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 "X86InstrBuilder.h"
19 #include "X86MachineFunctionInfo.h"
20 #include "X86TargetMachine.h"
21 #include "X86TargetObjectFile.h"
22 #include "llvm/ADT/SmallBitVector.h"
23 #include "llvm/ADT/SmallSet.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/ADT/StringExtras.h"
26 #include "llvm/ADT/StringSwitch.h"
27 #include "llvm/ADT/VariadicFunction.h"
28 #include "llvm/CodeGen/IntrinsicLowering.h"
29 #include "llvm/CodeGen/MachineFrameInfo.h"
30 #include "llvm/CodeGen/MachineFunction.h"
31 #include "llvm/CodeGen/MachineInstrBuilder.h"
32 #include "llvm/CodeGen/MachineJumpTableInfo.h"
33 #include "llvm/CodeGen/MachineModuleInfo.h"
34 #include "llvm/CodeGen/MachineRegisterInfo.h"
35 #include "llvm/IR/CallSite.h"
36 #include "llvm/IR/CallingConv.h"
37 #include "llvm/IR/Constants.h"
38 #include "llvm/IR/DerivedTypes.h"
39 #include "llvm/IR/Function.h"
40 #include "llvm/IR/GlobalAlias.h"
41 #include "llvm/IR/GlobalVariable.h"
42 #include "llvm/IR/Instructions.h"
43 #include "llvm/IR/Intrinsics.h"
44 #include "llvm/MC/MCAsmInfo.h"
45 #include "llvm/MC/MCContext.h"
46 #include "llvm/MC/MCExpr.h"
47 #include "llvm/MC/MCSymbol.h"
48 #include "llvm/Support/CommandLine.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/ErrorHandling.h"
51 #include "llvm/Support/MathExtras.h"
52 #include "llvm/Target/TargetOptions.h"
53 #include "X86IntrinsicsInfo.h"
59 #define DEBUG_TYPE "x86-isel"
61 STATISTIC(NumTailCalls, "Number of tail calls");
63 static cl::opt<bool> ExperimentalVectorWideningLegalization(
64 "x86-experimental-vector-widening-legalization", cl::init(false),
65 cl::desc("Enable an experimental vector type legalization through widening "
66 "rather than promotion."),
69 static cl::opt<bool> ExperimentalVectorShuffleLowering(
70 "x86-experimental-vector-shuffle-lowering", cl::init(false),
71 cl::desc("Enable an experimental vector shuffle lowering code path."),
74 // Forward declarations.
75 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1,
78 static SDValue ExtractSubVector(SDValue Vec, unsigned IdxVal,
79 SelectionDAG &DAG, SDLoc dl,
80 unsigned vectorWidth) {
81 assert((vectorWidth == 128 || vectorWidth == 256) &&
82 "Unsupported vector width");
83 EVT VT = Vec.getValueType();
84 EVT ElVT = VT.getVectorElementType();
85 unsigned Factor = VT.getSizeInBits()/vectorWidth;
86 EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT,
87 VT.getVectorNumElements()/Factor);
89 // Extract from UNDEF is UNDEF.
90 if (Vec.getOpcode() == ISD::UNDEF)
91 return DAG.getUNDEF(ResultVT);
93 // Extract the relevant vectorWidth bits. Generate an EXTRACT_SUBVECTOR
94 unsigned ElemsPerChunk = vectorWidth / ElVT.getSizeInBits();
96 // This is the index of the first element of the vectorWidth-bit chunk
98 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits()) / vectorWidth)
101 // If the input is a buildvector just emit a smaller one.
102 if (Vec.getOpcode() == ISD::BUILD_VECTOR)
103 return DAG.getNode(ISD::BUILD_VECTOR, dl, ResultVT,
104 makeArrayRef(Vec->op_begin()+NormalizedIdxVal,
107 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal);
108 SDValue Result = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec,
114 /// Generate a DAG to grab 128-bits from a vector > 128 bits. This
115 /// sets things up to match to an AVX VEXTRACTF128 / VEXTRACTI128
116 /// or AVX-512 VEXTRACTF32x4 / VEXTRACTI32x4
117 /// instructions or a simple subregister reference. Idx is an index in the
118 /// 128 bits we want. It need not be aligned to a 128-bit bounday. That makes
119 /// lowering EXTRACT_VECTOR_ELT operations easier.
120 static SDValue Extract128BitVector(SDValue Vec, unsigned IdxVal,
121 SelectionDAG &DAG, SDLoc dl) {
122 assert((Vec.getValueType().is256BitVector() ||
123 Vec.getValueType().is512BitVector()) && "Unexpected vector size!");
124 return ExtractSubVector(Vec, IdxVal, DAG, dl, 128);
127 /// Generate a DAG to grab 256-bits from a 512-bit vector.
128 static SDValue Extract256BitVector(SDValue Vec, unsigned IdxVal,
129 SelectionDAG &DAG, SDLoc dl) {
130 assert(Vec.getValueType().is512BitVector() && "Unexpected vector size!");
131 return ExtractSubVector(Vec, IdxVal, DAG, dl, 256);
134 static SDValue InsertSubVector(SDValue Result, SDValue Vec,
135 unsigned IdxVal, SelectionDAG &DAG,
136 SDLoc dl, unsigned vectorWidth) {
137 assert((vectorWidth == 128 || vectorWidth == 256) &&
138 "Unsupported vector width");
139 // Inserting UNDEF is Result
140 if (Vec.getOpcode() == ISD::UNDEF)
142 EVT VT = Vec.getValueType();
143 EVT ElVT = VT.getVectorElementType();
144 EVT ResultVT = Result.getValueType();
146 // Insert the relevant vectorWidth bits.
147 unsigned ElemsPerChunk = vectorWidth/ElVT.getSizeInBits();
149 // This is the index of the first element of the vectorWidth-bit chunk
151 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits())/vectorWidth)
154 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal);
155 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec,
158 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
159 /// sets things up to match to an AVX VINSERTF128/VINSERTI128 or
160 /// AVX-512 VINSERTF32x4/VINSERTI32x4 instructions or a
161 /// simple superregister reference. Idx is an index in the 128 bits
162 /// we want. It need not be aligned to a 128-bit bounday. That makes
163 /// lowering INSERT_VECTOR_ELT operations easier.
164 static SDValue Insert128BitVector(SDValue Result, SDValue Vec,
165 unsigned IdxVal, SelectionDAG &DAG,
167 assert(Vec.getValueType().is128BitVector() && "Unexpected vector size!");
168 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 128);
171 static SDValue Insert256BitVector(SDValue Result, SDValue Vec,
172 unsigned IdxVal, SelectionDAG &DAG,
174 assert(Vec.getValueType().is256BitVector() && "Unexpected vector size!");
175 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 256);
178 /// Concat two 128-bit vectors into a 256 bit vector using VINSERTF128
179 /// instructions. This is used because creating CONCAT_VECTOR nodes of
180 /// BUILD_VECTORS returns a larger BUILD_VECTOR while we're trying to lower
181 /// large BUILD_VECTORS.
182 static SDValue Concat128BitVectors(SDValue V1, SDValue V2, EVT VT,
183 unsigned NumElems, SelectionDAG &DAG,
185 SDValue V = Insert128BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
186 return Insert128BitVector(V, V2, NumElems/2, DAG, dl);
189 static SDValue Concat256BitVectors(SDValue V1, SDValue V2, EVT VT,
190 unsigned NumElems, SelectionDAG &DAG,
192 SDValue V = Insert256BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
193 return Insert256BitVector(V, V2, NumElems/2, DAG, dl);
196 static TargetLoweringObjectFile *createTLOF(const Triple &TT) {
197 if (TT.isOSBinFormatMachO()) {
198 if (TT.getArch() == Triple::x86_64)
199 return new X86_64MachoTargetObjectFile();
200 return new TargetLoweringObjectFileMachO();
204 return new X86LinuxTargetObjectFile();
205 if (TT.isOSBinFormatELF())
206 return new TargetLoweringObjectFileELF();
207 if (TT.isKnownWindowsMSVCEnvironment())
208 return new X86WindowsTargetObjectFile();
209 if (TT.isOSBinFormatCOFF())
210 return new TargetLoweringObjectFileCOFF();
211 llvm_unreachable("unknown subtarget type");
214 // FIXME: This should stop caching the target machine as soon as
215 // we can remove resetOperationActions et al.
216 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
217 : TargetLowering(TM, createTLOF(Triple(TM.getTargetTriple()))) {
218 Subtarget = &TM.getSubtarget<X86Subtarget>();
219 X86ScalarSSEf64 = Subtarget->hasSSE2();
220 X86ScalarSSEf32 = Subtarget->hasSSE1();
221 TD = getDataLayout();
223 resetOperationActions();
226 void X86TargetLowering::resetOperationActions() {
227 const TargetMachine &TM = getTargetMachine();
228 static bool FirstTimeThrough = true;
230 // If none of the target options have changed, then we don't need to reset the
231 // operation actions.
232 if (!FirstTimeThrough && TO == TM.Options) return;
234 if (!FirstTimeThrough) {
235 // Reinitialize the actions.
237 FirstTimeThrough = false;
242 // Set up the TargetLowering object.
243 static const MVT IntVTs[] = { MVT::i8, MVT::i16, MVT::i32, MVT::i64 };
245 // X86 is weird, it always uses i8 for shift amounts and setcc results.
246 setBooleanContents(ZeroOrOneBooleanContent);
247 // X86-SSE is even stranger. It uses -1 or 0 for vector masks.
248 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
250 // For 64-bit since we have so many registers use the ILP scheduler, for
251 // 32-bit code use the register pressure specific scheduling.
252 // For Atom, always use ILP scheduling.
253 if (Subtarget->isAtom())
254 setSchedulingPreference(Sched::ILP);
255 else if (Subtarget->is64Bit())
256 setSchedulingPreference(Sched::ILP);
258 setSchedulingPreference(Sched::RegPressure);
259 const X86RegisterInfo *RegInfo =
260 TM.getSubtarget<X86Subtarget>().getRegisterInfo();
261 setStackPointerRegisterToSaveRestore(RegInfo->getStackRegister());
263 // Bypass expensive divides on Atom when compiling with O2
264 if (Subtarget->hasSlowDivide() && TM.getOptLevel() >= CodeGenOpt::Default) {
265 addBypassSlowDiv(32, 8);
266 if (Subtarget->is64Bit())
267 addBypassSlowDiv(64, 16);
270 if (Subtarget->isTargetKnownWindowsMSVC()) {
271 // Setup Windows compiler runtime calls.
272 setLibcallName(RTLIB::SDIV_I64, "_alldiv");
273 setLibcallName(RTLIB::UDIV_I64, "_aulldiv");
274 setLibcallName(RTLIB::SREM_I64, "_allrem");
275 setLibcallName(RTLIB::UREM_I64, "_aullrem");
276 setLibcallName(RTLIB::MUL_I64, "_allmul");
277 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::X86_StdCall);
278 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::X86_StdCall);
279 setLibcallCallingConv(RTLIB::SREM_I64, CallingConv::X86_StdCall);
280 setLibcallCallingConv(RTLIB::UREM_I64, CallingConv::X86_StdCall);
281 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::X86_StdCall);
283 // The _ftol2 runtime function has an unusual calling conv, which
284 // is modeled by a special pseudo-instruction.
285 setLibcallName(RTLIB::FPTOUINT_F64_I64, nullptr);
286 setLibcallName(RTLIB::FPTOUINT_F32_I64, nullptr);
287 setLibcallName(RTLIB::FPTOUINT_F64_I32, nullptr);
288 setLibcallName(RTLIB::FPTOUINT_F32_I32, nullptr);
291 if (Subtarget->isTargetDarwin()) {
292 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
293 setUseUnderscoreSetJmp(false);
294 setUseUnderscoreLongJmp(false);
295 } else if (Subtarget->isTargetWindowsGNU()) {
296 // MS runtime is weird: it exports _setjmp, but longjmp!
297 setUseUnderscoreSetJmp(true);
298 setUseUnderscoreLongJmp(false);
300 setUseUnderscoreSetJmp(true);
301 setUseUnderscoreLongJmp(true);
304 // Set up the register classes.
305 addRegisterClass(MVT::i8, &X86::GR8RegClass);
306 addRegisterClass(MVT::i16, &X86::GR16RegClass);
307 addRegisterClass(MVT::i32, &X86::GR32RegClass);
308 if (Subtarget->is64Bit())
309 addRegisterClass(MVT::i64, &X86::GR64RegClass);
311 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
313 // We don't accept any truncstore of integer registers.
314 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
315 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
316 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
317 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
318 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
319 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
321 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
323 // SETOEQ and SETUNE require checking two conditions.
324 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
325 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
326 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
327 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
328 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
329 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
331 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
333 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
334 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
335 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
337 if (Subtarget->is64Bit()) {
338 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
339 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
340 } else if (!TM.Options.UseSoftFloat) {
341 // We have an algorithm for SSE2->double, and we turn this into a
342 // 64-bit FILD followed by conditional FADD for other targets.
343 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
344 // We have an algorithm for SSE2, and we turn this into a 64-bit
345 // FILD for other targets.
346 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
349 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
351 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
352 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
354 if (!TM.Options.UseSoftFloat) {
355 // SSE has no i16 to fp conversion, only i32
356 if (X86ScalarSSEf32) {
357 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
358 // f32 and f64 cases are Legal, f80 case is not
359 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
361 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
362 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
365 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
366 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
369 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
370 // are Legal, f80 is custom lowered.
