LOAD nodes.
25278 static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
25279 TargetLowering::DAGCombinerInfo &DCI,
25280 const X86Subtarget *Subtarget) {
25281 LoadSDNode *Ld = cast<LoadSDNode>(N);
25282 EVT RegVT = Ld->getValueType(0);
25283 EVT MemVT = Ld->getMemoryVT();
25285 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25287 // For chips with slow 32-byte unaligned loads, break the 32-byte operation
25288 // into two 16-byte operations.
25289 ISD::LoadExtType Ext = Ld->getExtensionType();
25291 unsigned AddressSpace = Ld->getAddressSpace();
25292 unsigned Alignment = Ld->getAlignment();
25293 if (RegVT.is256BitVector() && !DCI.isBeforeLegalizeOps() &&
25294 Ext == ISD::NON_EXTLOAD &&
25295 TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), RegVT,
25296 AddressSpace, Alignment, &Fast) && !Fast) {
25297 unsigned NumElems = RegVT.getVectorNumElements();
25301 SDValue Ptr = Ld->getBasePtr();
25302 SDValue Increment =
25303 DAG.getConstant(16, dl, TLI.getPointerTy(DAG.getDataLayout()));
25305 EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
25307 SDValue Load1 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
25308 Ld->getPointerInfo(), Ld->isVolatile(),
25309 Ld->isNonTemporal(), Ld->isInvariant(),
25311 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
25312 SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
25313 Ld->getPointerInfo(), Ld->isVolatile(),
25314 Ld->isNonTemporal(), Ld->isInvariant(),
25315 std::min(16U, Alignment));
25316 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
25318 Load2.getValue(1));
25320 SDValue NewVec = DAG.getUNDEF(RegVT);
25321 NewVec = Insert128BitVector(NewVec, Load1, 0, DAG, dl);
25322 NewVec = Insert128BitVector(NewVec, Load2, NumElems/2, DAG, dl);
25323 return DCI.CombineTo(N, NewVec, TF, true);
25329 /// PerformMLOADCombine - Resolve extending loads
25330 static SDValue PerformMLOADCombine(SDNode *N, SelectionDAG &DAG,
25331 TargetLowering::DAGCombinerInfo &DCI,
25332 const X86Subtarget *Subtarget) {
25333 MaskedLoadSDNode *Mld = cast<MaskedLoadSDNode>(N);
25334 if (Mld->getExtensionType() != ISD::SEXTLOAD)
25337 EVT VT = Mld->getValueType(0);
25338 unsigned NumElems = VT.getVectorNumElements();
25339 EVT LdVT = Mld->getMemoryVT();
25342 assert(LdVT != VT && "Cannot extend to the same type");
25343 unsigned ToSz = VT.getVectorElementType().getSizeInBits();
25344 unsigned FromSz = LdVT.getVectorElementType().getSizeInBits();
25345 // From, To sizes and ElemCount must be pow of two
25346 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
25347 "Unexpected size for extending masked load");
25349 unsigned SizeRatio = ToSz / FromSz;
25350 assert(SizeRatio * NumElems * FromSz == VT.getSizeInBits());
25352 // Create a type on which we perform the shuffle
25353 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
25354 LdVT.getScalarType(), NumElems*SizeRatio);
25355 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
25357 // Convert Src0 value
25358 SDValue WideSrc0 = DAG.getBitcast(WideVecVT, Mld->getSrc0());
25359 if (Mld->getSrc0().getOpcode() != ISD::UNDEF) {
25360 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
25361 for (unsigned i = 0; i != NumElems; ++i)
25362 ShuffleVec[i] = i * SizeRatio;
25364 // Can't shuffle using an illegal type.
25365 assert(DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) &&
25366 "WideVecVT should be legal");
25367 WideSrc0 = DAG.getVectorShuffle(WideVecVT, dl, WideSrc0,
25368 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
25370 // Prepare the new mask
25372 SDValue Mask = Mld->getMask();
25373 if (Mask.getValueType() == VT) {
25374 // Mask and original value have the same type
25375 NewMask = DAG.getBitcast(WideVecVT, Mask);
25376 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
25377 for (unsigned i = 0; i != NumElems; ++i)
25378 ShuffleVec[i] = i * SizeRatio;
25379 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
25380 ShuffleVec[i] = NumElems*SizeRatio;
25381 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
25382 DAG.getConstant(0, dl, WideVecVT),
25386 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
25387 unsigned WidenNumElts = NumElems*SizeRatio;
25388 unsigned MaskNumElts = VT.getVectorNumElements();
25389 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
25392 unsigned NumConcat = WidenNumElts / MaskNumElts;
25393 SmallVector<SDValue, 16> Ops(NumConcat);
25394 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
25396 for (unsigned i = 1; i != NumConcat; ++i)
25399 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
25402 SDValue WideLd = DAG.getMaskedLoad(WideVecVT, dl, Mld->getChain(),
25403 Mld->getBasePtr(), NewMask, WideSrc0,
25404 Mld->getMemoryVT(), Mld->getMemOperand(),
25406 SDValue NewVec = DAG.getNode(X86ISD::VSEXT, dl, VT, WideLd);
25407 return DCI.CombineTo(N, NewVec, WideLd.getValue(1), true);
25409 /// PerformMSTORECombine - Resolve truncating stores
25410 static SDValue PerformMSTORECombine(SDNode *N, SelectionDAG &DAG,
25411 const X86Subtarget *Subtarget) {
25412 MaskedStoreSDNode *Mst = cast<MaskedStoreSDNode>(N);
25413 if (!Mst->isTruncatingStore())
25416 EVT VT = Mst->getValue().getValueType();
25417 unsigned NumElems = VT.getVectorNumElements();
25418 EVT StVT = Mst->getMemoryVT();
25421 assert(StVT != VT && "Cannot truncate to the same type");
25422 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
25423 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
25425 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25427 // The truncating store is legal in some cases. For example
25428 // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
25429 // are designated for truncate store.
25430 // In this case we don't need any further transformations.
25431 if (TLI.isTruncStoreLegal(VT, StVT))
25434 // From, To sizes and ElemCount must be pow of two
25435 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
25436 "Unexpected size for truncating masked store");
25437 // We are going to use the original vector elt for storing.
25438 // Accumulated smaller vector elements must be a multiple of the store size.
25439 assert (((NumElems * FromSz) % ToSz) == 0 &&
25440 "Unexpected ratio for truncating masked store");
25442 unsigned SizeRatio = FromSz / ToSz;
25443 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
25445 // Create a type on which we perform the shuffle
25446 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
25447 StVT.getScalarType(), NumElems*SizeRatio);
25449 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
25451 SDValue WideVec = DAG.getBitcast(WideVecVT, Mst->getValue());
25452 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
25453 for (unsigned i = 0; i != NumElems; ++i)
25454 ShuffleVec[i] = i * SizeRatio;
25456 // Can't shuffle using an illegal type.
25457 assert(DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) &&
25458 "WideVecVT should be legal");
25460 SDValue TruncatedVal = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
25461 DAG.getUNDEF(WideVecVT),
25465 SDValue Mask = Mst->getMask();
25466 if (Mask.getValueType() == VT) {
25467 // Mask and original value have the same type
25468 NewMask = DAG.getBitcast(WideVecVT, Mask);
25469 for (unsigned i = 0; i != NumElems; ++i)
25470 ShuffleVec[i] = i * SizeRatio;
25471 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
25472 ShuffleVec[i] = NumElems*SizeRatio;
25473 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
25474 DAG.getConstant(0, dl, WideVecVT),
25478 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
25479 unsigned WidenNumElts = NumElems*SizeRatio;
25480 unsigned MaskNumElts = VT.getVectorNumElements();
25481 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
25484 unsigned NumConcat = WidenNumElts / MaskNumElts;
25485 SmallVector<SDValue, 16> Ops(NumConcat);
25486 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
25488 for (unsigned i = 1; i != NumConcat; ++i)
25491 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
25494 return DAG.getMaskedStore(Mst->getChain(), dl, TruncatedVal, Mst->getBasePtr(),
25495 NewMask, StVT, Mst->getMemOperand(), false);
25497 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
25498 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
25499 const X86Subtarget *Subtarget) {
25500 StoreSDNode *St = cast<StoreSDNode>(N);
25501 EVT VT = St->getValue().getValueType();
25502 EVT StVT = St->getMemoryVT();
25504 SDValue StoredVal = St->getOperand(1);
25505 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25507 // If we are saving a concatenation of two XMM registers and 32-byte stores
25508 // are slow, such as on Sandy Bridge, perform two 16-byte stores.
25510 unsigned AddressSpace = St->getAddressSpace();
25511 unsigned Alignment = St->getAlignment();
25512 if (VT.is256BitVector() && StVT == VT &&
25513 TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), VT,
25514 AddressSpace, Alignment, &Fast) && !Fast) {
25515 unsigned NumElems = VT.getVectorNumElements();
25519 SDValue Value0 = Extract128BitVector(StoredVal, 0, DAG, dl);
25520 SDValue Value1 = Extract128BitVector(StoredVal, NumElems/2, DAG, dl);
25523 DAG.getConstant(16, dl, TLI.getPointerTy(DAG.getDataLayout()));
25524 SDValue Ptr0 = St->getBasePtr();
25525 SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
25527 SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
25528 St->getPointerInfo(), St->isVolatile(),
25529 St->isNonTemporal(), Alignment);
25530 SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
25531 St->getPointerInfo(), St->isVolatile(),
25532 St->isNonTemporal(),
25533 std::min(16U, Alignment));
25534 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
25537 // Optimize trunc store (of multiple scalars) to shuffle and store.
25538 // First, pack all of the elements in one place. Next, store to memory
25539 // in fewer chunks.
25540 if (St->isTruncatingStore() && VT.isVector()) {
25541 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25542 unsigned NumElems = VT.getVectorNumElements();
25543 assert(StVT != VT && "Cannot truncate to the same type");
25544 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
25545 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
25547 // The truncating store is legal in some cases. For example
25548 // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
25549 // are designated for truncate store.
25550 // In this case we don't need any further transformations.
25551 if (TLI.isTruncStoreLegal(VT, StVT))
25554 // From, To sizes and ElemCount must be pow of two
25555 if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
25556 // We are going to use the original vector elt for storing.
25557 // Accumulated smaller vector elements must be a multiple of the store size.