371 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
372 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
374 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
376 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
377 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
379 if (X86ScalarSSEf32) {
380 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
381 // f32 and f64 cases are Legal, f80 case is not
382 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
384 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
385 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
388 // Handle FP_TO_UINT by promoting the destination to a larger signed
390 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
391 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
392 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
394 if (Subtarget->is64Bit()) {
395 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
396 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
397 } else if (!TM.Options.UseSoftFloat) {
398 // Since AVX is a superset of SSE3, only check for SSE here.
399 if (Subtarget->hasSSE1() && !Subtarget->hasSSE3())
400 // Expand FP_TO_UINT into a select.
401 // FIXME: We would like to use a Custom expander here eventually to do
402 // the optimal thing for SSE vs. the default expansion in the legalizer.
403 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
405 // With SSE3 we can use fisttpll to convert to a signed i64; without
406 // SSE, we're stuck with a fistpll.
407 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
410 if (isTargetFTOL()) {
411 // Use the _ftol2 runtime function, which has a pseudo-instruction
412 // to handle its weird calling convention.
413 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Custom);
416 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
417 if (!X86ScalarSSEf64) {
418 setOperationAction(ISD::BITCAST , MVT::f32 , Expand);
419 setOperationAction(ISD::BITCAST , MVT::i32 , Expand);
420 if (Subtarget->is64Bit()) {
421 setOperationAction(ISD::BITCAST , MVT::f64 , Expand);
422 // Without SSE, i64->f64 goes through memory.
423 setOperationAction(ISD::BITCAST , MVT::i64 , Expand);
427 // Scalar integer divide and remainder are lowered to use operations that
428 // produce two results, to match the available instructions. This exposes
429 // the two-result form to trivial CSE, which is able to combine x/y and x%y
430 // into a single instruction.
432 // Scalar integer multiply-high is also lowered to use two-result
433 // operations, to match the available instructions. However, plain multiply
434 // (low) operations are left as Legal, as there are single-result
435 // instructions for this in x86. Using the two-result multiply instructions
436 // when both high and low results are needed must be arranged by dagcombine.
437 for (unsigned i = 0; i != array_lengthof(IntVTs); ++i) {
439 setOperationAction(ISD::MULHS, VT, Expand);
440 setOperationAction(ISD::MULHU, VT, Expand);
441 setOperationAction(ISD::SDIV, VT, Expand);
442 setOperationAction(ISD::UDIV, VT, Expand);
443 setOperationAction(ISD::SREM, VT, Expand);
444 setOperationAction(ISD::UREM, VT, Expand);
446 // Add/Sub overflow ops with MVT::Glues are lowered to EFLAGS dependences.
447 setOperationAction(ISD::ADDC, VT, Custom);
448 setOperationAction(ISD::ADDE, VT, Custom);
449 setOperationAction(ISD::SUBC, VT, Custom);
450 setOperationAction(ISD::SUBE, VT, Custom);
453 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
454 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
455 setOperationAction(ISD::BR_CC , MVT::f32, Expand);
456 setOperationAction(ISD::BR_CC , MVT::f64, Expand);
457 setOperationAction(ISD::BR_CC , MVT::f80, Expand);
458 setOperationAction(ISD::BR_CC , MVT::i8, Expand);
459 setOperationAction(ISD::BR_CC , MVT::i16, Expand);
460 setOperationAction(ISD::BR_CC , MVT::i32, Expand);
461 setOperationAction(ISD::BR_CC , MVT::i64, Expand);
462 setOperationAction(ISD::SELECT_CC , MVT::f32, Expand);
463 setOperationAction(ISD::SELECT_CC , MVT::f64, Expand);
464 setOperationAction(ISD::SELECT_CC , MVT::f80, Expand);
465 setOperationAction(ISD::SELECT_CC , MVT::i8, Expand);
466 setOperationAction(ISD::SELECT_CC , MVT::i16, Expand);
467 setOperationAction(ISD::SELECT_CC , MVT::i32, Expand);
468 setOperationAction(ISD::SELECT_CC , MVT::i64, Expand);
469 if (Subtarget->is64Bit())
470 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
471 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
472 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
473 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
474 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
475 setOperationAction(ISD::FREM , MVT::f32 , Expand);
476 setOperationAction(ISD::FREM , MVT::f64 , Expand);
477 setOperationAction(ISD::FREM , MVT::f80 , Expand);
478 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
480 // Promote the i8 variants and force them on up to i32 which has a shorter
482 setOperationAction(ISD::CTTZ , MVT::i8 , Promote);
483 AddPromotedToType (ISD::CTTZ , MVT::i8 , MVT::i32);
484 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i8 , Promote);
485 AddPromotedToType (ISD::CTTZ_ZERO_UNDEF , MVT::i8 , MVT::i32);
486 if (Subtarget->hasBMI()) {
487 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i16 , Expand);
488 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32 , Expand);
489 if (Subtarget->is64Bit())
490 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
492 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
493 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
494 if (Subtarget->is64Bit())
495 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
498 if (Subtarget->hasLZCNT()) {
499 // When promoting the i8 variants, force them to i32 for a shorter
501 setOperationAction(ISD::CTLZ , MVT::i8 , Promote);
502 AddPromotedToType (ISD::CTLZ , MVT::i8 , MVT::i32);
503 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Promote);
504 AddPromotedToType (ISD::CTLZ_ZERO_UNDEF, MVT::i8 , MVT::i32);
505 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Expand);
506 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Expand);
507 if (Subtarget->is64Bit())
508 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
510 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
511 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
512 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
513 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Custom);
514 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Custom);
515 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Custom);
516 if (Subtarget->is64Bit()) {
517 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
518 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Custom);
522 // Special handling for half-precision floating point conversions.
523 // If we don't have F16C support, then lower half float conversions
524 // into library calls.
525 if (TM.Options.UseSoftFloat || !Subtarget->hasF16C()) {
526 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
527 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
530 // There's never any support for operations beyond MVT::f32.
531 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
532 setOperationAction(ISD::FP16_TO_FP, MVT::f80, Expand);
533 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
534 setOperationAction(ISD::FP_TO_FP16, MVT::f80, Expand);
536 setLoadExtAction(ISD::EXTLOAD, MVT::f16, Expand);
537 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
538 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
539 setTruncStoreAction(MVT::f80, MVT::f16, Expand);
541 if (Subtarget->hasPOPCNT()) {
542 setOperationAction(ISD::CTPOP , MVT::i8 , Promote);
544 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
545 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
546 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
547 if (Subtarget->is64Bit())
548 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
551 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
553 if (!Subtarget->hasMOVBE())
554 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
556 // These should be promoted to a larger select which is supported.
557 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
558 // X86 wants to expand cmov itself.
559 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
560 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
561 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
562 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
563 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
564 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
565 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
566 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
567 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
568 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
569 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
570 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
571 if (Subtarget->is64Bit()) {
572 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
573 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
575 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
576 // NOTE: EH_SJLJ_SETJMP/_LONGJMP supported here is NOT intended to support
577 // SjLj exception handling but a light-weight setjmp/longjmp replacement to
578 // support continuation, user-level threading, and etc.. As a result, no
579 // other SjLj exception interfaces are implemented and please don't build
580 // your own exception handling based on them.
581 // LLVM/Clang supports zero-cost DWARF exception handling.
582 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
583 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
586 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
587 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
588 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
589 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
590 if (Subtarget->is64Bit())
591 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
592 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
593 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
594 if (Subtarget->is64Bit()) {
595 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
596 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
597 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
598 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
599 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
601 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
602 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
603 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
604 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
605 if (Subtarget->is64Bit()) {
606 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
607 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
608 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
611 if (Subtarget->hasSSE1())
612 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
614 setOperationAction(ISD::ATOMIC_FENCE , MVT::Other, Custom);
616 // Expand certain atomics
617 for (unsigned i = 0; i != array_lengthof(IntVTs); ++i) {
619 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Custom);
620 setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
621 setOperationAction(ISD::ATOMIC_STORE, VT, Custom);
624 if (Subtarget->hasCmpxchg16b()) {
625 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i128, Custom);
628 // FIXME - use subtarget debug flags
629 if (!Subtarget->isTargetDarwin() && !Subtarget->isTargetELF() &&
630 !Subtarget->isTargetCygMing() && !Subtarget->isTargetWin64()) {
631 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
634 if (Subtarget->is64Bit()) {
635 setExceptionPointerRegister(X86::RAX);
636 setExceptionSelectorRegister(X86::RDX);
638 setExceptionPointerRegister(X86::EAX);
639 setExceptionSelectorRegister(X86::EDX);
641 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
642 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
644 setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
645 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
647 setOperationAction(ISD::TRAP, MVT::Other, Legal);
648 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Legal);
650 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
651 setOperationAction(ISD::VASTART , MVT::Other, Custom);
652 setOperationAction(ISD::VAEND , MVT::Other, Expand);
653 if (Subtarget->is64Bit() && !Subtarget->isTargetWin64()) {
654 // TargetInfo::X86_64ABIBuiltinVaList
655 setOperationAction(ISD::VAARG , MVT::Other, Custom);
656 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
658 // TargetInfo::CharPtrBuiltinVaList
659 setOperationAction(ISD::VAARG , MVT::Other, Expand);
660 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
663 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
664 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
666 setOperationAction(ISD::DYNAMIC_STACKALLOC, getPointerTy(), Custom);
668 if (!TM.Options.UseSoftFloat && X86ScalarSSEf64) {
669 // f32 and f64 use SSE.
670 // Set up the FP register classes.
671 addRegisterClass(MVT::f32, &X86::FR32RegClass);
672 addRegisterClass(MVT::f64, &X86::FR64RegClass);
674 // Use ANDPD to simulate FABS.
675 setOperationAction(ISD::FABS , MVT::f64, Custom);
676 setOperationAction(ISD::FABS , MVT::f32, Custom);
678 // Use XORP to simulate FNEG.
679 setOperationAction(ISD::FNEG , MVT::f64, Custom);
680 setOperationAction(ISD::FNEG , MVT::f32, Custom);
682 // Use ANDPD and ORPD to simulate FCOPYSIGN.
683 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
684 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
686 // Lower this to FGETSIGNx86 plus an AND.
687 setOperationAction(ISD::FGETSIGN, MVT::i64, Custom);
688 setOperationAction(ISD::FGETSIGN, MVT::i32, Custom);
690 // We don't support sin/cos/fmod
691 setOperationAction(ISD::FSIN , MVT::f64, Expand);
692 setOperationAction(ISD::FCOS , MVT::f64, Expand);
693 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
694 setOperationAction(ISD::FSIN , MVT::f32, Expand);
695 setOperationAction(ISD::FCOS , MVT::f32, Expand);
696 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
698 // Expand FP immediates into loads from the stack, except for the special
700 addLegalFPImmediate(APFloat(+0.0)); // xorpd
701 addLegalFPImmediate(APFloat(+0.0f)); // xorps
702 } else if (!TM.Options.UseSoftFloat && X86ScalarSSEf32) {
703 // Use SSE for f32, x87 for f64.
704 // Set up the FP register classes.
705 addRegisterClass(MVT::f32, &X86::FR32RegClass);
706 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
708 // Use ANDPS to simulate FABS.
709 setOperationAction(ISD::FABS , MVT::f32, Custom);
711 // Use XORP to simulate FNEG.
712 setOperationAction(ISD::FNEG , MVT::f32, Custom);
714 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
716 // Use ANDPS and ORPS to simulate FCOPYSIGN.
717 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
718 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
720 // We don't support sin/cos/fmod
721 setOperationAction(ISD::FSIN , MVT::f32, Expand);
722 setOperationAction(ISD::FCOS , MVT::f32, Expand);
723 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
725 // Special cases we handle for FP constants.
726 addLegalFPImmediate(APFloat(+0.0f)); // xorps
727 addLegalFPImmediate(APFloat(+0.0)); // FLD0
728 addLegalFPImmediate(APFloat(+1.0)); // FLD1
729 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
730 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
732 if (!TM.Options.UnsafeFPMath) {
733 setOperationAction(ISD::FSIN , MVT::f64, Expand);
734 setOperationAction(ISD::FCOS , MVT::f64, Expand);
735 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
737 } else if (!TM.Options.UseSoftFloat) {
738 // f32 and f64 in x87.
739 // Set up the FP register classes.
740 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
741 addRegisterClass(MVT::f32, &X86::RFP32RegClass);
743 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
744 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
745 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
746 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
748 if (!TM.Options.UnsafeFPMath) {
749 setOperationAction(ISD::FSIN , MVT::f64, Expand);
750 setOperationAction(ISD::FSIN , MVT::f32, Expand);
751 setOperationAction(ISD::FCOS , MVT::f64, Expand);
752 setOperationAction(ISD::FCOS , MVT::f32, Expand);
753 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
754 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
756 addLegalFPImmediate(APFloat(+0.0)); // FLD0
757 addLegalFPImmediate(APFloat(+1.0)); // FLD1
758 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
759 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
760 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
761 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
762 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
763 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
766 // We don't support FMA.
767 setOperationAction(ISD::FMA, MVT::f64, Expand);
768 setOperationAction(ISD::FMA, MVT::f32, Expand);
770 // Long double always uses X87.
771 if (!TM.Options.UseSoftFloat) {
772 addRegisterClass(MVT::f80, &X86::RFP80RegClass);
773 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
774 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
776 APFloat TmpFlt = APFloat::getZero(APFloat::x87DoubleExtended);
777 addLegalFPImmediate(TmpFlt); // FLD0
779 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
782 APFloat TmpFlt2(+1.0);
783 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
785 addLegalFPImmediate(TmpFlt2); // FLD1
786 TmpFlt2.changeSign();
787 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
790 if (!TM.Options.UnsafeFPMath) {
791 setOperationAction(ISD::FSIN , MVT::f80, Expand);
792 setOperationAction(ISD::FCOS , MVT::f80, Expand);
793 setOperationAction(ISD::FSINCOS, MVT::f80, Expand);
796 setOperationAction(ISD::FFLOOR, MVT::f80, Expand);
797 setOperationAction(ISD::FCEIL, MVT::f80, Expand);
798 setOperationAction(ISD::FTRUNC, MVT::f80, Expand);
799 setOperationAction(ISD::FRINT, MVT::f80, Expand);
800 setOperationAction(ISD::FNEARBYINT, MVT::f80, Expand);
801 setOperationAction(ISD::FMA, MVT::f80, Expand);
804 // Always use a library call for pow.
805 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
806 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
807 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
809 setOperationAction(ISD::FLOG, MVT::f80, Expand);
810 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
811 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
812 setOperationAction(ISD::FEXP, MVT::f80, Expand);
813 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
815 // First set operation action for all vector types to either promote
816 // (for widening) or expand (for scalarization). Then we will selectively
817 // turn on ones that can be effectively codegen'd.
818 for (int i = MVT::FIRST_VECTOR_VALUETYPE;
819 i <= MVT::LAST_VECTOR_VALUETYPE; ++i) {
820 MVT VT = (MVT::SimpleValueType)i;
821 setOperationAction(ISD::ADD , VT, Expand);
822 setOperationAction(ISD::SUB , VT, Expand);
823 setOperationAction(ISD::FADD, VT, Expand);
824 setOperationAction(ISD::FNEG, VT, Expand);
825 setOperationAction(ISD::FSUB, VT, Expand);
826 setOperationAction(ISD::MUL , VT, Expand);
827 setOperationAction(ISD::FMUL, VT, Expand);
828 setOperationAction(ISD::SDIV, VT, Expand);
829 setOperationAction(ISD::UDIV, VT, Expand);
830 setOperationAction(ISD::FDIV, VT, Expand);
831 setOperationAction(ISD::SREM, VT, Expand);
832 setOperationAction(ISD::UREM, VT, Expand);
833 setOperationAction(ISD::LOAD, VT, Expand);
834 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
835 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT,Expand);
836 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
837 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT,Expand);
838 setOperationAction(ISD::INSERT_SUBVECTOR, VT,Expand);
839 setOperationAction(ISD::FABS, VT, Expand);
840 setOperationAction(ISD::FSIN, VT, Expand);
841 setOperationAction(ISD::FSINCOS, VT, Expand);
842 setOperationAction(ISD::FCOS, VT, Expand);
843 setOperationAction(ISD::FSINCOS, VT, Expand);
844 setOperationAction(ISD::FREM, VT, Expand);
845 setOperationAction(ISD::FMA, VT, Expand);
846 setOperationAction(ISD::FPOWI, VT, Expand);
847 setOperationAction(ISD::FSQRT, VT, Expand);
848 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
849 setOperationAction(ISD::FFLOOR, VT, Expand);
850 setOperationAction(ISD::FCEIL, VT, Expand);
851 setOperationAction(ISD::FTRUNC, VT, Expand);
852 setOperationAction(ISD::FRINT, VT, Expand);
853 setOperationAction(ISD::FNEARBYINT, VT, Expand);
854 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
855 setOperationAction(ISD::MULHS, VT, Expand);
856 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
857 setOperationAction(ISD::MULHU, VT, Expand);
858 setOperationAction(ISD::SDIVREM, VT, Expand);
859 setOperationAction(ISD::UDIVREM, VT, Expand);
860 setOperationAction(ISD::FPOW, VT, Expand);
861 setOperationAction(ISD::CTPOP, VT, Expand);
862 setOperationAction(ISD::CTTZ, VT, Expand);
863 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
864 setOperationAction(ISD::CTLZ, VT, Expand);
865 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
866 setOperationAction(ISD::SHL, VT, Expand);
867 setOperationAction(ISD::SRA, VT, Expand);
868 setOperationAction(ISD::SRL, VT, Expand);
869 setOperationAction(ISD::ROTL, VT, Expand);
870 setOperationAction(ISD::ROTR, VT, Expand);
871 setOperationAction(ISD::BSWAP, VT, Expand);
872 setOperationAction(ISD::SETCC, VT, Expand);
873 setOperationAction(ISD::FLOG, VT, Expand);
874 setOperationAction(ISD::FLOG2, VT, Expand);
875 setOperationAction(ISD::FLOG10, VT, Expand);
876 setOperationAction(ISD::FEXP, VT, Expand);
877 setOperationAction(ISD::FEXP2, VT, Expand);
878 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
879 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
880 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
881 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
882 setOperationAction(ISD::SIGN_EXTEND_INREG, VT,Expand);
883 setOperationAction(ISD::TRUNCATE, VT, Expand);
884 setOperationAction(ISD::SIGN_EXTEND, VT, Expand);
885 setOperationAction(ISD::ZERO_EXTEND, VT, Expand);
886 setOperationAction(ISD::ANY_EXTEND, VT, Expand);
887 setOperationAction(ISD::VSELECT, VT, Expand);
888 setOperationAction(ISD::SELECT_CC, VT, Expand);
889 for (int InnerVT = MVT::FIRST_VECTOR_VALUETYPE;
890 InnerVT <= MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
891 setTruncStoreAction(VT,
892 (MVT::SimpleValueType)InnerVT, Expand);
893 setLoadExtAction(ISD::SEXTLOAD, VT, Expand);
894 setLoadExtAction(ISD::ZEXTLOAD, VT, Expand);
896 // N.b. ISD::EXTLOAD legality is basically ignored except for i1-like types,
897 // we have to deal with them whether we ask for Expansion or not. Setting
898 // Expand causes its own optimisation problems though, so leave them legal.
899 if (VT.getVectorElementType() == MVT::i1)
900 setLoadExtAction(ISD::EXTLOAD, VT, Expand);
903 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
904 // with -msoft-float, disable use of MMX as well.
905 if (!TM.Options.UseSoftFloat && Subtarget->hasMMX()) {
906 addRegisterClass(MVT::x86mmx, &X86::VR64RegClass);
907 // No operations on x86mmx supported, everything uses intrinsics.
910 // MMX-sized vectors (other than x86mmx) are expected to be expanded
911 // into smaller operations.
912 setOperationAction(ISD::MULHS, MVT::v8i8, Expand);
913 setOperationAction(ISD::MULHS, MVT::v4i16, Expand);
914 setOperationAction(ISD::MULHS, MVT::v2i32, Expand);
915 setOperationAction(ISD::MULHS, MVT::v1i64, Expand);
916 setOperationAction(ISD::AND, MVT::v8i8, Expand);
917 setOperationAction(ISD::AND, MVT::v4i16, Expand);
918 setOperationAction(ISD::AND, MVT::v2i32, Expand);
919 setOperationAction(ISD::AND, MVT::v1i64, Expand);
920 setOperationAction(ISD::OR, MVT::v8i8, Expand);
921 setOperationAction(ISD::OR, MVT::v4i16, Expand);
922 setOperationAction(ISD::OR, MVT::v2i32, Expand);
923 setOperationAction(ISD::OR, MVT::v1i64, Expand);
924 setOperationAction(ISD::XOR, MVT::v8i8, Expand);
925 setOperationAction(ISD::XOR, MVT::v4i16, Expand);
926 setOperationAction(ISD::XOR, MVT::v2i32, Expand);
927 setOperationAction(ISD::XOR, MVT::v1i64, Expand);
928 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Expand);
929 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Expand);
930 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2i32, Expand);
931 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Expand);
932 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v1i64, Expand);
933 setOperationAction(ISD::SELECT, MVT::v8i8, Expand);
934 setOperationAction(ISD::SELECT, MVT::v4i16, Expand);
935 setOperationAction(ISD::SELECT, MVT::v2i32, Expand);
936 setOperationAction(ISD::SELECT, MVT::v1i64, Expand);
937 setOperationAction(ISD::BITCAST, MVT::v8i8, Expand);
938 setOperationAction(ISD::BITCAST, MVT::v4i16, Expand);
939 setOperationAction(ISD::BITCAST, MVT::v2i32, Expand);
940 setOperationAction(ISD::BITCAST, MVT::v1i64, Expand);
942 if (!TM.Options.UseSoftFloat && Subtarget->hasSSE1()) {
943 addRegisterClass(MVT::v4f32, &X86::VR128RegClass);
945 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
946 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
947 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
948 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
949 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
950 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
951 setOperationAction(ISD::FABS, MVT::v4f32, Custom);
952 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
953 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
954 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
955 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
956 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
959 if (!TM.Options.UseSoftFloat && Subtarget->hasSSE2()) {
960 addRegisterClass(MVT::v2f64, &X86::VR128RegClass);
962 // FIXME: Unfortunately, -soft-float and -no-implicit-float mean XMM
963 // registers cannot be used even for integer operations.
964 addRegisterClass(MVT::v16i8, &X86::VR128RegClass);
965 addRegisterClass(MVT::v8i16, &X86::VR128RegClass);
966 addRegisterClass(MVT::v4i32, &X86::VR128RegClass);
967 addRegisterClass(MVT::v2i64, &X86::VR128RegClass);
969 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
970 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
971 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
972 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
973 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
974 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
975 setOperationAction(ISD::UMUL_LOHI, MVT::v4i32, Custom);
976 setOperationAction(ISD::SMUL_LOHI, MVT::v4i32, Custom);
977 setOperationAction(ISD::MULHU, MVT::v8i16, Legal);
978 setOperationAction(ISD::MULHS, MVT::v8i16, Legal);
979 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
980 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
981 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
982 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
983 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
984 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
985 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
986 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
987 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
988 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
989 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
990 setOperationAction(ISD::FABS, MVT::v2f64, Custom);
992 setOperationAction(ISD::SETCC, MVT::v2i64, Custom);
993 setOperationAction(ISD::SETCC, MVT::v16i8, Custom);
994 setOperationAction(ISD::SETCC, MVT::v8i16, Custom);
995 setOperationAction(ISD::SETCC, MVT::v4i32, Custom);
997 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
998 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
999 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
1000 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
1001 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
1003 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
1004 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
1005 MVT VT = (MVT::SimpleValueType)i;
1006 // Do not attempt to custom lower non-power-of-2 vectors
1007 if (!isPowerOf2_32(VT.getVectorNumElements()))
1009 // Do not attempt to custom lower non-128-bit vectors
1010 if (!VT.is128BitVector())
1012 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1013 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1014 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1017 // We support custom legalizing of sext and anyext loads for specific
1018 // memory vector types which we can load as a scalar (or sequence of
1019 // scalars) and extend in-register to a legal 128-bit vector type. For sext
1020 // loads these must work with a single scalar load.
1021 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i8, Custom);
1022 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i16, Custom);
1023 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i8, Custom);
1024 setLoadExtAction(ISD::EXTLOAD, MVT::v2i8, Custom);
1025 setLoadExtAction(ISD::EXTLOAD, MVT::v2i16, Custom);
1026 setLoadExtAction(ISD::EXTLOAD, MVT::v2i32, Custom);
1027 setLoadExtAction(ISD::EXTLOAD, MVT::v4i8, Custom);
1028 setLoadExtAction(ISD::EXTLOAD, MVT::v4i16, Custom);
1029 setLoadExtAction(ISD::EXTLOAD, MVT::v8i8, Custom);
1031 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
1032 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
1033 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
1034 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
1035 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
1036 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
1038 if (Subtarget->is64Bit()) {
1039 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
1040 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
1043 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
1044 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
1045 MVT VT = (MVT::SimpleValueType)i;
1047 // Do not attempt to promote non-128-bit vectors
1048 if (!VT.is128BitVector())
1051 setOperationAction(ISD::AND, VT, Promote);
1052 AddPromotedToType (ISD::AND, VT, MVT::v2i64);
1053 setOperationAction(ISD::OR, VT, Promote);
1054 AddPromotedToType (ISD::OR, VT, MVT::v2i64);
1055 setOperationAction(ISD::XOR, VT, Promote);
1056 AddPromotedToType (ISD::XOR, VT, MVT::v2i64);
1057 setOperationAction(ISD::LOAD, VT, Promote);
1058 AddPromotedToType (ISD::LOAD, VT, MVT::v2i64);
1059 setOperationAction(ISD::SELECT, VT, Promote);
1060 AddPromotedToType (ISD::SELECT, VT, MVT::v2i64);
1063 // Custom lower v2i64 and v2f64 selects.
1064 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
1065 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
1066 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
1067 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
1069 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
1070 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
1072 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Custom);
1073 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
1074 // As there is no 64-bit GPR available, we need build a special custom
1075 // sequence to convert from v2i32 to v2f32.