25558 if (0 != (NumElems * FromSz) % ToSz) return SDValue();
25560 unsigned SizeRatio = FromSz / ToSz;
25562 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
25564 // Create a type on which we perform the shuffle
25565 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
25566 StVT.getScalarType(), NumElems*SizeRatio);
25568 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
25570 SDValue WideVec = DAG.getBitcast(WideVecVT, St->getValue());
25571 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
25572 for (unsigned i = 0; i != NumElems; ++i)
25573 ShuffleVec[i] = i * SizeRatio;
25575 // Can't shuffle using an illegal type.
25576 if (!TLI.isTypeLegal(WideVecVT))
25579 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
25580 DAG.getUNDEF(WideVecVT),
25582 // At this point all of the data is stored at the bottom of the
25583 // register. We now need to save it to mem.
25585 // Find the largest store unit
25586 MVT StoreType = MVT::i8;
25587 for (MVT Tp : MVT::integer_valuetypes()) {
25588 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToSz)
25592 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
25593 if (TLI.isTypeLegal(MVT::f64) && StoreType.getSizeInBits() < 64 &&
25594 (64 <= NumElems * ToSz))
25595 StoreType = MVT::f64;
25597 // Bitcast the original vector into a vector of store-size units
25598 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
25599 StoreType, VT.getSizeInBits()/StoreType.getSizeInBits());
25600 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
25601 SDValue ShuffWide = DAG.getBitcast(StoreVecVT, Shuff);
25602 SmallVector<SDValue, 8> Chains;
25603 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits() / 8, dl,
25604 TLI.getPointerTy(DAG.getDataLayout()));
25605 SDValue Ptr = St->getBasePtr();
25607 // Perform one or more big stores into memory.
25608 for (unsigned i=0, e=(ToSz*NumElems)/StoreType.getSizeInBits(); i!=e; ++i) {
25609 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
25610 StoreType, ShuffWide,
25611 DAG.getIntPtrConstant(i, dl));
25612 SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr,
25613 St->getPointerInfo(), St->isVolatile(),
25614 St->isNonTemporal(), St->getAlignment());
25615 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
25616 Chains.push_back(Ch);
25619 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
25622 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
25623 // the FP state in cases where an emms may be missing.
25624 // A preferable solution to the general problem is to figure out the right
25625 // places to insert EMMS. This qualifies as a quick hack.
25627 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
25628 if (VT.getSizeInBits() != 64)
25631 const Function *F = DAG.getMachineFunction().getFunction();
25632 bool NoImplicitFloatOps = F->hasFnAttribute(Attribute::NoImplicitFloat);
25634 !Subtarget->useSoftFloat() && !NoImplicitFloatOps && Subtarget->hasSSE2();
25635 if ((VT.isVector() ||
25636 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
25637 isa<LoadSDNode>(St->getValue()) &&
25638 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
25639 St->getChain().hasOneUse() && !St->isVolatile()) {
25640 SDNode* LdVal = St->getValue().getNode();
25641 LoadSDNode *Ld = nullptr;
25642 int TokenFactorIndex = -1;
25643 SmallVector<SDValue, 8> Ops;
25644 SDNode* ChainVal = St->getChain().getNode();
25645 // Must be a store of a load. We currently handle two cases: the load
25646 // is a direct child, and it's under an intervening TokenFactor. It is
25647 // possible to dig deeper under nested TokenFactors.
25648 if (ChainVal == LdVal)
25649 Ld = cast<LoadSDNode>(St->getChain());
25650 else if (St->getValue().hasOneUse() &&
25651 ChainVal->getOpcode() == ISD::TokenFactor) {
25652 for (unsigned i = 0, e = ChainVal->getNumOperands(); i != e; ++i) {
25653 if (ChainVal->getOperand(i).getNode() == LdVal) {
25654 TokenFactorIndex = i;
25655 Ld = cast<LoadSDNode>(St->getValue());
25657 Ops.push_back(ChainVal->getOperand(i));
25661 if (!Ld || !ISD::isNormalLoad(Ld))
25664 // If this is not the MMX case, i.e. we are just turning i64 load/store
25665 // into f64 load/store, avoid the transformation if there are multiple
25666 // uses of the loaded value.
25667 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
25672 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
25673 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
25675 if (Subtarget->is64Bit() || F64IsLegal) {
25676 MVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
25677 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
25678 Ld->getPointerInfo(), Ld->isVolatile(),
25679 Ld->isNonTemporal(), Ld->isInvariant(),
25680 Ld->getAlignment());
25681 SDValue NewChain = NewLd.getValue(1);
25682 if (TokenFactorIndex != -1) {
25683 Ops.push_back(NewChain);
25684 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
25686 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
25687 St->getPointerInfo(),
25688 St->isVolatile(), St->isNonTemporal(),
25689 St->getAlignment());
25692 // Otherwise, lower to two pairs of 32-bit loads / stores.
25693 SDValue LoAddr = Ld->getBasePtr();
25694 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
25695 DAG.getConstant(4, LdDL, MVT::i32));
25697 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
25698 Ld->getPointerInfo(),
25699 Ld->isVolatile(), Ld->isNonTemporal(),
25700 Ld->isInvariant(), Ld->getAlignment());
25701 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
25702 Ld->getPointerInfo().getWithOffset(4),
25703 Ld->isVolatile(), Ld->isNonTemporal(),
25705 MinAlign(Ld->getAlignment(), 4));
25707 SDValue NewChain = LoLd.getValue(1);
25708 if (TokenFactorIndex != -1) {
25709 Ops.push_back(LoLd);
25710 Ops.push_back(HiLd);
25711 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
25714 LoAddr = St->getBasePtr();
25715 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
25716 DAG.getConstant(4, StDL, MVT::i32));
25718 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
25719 St->getPointerInfo(),
25720 St->isVolatile(), St->isNonTemporal(),
25721 St->getAlignment());
25722 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
25723 St->getPointerInfo().getWithOffset(4),
25725 St->isNonTemporal(),
25726 MinAlign(St->getAlignment(), 4));
25727 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
25730 // This is similar to the above case, but here we handle a scalar 64-bit
25731 // integer store that is extracted from a vector on a 32-bit target.
25732 // If we have SSE2, then we can treat it like a floating-point double
25733 // to get past legalization. The execution dependencies fixup pass will
25734 // choose the optimal machine instruction for the store if this really is
25735 // an integer or v2f32 rather than an f64.
25736 if (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit() &&
25737 St->getOperand(1).getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
25738 SDValue OldExtract = St->getOperand(1);
25739 SDValue ExtOp0 = OldExtract.getOperand(0);
25740 unsigned VecSize = ExtOp0.getValueSizeInBits();
25741 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, VecSize / 64);
25742 SDValue BitCast = DAG.getBitcast(VecVT, ExtOp0);
25743 SDValue NewExtract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
25744 BitCast, OldExtract.getOperand(1));
25745 return DAG.getStore(St->getChain(), dl, NewExtract, St->getBasePtr(),
25746 St->getPointerInfo(), St->isVolatile(),
25747 St->isNonTemporal(), St->getAlignment());
25753 /// Return 'true' if this vector operation is "horizontal"
25754 /// and return the operands for the horizontal operation in LHS and RHS. A
25755 /// horizontal operation performs the binary operation on successive elements
25756 /// of its first operand, then on successive elements of its second operand,
25757 /// returning the resulting values in a vector. For example, if
25758 /// A = < float a0, float a1, float a2, float a3 >
25760 /// B = < float b0, float b1, float b2, float b3 >
25761 /// then the result of doing a horizontal operation on A and B is
25762 /// A horizontal-op B = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >.
25763 /// In short, LHS and RHS are inspected to see if LHS op RHS is of the form
25764 /// A horizontal-op B, for some already available A and B, and if so then LHS is
25765 /// set to A, RHS to B, and the routine returns 'true'.
25766 /// Note that the binary operation should have the property that if one of the
25767 /// operands is UNDEF then the result is UNDEF.
25768 static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool IsCommutative) {
25769 // Look for the following pattern: if
25770 // A = < float a0, float a1, float a2, float a3 >
25771 // B = < float b0, float b1, float b2, float b3 >
25773 // LHS = VECTOR_SHUFFLE A, B, <0, 2, 4, 6>
25774 // RHS = VECTOR_SHUFFLE A, B, <1, 3, 5, 7>
25775 // then LHS op RHS = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >
25776 // which is A horizontal-op B.
25778 // At least one of the operands should be a vector shuffle.
25779 if (LHS.getOpcode() != ISD::VECTOR_SHUFFLE &&
25780 RHS.getOpcode() != ISD::VECTOR_SHUFFLE)
25783 MVT VT = LHS.getSimpleValueType();
25785 assert((VT.is128BitVector() || VT.is256BitVector()) &&
25786 "Unsupported vector type for horizontal add/sub");
25788 // Handle 128 and 256-bit vector lengths. AVX defines horizontal add/sub to
25789 // operate independently on 128-bit lanes.
25790 unsigned NumElts = VT.getVectorNumElements();
25791 unsigned NumLanes = VT.getSizeInBits()/128;
25792 unsigned NumLaneElts = NumElts / NumLanes;
25793 assert((NumLaneElts % 2 == 0) &&
25794 "Vector type should have an even number of elements in each lane");
25795 unsigned HalfLaneElts = NumLaneElts/2;
25797 // View LHS in the form
25798 // LHS = VECTOR_SHUFFLE A, B, LMask
25799 // If LHS is not a shuffle then pretend it is the shuffle
25800 // LHS = VECTOR_SHUFFLE LHS, undef, <0, 1, ..., N-1>
25801 // NOTE: in what follows a default initialized SDValue represents an UNDEF of
25804 SmallVector<int, 16> LMask(NumElts);
25805 if (LHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
25806 if (LHS.getOperand(0).getOpcode() != ISD::UNDEF)
25807 A = LHS.getOperand(0);
25808 if (LHS.getOperand(1).getOpcode() != ISD::UNDEF)
25809 B = LHS.getOperand(1);
25810 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(LHS.getNode())->getMask();
25811 std::copy(Mask.begin(), Mask.end(), LMask.begin());
25813 if (LHS.getOpcode() != ISD::UNDEF)
25815 for (unsigned i = 0; i != NumElts; ++i)
25819 // Likewise, view RHS in the form
25820 // RHS = VECTOR_SHUFFLE C, D, RMask
25822 SmallVector<int, 16> RMask(NumElts);
25823 if (RHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
25824 if (RHS.getOperand(0).getOpcode() != ISD::UNDEF)
25825 C = RHS.getOperand(0);
25826 if (RHS.getOperand(1).getOpcode() != ISD::UNDEF)
25827 D = RHS.getOperand(1);
25828 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(RHS.getNode())->getMask();
25829 std::copy(Mask.begin(), Mask.end(), RMask.begin());
25831 if (RHS.getOpcode() != ISD::UNDEF)
25833 for (unsigned i = 0; i != NumElts; ++i)
25837 // Check that the shuffles are both shuffling the same vectors.