1076 if (!Subtarget->is64Bit())
1077 setOperationAction(ISD::UINT_TO_FP, MVT::v2f32, Custom);
1079 setOperationAction(ISD::FP_EXTEND, MVT::v2f32, Custom);
1080 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Custom);
1082 setLoadExtAction(ISD::EXTLOAD, MVT::v2f32, Legal);
1084 setOperationAction(ISD::BITCAST, MVT::v2i32, Custom);
1085 setOperationAction(ISD::BITCAST, MVT::v4i16, Custom);
1086 setOperationAction(ISD::BITCAST, MVT::v8i8, Custom);
1089 if (!TM.Options.UseSoftFloat && Subtarget->hasSSE41()) {
1090 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
1091 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
1092 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
1093 setOperationAction(ISD::FRINT, MVT::f32, Legal);
1094 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
1095 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
1096 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
1097 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
1098 setOperationAction(ISD::FRINT, MVT::f64, Legal);
1099 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
1101 setOperationAction(ISD::FFLOOR, MVT::v4f32, Legal);
1102 setOperationAction(ISD::FCEIL, MVT::v4f32, Legal);
1103 setOperationAction(ISD::FTRUNC, MVT::v4f32, Legal);
1104 setOperationAction(ISD::FRINT, MVT::v4f32, Legal);
1105 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Legal);
1106 setOperationAction(ISD::FFLOOR, MVT::v2f64, Legal);
1107 setOperationAction(ISD::FCEIL, MVT::v2f64, Legal);
1108 setOperationAction(ISD::FTRUNC, MVT::v2f64, Legal);
1109 setOperationAction(ISD::FRINT, MVT::v2f64, Legal);
1110 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Legal);
1112 // FIXME: Do we need to handle scalar-to-vector here?
1113 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
1115 setOperationAction(ISD::VSELECT, MVT::v2f64, Custom);
1116 setOperationAction(ISD::VSELECT, MVT::v2i64, Custom);
1117 setOperationAction(ISD::VSELECT, MVT::v4i32, Custom);
1118 setOperationAction(ISD::VSELECT, MVT::v4f32, Custom);
1119 setOperationAction(ISD::VSELECT, MVT::v8i16, Custom);
1120 // There is no BLENDI for byte vectors. We don't need to custom lower
1121 // some vselects for now.
1122 setOperationAction(ISD::VSELECT, MVT::v16i8, Legal);
1124 // SSE41 brings specific instructions for doing vector sign extend even in
1125 // cases where we don't have SRA.
1126 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i8, Custom);
1127 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i16, Custom);
1128 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i32, Custom);
1130 // i8 and i16 vectors are custom because the source register and source
1131 // source memory operand types are not the same width. f32 vectors are
1132 // custom since the immediate controlling the insert encodes additional
1134 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
1135 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
1136 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
1137 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
1139 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
1140 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
1141 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
1142 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
1144 // FIXME: these should be Legal, but that's only for the case where
1145 // the index is constant. For now custom expand to deal with that.
1146 if (Subtarget->is64Bit()) {
1147 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
1148 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
1152 if (Subtarget->hasSSE2()) {
1153 setOperationAction(ISD::SRL, MVT::v8i16, Custom);
1154 setOperationAction(ISD::SRL, MVT::v16i8, Custom);
1156 setOperationAction(ISD::SHL, MVT::v8i16, Custom);
1157 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
1159 setOperationAction(ISD::SRA, MVT::v8i16, Custom);
1160 setOperationAction(ISD::SRA, MVT::v16i8, Custom);
1162 // In the customized shift lowering, the legal cases in AVX2 will be
1164 setOperationAction(ISD::SRL, MVT::v2i64, Custom);
1165 setOperationAction(ISD::SRL, MVT::v4i32, Custom);
1167 setOperationAction(ISD::SHL, MVT::v2i64, Custom);
1168 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
1170 setOperationAction(ISD::SRA, MVT::v4i32, Custom);
1173 if (!TM.Options.UseSoftFloat && Subtarget->hasFp256()) {
1174 addRegisterClass(MVT::v32i8, &X86::VR256RegClass);
1175 addRegisterClass(MVT::v16i16, &X86::VR256RegClass);
1176 addRegisterClass(MVT::v8i32, &X86::VR256RegClass);
1177 addRegisterClass(MVT::v8f32, &X86::VR256RegClass);
1178 addRegisterClass(MVT::v4i64, &X86::VR256RegClass);
1179 addRegisterClass(MVT::v4f64, &X86::VR256RegClass);
1181 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
1182 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
1183 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
1185 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
1186 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
1187 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
1188 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
1189 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
1190 setOperationAction(ISD::FFLOOR, MVT::v8f32, Legal);
1191 setOperationAction(ISD::FCEIL, MVT::v8f32, Legal);
1192 setOperationAction(ISD::FTRUNC, MVT::v8f32, Legal);
1193 setOperationAction(ISD::FRINT, MVT::v8f32, Legal);
1194 setOperationAction(ISD::FNEARBYINT, MVT::v8f32, Legal);
1195 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
1196 setOperationAction(ISD::FABS, MVT::v8f32, Custom);
1198 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
1199 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
1200 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
1201 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
1202 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
1203 setOperationAction(ISD::FFLOOR, MVT::v4f64, Legal);
1204 setOperationAction(ISD::FCEIL, MVT::v4f64, Legal);
1205 setOperationAction(ISD::FTRUNC, MVT::v4f64, Legal);
1206 setOperationAction(ISD::FRINT, MVT::v4f64, Legal);
1207 setOperationAction(ISD::FNEARBYINT, MVT::v4f64, Legal);
1208 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
1209 setOperationAction(ISD::FABS, MVT::v4f64, Custom);
1211 // (fp_to_int:v8i16 (v8f32 ..)) requires the result type to be promoted
1212 // even though v8i16 is a legal type.
1213 setOperationAction(ISD::FP_TO_SINT, MVT::v8i16, Promote);
1214 setOperationAction(ISD::FP_TO_UINT, MVT::v8i16, Promote);
1215 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1217 setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Promote);
1218 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1219 setOperationAction(ISD::FP_ROUND, MVT::v4f32, Legal);
1221 setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Custom);
1222 setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Custom);
1224 setLoadExtAction(ISD::EXTLOAD, MVT::v4f32, Legal);
1226 setOperationAction(ISD::SRL, MVT::v16i16, Custom);
1227 setOperationAction(ISD::SRL, MVT::v32i8, Custom);
1229 setOperationAction(ISD::SHL, MVT::v16i16, Custom);
1230 setOperationAction(ISD::SHL, MVT::v32i8, Custom);
1232 setOperationAction(ISD::SRA, MVT::v16i16, Custom);
1233 setOperationAction(ISD::SRA, MVT::v32i8, Custom);
1235 setOperationAction(ISD::SETCC, MVT::v32i8, Custom);
1236 setOperationAction(ISD::SETCC, MVT::v16i16, Custom);
1237 setOperationAction(ISD::SETCC, MVT::v8i32, Custom);
1238 setOperationAction(ISD::SETCC, MVT::v4i64, Custom);
1240 setOperationAction(ISD::SELECT, MVT::v4f64, Custom);
1241 setOperationAction(ISD::SELECT, MVT::v4i64, Custom);
1242 setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
1244 setOperationAction(ISD::VSELECT, MVT::v4f64, Custom);
1245 setOperationAction(ISD::VSELECT, MVT::v4i64, Custom);
1246 setOperationAction(ISD::VSELECT, MVT::v8i32, Custom);
1247 setOperationAction(ISD::VSELECT, MVT::v8f32, Custom);
1249 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i64, Custom);
1250 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i32, Custom);
1251 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1252 setOperationAction(ISD::ZERO_EXTEND, MVT::v4i64, Custom);
1253 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i32, Custom);
1254 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i16, Custom);
1255 setOperationAction(ISD::ANY_EXTEND, MVT::v4i64, Custom);
1256 setOperationAction(ISD::ANY_EXTEND, MVT::v8i32, Custom);
1257 setOperationAction(ISD::ANY_EXTEND, MVT::v16i16, Custom);
1258 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1259 setOperationAction(ISD::TRUNCATE, MVT::v8i16, Custom);
1260 setOperationAction(ISD::TRUNCATE, MVT::v4i32, Custom);
1262 if (Subtarget->hasFMA() || Subtarget->hasFMA4()) {
1263 setOperationAction(ISD::FMA, MVT::v8f32, Legal);
1264 setOperationAction(ISD::FMA, MVT::v4f64, Legal);
1265 setOperationAction(ISD::FMA, MVT::v4f32, Legal);
1266 setOperationAction(ISD::FMA, MVT::v2f64, Legal);
1267 setOperationAction(ISD::FMA, MVT::f32, Legal);
1268 setOperationAction(ISD::FMA, MVT::f64, Legal);
1271 if (Subtarget->hasInt256()) {
1272 setOperationAction(ISD::ADD, MVT::v4i64, Legal);
1273 setOperationAction(ISD::ADD, MVT::v8i32, Legal);
1274 setOperationAction(ISD::ADD, MVT::v16i16, Legal);
1275 setOperationAction(ISD::ADD, MVT::v32i8, Legal);
1277 setOperationAction(ISD::SUB, MVT::v4i64, Legal);
1278 setOperationAction(ISD::SUB, MVT::v8i32, Legal);
1279 setOperationAction(ISD::SUB, MVT::v16i16, Legal);
1280 setOperationAction(ISD::SUB, MVT::v32i8, Legal);
1282 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1283 setOperationAction(ISD::MUL, MVT::v8i32, Legal);
1284 setOperationAction(ISD::MUL, MVT::v16i16, Legal);
1285 // Don't lower v32i8 because there is no 128-bit byte mul
1287 setOperationAction(ISD::UMUL_LOHI, MVT::v8i32, Custom);
1288 setOperationAction(ISD::SMUL_LOHI, MVT::v8i32, Custom);
1289 setOperationAction(ISD::MULHU, MVT::v16i16, Legal);
1290 setOperationAction(ISD::MULHS, MVT::v16i16, Legal);
1292 setOperationAction(ISD::VSELECT, MVT::v16i16, Custom);
1293 setOperationAction(ISD::VSELECT, MVT::v32i8, Legal);
1295 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
1296 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
1297 setOperationAction(ISD::ADD, MVT::v16i16, Custom);
1298 setOperationAction(ISD::ADD, MVT::v32i8, Custom);
1300 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
1301 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
1302 setOperationAction(ISD::SUB, MVT::v16i16, Custom);
1303 setOperationAction(ISD::SUB, MVT::v32i8, Custom);
1305 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1306 setOperationAction(ISD::MUL, MVT::v8i32, Custom);
1307 setOperationAction(ISD::MUL, MVT::v16i16, Custom);
1308 // Don't lower v32i8 because there is no 128-bit byte mul
1311 // In the customized shift lowering, the legal cases in AVX2 will be
1313 setOperationAction(ISD::SRL, MVT::v4i64, Custom);
1314 setOperationAction(ISD::SRL, MVT::v8i32, Custom);
1316 setOperationAction(ISD::SHL, MVT::v4i64, Custom);
1317 setOperationAction(ISD::SHL, MVT::v8i32, Custom);
1319 setOperationAction(ISD::SRA, MVT::v8i32, Custom);
1321 // Custom lower several nodes for 256-bit types.
1322 for (int i = MVT::FIRST_VECTOR_VALUETYPE;
1323 i <= MVT::LAST_VECTOR_VALUETYPE; ++i) {
1324 MVT VT = (MVT::SimpleValueType)i;
1326 // Extract subvector is special because the value type
1327 // (result) is 128-bit but the source is 256-bit wide.