25838 if (!(A == C && B == D) && !(A == D && B == C))
25841 // If everything is UNDEF then bail out: it would be better to fold to UNDEF.
25842 if (!A.getNode() && !B.getNode())
25845 // If A and B occur in reverse order in RHS, then "swap" them (which means
25846 // rewriting the mask).
25848 ShuffleVectorSDNode::commuteMask(RMask);
25850 // At this point LHS and RHS are equivalent to
25851 // LHS = VECTOR_SHUFFLE A, B, LMask
25852 // RHS = VECTOR_SHUFFLE A, B, RMask
25853 // Check that the masks correspond to performing a horizontal operation.
25854 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
25855 for (unsigned i = 0; i != NumLaneElts; ++i) {
25856 int LIdx = LMask[i+l], RIdx = RMask[i+l];
25858 // Ignore any UNDEF components.
25859 if (LIdx < 0 || RIdx < 0 ||
25860 (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) ||
25861 (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts)))
25864 // Check that successive elements are being operated on. If not, this is
25865 // not a horizontal operation.
25866 unsigned Src = (i/HalfLaneElts); // each lane is split between srcs
25867 int Index = 2*(i%HalfLaneElts) + NumElts*Src + l;
25868 if (!(LIdx == Index && RIdx == Index + 1) &&
25869 !(IsCommutative && LIdx == Index + 1 && RIdx == Index))
25874 LHS = A.getNode() ? A : B; // If A is 'UNDEF', use B for it.
25875 RHS = B.getNode() ? B : A; // If B is 'UNDEF', use A for it.
25879 /// Do target-specific dag combines on floating point adds.
25880 static SDValue PerformFADDCombine(SDNode *N, SelectionDAG &DAG,
25881 const X86Subtarget *Subtarget) {
25882 EVT VT = N->getValueType(0);
25883 SDValue LHS = N->getOperand(0);
25884 SDValue RHS = N->getOperand(1);
25886 // Try to synthesize horizontal adds from adds of shuffles.
25887 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
25888 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
25889 isHorizontalBinOp(LHS, RHS, true))
25890 return DAG.getNode(X86ISD::FHADD, SDLoc(N), VT, LHS, RHS);
25894 /// Do target-specific dag combines on floating point subs.
25895 static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG,
25896 const X86Subtarget *Subtarget) {
25897 EVT VT = N->getValueType(0);
25898 SDValue LHS = N->getOperand(0);
25899 SDValue RHS = N->getOperand(1);
25901 // Try to synthesize horizontal subs from subs of shuffles.
25902 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
25903 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
25904 isHorizontalBinOp(LHS, RHS, false))
25905 return DAG.getNode(X86ISD::FHSUB, SDLoc(N), VT, LHS, RHS);
25909 /// Do target-specific dag combines on X86ISD::FOR and X86ISD::FXOR nodes.
25910 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG,
25911 const X86Subtarget *Subtarget) {
25912 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
25914 // F[X]OR(0.0, x) -> x
25915 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25916 if (C->getValueAPF().isPosZero())
25917 return N->getOperand(1);
25919 // F[X]OR(x, 0.0) -> x
25920 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25921 if (C->getValueAPF().isPosZero())
25922 return N->getOperand(0);
25924 EVT VT = N->getValueType(0);
25925 if (VT.is512BitVector() && !Subtarget->hasDQI()) {
25927 MVT IntScalar = MVT::getIntegerVT(VT.getScalarSizeInBits());
25928 MVT IntVT = MVT::getVectorVT(IntScalar, VT.getVectorNumElements());
25930 SDValue Op0 = DAG.getNode(ISD::BITCAST, dl, IntVT, N->getOperand(0));
25931 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, IntVT, N->getOperand(1));
25932 unsigned IntOpcode = (N->getOpcode() == X86ISD::FOR) ? ISD::OR : ISD::XOR;
25933 SDValue IntOp = DAG.getNode(IntOpcode, dl, IntVT, Op0, Op1);
25934 return DAG.getNode(ISD::BITCAST, dl, VT, IntOp);
25939 /// Do target-specific dag combines on X86ISD::FMIN and X86ISD::FMAX nodes.
25940 static SDValue PerformFMinFMaxCombine(SDNode *N, SelectionDAG &DAG) {
25941 assert(N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD::FMAX);
25943 // Only perform optimizations if UnsafeMath is used.
25944 if (!DAG.getTarget().Options.UnsafeFPMath)
25947 // If we run in unsafe-math mode, then convert the FMAX and FMIN nodes
25948 // into FMINC and FMAXC, which are Commutative operations.
25949 unsigned NewOp = 0;
25950 switch (N->getOpcode()) {
25951 default: llvm_unreachable("unknown opcode");
25952 case X86ISD::FMIN: NewOp = X86ISD::FMINC; break;
25953 case X86ISD::FMAX: NewOp = X86ISD::FMAXC; break;
25956 return DAG.getNode(NewOp, SDLoc(N), N->getValueType(0),
25957 N->getOperand(0), N->getOperand(1));
25960 /// Do target-specific dag combines on X86ISD::FAND nodes.
25961 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
25962 // FAND(0.0, x) -> 0.0
25963 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25964 if (C->getValueAPF().isPosZero())
25965 return N->getOperand(0);
25967 // FAND(x, 0.0) -> 0.0
25968 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25969 if (C->getValueAPF().isPosZero())
25970 return N->getOperand(1);
25975 /// Do target-specific dag combines on X86ISD::FANDN nodes
25976 static SDValue PerformFANDNCombine(SDNode *N, SelectionDAG &DAG) {
25977 // FANDN(0.0, x) -> x
25978 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25979 if (C->getValueAPF().isPosZero())
25980 return N->getOperand(1);
25982 // FANDN(x, 0.0) -> 0.0
25983 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25984 if (C->getValueAPF().isPosZero())
25985 return N->getOperand(1);
25990 static SDValue PerformBTCombine(SDNode *N,
25992 TargetLowering::DAGCombinerInfo &DCI) {
25993 // BT ignores high bits in the bit index operand.
25994 SDValue Op1 = N->getOperand(1);
25995 if (Op1.hasOneUse()) {
25996 unsigned BitWidth = Op1.getValueSizeInBits();
25997 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
25998 APInt KnownZero, KnownOne;
25999 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
26000 !DCI.isBeforeLegalizeOps());
26001 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
26002 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
26003 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
26004 DCI.CommitTargetLoweringOpt(TLO);
26009 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
26010 SDValue Op = N->getOperand(0);
26011 if (Op.getOpcode() == ISD::BITCAST)
26012 Op = Op.getOperand(0);
26013 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
26014 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
26015 VT.getVectorElementType().getSizeInBits() ==
26016 OpVT.getVectorElementType().getSizeInBits()) {
26017 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
26022 static SDValue PerformSIGN_EXTEND_INREGCombine(SDNode *N, SelectionDAG &DAG,
26023 const X86Subtarget *Subtarget) {
26024 EVT VT = N->getValueType(0);
26025 if (!VT.isVector())
26028 SDValue N0 = N->getOperand(0);
26029 SDValue N1 = N->getOperand(1);
26030 EVT ExtraVT = cast<VTSDNode>(N1)->getVT();
26033 // The SIGN_EXTEND_INREG to v4i64 is expensive operation on the
26034 // both SSE and AVX2 since there is no sign-extended shift right
26035 // operation on a vector with 64-bit elements.
26036 //(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) ->
26037 // (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT)))
26038 if (VT == MVT::v4i64 && (N0.getOpcode() == ISD::ANY_EXTEND ||
26039 N0.getOpcode() == ISD::SIGN_EXTEND)) {
26040 SDValue N00 = N0.getOperand(0);
26042 // EXTLOAD has a better solution on AVX2,
26043 // it may be replaced with X86ISD::VSEXT node.
26044 if (N00.getOpcode() == ISD::LOAD && Subtarget->hasInt256())
26045 if (!ISD::isNormalLoad(N00.getNode()))
26048 if (N00.getValueType() == MVT::v4i32 && ExtraVT.getSizeInBits() < 128) {
26049 SDValue Tmp = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32,
26051 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i64, Tmp);
26057 /// sext(add_nsw(x, C)) --> add(sext(x), C_sext)
26058 /// Promoting a sign extension ahead of an 'add nsw' exposes opportunities
26059 /// to combine math ops, use an LEA, or use a complex addressing mode. This can
26060 /// eliminate extend, add, and shift instructions.
26061 static SDValue promoteSextBeforeAddNSW(SDNode *Sext, SelectionDAG &DAG,
26062 const X86Subtarget *Subtarget) {
26063 // TODO: This should be valid for other integer types.
26064 EVT VT = Sext->getValueType(0);
26065 if (VT != MVT::i64)
26068 // We need an 'add nsw' feeding into the 'sext'.
26069 SDValue Add = Sext->getOperand(0);
26070 if (Add.getOpcode() != ISD::ADD || !Add->getFlags()->hasNoSignedWrap())
26073 // Having a constant operand to the 'add' ensures that we are not increasing
26074 // the instruction count because the constant is extended for free below.
26075 // A constant operand can also become the displacement field of an LEA.
26076 auto *AddOp1 = dyn_cast<ConstantSDNode>(Add.getOperand(1));
26080 // Don't make the 'add' bigger if there's no hope of combining it with some
26081 // other 'add' or 'shl' instruction.
26082 // TODO: It may be profitable to generate simpler LEA instructions in place
26083 // of single 'add' instructions, but the cost model for selecting an LEA
26084 // currently has a high threshold.
26085 bool HasLEAPotential = false;
26086 for (auto *User : Sext->uses()) {
26087 if (User->getOpcode() == ISD::ADD || User->getOpcode() == ISD::SHL) {
26088 HasLEAPotential = true;
26092 if (!HasLEAPotential)
26095 // Everything looks good, so pull the 'sext' ahead of the 'add'.
26096 int64_t AddConstant = AddOp1->getSExtValue();
26097 SDValue AddOp0 = Add.getOperand(0);
26098 SDValue NewSext = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(Sext), VT, AddOp0);
26099 SDValue NewConstant = DAG.getConstant(AddConstant, SDLoc(Add), VT);
26101 // The wider add is guaranteed to not wrap because both operands are
26104 Flags.setNoSignedWrap(true);
26105 return DAG.getNode(ISD::ADD, SDLoc(Add), VT, NewSext, NewConstant, &Flags);
26108 static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
26109 TargetLowering::DAGCombinerInfo &DCI,
26110 const X86Subtarget *Subtarget) {
26111 SDValue N0 = N->getOperand(0);
26112 EVT VT = N->getValueType(0);
26113 EVT SVT = VT.getScalarType();
26114 EVT InVT = N0.getValueType();
26115 EVT InSVT = InVT.getScalarType();
26118 // (i8,i32 sext (sdivrem (i8 x, i8 y)) ->
26119 // (i8,i32 (sdivrem_sext_hreg (i8 x, i8 y)
26120 // This exposes the sext to the sdivrem lowering, so that it directly extends
26121 // from AH (which we otherwise need to do contortions to access).
26122 if (N0.getOpcode() == ISD::SDIVREM && N0.getResNo() == 1 &&
26123 InVT == MVT::i8 && VT == MVT::i32) {
26124 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
26125 SDValue R = DAG.getNode(X86ISD::SDIVREM8_SEXT_HREG, DL, NodeTys,
26126 N0.getOperand(0), N0.getOperand(1));
26127 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
26128 return R.getValue(1);
26131 if (!DCI.isBeforeLegalizeOps()) {
26132 if (InVT == MVT::i1) {
26133 SDValue Zero = DAG.getConstant(0, DL, VT);
26135 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), DL, VT);
26136 return DAG.getNode(ISD::SELECT, DL, VT, N0, AllOnes, Zero);
26141 if (VT.isVector() && Subtarget->hasSSE2()) {
26142 auto ExtendVecSize = [&DAG](SDLoc DL, SDValue N, unsigned Size) {
26143 EVT InVT = N.getValueType();
26144 EVT OutVT = EVT::getVectorVT(*DAG.getContext(), InVT.getScalarType(),
26145 Size / InVT.getScalarSizeInBits());
26146 SmallVector<SDValue, 8> Opnds(Size / InVT.getSizeInBits(),
26147 DAG.getUNDEF(InVT));
26149 return DAG.getNode(ISD::CONCAT_VECTORS, DL, OutVT, Opnds);
26152 // If target-size is less than 128-bits, extend to a type that would extend
26153 // to 128 bits, extend that and extract the original target vector.
26154 if (VT.getSizeInBits() < 128 && !(128 % VT.getSizeInBits()) &&
26155 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
26156 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
26157 unsigned Scale = 128 / VT.getSizeInBits();
26159 EVT::getVectorVT(*DAG.getContext(), SVT, 128 / SVT.getSizeInBits());
26160 SDValue Ex = ExtendVecSize(DL, N0, Scale * InVT.getSizeInBits());
26161 SDValue SExt = DAG.getNode(ISD::SIGN_EXTEND, DL, ExVT, Ex);
26162 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, SExt,
26163 DAG.getIntPtrConstant(0, DL));
26166 // If target-size is 128-bits, then convert to ISD::SIGN_EXTEND_VECTOR_INREG
26167 // which ensures lowering to X86ISD::VSEXT (pmovsx*).
26168 if (VT.getSizeInBits() == 128 &&
26169 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
26170 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
26171 SDValue ExOp = ExtendVecSize(DL, N0, 128);
26172 return DAG.getSignExtendVectorInReg(ExOp, DL, VT);
26175 // On pre-AVX2 targets, split into 128-bit nodes of
26176 // ISD::SIGN_EXTEND_VECTOR_INREG.
26177 if (!Subtarget->hasInt256() && !(VT.getSizeInBits() % 128) &&
26178 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
26179 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
26180 unsigned NumVecs = VT.getSizeInBits() / 128;
26181 unsigned NumSubElts = 128 / SVT.getSizeInBits();
26182 EVT SubVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumSubElts);
26183 EVT InSubVT = EVT::getVectorVT(*DAG.getContext(), InSVT, NumSubElts);
26185 SmallVector<SDValue, 8> Opnds;
26186 for (unsigned i = 0, Offset = 0; i != NumVecs;
26187 ++i, Offset += NumSubElts) {
26188 SDValue SrcVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InSubVT, N0,
26189 DAG.getIntPtrConstant(Offset, DL));
26190 SrcVec = ExtendVecSize(DL, SrcVec, 128);
26191 SrcVec = DAG.getSignExtendVectorInReg(SrcVec, DL, SubVT);
26192 Opnds.push_back(SrcVec);
26194 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Opnds);
26198 if (Subtarget->hasAVX() && VT.is256BitVector())
26199 if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
26202 if (SDValue NewAdd = promoteSextBeforeAddNSW(N, DAG, Subtarget))
26208 static SDValue PerformFMACombine(SDNode *N, SelectionDAG &DAG,
26209 const X86Subtarget* Subtarget) {
26211 EVT VT = N->getValueType(0);
26213 // Let legalize expand this if it isn't a legal type yet.
26214 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
26217 EVT ScalarVT = VT.getScalarType();
26218 if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) ||
26219 (!Subtarget->hasFMA() && !Subtarget->hasFMA4() &&
26220 !Subtarget->hasAVX512()))
26223 SDValue A = N->getOperand(0);
26224 SDValue B = N->getOperand(1);
26225 SDValue C = N->getOperand(2);
26227 bool NegA = (A.getOpcode() == ISD::FNEG);
26228 bool NegB = (B.getOpcode() == ISD::FNEG);
26229 bool NegC = (C.getOpcode() == ISD::FNEG);
26231 // Negative multiplication when NegA xor NegB
26232 bool NegMul = (NegA != NegB);
26234 A = A.getOperand(0);
26236 B = B.getOperand(0);
26238 C = C.getOperand(0);
26242 Opcode = (!NegC) ? X86ISD::FMADD : X86ISD::FMSUB;
26244 Opcode = (!NegC) ? X86ISD::FNMADD : X86ISD::FNMSUB;
26246 return DAG.getNode(Opcode, dl, VT, A, B, C);
26249 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG,
26250 TargetLowering::DAGCombinerInfo &DCI,
26251 const X86Subtarget *Subtarget) {
26252 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
26253 // (and (i32 x86isd::setcc_carry), 1)
26254 // This eliminates the zext. This transformation is necessary because
26255 // ISD::SETCC is always legalized to i8.
26257 SDValue N0 = N->getOperand(0);
26258 EVT VT = N->getValueType(0);
26260 if (N0.getOpcode() == ISD::AND &&
26262 N0.getOperand(0).hasOneUse()) {
26263 SDValue N00 = N0.getOperand(0);
26264 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
26265 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
26266 if (!C || C->getZExtValue() != 1)
26268 return DAG.getNode(ISD::AND, dl, VT,
26269 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
26270 N00.getOperand(0), N00.getOperand(1)),
26271 DAG.getConstant(1, dl, VT));
26275 if (N0.getOpcode() == ISD::TRUNCATE &&
26277 N0.getOperand(0).hasOneUse()) {
26278 SDValue N00 = N0.getOperand(0);
26279 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
26280 return DAG.getNode(ISD::AND, dl, VT,
26281 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
26282 N00.getOperand(0), N00.getOperand(1)),
26283 DAG.getConstant(1, dl, VT));
26287 if (VT.is256BitVector())
26288 if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
26291 // (i8,i32 zext (udivrem (i8 x, i8 y)) ->
26292 // (i8,i32 (udivrem_zext_hreg (i8 x, i8 y)
26293 // This exposes the zext to the udivrem lowering, so that it directly extends
26294 // from AH (which we otherwise need to do contortions to access).
26295 if (N0.getOpcode() == ISD::UDIVREM &&
26296 N0.getResNo() == 1 && N0.getValueType() == MVT::i8 &&
26297 (VT == MVT::i32 || VT == MVT::i64)) {
26298 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
26299 SDValue R = DAG.getNode(X86ISD::UDIVREM8_ZEXT_HREG, dl, NodeTys,
26300 N0.getOperand(0), N0.getOperand(1));
26301 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
26302 return R.getValue(1);
26308 // Optimize x == -y --> x+y == 0
26309 // x != -y --> x+y != 0
26310 static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG,
26311 const X86Subtarget* Subtarget) {
26312 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
26313 SDValue LHS = N->getOperand(0);
26314 SDValue RHS = N->getOperand(1);
26315 EVT VT = N->getValueType(0);
26318 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB)
26319 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(LHS.getOperand(0)))
26320 if (C->getAPIntValue() == 0 && LHS.hasOneUse()) {
26321 SDValue addV = DAG.getNode(ISD::ADD, DL, LHS.getValueType(), RHS,
26322 LHS.getOperand(1));
26323 return DAG.getSetCC(DL, N->getValueType(0), addV,
26324 DAG.getConstant(0, DL, addV.getValueType()), CC);
26326 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && RHS.getOpcode() == ISD::SUB)
26327 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS.getOperand(0)))
26328 if (C->getAPIntValue() == 0 && RHS.hasOneUse()) {
26329 SDValue addV = DAG.getNode(ISD::ADD, DL, RHS.getValueType(), LHS,
26330 RHS.getOperand(1));
26331 return DAG.getSetCC(DL, N->getValueType(0), addV,
26332 DAG.getConstant(0, DL, addV.getValueType()), CC);
26335 if (VT.getScalarType() == MVT::i1 &&
26336 (CC == ISD::SETNE || CC == ISD::SETEQ || ISD::isSignedIntSetCC(CC))) {
26338 (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
26339 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
26340 bool IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
26342 if (!IsSEXT0 || !IsVZero1) {
26343 // Swap the operands and update the condition code.