1328 if (VT.is128BitVector())
1329 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1331 // Do not attempt to custom lower other non-256-bit vectors
1332 if (!VT.is256BitVector())
1335 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1336 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1337 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1338 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1339 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1340 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1341 setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
1344 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
1345 for (int i = MVT::v32i8; i != MVT::v4i64; ++i) {
1346 MVT VT = (MVT::SimpleValueType)i;
1348 // Do not attempt to promote non-256-bit vectors
1349 if (!VT.is256BitVector())
1352 setOperationAction(ISD::AND, VT, Promote);
1353 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
1354 setOperationAction(ISD::OR, VT, Promote);
1355 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
1356 setOperationAction(ISD::XOR, VT, Promote);
1357 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
1358 setOperationAction(ISD::LOAD, VT, Promote);
1359 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
1360 setOperationAction(ISD::SELECT, VT, Promote);
1361 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
1365 if (!TM.Options.UseSoftFloat && Subtarget->hasAVX512()) {
1366 addRegisterClass(MVT::v16i32, &X86::VR512RegClass);
1367 addRegisterClass(MVT::v16f32, &X86::VR512RegClass);
1368 addRegisterClass(MVT::v8i64, &X86::VR512RegClass);
1369 addRegisterClass(MVT::v8f64, &X86::VR512RegClass);
1371 addRegisterClass(MVT::i1, &X86::VK1RegClass);
1372 addRegisterClass(MVT::v8i1, &X86::VK8RegClass);
1373 addRegisterClass(MVT::v16i1, &X86::VK16RegClass);
1375 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
1376 setOperationAction(ISD::SETCC, MVT::i1, Custom);
1377 setOperationAction(ISD::XOR, MVT::i1, Legal);
1378 setOperationAction(ISD::OR, MVT::i1, Legal);
1379 setOperationAction(ISD::AND, MVT::i1, Legal);
1380 setLoadExtAction(ISD::EXTLOAD, MVT::v8f32, Legal);
1381 setOperationAction(ISD::LOAD, MVT::v16f32, Legal);
1382 setOperationAction(ISD::LOAD, MVT::v8f64, Legal);
1383 setOperationAction(ISD::LOAD, MVT::v8i64, Legal);
1384 setOperationAction(ISD::LOAD, MVT::v16i32, Legal);
1385 setOperationAction(ISD::LOAD, MVT::v16i1, Legal);
1387 setOperationAction(ISD::FADD, MVT::v16f32, Legal);
1388 setOperationAction(ISD::FSUB, MVT::v16f32, Legal);
1389 setOperationAction(ISD::FMUL, MVT::v16f32, Legal);
1390 setOperationAction(ISD::FDIV, MVT::v16f32, Legal);
1391 setOperationAction(ISD::FSQRT, MVT::v16f32, Legal);
1392 setOperationAction(ISD::FNEG, MVT::v16f32, Custom);
1394 setOperationAction(ISD::FADD, MVT::v8f64, Legal);
1395 setOperationAction(ISD::FSUB, MVT::v8f64, Legal);
1396 setOperationAction(ISD::FMUL, MVT::v8f64, Legal);
1397 setOperationAction(ISD::FDIV, MVT::v8f64, Legal);
1398 setOperationAction(ISD::FSQRT, MVT::v8f64, Legal);
1399 setOperationAction(ISD::FNEG, MVT::v8f64, Custom);
1400 setOperationAction(ISD::FMA, MVT::v8f64, Legal);
1401 setOperationAction(ISD::FMA, MVT::v16f32, Legal);
1403 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Legal);
1404 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Legal);
1405 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Legal);
1406 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Legal);
1407 if (Subtarget->is64Bit()) {
1408 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Legal);
1409 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Legal);
1410 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Legal);
1411 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Legal);
1413 setOperationAction(ISD::FP_TO_SINT, MVT::v16i32, Legal);
1414 setOperationAction(ISD::FP_TO_UINT, MVT::v16i32, Legal);
1415 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1416 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1417 setOperationAction(ISD::SINT_TO_FP, MVT::v16i32, Legal);
1418 setOperationAction(ISD::UINT_TO_FP, MVT::v16i32, Legal);
1419 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1420 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1421 setOperationAction(ISD::FP_ROUND, MVT::v8f32, Legal);
1422 setOperationAction(ISD::FP_EXTEND, MVT::v8f32, Legal);
1424 setOperationAction(ISD::TRUNCATE, MVT::i1, Custom);
1425 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1426 setOperationAction(ISD::TRUNCATE, MVT::v8i32, Custom);
1427 setOperationAction(ISD::TRUNCATE, MVT::v8i1, Custom);
1428 setOperationAction(ISD::TRUNCATE, MVT::v16i1, Custom);
1429 setOperationAction(ISD::TRUNCATE, MVT::v16i16, Custom);
1430 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i32, Custom);
1431 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i64, Custom);
1432 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i32, Custom);
1433 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i64, Custom);
1434 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i8, Custom);
1435 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i16, Custom);
1436 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1438 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f64, Custom);
1439 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i64, Custom);
1440 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16f32, Custom);
1441 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i32, Custom);
1442 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i1, Custom);
1443 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i1, Legal);
1445 setOperationAction(ISD::SETCC, MVT::v16i1, Custom);
1446 setOperationAction(ISD::SETCC, MVT::v8i1, Custom);
1448 setOperationAction(ISD::MUL, MVT::v8i64, Custom);
1450 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i1, Custom);
1451 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i1, Custom);
1452 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i1, Custom);
1453 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i1, Custom);
1454 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i1, Custom);
1455 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i1, Custom);
1456 setOperationAction(ISD::SELECT, MVT::v8f64, Custom);
1457 setOperationAction(ISD::SELECT, MVT::v8i64, Custom);
1458 setOperationAction(ISD::SELECT, MVT::v16f32, Custom);
1460 setOperationAction(ISD::ADD, MVT::v8i64, Legal);
1461 setOperationAction(ISD::ADD, MVT::v16i32, Legal);
1463 setOperationAction(ISD::SUB, MVT::v8i64, Legal);
1464 setOperationAction(ISD::SUB, MVT::v16i32, Legal);
1466 setOperationAction(ISD::MUL, MVT::v16i32, Legal);
1468 setOperationAction(ISD::SRL, MVT::v8i64, Custom);
1469 setOperationAction(ISD::SRL, MVT::v16i32, Custom);
1471 setOperationAction(ISD::SHL, MVT::v8i64, Custom);
1472 setOperationAction(ISD::SHL, MVT::v16i32, Custom);
1474 setOperationAction(ISD::SRA, MVT::v8i64, Custom);
1475 setOperationAction(ISD::SRA, MVT::v16i32, Custom);
1477 setOperationAction(ISD::AND, MVT::v8i64, Legal);
1478 setOperationAction(ISD::OR, MVT::v8i64, Legal);
1479 setOperationAction(ISD::XOR, MVT::v8i64, Legal);
1480 setOperationAction(ISD::AND, MVT::v16i32, Legal);
1481 setOperationAction(ISD::OR, MVT::v16i32, Legal);
1482 setOperationAction(ISD::XOR, MVT::v16i32, Legal);
1484 if (Subtarget->hasCDI()) {
1485 setOperationAction(ISD::CTLZ, MVT::v8i64, Legal);
1486 setOperationAction(ISD::CTLZ, MVT::v16i32, Legal);
1489 // Custom lower several nodes.
1490 for (int i = MVT::FIRST_VECTOR_VALUETYPE;
1491 i <= MVT::LAST_VECTOR_VALUETYPE; ++i) {
1492 MVT VT = (MVT::SimpleValueType)i;
1494 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1495 // Extract subvector is special because the value type
1496 // (result) is 256/128-bit but the source is 512-bit wide.
1497 if (VT.is128BitVector() || VT.is256BitVector())
1498 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1500 if (VT.getVectorElementType() == MVT::i1)
1501 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
1503 // Do not attempt to custom lower other non-512-bit vectors
1504 if (!VT.is512BitVector())
1507 if ( EltSize >= 32) {
1508 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1509 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1510 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1511 setOperationAction(ISD::VSELECT, VT, Legal);
1512 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1513 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1514 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1517 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1518 MVT VT = (MVT::SimpleValueType)i;
1520 // Do not attempt to promote non-256-bit vectors
1521 if (!VT.is512BitVector())
1524 setOperationAction(ISD::SELECT, VT, Promote);
1525 AddPromotedToType (ISD::SELECT, VT, MVT::v8i64);
1529 if (!TM.Options.UseSoftFloat && Subtarget->hasBWI()) {
1530 addRegisterClass(MVT::v32i16, &X86::VR512RegClass);
1531 addRegisterClass(MVT::v64i8, &X86::VR512RegClass);
1533 addRegisterClass(MVT::v32i1, &X86::VK32RegClass);
1534 addRegisterClass(MVT::v64i1, &X86::VK64RegClass);
1536 setOperationAction(ISD::LOAD, MVT::v32i16, Legal);
1537 setOperationAction(ISD::LOAD, MVT::v64i8, Legal);
1538 setOperationAction(ISD::SETCC, MVT::v32i1, Custom);
1539 setOperationAction(ISD::SETCC, MVT::v64i1, Custom);
1541 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1542 const MVT VT = (MVT::SimpleValueType)i;
1544 const unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1546 // Do not attempt to promote non-256-bit vectors
1547 if (!VT.is512BitVector())
1550 if ( EltSize < 32) {
1551 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1552 setOperationAction(ISD::VSELECT, VT, Legal);
1557 if (!TM.Options.UseSoftFloat && Subtarget->hasVLX()) {
1558 addRegisterClass(MVT::v4i1, &X86::VK4RegClass);
1559 addRegisterClass(MVT::v2i1, &X86::VK2RegClass);
1561 setOperationAction(ISD::SETCC, MVT::v4i1, Custom);
1562 setOperationAction(ISD::SETCC, MVT::v2i1, Custom);
1565 // SIGN_EXTEND_INREGs are evaluated by the extend type. Handle the expansion
1566 // of this type with custom code.
1567 for (int VT = MVT::FIRST_VECTOR_VALUETYPE;
1568 VT != MVT::LAST_VECTOR_VALUETYPE; VT++) {
1569 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,
1573 // We want to custom lower some of our intrinsics.
1574 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1575 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
1576 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
1577 if (!Subtarget->is64Bit())
1578 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
1580 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1581 // handle type legalization for these operations here.
1583 // FIXME: We really should do custom legalization for addition and
1584 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1585 // than generic legalization for 64-bit multiplication-with-overflow, though.
1586 for (unsigned i = 0, e = 3+Subtarget->is64Bit(); i != e; ++i) {
1587 // Add/Sub/Mul with overflow operations are custom lowered.
1589 setOperationAction(ISD::SADDO, VT, Custom);
1590 setOperationAction(ISD::UADDO, VT, Custom);
1591 setOperationAction(ISD::SSUBO, VT, Custom);
1592 setOperationAction(ISD::USUBO, VT, Custom);
1593 setOperationAction(ISD::SMULO, VT, Custom);
1594 setOperationAction(ISD::UMULO, VT, Custom);
1597 // There are no 8-bit 3-address imul/mul instructions
1598 setOperationAction(ISD::SMULO, MVT::i8, Expand);
1599 setOperationAction(ISD::UMULO, MVT::i8, Expand);
1601 if (!Subtarget->is64Bit()) {
1602 // These libcalls are not available in 32-bit.
1603 setLibcallName(RTLIB::SHL_I128, nullptr);
1604 setLibcallName(RTLIB::SRL_I128, nullptr);
1605 setLibcallName(RTLIB::SRA_I128, nullptr);
1608 // Combine sin / cos into one node or libcall if possible.
1609 if (Subtarget->hasSinCos()) {
1610 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
1611 setLibcallName(RTLIB::SINCOS_F64, "sincos");
1612 if (Subtarget->isTargetDarwin()) {
1613 // For MacOSX, we don't want to the normal expansion of a libcall to
1614 // sincos. We want to issue a libcall to __sincos_stret to avoid memory
1616 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
1617 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
1621 if (Subtarget->isTargetWin64()) {
1622 setOperationAction(ISD::SDIV, MVT::i128, Custom);
1623 setOperationAction(ISD::UDIV, MVT::i128, Custom);
1624 setOperationAction(ISD::SREM, MVT::i128, Custom);
1625 setOperationAction(ISD::UREM, MVT::i128, Custom);
1626 setOperationAction(ISD::SDIVREM, MVT::i128, Custom);
1627 setOperationAction(ISD::UDIVREM, MVT::i128, Custom);
1630 // We have target-specific dag combine patterns for the following nodes:
1631 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1632 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1633 setTargetDAGCombine(ISD::VSELECT);
1634 setTargetDAGCombine(ISD::SELECT);
1635 setTargetDAGCombine(ISD::SHL);
1636 setTargetDAGCombine(ISD::SRA);
1637 setTargetDAGCombine(ISD::SRL);
1638 setTargetDAGCombine(ISD::OR);
1639 setTargetDAGCombine(ISD::AND);
1640 setTargetDAGCombine(ISD::ADD);
1641 setTargetDAGCombine(ISD::FADD);
1642 setTargetDAGCombine(ISD::FSUB);
1643 setTargetDAGCombine(ISD::FMA);
1644 setTargetDAGCombine(ISD::SUB);
1645 setTargetDAGCombine(ISD::LOAD);
1646 setTargetDAGCombine(ISD::STORE);
1647 setTargetDAGCombine(ISD::ZERO_EXTEND);
1648 setTargetDAGCombine(ISD::ANY_EXTEND);
1649 setTargetDAGCombine(ISD::SIGN_EXTEND);
1650 setTargetDAGCombine(ISD::SIGN_EXTEND_INREG);
1651 setTargetDAGCombine(ISD::TRUNCATE);
1652 setTargetDAGCombine(ISD::SINT_TO_FP);
1653 setTargetDAGCombine(ISD::SETCC);
1654 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
1655 setTargetDAGCombine(ISD::BUILD_VECTOR);
1656 if (Subtarget->is64Bit())
1657 setTargetDAGCombine(ISD::MUL);
1658 setTargetDAGCombine(ISD::XOR);
1660 computeRegisterProperties();
1662 // On Darwin, -Os means optimize for size without hurting performance,
1663 // do not reduce the limit.
1664 MaxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1665 MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 16 : 8;
1666 MaxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1667 MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1668 MaxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1669 MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1670 setPrefLoopAlignment(4); // 2^4 bytes.
1672 // Predictable cmov don't hurt on atom because it's in-order.
1673 PredictableSelectIsExpensive = !Subtarget->isAtom();
1675 setPrefFunctionAlignment(4); // 2^4 bytes.
1677 verifyIntrinsicTables();
1680 // This has so far only been implemented for 64-bit MachO.
1681 bool X86TargetLowering::useLoadStackGuardNode() const {
1682 return Subtarget->getTargetTriple().getObjectFormat() == Triple::MachO &&
1683 Subtarget->is64Bit();
1686 TargetLoweringBase::LegalizeTypeAction
1687 X86TargetLowering::getPreferredVectorAction(EVT VT) const {
1688 if (ExperimentalVectorWideningLegalization &&
1689 VT.getVectorNumElements() != 1 &&
1690 VT.getVectorElementType().getSimpleVT() != MVT::i1)
1691 return TypeWidenVector;
1693 return TargetLoweringBase::getPreferredVectorAction(VT);
1696 EVT X86TargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
1698 return Subtarget->hasAVX512() ? MVT::i1: MVT::i8;
1700 const unsigned NumElts = VT.getVectorNumElements();
1701 const EVT EltVT = VT.getVectorElementType();
1702 if (VT.is512BitVector()) {
1703 if (Subtarget->hasAVX512())
1704 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1705 EltVT == MVT::f32 || EltVT == MVT::f64)
1707 case 8: return MVT::v8i1;
1708 case 16: return MVT::v16i1;
1710 if (Subtarget->hasBWI())
1711 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1713 case 32: return MVT::v32i1;
1714 case 64: return MVT::v64i1;
1718 if (VT.is256BitVector() || VT.is128BitVector()) {
1719 if (Subtarget->hasVLX())
1720 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1721 EltVT == MVT::f32 || EltVT == MVT::f64)
1723 case 2: return MVT::v2i1;
1724 case 4: return MVT::v4i1;
1725 case 8: return MVT::v8i1;
1727 if (Subtarget->hasBWI() && Subtarget->hasVLX())
1728 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1730 case 8: return MVT::v8i1;
1731 case 16: return MVT::v16i1;
1732 case 32: return MVT::v32i1;
1736 return VT.changeVectorElementTypeToInteger();
1739 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1740 /// the desired ByVal argument alignment.
1741 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1744 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1745 if (VTy->getBitWidth() == 128)
1747 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1748 unsigned EltAlign = 0;
1749 getMaxByValAlign(ATy->getElementType(), EltAlign);
1750 if (EltAlign > MaxAlign)
1751 MaxAlign = EltAlign;
1752 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1753 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1754 unsigned EltAlign = 0;
1755 getMaxByValAlign(STy->getElementType(i), EltAlign);
1756 if (EltAlign > MaxAlign)
1757 MaxAlign = EltAlign;
1764 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1765 /// function arguments in the caller parameter area. For X86, aggregates
1766 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1767 /// are at 4-byte boundaries.
1768 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty) const {
1769 if (Subtarget->is64Bit()) {
1770 // Max of 8 and alignment of type.
1771 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1778 if (Subtarget->hasSSE1())
1779 getMaxByValAlign(Ty, Align);
1783 /// getOptimalMemOpType - Returns the target specific optimal type for load
1784 /// and store operations as a result of memset, memcpy, and memmove
1785 /// lowering. If DstAlign is zero that means it's safe to destination
1786 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1787 /// means there isn't a need to check it against alignment requirement,
1788 /// probably because the source does not need to be loaded. If 'IsMemset' is
1789 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
1790 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
1791 /// source is constant so it does not need to be loaded.
1792 /// It returns EVT::Other if the type should be determined using generic
1793 /// target-independent logic.
1795 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1796 unsigned DstAlign, unsigned SrcAlign,
1797 bool IsMemset, bool ZeroMemset,
1799 MachineFunction &MF) const {
1800 const Function *F = MF.getFunction();
1801 if ((!IsMemset || ZeroMemset) &&
1802 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
1803 Attribute::NoImplicitFloat)) {
1805 (Subtarget->isUnalignedMemAccessFast() ||
1806 ((DstAlign == 0 || DstAlign >= 16) &&
1807 (SrcAlign == 0 || SrcAlign >= 16)))) {
1809 if (Subtarget->hasInt256())
1811 if (Subtarget->hasFp256())
1814 if (Subtarget->hasSSE2())
1816 if (Subtarget->hasSSE1())
1818 } else if (!MemcpyStrSrc && Size >= 8 &&
1819 !Subtarget->is64Bit() &&
1820 Subtarget->hasSSE2()) {
1821 // Do not use f64 to lower memcpy if source is string constant. It's
1822 // better to use i32 to avoid the loads.
1826 if (Subtarget->is64Bit() && Size >= 8)
1831 bool X86TargetLowering::isSafeMemOpType(MVT VT) const {
1833 return X86ScalarSSEf32;
1834 else if (VT == MVT::f64)
1835 return X86ScalarSSEf64;
1840 X86TargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
1845 *Fast = Subtarget->isUnalignedMemAccessFast();
1849 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1850 /// current function. The returned value is a member of the
1851 /// MachineJumpTableInfo::JTEntryKind enum.
1852 unsigned X86TargetLowering::getJumpTableEncoding() const {
1853 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1855 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1856 Subtarget->isPICStyleGOT())
1857 return MachineJumpTableInfo::EK_Custom32;
1859 // Otherwise, use the normal jump table encoding heuristics.
1860 return TargetLowering::getJumpTableEncoding();
1864 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1865 const MachineBasicBlock *MBB,
1866 unsigned uid,MCContext &Ctx) const{
1867 assert(MBB->getParent()->getTarget().getRelocationModel() == Reloc::PIC_ &&
1868 Subtarget->isPICStyleGOT());
1869 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1871 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1872 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1875 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1877 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1878 SelectionDAG &DAG) const {
1879 if (!Subtarget->is64Bit())
1880 // This doesn't have SDLoc associated with it, but is not really the
1881 // same as a Register.
1882 return DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), getPointerTy());
1886 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1887 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1889 const MCExpr *X86TargetLowering::
1890 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1891 MCContext &Ctx) const {
1892 // X86-64 uses RIP relative addressing based on the jump table label.
1893 if (Subtarget->isPICStyleRIPRel())
1894 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1896 // Otherwise, the reference is relative to the PIC base.
1897 return MCSymbolRefExpr::Create(MF->getPICBaseSymbol(), Ctx);
1900 // FIXME: Why this routine is here? Move to RegInfo!
1901 std::pair<const TargetRegisterClass*, uint8_t>
1902 X86TargetLowering::findRepresentativeClass(MVT VT) const{
1903 const TargetRegisterClass *RRC = nullptr;
1905 switch (VT.SimpleTy) {
1907 return TargetLowering::findRepresentativeClass(VT);
1908 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
1909 RRC = Subtarget->is64Bit() ? &X86::GR64RegClass : &X86::GR32RegClass;
1912 RRC = &X86::VR64RegClass;
1914 case MVT::f32: case MVT::f64:
1915 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1916 case MVT::v4f32: case MVT::v2f64:
1917 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
1919 RRC = &X86::VR128RegClass;
1922 return std::make_pair(RRC, Cost);
1925 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1926 unsigned &Offset) const {
1927 if (!Subtarget->isTargetLinux())
1930 if (Subtarget->is64Bit()) {
1931 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1933 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1945 bool X86TargetLowering::isNoopAddrSpaceCast(unsigned SrcAS,
1946 unsigned DestAS) const {
1947 assert(SrcAS != DestAS && "Expected different address spaces!");
1949 return SrcAS < 256 && DestAS < 256;
1952 //===----------------------------------------------------------------------===//
1953 // Return Value Calling Convention Implementation
1954 //===----------------------------------------------------------------------===//
1956 #include "X86GenCallingConv.inc"
1959 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv,
1960 MachineFunction &MF, bool isVarArg,
1961 const SmallVectorImpl<ISD::OutputArg> &Outs,
1962 LLVMContext &Context) const {
1963 SmallVector<CCValAssign, 16> RVLocs;
1964 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
1965 return CCInfo.CheckReturn(Outs, RetCC_X86);
1968 const MCPhysReg *X86TargetLowering::getScratchRegisters(CallingConv::ID) const {
1969 static const MCPhysReg ScratchRegs[] = { X86::R11, 0 };
1974 X86TargetLowering::LowerReturn(SDValue Chain,
1975 CallingConv::ID CallConv, bool isVarArg,
1976 const SmallVectorImpl<ISD::OutputArg> &Outs,
1977 const SmallVectorImpl<SDValue> &OutVals,
1978 SDLoc dl, SelectionDAG &DAG) const {
1979 MachineFunction &MF = DAG.getMachineFunction();
1980 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1982 SmallVector<CCValAssign, 16> RVLocs;
1983 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, *DAG.getContext());
1984 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1987 SmallVector<SDValue, 6> RetOps;
1988 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1989 // Operand #1 = Bytes To Pop
1990 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
1993 // Copy the result values into the output registers.
1994 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1995 CCValAssign &VA = RVLocs[i];
1996 assert(VA.isRegLoc() && "Can only return in registers!");
1997 SDValue ValToCopy = OutVals[i];
1998 EVT ValVT = ValToCopy.getValueType();
2000 // Promote values to the appropriate types
2001 if (VA.getLocInfo() == CCValAssign::SExt)
2002 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
2003 else if (VA.getLocInfo() == CCValAssign::ZExt)
2004 ValToCopy = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), ValToCopy);
2005 else if (VA.getLocInfo() == CCValAssign::AExt)
2006 ValToCopy = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), ValToCopy);
2007 else if (VA.getLocInfo() == CCValAssign::BCvt)
2008 ValToCopy = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), ValToCopy);
2010 assert(VA.getLocInfo() != CCValAssign::FPExt &&
2011 "Unexpected FP-extend for return value.");
2013 // If this is x86-64, and we disabled SSE, we can't return FP values,
2014 // or SSE or MMX vectors.
2015 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
2016 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
2017 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
2018 report_fatal_error("SSE register return with SSE disabled");
2020 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
2021 // llvm-gcc has never done it right and no one has noticed, so this
2022 // should be OK for now.
2023 if (ValVT == MVT::f64 &&
2024 (Subtarget->is64Bit() && !Subtarget->hasSSE2()))
2025 report_fatal_error("SSE2 register return with SSE2 disabled");
2027 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
2028 // the RET instruction and handled by the FP Stackifier.
2029 if (VA.getLocReg() == X86::FP0 ||
2030 VA.getLocReg() == X86::FP1) {
2031 // If this is a copy from an xmm register to ST(0), use an FPExtend to
2032 // change the value to the FP stack register class.
2033 if (isScalarFPTypeInSSEReg(VA.getValVT()))
2034 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
2035 RetOps.push_back(ValToCopy);
2036 // Don't emit a copytoreg.
2040 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
2041 // which is returned in RAX / RDX.
2042 if (Subtarget->is64Bit()) {
2043 if (ValVT == MVT::x86mmx) {
2044 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
2045 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ValToCopy);
2046 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
2048 // If we don't have SSE2 available, convert to v4f32 so the generated
2049 // register is legal.
2050 if (!Subtarget->hasSSE2())
2051 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32,ValToCopy);
2056 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
2057 Flag = Chain.getValue(1);
2058 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2061 // The x86-64 ABIs require that for returning structs by value we copy
2062 // the sret argument into %rax/%eax (depending on ABI) for the return.
2063 // Win32 requires us to put the sret argument to %eax as well.
2064 // We saved the argument into a virtual register in the entry block,
2065 // so now we copy the value out and into %rax/%eax.
2066 if (DAG.getMachineFunction().getFunction()->hasStructRetAttr() &&
2067 (Subtarget->is64Bit() || Subtarget->isTargetKnownWindowsMSVC())) {
2068 MachineFunction &MF = DAG.getMachineFunction();
2069 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2070 unsigned Reg = FuncInfo->getSRetReturnReg();
2072 "SRetReturnReg should have been set in LowerFormalArguments().");
2073 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
2076 = (Subtarget->is64Bit() && !Subtarget->isTarget64BitILP32()) ?
2077 X86::RAX : X86::EAX;
2078 Chain = DAG.getCopyToReg(Chain, dl, RetValReg, Val, Flag);
2079 Flag = Chain.getValue(1);
2081 // RAX/EAX now acts like a return value.
2082 RetOps.push_back(DAG.getRegister(RetValReg, getPointerTy()));
2085 RetOps[0] = Chain; // Update chain.
2087 // Add the flag if we have it.
2089 RetOps.push_back(Flag);
2091 return DAG.getNode(X86ISD::RET_FLAG, dl, MVT::Other, RetOps);
2094 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2095 if (N->getNumValues() != 1)
2097 if (!N->hasNUsesOfValue(1, 0))
2100 SDValue TCChain = Chain;
2101 SDNode *Copy = *N->use_begin();
2102 if (Copy->getOpcode() == ISD::CopyToReg) {
2103 // If the copy has a glue operand, we conservatively assume it isn't safe to
2104 // perform a tail call.
2105 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2107 TCChain = Copy->getOperand(0);
2108 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
2111 bool HasRet = false;
2112 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2114 if (UI->getOpcode() != X86ISD::RET_FLAG)
2116 // If we are returning more than one value, we can definitely
2117 // not make a tail call see PR19530
2118 if (UI->getNumOperands() > 4)
2120 if (UI->getNumOperands() == 4 &&
2121 UI->getOperand(UI->getNumOperands()-1).getValueType() != MVT::Glue)
2134 X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
2135 ISD::NodeType ExtendKind) const {
2137 // TODO: Is this also valid on 32-bit?
2138 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
2139 ReturnMVT = MVT::i8;
2141 ReturnMVT = MVT::i32;
2143 EVT MinVT = getRegisterType(Context, ReturnMVT);
2144 return VT.bitsLT(MinVT) ? MinVT : VT;
2147 /// LowerCallResult - Lower the result values of a call into the
2148 /// appropriate copies out of appropriate physical registers.