26344 std::swap(LHS, RHS);
26345 CC = ISD::getSetCCSwappedOperands(CC);
26347 IsSEXT0 = (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
26348 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
26349 IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
26352 if (IsSEXT0 && IsVZero1) {
26353 assert(VT == LHS.getOperand(0).getValueType() &&
26354 "Uexpected operand type");
26355 if (CC == ISD::SETGT)
26356 return DAG.getConstant(0, DL, VT);
26357 if (CC == ISD::SETLE)
26358 return DAG.getConstant(1, DL, VT);
26359 if (CC == ISD::SETEQ || CC == ISD::SETGE)
26360 return DAG.getNOT(DL, LHS.getOperand(0), VT);
26362 assert((CC == ISD::SETNE || CC == ISD::SETLT) &&
26363 "Unexpected condition code!");
26364 return LHS.getOperand(0);
26371 static SDValue PerformBLENDICombine(SDNode *N, SelectionDAG &DAG) {
26372 SDValue V0 = N->getOperand(0);
26373 SDValue V1 = N->getOperand(1);
26375 EVT VT = N->getValueType(0);
26377 // Canonicalize a v2f64 blend with a mask of 2 by swapping the vector
26378 // operands and changing the mask to 1. This saves us a bunch of
26379 // pattern-matching possibilities related to scalar math ops in SSE/AVX.
26380 // x86InstrInfo knows how to commute this back after instruction selection
26381 // if it would help register allocation.
26383 // TODO: If optimizing for size or a processor that doesn't suffer from
26384 // partial register update stalls, this should be transformed into a MOVSD
26385 // instruction because a MOVSD is 1-2 bytes smaller than a BLENDPD.
26387 if (VT == MVT::v2f64)
26388 if (auto *Mask = dyn_cast<ConstantSDNode>(N->getOperand(2)))
26389 if (Mask->getZExtValue() == 2 && !isShuffleFoldableLoad(V0)) {
26390 SDValue NewMask = DAG.getConstant(1, DL, MVT::i8);
26391 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V0, NewMask);
26397 // Helper function of PerformSETCCCombine. It is to materialize "setb reg"
26398 // as "sbb reg,reg", since it can be extended without zext and produces
26399 // an all-ones bit which is more useful than 0/1 in some cases.
26400 static SDValue MaterializeSETB(SDLoc DL, SDValue EFLAGS, SelectionDAG &DAG,
26403 return DAG.getNode(ISD::AND, DL, VT,
26404 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
26405 DAG.getConstant(X86::COND_B, DL, MVT::i8),
26407 DAG.getConstant(1, DL, VT));
26408 assert (VT == MVT::i1 && "Unexpected type for SECCC node");
26409 return DAG.getNode(ISD::TRUNCATE, DL, MVT::i1,
26410 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
26411 DAG.getConstant(X86::COND_B, DL, MVT::i8),
26415 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
26416 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG,
26417 TargetLowering::DAGCombinerInfo &DCI,
26418 const X86Subtarget *Subtarget) {
26420 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0));
26421 SDValue EFLAGS = N->getOperand(1);
26423 if (CC == X86::COND_A) {
26424 // Try to convert COND_A into COND_B in an attempt to facilitate
26425 // materializing "setb reg".
26427 // Do not flip "e > c", where "c" is a constant, because Cmp instruction
26428 // cannot take an immediate as its first operand.
26430 if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() &&
26431 EFLAGS.getValueType().isInteger() &&
26432 !isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
26433 SDValue NewSub = DAG.getNode(X86ISD::SUB, SDLoc(EFLAGS),
26434 EFLAGS.getNode()->getVTList(),
26435 EFLAGS.getOperand(1), EFLAGS.getOperand(0));
26436 SDValue NewEFLAGS = SDValue(NewSub.getNode(), EFLAGS.getResNo());
26437 return MaterializeSETB(DL, NewEFLAGS, DAG, N->getSimpleValueType(0));
26441 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
26442 // a zext and produces an all-ones bit which is more useful than 0/1 in some
26444 if (CC == X86::COND_B)
26445 return MaterializeSETB(DL, EFLAGS, DAG, N->getSimpleValueType(0));
26447 if (SDValue Flags = checkBoolTestSetCCCombine(EFLAGS, CC)) {
26448 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
26449 return DAG.getNode(X86ISD::SETCC, DL, N->getVTList(), Cond, Flags);
26455 // Optimize branch condition evaluation.
26457 static SDValue PerformBrCondCombine(SDNode *N, SelectionDAG &DAG,
26458 TargetLowering::DAGCombinerInfo &DCI,
26459 const X86Subtarget *Subtarget) {
26461 SDValue Chain = N->getOperand(0);
26462 SDValue Dest = N->getOperand(1);
26463 SDValue EFLAGS = N->getOperand(3);
26464 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(2));
26466 if (SDValue Flags = checkBoolTestSetCCCombine(EFLAGS, CC)) {
26467 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
26468 return DAG.getNode(X86ISD::BRCOND, DL, N->getVTList(), Chain, Dest, Cond,
26475 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
26476 SelectionDAG &DAG) {
26477 // Take advantage of vector comparisons producing 0 or -1 in each lane to
26478 // optimize away operation when it's from a constant.
26480 // The general transformation is:
26481 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
26482 // AND(VECTOR_CMP(x,y), constant2)
26483 // constant2 = UNARYOP(constant)
26485 // Early exit if this isn't a vector operation, the operand of the
26486 // unary operation isn't a bitwise AND, or if the sizes of the operations
26487 // aren't the same.
26488 EVT VT = N->getValueType(0);
26489 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
26490 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
26491 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
26494 // Now check that the other operand of the AND is a constant. We could
26495 // make the transformation for non-constant splats as well, but it's unclear
26496 // that would be a benefit as it would not eliminate any operations, just
26497 // perform one more step in scalar code before moving to the vector unit.
26498 if (BuildVectorSDNode *BV =
26499 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
26500 // Bail out if the vector isn't a constant.
26501 if (!BV->isConstant())
26504 // Everything checks out. Build up the new and improved node.
26506 EVT IntVT = BV->getValueType(0);
26507 // Create a new constant of the appropriate type for the transformed
26509 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
26510 // The AND node needs bitcasts to/from an integer vector type around it.
26511 SDValue MaskConst = DAG.getBitcast(IntVT, SourceConst);
26512 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
26513 N->getOperand(0)->getOperand(0), MaskConst);
26514 SDValue Res = DAG.getBitcast(VT, NewAnd);
26521 static SDValue PerformUINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
26522 const X86Subtarget *Subtarget) {
26523 SDValue Op0 = N->getOperand(0);
26524 EVT VT = N->getValueType(0);
26525 EVT InVT = Op0.getValueType();
26526 EVT InSVT = InVT.getScalarType();
26527 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
26529 // UINT_TO_FP(vXi8) -> SINT_TO_FP(ZEXT(vXi8 to vXi32))
26530 // UINT_TO_FP(vXi16) -> SINT_TO_FP(ZEXT(vXi16 to vXi32))
26531 if (InVT.isVector() && (InSVT == MVT::i8 || InSVT == MVT::i16)) {
26533 EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
26534 InVT.getVectorNumElements());
26535 SDValue P = DAG.getNode(ISD::ZERO_EXTEND, dl, DstVT, Op0);
26537 if (TLI.isOperationLegal(ISD::UINT_TO_FP, DstVT))
26538 return DAG.getNode(ISD::UINT_TO_FP, dl, VT, P);
26540 return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P);
26546 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
26547 const X86Subtarget *Subtarget) {
26548 // First try to optimize away the conversion entirely when it's
26549 // conditionally from a constant. Vectors only.
26550 if (SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG))
26553 // Now move on to more general possibilities.
26554 SDValue Op0 = N->getOperand(0);
26555 EVT VT = N->getValueType(0);
26556 EVT InVT = Op0.getValueType();
26557 EVT InSVT = InVT.getScalarType();
26559 // SINT_TO_FP(vXi8) -> SINT_TO_FP(SEXT(vXi8 to vXi32))
26560 // SINT_TO_FP(vXi16) -> SINT_TO_FP(SEXT(vXi16 to vXi32))
26561 if (InVT.isVector() && (InSVT == MVT::i8 || InSVT == MVT::i16)) {
26563 EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
26564 InVT.getVectorNumElements());
26565 SDValue P = DAG.getNode(ISD::SIGN_EXTEND, dl, DstVT, Op0);
26566 return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P);
26569 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
26570 // a 32-bit target where SSE doesn't support i64->FP operations.
26571 if (!Subtarget->useSoftFloat() && Op0.getOpcode() == ISD::LOAD) {
26572 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
26573 EVT LdVT = Ld->getValueType(0);
26575 // This transformation is not supported if the result type is f16
26576 if (VT == MVT::f16)
26579 if (!Ld->isVolatile() && !VT.isVector() &&
26580 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
26581 !Subtarget->is64Bit() && LdVT == MVT::i64) {
26582 SDValue FILDChain = Subtarget->getTargetLowering()->BuildFILD(
26583 SDValue(N, 0), LdVT, Ld->getChain(), Op0, DAG);
26584 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
26591 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
26592 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
26593 X86TargetLowering::DAGCombinerInfo &DCI) {
26594 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
26595 // the result is either zero or one (depending on the input carry bit).
26596 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
26597 if (X86::isZeroNode(N->getOperand(0)) &&
26598 X86::isZeroNode(N->getOperand(1)) &&
26599 // We don't have a good way to replace an EFLAGS use, so only do this when
26601 SDValue(N, 1).use_empty()) {
26603 EVT VT = N->getValueType(0);
26604 SDValue CarryOut = DAG.getConstant(0, DL, N->getValueType(1));
26605 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
26606 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
26607 DAG.getConstant(X86::COND_B, DL,
26610 DAG.getConstant(1, DL, VT));
26611 return DCI.CombineTo(N, Res1, CarryOut);
26617 // fold (add Y, (sete X, 0)) -> adc 0, Y
26618 // (add Y, (setne X, 0)) -> sbb -1, Y
26619 // (sub (sete X, 0), Y) -> sbb 0, Y
26620 // (sub (setne X, 0), Y) -> adc -1, Y
26621 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
26624 // Look through ZExts.
26625 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
26626 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
26629 SDValue SetCC = Ext.getOperand(0);
26630 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
26633 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
26634 if (CC != X86::COND_E && CC != X86::COND_NE)
26637 SDValue Cmp = SetCC.getOperand(1);
26638 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
26639 !X86::isZeroNode(Cmp.getOperand(1)) ||
26640 !Cmp.getOperand(0).getValueType().isInteger())
26643 SDValue CmpOp0 = Cmp.getOperand(0);
26644 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
26645 DAG.getConstant(1, DL, CmpOp0.getValueType()));
26647 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
26648 if (CC == X86::COND_NE)
26649 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
26650 DL, OtherVal.getValueType(), OtherVal,
26651 DAG.getConstant(-1ULL, DL, OtherVal.getValueType()),
26653 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
26654 DL, OtherVal.getValueType(), OtherVal,
26655 DAG.getConstant(0, DL, OtherVal.getValueType()), NewCmp);
26658 /// PerformADDCombine - Do target-specific dag combines on integer adds.
26659 static SDValue PerformAddCombine(SDNode *N, SelectionDAG &DAG,
26660 const X86Subtarget *Subtarget) {
26661 EVT VT = N->getValueType(0);
26662 SDValue Op0 = N->getOperand(0);
26663 SDValue Op1 = N->getOperand(1);
26665 // Try to synthesize horizontal adds from adds of shuffles.
26666 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
26667 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
26668 isHorizontalBinOp(Op0, Op1, true))
26669 return DAG.getNode(X86ISD::HADD, SDLoc(N), VT, Op0, Op1);
26671 return OptimizeConditionalInDecrement(N, DAG);
26674 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG,
26675 const X86Subtarget *Subtarget) {
26676 SDValue Op0 = N->getOperand(0);
26677 SDValue Op1 = N->getOperand(1);
26679 // X86 can't encode an immediate LHS of a sub. See if we can push the
26680 // negation into a preceding instruction.
26681 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
26682 // If the RHS of the sub is a XOR with one use and a constant, invert the
26683 // immediate. Then add one to the LHS of the sub so we can turn
26684 // X-Y -> X+~Y+1, saving one register.
26685 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
26686 isa<ConstantSDNode>(Op1.getOperand(1))) {
26687 APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue();
26688 EVT VT = Op0.getValueType();
26689 SDValue NewXor = DAG.getNode(ISD::XOR, SDLoc(Op1), VT,
26691 DAG.getConstant(~XorC, SDLoc(Op1), VT));
26692 return DAG.getNode(ISD::ADD, SDLoc(N), VT, NewXor,
26693 DAG.getConstant(C->getAPIntValue() + 1, SDLoc(N), VT));
26697 // Try to synthesize horizontal adds from adds of shuffles.
26698 EVT VT = N->getValueType(0);
26699 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
26700 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
26701 isHorizontalBinOp(Op0, Op1, true))
26702 return DAG.getNode(X86ISD::HSUB, SDLoc(N), VT, Op0, Op1);
26704 return OptimizeConditionalInDecrement(N, DAG);
26707 /// performVZEXTCombine - Performs build vector combines
26708 static SDValue performVZEXTCombine(SDNode *N, SelectionDAG &DAG,
26709 TargetLowering::DAGCombinerInfo &DCI,
26710 const X86Subtarget *Subtarget) {
26712 MVT VT = N->getSimpleValueType(0);
26713 SDValue Op = N->getOperand(0);
26714 MVT OpVT = Op.getSimpleValueType();
26715 MVT OpEltVT = OpVT.getVectorElementType();
26716 unsigned InputBits = OpEltVT.getSizeInBits() * VT.getVectorNumElements();
26718 // (vzext (bitcast (vzext (x)) -> (vzext x)
26720 while (V.getOpcode() == ISD::BITCAST)
26721 V = V.getOperand(0);
26723 if (V != Op && V.getOpcode() == X86ISD::VZEXT) {
26724 MVT InnerVT = V.getSimpleValueType();
26725 MVT InnerEltVT = InnerVT.getVectorElementType();
26727 // If the element sizes match exactly, we can just do one larger vzext. This
26728 // is always an exact type match as vzext operates on integer types.
26729 if (OpEltVT == InnerEltVT) {
26730 assert(OpVT == InnerVT && "Types must match for vzext!");
26731 return DAG.getNode(X86ISD::VZEXT, DL, VT, V.getOperand(0));
26734 // The only other way we can combine them is if only a single element of the
26735 // inner vzext is used in the input to the outer vzext.
26736 if (InnerEltVT.getSizeInBits() < InputBits)
26739 // In this case, the inner vzext is completely dead because we're going to
26740 // only look at bits inside of the low element. Just do the outer vzext on
26741 // a bitcast of the input to the inner.
26742 return DAG.getNode(X86ISD::VZEXT, DL, VT, DAG.getBitcast(OpVT, V));
26745 // Check if we can bypass extracting and re-inserting an element of an input
26746 // vector. Essentially:
26747 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast x)
26748 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR &&
26749 V.getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
26750 V.getOperand(0).getSimpleValueType().getSizeInBits() == InputBits) {
26751 SDValue ExtractedV = V.getOperand(0);
26752 SDValue OrigV = ExtractedV.getOperand(0);
26753 if (auto *ExtractIdx = dyn_cast<ConstantSDNode>(ExtractedV.getOperand(1)))
26754 if (ExtractIdx->getZExtValue() == 0) {
26755 MVT OrigVT = OrigV.getSimpleValueType();
26756 // Extract a subvector if necessary...
26757 if (OrigVT.getSizeInBits() > OpVT.getSizeInBits()) {
26758 int Ratio = OrigVT.getSizeInBits() / OpVT.getSizeInBits();
26759 OrigVT = MVT::getVectorVT(OrigVT.getVectorElementType(),
26760 OrigVT.getVectorNumElements() / Ratio);
26761 OrigV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigVT, OrigV,
26762 DAG.getIntPtrConstant(0, DL));
26764 Op = DAG.getBitcast(OpVT, OrigV);
26765 return DAG.getNode(X86ISD::VZEXT, DL, VT, Op);
26772 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
26773 DAGCombinerInfo &DCI) const {
26774 SelectionDAG &DAG = DCI.DAG;
26775 switch (N->getOpcode()) {
26777 case ISD::EXTRACT_VECTOR_ELT:
26778 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, DCI);
26781 case X86ISD::SHRUNKBLEND:
26782 return PerformSELECTCombine(N, DAG, DCI, Subtarget);
26783 case ISD::BITCAST: return PerformBITCASTCombine(N, DAG, Subtarget);
26784 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI, Subtarget);
26785 case ISD::ADD: return PerformAddCombine(N, DAG, Subtarget);
26786 case ISD::SUB: return PerformSubCombine(N, DAG, Subtarget);
26787 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
26788 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
26791 case ISD::SRL: return PerformShiftCombine(N, DAG, DCI, Subtarget);
26792 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
26793 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
26794 case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget);
26795 case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget);
26796 case ISD::MLOAD: return PerformMLOADCombine(N, DAG, DCI, Subtarget);
26797 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
26798 case ISD::MSTORE: return PerformMSTORECombine(N, DAG, Subtarget);
26799 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, Subtarget);
26800 case ISD::UINT_TO_FP: return PerformUINT_TO_FPCombine(N, DAG, Subtarget);
26801 case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
26802 case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
26804 case X86ISD::FOR: return PerformFORCombine(N, DAG, Subtarget);
26806 case X86ISD::FMAX: return PerformFMinFMaxCombine(N, DAG);
26807 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
26808 case X86ISD::FANDN: return PerformFANDNCombine(N, DAG);
26809 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
26810 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
26811 case ISD::ANY_EXTEND:
26812 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG, DCI, Subtarget);
26813 case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
26814 case ISD::SIGN_EXTEND_INREG:
26815 return PerformSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
26816 case ISD::SETCC: return PerformISDSETCCCombine(N, DAG, Subtarget);
26817 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
26818 case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
26819 case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);
26820 case X86ISD::SHUFP: // Handle all target specific shuffles
26821 case X86ISD::PALIGNR:
26822 case X86ISD::UNPCKH:
26823 case X86ISD::UNPCKL:
26824 case X86ISD::MOVHLPS:
26825 case X86ISD::MOVLHPS:
26826 case X86ISD::PSHUFB:
26827 case X86ISD::PSHUFD:
26828 case X86ISD::PSHUFHW:
26829 case X86ISD::PSHUFLW:
26830 case X86ISD::MOVSS:
26831 case X86ISD::MOVSD:
26832 case X86ISD::VPERMILPI:
26833 case X86ISD::VPERM2X128:
26834 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
26835 case ISD::FMA: return PerformFMACombine(N, DAG, Subtarget);
26836 case X86ISD::BLENDI: return PerformBLENDICombine(N, DAG);
26842 /// isTypeDesirableForOp - Return true if the target has native support for
26843 /// the specified value type and it is 'desirable' to use the type for the
26844 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
26845 /// instruction encodings are longer and some i16 instructions are slow.
26846 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
26847 if (!isTypeLegal(VT))
26849 if (VT != MVT::i16)
26856 case ISD::SIGN_EXTEND:
26857 case ISD::ZERO_EXTEND:
26858 case ISD::ANY_EXTEND:
26871 /// IsDesirableToPromoteOp - This method query the target whether it is
26872 /// beneficial for dag combiner to promote the specified node. If true, it
26873 /// should return the desired promotion type by reference.
26874 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
26875 EVT VT = Op.getValueType();
26876 if (VT != MVT::i16)
26879 bool Promote = false;
26880 bool Commute = false;
26881 switch (Op.getOpcode()) {
26884 LoadSDNode *LD = cast<LoadSDNode>(Op);
26885 // If the non-extending load has a single use and it's not live out, then it
26886 // might be folded.
26887 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
26888 Op.hasOneUse()*/) {
26889 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
26890 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
26891 // The only case where we'd want to promote LOAD (rather then it being
26892 // promoted as an operand is when it's only use is liveout.
26893 if (UI->getOpcode() != ISD::CopyToReg)
26900 case ISD::SIGN_EXTEND:
26901 case ISD::ZERO_EXTEND:
26902 case ISD::ANY_EXTEND:
26907 SDValue N0 = Op.getOperand(0);
26908 // Look out for (store (shl (load), x)).
26909 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
26922 SDValue N0 = Op.getOperand(0);
26923 SDValue N1 = Op.getOperand(1);
26924 if (!Commute && MayFoldLoad(N1))
26926 // Avoid disabling potential load folding opportunities.
26927 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
26929 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
26939 //===----------------------------------------------------------------------===//
26940 // X86 Inline Assembly Support
26941 //===----------------------------------------------------------------------===//
26943 // Helper to match a string separated by whitespace.
26944 static bool matchAsm(StringRef S, ArrayRef<const char *> Pieces) {
26945 S = S.substr(S.find_first_not_of(" \t")); // Skip leading whitespace.
26947 for (StringRef Piece : Pieces) {
26948 if (!S.startswith(Piece)) // Check if the piece matches.