2151 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2152 CallingConv::ID CallConv, bool isVarArg,
2153 const SmallVectorImpl<ISD::InputArg> &Ins,
2154 SDLoc dl, SelectionDAG &DAG,
2155 SmallVectorImpl<SDValue> &InVals) const {
2157 // Assign locations to each value returned by this call.
2158 SmallVector<CCValAssign, 16> RVLocs;
2159 bool Is64Bit = Subtarget->is64Bit();
2160 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2162 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2164 // Copy all of the result registers out of their specified physreg.
2165 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2166 CCValAssign &VA = RVLocs[i];
2167 EVT CopyVT = VA.getValVT();
2169 // If this is x86-64, and we disabled SSE, we can't return FP values
2170 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
2171 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
2172 report_fatal_error("SSE register return with SSE disabled");
2175 // If we prefer to use the value in xmm registers, copy it out as f80 and
2176 // use a truncate to move it from fp stack reg to xmm reg.
2177 if ((VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) &&
2178 isScalarFPTypeInSSEReg(VA.getValVT()))
2181 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
2182 CopyVT, InFlag).getValue(1);
2183 SDValue Val = Chain.getValue(0);
2185 if (CopyVT != VA.getValVT())
2186 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
2187 // This truncation won't change the value.
2188 DAG.getIntPtrConstant(1));
2190 InFlag = Chain.getValue(2);
2191 InVals.push_back(Val);
2197 //===----------------------------------------------------------------------===//
2198 // C & StdCall & Fast Calling Convention implementation
2199 //===----------------------------------------------------------------------===//
2200 // StdCall calling convention seems to be standard for many Windows' API
2201 // routines and around. It differs from C calling convention just a little:
2202 // callee should clean up the stack, not caller. Symbols should be also
2203 // decorated in some fancy way :) It doesn't support any vector arguments.
2204 // For info on fast calling convention see Fast Calling Convention (tail call)
2205 // implementation LowerX86_32FastCCCallTo.
2207 /// CallIsStructReturn - Determines whether a call uses struct return
2209 enum StructReturnType {
2214 static StructReturnType
2215 callIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
2217 return NotStructReturn;
2219 const ISD::ArgFlagsTy &Flags = Outs[0].Flags;
2220 if (!Flags.isSRet())
2221 return NotStructReturn;
2222 if (Flags.isInReg())
2223 return RegStructReturn;
2224 return StackStructReturn;
2227 /// ArgsAreStructReturn - Determines whether a function uses struct
2228 /// return semantics.
2229 static StructReturnType
2230 argsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
2232 return NotStructReturn;
2234 const ISD::ArgFlagsTy &Flags = Ins[0].Flags;
2235 if (!Flags.isSRet())
2236 return NotStructReturn;
2237 if (Flags.isInReg())
2238 return RegStructReturn;
2239 return StackStructReturn;
2242 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
2243 /// by "Src" to address "Dst" with size and alignment information specified by
2244 /// the specific parameter attribute. The copy will be passed as a byval
2245 /// function parameter.
2247 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2248 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2250 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
2252 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2253 /*isVolatile*/false, /*AlwaysInline=*/true,
2254 MachinePointerInfo(), MachinePointerInfo());
2257 /// IsTailCallConvention - Return true if the calling convention is one that
2258 /// supports tail call optimization.
2259 static bool IsTailCallConvention(CallingConv::ID CC) {
2260 return (CC == CallingConv::Fast || CC == CallingConv::GHC ||
2261 CC == CallingConv::HiPE);
2264 /// \brief Return true if the calling convention is a C calling convention.
2265 static bool IsCCallConvention(CallingConv::ID CC) {
2266 return (CC == CallingConv::C || CC == CallingConv::X86_64_Win64 ||
2267 CC == CallingConv::X86_64_SysV);
2270 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2271 if (!CI->isTailCall() || getTargetMachine().Options.DisableTailCalls)
2275 CallingConv::ID CalleeCC = CS.getCallingConv();
2276 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
2282 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
2283 /// a tailcall target by changing its ABI.
2284 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC,
2285 bool GuaranteedTailCallOpt) {
2286 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
2290 X86TargetLowering::LowerMemArgument(SDValue Chain,
2291 CallingConv::ID CallConv,
2292 const SmallVectorImpl<ISD::InputArg> &Ins,
2293 SDLoc dl, SelectionDAG &DAG,
2294 const CCValAssign &VA,
2295 MachineFrameInfo *MFI,
2297 // Create the nodes corresponding to a load from this parameter slot.
2298 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2299 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(
2300 CallConv, DAG.getTarget().Options.GuaranteedTailCallOpt);
2301 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
2304 // If value is passed by pointer we have address passed instead of the value
2306 if (VA.getLocInfo() == CCValAssign::Indirect)
2307 ValVT = VA.getLocVT();
2309 ValVT = VA.getValVT();
2311 // FIXME: For now, all byval parameter objects are marked mutable. This can be
2312 // changed with more analysis.
2313 // In case of tail call optimization mark all arguments mutable. Since they
2314 // could be overwritten by lowering of arguments in case of a tail call.
2315 if (Flags.isByVal()) {
2316 unsigned Bytes = Flags.getByValSize();
2317 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2318 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
2319 return DAG.getFrameIndex(FI, getPointerTy());
2321 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2322 VA.getLocMemOffset(), isImmutable);
2323 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2324 return DAG.getLoad(ValVT, dl, Chain, FIN,
2325 MachinePointerInfo::getFixedStack(FI),
2326 false, false, false, 0);
2330 // FIXME: Get this from tablegen.
2331 static ArrayRef<MCPhysReg> get64BitArgumentGPRs(CallingConv::ID CallConv,
2332 const X86Subtarget *Subtarget) {
2333 assert(Subtarget->is64Bit());
2335 if (Subtarget->isCallingConvWin64(CallConv)) {
2336 static const MCPhysReg GPR64ArgRegsWin64[] = {
2337 X86::RCX, X86::RDX, X86::R8, X86::R9
2339 return makeArrayRef(std::begin(GPR64ArgRegsWin64), std::end(GPR64ArgRegsWin64));
2342 static const MCPhysReg GPR64ArgRegs64Bit[] = {
2343 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
2345 return makeArrayRef(std::begin(GPR64ArgRegs64Bit), std::end(GPR64ArgRegs64Bit));
2348 // FIXME: Get this from tablegen.
2349 static ArrayRef<MCPhysReg> get64BitArgumentXMMs(MachineFunction &MF,
2350 CallingConv::ID CallConv,
2351 const X86Subtarget *Subtarget) {
2352 assert(Subtarget->is64Bit());
2353 if (Subtarget->isCallingConvWin64(CallConv)) {
2354 // The XMM registers which might contain var arg parameters are shadowed
2355 // in their paired GPR. So we only need to save the GPR to their home
2357 // TODO: __vectorcall will change this.
2361 const Function *Fn = MF.getFunction();
2362 bool NoImplicitFloatOps = Fn->getAttributes().
2363 hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat);
2364 assert(!(MF.getTarget().Options.UseSoftFloat && NoImplicitFloatOps) &&
2365 "SSE register cannot be used when SSE is disabled!");
2366 if (MF.getTarget().Options.UseSoftFloat || NoImplicitFloatOps ||
2367 !Subtarget->hasSSE1())
2368 // Kernel mode asks for SSE to be disabled, so there are no XMM argument
2372 static const MCPhysReg XMMArgRegs64Bit[] = {
2373 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2374 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2376 return makeArrayRef(std::begin(XMMArgRegs64Bit), std::end(XMMArgRegs64Bit));
2380 X86TargetLowering::LowerFormalArguments(SDValue Chain,
2381 CallingConv::ID CallConv,
2383 const SmallVectorImpl<ISD::InputArg> &Ins,
2386 SmallVectorImpl<SDValue> &InVals)
2388 MachineFunction &MF = DAG.getMachineFunction();
2389 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2391 const Function* Fn = MF.getFunction();
2392 if (Fn->hasExternalLinkage() &&
2393 Subtarget->isTargetCygMing() &&
2394 Fn->getName() == "main")
2395 FuncInfo->setForceFramePointer(true);
2397 MachineFrameInfo *MFI = MF.getFrameInfo();
2398 bool Is64Bit = Subtarget->is64Bit();
2399 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2401 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2402 "Var args not supported with calling convention fastcc, ghc or hipe");
2404 // Assign locations to all of the incoming arguments.
2405 SmallVector<CCValAssign, 16> ArgLocs;
2406 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2408 // Allocate shadow area for Win64
2410 CCInfo.AllocateStack(32, 8);
2412 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
2414 unsigned LastVal = ~0U;
2416 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2417 CCValAssign &VA = ArgLocs[i];
2418 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
2420 assert(VA.getValNo() != LastVal &&
2421 "Don't support value assigned to multiple locs yet");
2423 LastVal = VA.getValNo();
2425 if (VA.isRegLoc()) {
2426 EVT RegVT = VA.getLocVT();
2427 const TargetRegisterClass *RC;
2428 if (RegVT == MVT::i32)
2429 RC = &X86::GR32RegClass;
2430 else if (Is64Bit && RegVT == MVT::i64)
2431 RC = &X86::GR64RegClass;
2432 else if (RegVT == MVT::f32)
2433 RC = &X86::FR32RegClass;
2434 else if (RegVT == MVT::f64)
2435 RC = &X86::FR64RegClass;
2436 else if (RegVT.is512BitVector())
2437 RC = &X86::VR512RegClass;
2438 else if (RegVT.is256BitVector())
2439 RC = &X86::VR256RegClass;
2440 else if (RegVT.is128BitVector())
2441 RC = &X86::VR128RegClass;
2442 else if (RegVT == MVT::x86mmx)
2443 RC = &X86::VR64RegClass;
2444 else if (RegVT == MVT::i1)
2445 RC = &X86::VK1RegClass;
2446 else if (RegVT == MVT::v8i1)
2447 RC = &X86::VK8RegClass;
2448 else if (RegVT == MVT::v16i1)
2449 RC = &X86::VK16RegClass;
2450 else if (RegVT == MVT::v32i1)
2451 RC = &X86::VK32RegClass;
2452 else if (RegVT == MVT::v64i1)
2453 RC = &X86::VK64RegClass;
2455 llvm_unreachable("Unknown argument type!");
2457 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2458 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2460 // If this is an 8 or 16-bit value, it is really passed promoted to 32
2461 // bits. Insert an assert[sz]ext to capture this, then truncate to the
2463 if (VA.getLocInfo() == CCValAssign::SExt)
2464 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2465 DAG.getValueType(VA.getValVT()));
2466 else if (VA.getLocInfo() == CCValAssign::ZExt)
2467 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2468 DAG.getValueType(VA.getValVT()));
2469 else if (VA.getLocInfo() == CCValAssign::BCvt)
2470 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2472 if (VA.isExtInLoc()) {
2473 // Handle MMX values passed in XMM regs.
2474 if (RegVT.isVector())
2475 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(), ArgValue);
2477 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2480 assert(VA.isMemLoc());
2481 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
2484 // If value is passed via pointer - do a load.
2485 if (VA.getLocInfo() == CCValAssign::Indirect)
2486 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
2487 MachinePointerInfo(), false, false, false, 0);
2489 InVals.push_back(ArgValue);
2492 if (Subtarget->is64Bit() || Subtarget->isTargetKnownWindowsMSVC()) {
2493 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2494 // The x86-64 ABIs require that for returning structs by value we copy
2495 // the sret argument into %rax/%eax (depending on ABI) for the return.
2496 // Win32 requires us to put the sret argument to %eax as well.
2497 // Save the argument into a virtual register so that we can access it
2498 // from the return points.
2499 if (Ins[i].Flags.isSRet()) {
2500 unsigned Reg = FuncInfo->getSRetReturnReg();
2502 MVT PtrTy = getPointerTy();
2503 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrTy));
2504 FuncInfo->setSRetReturnReg(Reg);
2506 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[i]);
2507 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
2513 unsigned StackSize = CCInfo.getNextStackOffset();
2514 // Align stack specially for tail calls.
2515 if (FuncIsMadeTailCallSafe(CallConv,
2516 MF.getTarget().Options.GuaranteedTailCallOpt))
2517 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
2519 // If the function takes variable number of arguments, make a frame index for
2520 // the start of the first vararg value... for expansion of llvm.va_start. We
2521 // can skip this if there are no va_start calls.
2522 if (MFI->hasVAStart() &&
2523 (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
2524 CallConv != CallingConv::X86_ThisCall))) {
2525 FuncInfo->setVarArgsFrameIndex(
2526 MFI->CreateFixedObject(1, StackSize, true));
2529 // 64-bit calling conventions support varargs and register parameters, so we
2530 // have to do extra work to spill them in the prologue or forward them to
2532 if (Is64Bit && isVarArg &&
2533 (MFI->hasVAStart() || MFI->hasMustTailInVarArgFunc())) {
2534 // Find the first unallocated argument registers.
2535 ArrayRef<MCPhysReg> ArgGPRs = get64BitArgumentGPRs(CallConv, Subtarget);
2536 ArrayRef<MCPhysReg> ArgXMMs = get64BitArgumentXMMs(MF, CallConv, Subtarget);
2537 unsigned NumIntRegs =
2538 CCInfo.getFirstUnallocated(ArgGPRs.data(), ArgGPRs.size());
2539 unsigned NumXMMRegs =
2540 CCInfo.getFirstUnallocated(ArgXMMs.data(), ArgXMMs.size());
2541 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
2542 "SSE register cannot be used when SSE is disabled!");
2544 // Gather all the live in physical registers.