26951 S = S.substr(Piece.size());
26952 StringRef::size_type Pos = S.find_first_not_of(" \t");
26953 if (Pos == 0) // We matched a prefix.
26962 static bool clobbersFlagRegisters(const SmallVector<StringRef, 4> &AsmPieces) {
26964 if (AsmPieces.size() == 3 || AsmPieces.size() == 4) {
26965 if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{cc}") &&
26966 std::count(AsmPieces.begin(), AsmPieces.end(), "~{flags}") &&
26967 std::count(AsmPieces.begin(), AsmPieces.end(), "~{fpsr}")) {
26969 if (AsmPieces.size() == 3)
26971 else if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{dirflag}"))
26978 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
26979 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
26981 std::string AsmStr = IA->getAsmString();
26983 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
26984 if (!Ty || Ty->getBitWidth() % 16 != 0)
26987 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
26988 SmallVector<StringRef, 4> AsmPieces;
26989 SplitString(AsmStr, AsmPieces, ";\n");
26991 switch (AsmPieces.size()) {
26992 default: return false;
26994 // FIXME: this should verify that we are targeting a 486 or better. If not,
26995 // we will turn this bswap into something that will be lowered to logical
26996 // ops instead of emitting the bswap asm. For now, we don't support 486 or
26997 // lower so don't worry about this.
26999 if (matchAsm(AsmPieces[0], {"bswap", "$0"}) ||
27000 matchAsm(AsmPieces[0], {"bswapl", "$0"}) ||
27001 matchAsm(AsmPieces[0], {"bswapq", "$0"}) ||
27002 matchAsm(AsmPieces[0], {"bswap", "${0:q}"}) ||
27003 matchAsm(AsmPieces[0], {"bswapl", "${0:q}"}) ||
27004 matchAsm(AsmPieces[0], {"bswapq", "${0:q}"})) {
27005 // No need to check constraints, nothing other than the equivalent of
27006 // "=r,0" would be valid here.
27007 return IntrinsicLowering::LowerToByteSwap(CI);
27010 // rorw $$8, ${0:w} --> llvm.bswap.i16
27011 if (CI->getType()->isIntegerTy(16) &&
27012 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
27013 (matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) ||
27014 matchAsm(AsmPieces[0], {"rolw", "$$8,", "${0:w}"}))) {
27016 StringRef ConstraintsStr = IA->getConstraintString();
27017 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
27018 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
27019 if (clobbersFlagRegisters(AsmPieces))
27020 return IntrinsicLowering::LowerToByteSwap(CI);
27024 if (CI->getType()->isIntegerTy(32) &&
27025 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
27026 matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) &&
27027 matchAsm(AsmPieces[1], {"rorl", "$$16,", "$0"}) &&
27028 matchAsm(AsmPieces[2], {"rorw", "$$8,", "${0:w}"})) {
27030 StringRef ConstraintsStr = IA->getConstraintString();
27031 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
27032 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
27033 if (clobbersFlagRegisters(AsmPieces))
27034 return IntrinsicLowering::LowerToByteSwap(CI);
27037 if (CI->getType()->isIntegerTy(64)) {
27038 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
27039 if (Constraints.size() >= 2 &&
27040 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
27041 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
27042 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
27043 if (matchAsm(AsmPieces[0], {"bswap", "%eax"}) &&
27044 matchAsm(AsmPieces[1], {"bswap", "%edx"}) &&
27045 matchAsm(AsmPieces[2], {"xchgl", "%eax,", "%edx"}))
27046 return IntrinsicLowering::LowerToByteSwap(CI);
27054 /// getConstraintType - Given a constraint letter, return the type of
27055 /// constraint it is for this target.
27056 X86TargetLowering::ConstraintType
27057 X86TargetLowering::getConstraintType(StringRef Constraint) const {
27058 if (Constraint.size() == 1) {
27059 switch (Constraint[0]) {
27070 return C_RegisterClass;
27094 return TargetLowering::getConstraintType(Constraint);
27097 /// Examine constraint type and operand type and determine a weight value.
27098 /// This object must already have been set up with the operand type
27099 /// and the current alternative constraint selected.
27100 TargetLowering::ConstraintWeight
27101 X86TargetLowering::getSingleConstraintMatchWeight(
27102 AsmOperandInfo &info, const char *constraint) const {
27103 ConstraintWeight weight = CW_Invalid;
27104 Value *CallOperandVal = info.CallOperandVal;
27105 // If we don't have a value, we can't do a match,
27106 // but allow it at the lowest weight.
27107 if (!CallOperandVal)
27109 Type *type = CallOperandVal->getType();
27110 // Look at the constraint type.
27111 switch (*constraint) {
27113 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
27124 if (CallOperandVal->getType()->isIntegerTy())
27125 weight = CW_SpecificReg;
27130 if (type->isFloatingPointTy())
27131 weight = CW_SpecificReg;
27134 if (type->isX86_MMXTy() && Subtarget->hasMMX())
27135 weight = CW_SpecificReg;
27139 if (((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasSSE1()) ||
27140 ((type->getPrimitiveSizeInBits() == 256) && Subtarget->hasFp256()))
27141 weight = CW_Register;
27144 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
27145 if (C->getZExtValue() <= 31)
27146 weight = CW_Constant;
27150 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27151 if (C->getZExtValue() <= 63)
27152 weight = CW_Constant;
27156 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27157 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
27158 weight = CW_Constant;
27162 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27163 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
27164 weight = CW_Constant;
27168 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27169 if (C->getZExtValue() <= 3)
27170 weight = CW_Constant;
27174 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27175 if (C->getZExtValue() <= 0xff)
27176 weight = CW_Constant;
27181 if (isa<ConstantFP>(CallOperandVal)) {
27182 weight = CW_Constant;
27186 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27187 if ((C->getSExtValue() >= -0x80000000LL) &&
27188 (C->getSExtValue() <= 0x7fffffffLL))
27189 weight = CW_Constant;
27193 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27194 if (C->getZExtValue() <= 0xffffffff)
27195 weight = CW_Constant;
27202 /// LowerXConstraint - try to replace an X constraint, which matches anything,
27203 /// with another that has more specific requirements based on the type of the
27204 /// corresponding operand.
27205 const char *X86TargetLowering::
27206 LowerXConstraint(EVT ConstraintVT) const {
27207 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
27208 // 'f' like normal targets.
27209 if (ConstraintVT.isFloatingPoint()) {
27210 if (Subtarget->hasSSE2())
27212 if (Subtarget->hasSSE1())
27216 return TargetLowering::LowerXConstraint(ConstraintVT);
27219 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
27220 /// vector. If it is invalid, don't add anything to Ops.
27221 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
27222 std::string &Constraint,
27223 std::vector<SDValue>&Ops,
27224 SelectionDAG &DAG) const {
27227 // Only support length 1 constraints for now.
27228 if (Constraint.length() > 1) return;
27230 char ConstraintLetter = Constraint[0];
27231 switch (ConstraintLetter) {
27234 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27235 if (C->getZExtValue() <= 31) {
27236 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27237 Op.getValueType());
27243 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27244 if (C->getZExtValue() <= 63) {
27245 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27246 Op.getValueType());
27252 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27253 if (isInt<8>(C->getSExtValue())) {
27254 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27255 Op.getValueType());
27261 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27262 if (C->getZExtValue() == 0xff || C->getZExtValue() == 0xffff ||
27263 (Subtarget->is64Bit() && C->getZExtValue() == 0xffffffff)) {
27264 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op),
27265 Op.getValueType());
27271 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27272 if (C->getZExtValue() <= 3) {
27273 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27274 Op.getValueType());
27280 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27281 if (C->getZExtValue() <= 255) {
27282 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27283 Op.getValueType());
27289 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27290 if (C->getZExtValue() <= 127) {
27291 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27292 Op.getValueType());
27298 // 32-bit signed value
27299 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27300 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
27301 C->getSExtValue())) {
27302 // Widen to 64 bits here to get it sign extended.
27303 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op), MVT::i64);
27306 // FIXME gcc accepts some relocatable values here too, but only in certain
27307 // memory models; it's complicated.
27312 // 32-bit unsigned value
27313 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27314 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
27315 C->getZExtValue())) {
27316 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27317 Op.getValueType());
27321 // FIXME gcc accepts some relocatable values here too, but only in certain
27322 // memory models; it's complicated.
27326 // Literal immediates are always ok.
27327 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
27328 // Widen to 64 bits here to get it sign extended.
27329 Result = DAG.getTargetConstant(CST->getSExtValue(), SDLoc(Op), MVT::i64);
27333 // In any sort of PIC mode addresses need to be computed at runtime by
27334 // adding in a register or some sort of table lookup. These can't
27335 // be used as immediates.
27336 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
27339 // If we are in non-pic codegen mode, we allow the address of a global (with
27340 // an optional displacement) to be used with 'i'.
27341 GlobalAddressSDNode *GA = nullptr;
27342 int64_t Offset = 0;
27344 // Match either (GA), (GA+C), (GA+C1+C2), etc.
27346 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
27347 Offset += GA->getOffset();
27349 } else if (Op.getOpcode() == ISD::ADD) {
27350 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
27351 Offset += C->getZExtValue();
27352 Op = Op.getOperand(0);
27355 } else if (Op.getOpcode() == ISD::SUB) {
27356 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
27357 Offset += -C->getZExtValue();
27358 Op = Op.getOperand(0);
27363 // Otherwise, this isn't something we can handle, reject it.
27367 const GlobalValue *GV = GA->getGlobal();
27368 // If we require an extra load to get this address, as in PIC mode, we
27369 // can't accept it.
27370 if (isGlobalStubReference(
27371 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget())))
27374 Result = DAG.getTargetGlobalAddress(GV, SDLoc(Op),
27375 GA->getValueType(0), Offset);
27380 if (Result.getNode()) {
27381 Ops.push_back(Result);
27384 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
27387 std::pair<unsigned, const TargetRegisterClass *>
27388 X86TargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
27389 StringRef Constraint,
27391 // First, see if this is a constraint that directly corresponds to an LLVM
27393 if (Constraint.size() == 1) {
27394 // GCC Constraint Letters
27395 switch (Constraint[0]) {
27397 // TODO: Slight differences here in allocation order and leaving
27398 // RIP in the class. Do they matter any more here than they do
27399 // in the normal allocation?
27400 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
27401 if (Subtarget->is64Bit()) {
27402 if (VT == MVT::i32 || VT == MVT::f32)
27403 return std::make_pair(0U, &X86::GR32RegClass);
27404 if (VT == MVT::i16)
27405 return std::make_pair(0U, &X86::GR16RegClass);
27406 if (VT == MVT::i8 || VT == MVT::i1)
27407 return std::make_pair(0U, &X86::GR8RegClass);
27408 if (VT == MVT::i64 || VT == MVT::f64)
27409 return std::make_pair(0U, &X86::GR64RegClass);
27412 // 32-bit fallthrough
27413 case 'Q': // Q_REGS
27414 if (VT == MVT::i32 || VT == MVT::f32)
27415 return std::make_pair(0U, &X86::GR32_ABCDRegClass);
27416 if (VT == MVT::i16)
27417 return std::make_pair(0U, &X86::GR16_ABCDRegClass);
27418 if (VT == MVT::i8 || VT == MVT::i1)
27419 return std::make_pair(0U, &X86::GR8_ABCD_LRegClass);
27420 if (VT == MVT::i64)
27421 return std::make_pair(0U, &X86::GR64_ABCDRegClass);
27423 case 'r': // GENERAL_REGS
27424 case 'l': // INDEX_REGS
27425 if (VT == MVT::i8 || VT == MVT::i1)
27426 return std::make_pair(0U, &X86::GR8RegClass);
27427 if (VT == MVT::i16)
27428 return std::make_pair(0U, &X86::GR16RegClass);
27429 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
27430 return std::make_pair(0U, &X86::GR32RegClass);
27431 return std::make_pair(0U, &X86::GR64RegClass);
27432 case 'R': // LEGACY_REGS
27433 if (VT == MVT::i8 || VT == MVT::i1)
27434 return std::make_pair(0U, &X86::GR8_NOREXRegClass);
27435 if (VT == MVT::i16)
27436 return std::make_pair(0U, &X86::GR16_NOREXRegClass);
27437 if (VT == MVT::i32 || !Subtarget->is64Bit())
27438 return std::make_pair(0U, &X86::GR32_NOREXRegClass);
27439 return std::make_pair(0U, &X86::GR64_NOREXRegClass);
27440 case 'f': // FP Stack registers.
27441 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
27442 // value to the correct fpstack register class.
27443 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
27444 return std::make_pair(0U, &X86::RFP32RegClass);
27445 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
27446 return std::make_pair(0U, &X86::RFP64RegClass);
27447 return std::make_pair(0U, &X86::RFP80RegClass);
27448 case 'y': // MMX_REGS if MMX allowed.
27449 if (!Subtarget->hasMMX()) break;
27450 return std::make_pair(0U, &X86::VR64RegClass);
27451 case 'Y': // SSE_REGS if SSE2 allowed
27452 if (!Subtarget->hasSSE2()) break;
27454 case 'x': // SSE_REGS if SSE1 allowed or AVX_REGS if AVX allowed
27455 if (!Subtarget->hasSSE1()) break;
27457 switch (VT.SimpleTy) {
27459 // Scalar SSE types.
27462 return std::make_pair(0U, &X86::FR32RegClass);
27465 return std::make_pair(0U, &X86::FR64RegClass);
27473 return std::make_pair(0U, &X86::VR128RegClass);
27481 return std::make_pair(0U, &X86::VR256RegClass);
27486 return std::make_pair(0U, &X86::VR512RegClass);
27492 // Use the default implementation in TargetLowering to convert the register
27493 // constraint into a member of a register class.
27494 std::pair<unsigned, const TargetRegisterClass*> Res;
27495 Res = TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
27497 // Not found as a standard register?
27499 // Map st(0) -> st(7) -> ST0
27500 if (Constraint.size() == 7 && Constraint[0] == '{' &&
27501 tolower(Constraint[1]) == 's' &&
27502 tolower(Constraint[2]) == 't' &&
27503 Constraint[3] == '(' &&
27504 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
27505 Constraint[5] == ')' &&
27506 Constraint[6] == '}') {
27508 Res.first = X86::FP0+Constraint[4]-'0';
27509 Res.second = &X86::RFP80RegClass;
27513 // GCC allows "st(0)" to be called just plain "st".
27514 if (StringRef("{st}").equals_lower(Constraint)) {
27515 Res.first = X86::FP0;
27516 Res.second = &X86::RFP80RegClass;
27521 if (StringRef("{flags}").equals_lower(Constraint)) {
27522 Res.first = X86::EFLAGS;
27523 Res.second = &X86::CCRRegClass;
27527 // 'A' means EAX + EDX.
27528 if (Constraint == "A") {
27529 Res.first = X86::EAX;
27530 Res.second = &X86::GR32_ADRegClass;
27536 // Otherwise, check to see if this is a register class of the wrong value
27537 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
27538 // turn into {ax},{dx}.
27539 // MVT::Other is used to specify clobber names.
27540 if (Res.second->hasType(VT) || VT == MVT::Other)
27541 return Res; // Correct type already, nothing to do.
27543 // Get a matching integer of the correct size. i.e. "ax" with MVT::32 should
27544 // return "eax". This should even work for things like getting 64bit integer
27545 // registers when given an f64 type.
27546 const TargetRegisterClass *Class = Res.second;
27547 if (Class == &X86::GR8RegClass || Class == &X86::GR16RegClass ||
27548 Class == &X86::GR32RegClass || Class == &X86::GR64RegClass) {
27549 unsigned Size = VT.getSizeInBits();
27550 MVT::SimpleValueType SimpleTy = Size == 1 || Size == 8 ? MVT::i8
27551 : Size == 16 ? MVT::i16
27552 : Size == 32 ? MVT::i32
27553 : Size == 64 ? MVT::i64
27555 unsigned DestReg = getX86SubSuperRegisterOrZero(Res.first, SimpleTy);
27557 Res.first = DestReg;
27558 Res.second = SimpleTy == MVT::i8 ? &X86::GR8RegClass
27559 : SimpleTy == MVT::i16 ? &X86::GR16RegClass
27560 : SimpleTy == MVT::i32 ? &X86::GR32RegClass
27561 : &X86::GR64RegClass;
27562 assert(Res.second->contains(Res.first) && "Register in register class");
27564 // No register found/type mismatch.
27566 Res.second = nullptr;
27568 } else if (Class == &X86::FR32RegClass || Class == &X86::FR64RegClass ||
27569 Class == &X86::VR128RegClass || Class == &X86::VR256RegClass ||
27570 Class == &X86::FR32XRegClass || Class == &X86::FR64XRegClass ||
27571 Class == &X86::VR128XRegClass || Class == &X86::VR256XRegClass ||
27572 Class == &X86::VR512RegClass) {
27573 // Handle references to XMM physical registers that got mapped into the
27574 // wrong class. This can happen with constraints like {xmm0} where the
27575 // target independent register mapper will just pick the first match it can
27576 // find, ignoring the required type.
27578 if (VT == MVT::f32 || VT == MVT::i32)
27579 Res.second = &X86::FR32RegClass;
27580 else if (VT == MVT::f64 || VT == MVT::i64)
27581 Res.second = &X86::FR64RegClass;
27582 else if (X86::VR128RegClass.hasType(VT))
27583 Res.second = &X86::VR128RegClass;
27584 else if (X86::VR256RegClass.hasType(VT))
27585 Res.second = &X86::VR256RegClass;
27586 else if (X86::VR512RegClass.hasType(VT))
27587 Res.second = &X86::VR512RegClass;
27589 // Type mismatch and not a clobber: Return an error;
27591 Res.second = nullptr;
27598 int X86TargetLowering::getScalingFactorCost(const DataLayout &DL,
27599 const AddrMode &AM, Type *Ty,
27600 unsigned AS) const {
27601 // Scaling factors are not free at all.
27602 // An indexed folded instruction, i.e., inst (reg1, reg2, scale),
27603 // will take 2 allocations in the out of order engine instead of 1
27604 // for plain addressing mode, i.e. inst (reg1).
27606 // vaddps (%rsi,%drx), %ymm0, %ymm1
27607 // Requires two allocations (one for the load, one for the computation)
27609 // vaddps (%rsi), %ymm0, %ymm1
27610 // Requires just 1 allocation, i.e., freeing allocations for other operations
27611 // and having less micro operations to execute.
27613 // For some X86 architectures, this is even worse because for instance for
27614 // stores, the complex addressing mode forces the instruction to use the
27615 // "load" ports instead of the dedicated "store" port.
27616 // E.g., on Haswell:
27617 // vmovaps %ymm1, (%r8, %rdi) can use port 2 or 3.
27618 // vmovaps %ymm1, (%r8) can use port 2, 3, or 7.
27619 if (isLegalAddressingMode(DL, AM, Ty, AS))
27620 // Scale represents reg2 * scale, thus account for 1
27621 // as soon as we use a second register.
27622 return AM.Scale != 0;
27626 bool X86TargetLowering::isIntDivCheap(EVT VT, AttributeSet Attr) const {
27627 // Integer division on x86 is expensive. However, when aggressively optimizing
27628 // for code size, we prefer to use a div instruction, as it is usually smaller
27629 // than the alternative sequence.
27630 // The exception to this is vector division. Since x86 doesn't have vector
27631 // integer division, leaving the division as-is is a loss even in terms of
27632 // size, because it will have to be scalarized, while the alternative code
27633 // sequence can be performed in vector form.
27634 bool OptSize = Attr.hasAttribute(AttributeSet::FunctionIndex,
27635 Attribute::MinSize);
27636 return OptSize && !VT.isVector();
27639 void X86TargetLowering::markInRegArguments(SelectionDAG &DAG,
27640 TargetLowering::ArgListTy& Args) const {
27641 // The MCU psABI requires some arguments to be passed in-register.
27642 // For regular calls, the inreg arguments are marked by the front-end.
27643 // However, for compiler generated library calls, we have to patch this
27645 if (!Subtarget->isTargetMCU() || !Args.size())
27648 unsigned FreeRegs = 3;
27649 for (auto &Arg : Args) {
27650 // For library functions, we do not expect any fancy types.
27651 unsigned Size = DAG.getDataLayout().getTypeSizeInBits(Arg.Ty);
27652 unsigned SizeInRegs = (Size + 31) / 32;
27653 if (SizeInRegs > 2 || SizeInRegs > FreeRegs)
27656 Arg.isInReg = true;
27657 FreeRegs -= SizeInRegs;