2545 SmallVector<SDValue, 6> LiveGPRs;
2546 SmallVector<SDValue, 8> LiveXMMRegs;
2548 for (MCPhysReg Reg : ArgGPRs.slice(NumIntRegs)) {
2549 unsigned GPR = MF.addLiveIn(Reg, &X86::GR64RegClass);
2551 DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64));
2553 if (!ArgXMMs.empty()) {
2554 unsigned AL = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2555 ALVal = DAG.getCopyFromReg(Chain, dl, AL, MVT::i8);
2556 for (MCPhysReg Reg : ArgXMMs.slice(NumXMMRegs)) {
2557 unsigned XMMReg = MF.addLiveIn(Reg, &X86::VR128RegClass);
2558 LiveXMMRegs.push_back(
2559 DAG.getCopyFromReg(Chain, dl, XMMReg, MVT::v4f32));
2563 // Store them to the va_list returned by va_start.
2564 if (MFI->hasVAStart()) {
2566 const TargetFrameLowering &TFI = *MF.getSubtarget().getFrameLowering();
2567 // Get to the caller-allocated home save location. Add 8 to account
2568 // for the return address.
2569 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2570 FuncInfo->setRegSaveFrameIndex(
2571 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
2572 // Fixup to set vararg frame on shadow area (4 x i64).
2574 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
2576 // For X86-64, if there are vararg parameters that are passed via
2577 // registers, then we must store them to their spots on the stack so
2578 // they may be loaded by deferencing the result of va_next.
2579 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
2580 FuncInfo->setVarArgsFPOffset(ArgGPRs.size() * 8 + NumXMMRegs * 16);
2581 FuncInfo->setRegSaveFrameIndex(MFI->CreateStackObject(
2582 ArgGPRs.size() * 8 + ArgXMMs.size() * 16, 16, false));
2585 // Store the integer parameter registers.
2586 SmallVector<SDValue, 8> MemOps;
2587 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
2589 unsigned Offset = FuncInfo->getVarArgsGPOffset();
2590 for (SDValue Val : LiveGPRs) {
2591 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
2592 DAG.getIntPtrConstant(Offset));
2594 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2595 MachinePointerInfo::getFixedStack(
2596 FuncInfo->getRegSaveFrameIndex(), Offset),
2598 MemOps.push_back(Store);
2602 if (!ArgXMMs.empty() && NumXMMRegs != ArgXMMs.size()) {
2603 // Now store the XMM (fp + vector) parameter registers.
2604 SmallVector<SDValue, 12> SaveXMMOps;
2605 SaveXMMOps.push_back(Chain);
2606 SaveXMMOps.push_back(ALVal);
2607 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2608 FuncInfo->getRegSaveFrameIndex()));
2609 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2610 FuncInfo->getVarArgsFPOffset()));
2611 SaveXMMOps.insert(SaveXMMOps.end(), LiveXMMRegs.begin(),
2613 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
2614 MVT::Other, SaveXMMOps));
2617 if (!MemOps.empty())
2618 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2620 // Add all GPRs, al, and XMMs to the list of forwards. We will add then
2621 // to the liveout set on a musttail call.
2622 assert(MFI->hasMustTailInVarArgFunc());
2623 auto &Forwards = FuncInfo->getForwardedMustTailRegParms();
2624 typedef X86MachineFunctionInfo::Forward Forward;
2626 for (unsigned I = 0, E = LiveGPRs.size(); I != E; ++I) {
2628 MF.getRegInfo().createVirtualRegister(&X86::GR64RegClass);
2629 Chain = DAG.getCopyToReg(Chain, dl, VReg, LiveGPRs[I]);
2630 Forwards.push_back(Forward(VReg, ArgGPRs[NumIntRegs + I], MVT::i64));
2633 if (!ArgXMMs.empty()) {
2635 MF.getRegInfo().createVirtualRegister(&X86::GR8RegClass);
2636 Chain = DAG.getCopyToReg(Chain, dl, ALVReg, ALVal);
2637 Forwards.push_back(Forward(ALVReg, X86::AL, MVT::i8));
2639 for (unsigned I = 0, E = LiveXMMRegs.size(); I != E; ++I) {
2641 MF.getRegInfo().createVirtualRegister(&X86::VR128RegClass);
2642 Chain = DAG.getCopyToReg(Chain, dl, VReg, LiveXMMRegs[I]);
2644 Forward(VReg, ArgXMMs[NumXMMRegs + I], MVT::v4f32));
2650 // Some CCs need callee pop.
2651 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
2652 MF.getTarget().Options.GuaranteedTailCallOpt)) {
2653 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
2655 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
2656 // If this is an sret function, the return should pop the hidden pointer.
2657 if (!Is64Bit && !IsTailCallConvention(CallConv) &&
2658 !Subtarget->getTargetTriple().isOSMSVCRT() &&
2659 argsAreStructReturn(Ins) == StackStructReturn)
2660 FuncInfo->setBytesToPopOnReturn(4);
2664 // RegSaveFrameIndex is X86-64 only.
2665 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
2666 if (CallConv == CallingConv::X86_FastCall ||
2667 CallConv == CallingConv::X86_ThisCall)
2668 // fastcc functions can't have varargs.
2669 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
2672 FuncInfo->setArgumentStackSize(StackSize);
2678 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
2679 SDValue StackPtr, SDValue Arg,
2680 SDLoc dl, SelectionDAG &DAG,
2681 const CCValAssign &VA,
2682 ISD::ArgFlagsTy Flags) const {
2683 unsigned LocMemOffset = VA.getLocMemOffset();
2684 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
2685 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
2686 if (Flags.isByVal())
2687 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
2689 return DAG.getStore(Chain, dl, Arg, PtrOff,
2690 MachinePointerInfo::getStack(LocMemOffset),
2694 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
2695 /// optimization is performed and it is required.
2697 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
2698 SDValue &OutRetAddr, SDValue Chain,
2699 bool IsTailCall, bool Is64Bit,
2700 int FPDiff, SDLoc dl) const {
2701 // Adjust the Return address stack slot.
2702 EVT VT = getPointerTy();
2703 OutRetAddr = getReturnAddressFrameIndex(DAG);
2705 // Load the "old" Return address.
2706 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
2707 false, false, false, 0);
2708 return SDValue(OutRetAddr.getNode(), 1);
2711 /// EmitTailCallStoreRetAddr - Emit a store of the return address if tail call
2712 /// optimization is performed and it is required (FPDiff!=0).
2713 static SDValue EmitTailCallStoreRetAddr(SelectionDAG &DAG, MachineFunction &MF,
2714 SDValue Chain, SDValue RetAddrFrIdx,
2715 EVT PtrVT, unsigned SlotSize,
2716 int FPDiff, SDLoc dl) {
2717 // Store the return address to the appropriate stack slot.
2718 if (!FPDiff) return Chain;
2719 // Calculate the new stack slot for the return address.
2720 int NewReturnAddrFI =
2721 MF.getFrameInfo()->CreateFixedObject(SlotSize, (int64_t)FPDiff - SlotSize,
2723 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, PtrVT);
2724 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
2725 MachinePointerInfo::getFixedStack(NewReturnAddrFI),
2731 X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2732 SmallVectorImpl<SDValue> &InVals) const {
2733 SelectionDAG &DAG = CLI.DAG;
2735 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
2736 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
2737 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
2738 SDValue Chain = CLI.Chain;
2739 SDValue Callee = CLI.Callee;
2740 CallingConv::ID CallConv = CLI.CallConv;
2741 bool &isTailCall = CLI.IsTailCall;
2742 bool isVarArg = CLI.IsVarArg;
2744 MachineFunction &MF = DAG.getMachineFunction();
2745 bool Is64Bit = Subtarget->is64Bit();
2746 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2747 StructReturnType SR = callIsStructReturn(Outs);
2748 bool IsSibcall = false;
2749 X86MachineFunctionInfo *X86Info = MF.getInfo<X86MachineFunctionInfo>();
2751 if (MF.getTarget().Options.DisableTailCalls)
2754 bool IsMustTail = CLI.CS && CLI.CS->isMustTailCall();
2756 // Force this to be a tail call. The verifier rules are enough to ensure
2757 // that we can lower this successfully without moving the return address
2760 } else if (isTailCall) {
2761 // Check if it's really possible to do a tail call.
2762 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
2763 isVarArg, SR != NotStructReturn,
2764 MF.getFunction()->hasStructRetAttr(), CLI.RetTy,
2765 Outs, OutVals, Ins, DAG);
2767 // Sibcalls are automatically detected tailcalls which do not require
2769 if (!MF.getTarget().Options.GuaranteedTailCallOpt && isTailCall)
2776 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2777 "Var args not supported with calling convention fastcc, ghc or hipe");
2779 // Analyze operands of the call, assigning locations to each operand.
2780 SmallVector<CCValAssign, 16> ArgLocs;
2781 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2783 // Allocate shadow area for Win64
2785 CCInfo.AllocateStack(32, 8);
2787 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2789 // Get a count of how many bytes are to be pushed on the stack.
2790 unsigned NumBytes = CCInfo.getNextStackOffset();
2792 // This is a sibcall. The memory operands are available in caller's
2793 // own caller's stack.
2795 else if (MF.getTarget().Options.GuaranteedTailCallOpt &&
2796 IsTailCallConvention(CallConv))
2797 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
2800 if (isTailCall && !IsSibcall && !IsMustTail) {
2801 // Lower arguments at fp - stackoffset + fpdiff.
2802 unsigned NumBytesCallerPushed = X86Info->getBytesToPopOnReturn();
2804 FPDiff = NumBytesCallerPushed - NumBytes;
2806 // Set the delta of movement of the returnaddr stackslot.
2807 // But only set if delta is greater than previous delta.
2808 if (FPDiff < X86Info->getTCReturnAddrDelta())
2809 X86Info->setTCReturnAddrDelta(FPDiff);
2812 unsigned NumBytesToPush = NumBytes;
2813 unsigned NumBytesToPop = NumBytes;
2815 // If we have an inalloca argument, all stack space has already been allocated
2816 // for us and be right at the top of the stack. We don't support multiple
2817 // arguments passed in memory when using inalloca.
2818 if (!Outs.empty() && Outs.back().Flags.isInAlloca()) {
2820 if (!ArgLocs.back().isMemLoc())
2821 report_fatal_error("cannot use inalloca attribute on a register "
2823 if (ArgLocs.back().getLocMemOffset() != 0)
2824 report_fatal_error("any parameter with the inalloca attribute must be "
2825 "the only memory argument");
2829 Chain = DAG.getCALLSEQ_START(
2830 Chain, DAG.getIntPtrConstant(NumBytesToPush, true), dl);
2832 SDValue RetAddrFrIdx;
2833 // Load return address for tail calls.
2834 if (isTailCall && FPDiff)
2835 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
2836 Is64Bit, FPDiff, dl);
2838 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2839 SmallVector<SDValue, 8> MemOpChains;
2842 // Walk the register/memloc assignments, inserting copies/loads. In the case
2843 // of tail call optimization arguments are handle later.
2844 const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
2845 DAG.getSubtarget().getRegisterInfo());
2846 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2847 // Skip inalloca arguments, they have already been written.
2848 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2849 if (Flags.isInAlloca())
2852 CCValAssign &VA = ArgLocs[i];
2853 EVT RegVT = VA.getLocVT();
2854 SDValue Arg = OutVals[i];
2855 bool isByVal = Flags.isByVal();
2857 // Promote the value if needed.
2858 switch (VA.getLocInfo()) {
2859 default: llvm_unreachable("Unknown loc info!");
2860 case CCValAssign::Full: break;
2861 case CCValAssign::SExt:
2862 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
2864 case CCValAssign::ZExt:
2865 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
2867 case CCValAssign::AExt:
2868 if (RegVT.is128BitVector()) {
2869 // Special case: passing MMX values in XMM registers.
2870 Arg = DAG.getNode(ISD::BITCAST, dl, MVT::i64, Arg);
2871 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
2872 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
2874 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
2876 case CCValAssign::BCvt:
2877 Arg = DAG.getNode(ISD::BITCAST, dl, RegVT, Arg);
2879 case CCValAssign::Indirect: {
2880 // Store the argument.
2881 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
2882 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
2883 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
2884 MachinePointerInfo::getFixedStack(FI),
2891 if (VA.isRegLoc()) {
2892 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2893 if (isVarArg && IsWin64) {
2894 // Win64 ABI requires argument XMM reg to be copied to the corresponding
2895 // shadow reg if callee is a varargs function.
2896 unsigned ShadowReg = 0;
2897 switch (VA.getLocReg()) {
2898 case X86::XMM0: ShadowReg = X86::RCX; break;
2899 case X86::XMM1: ShadowReg = X86::RDX; break;
2900 case X86::XMM2: ShadowReg = X86::R8; break;
2901 case X86::XMM3: ShadowReg = X86::R9; break;
2904 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
2906 } else if (!IsSibcall && (!isTailCall || isByVal)) {
2907 assert(VA.isMemLoc());
2908 if (!StackPtr.getNode())
2909 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
2911 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
2912 dl, DAG, VA, Flags));
2916 if (!MemOpChains.empty())
2917 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
2919 if (Subtarget->isPICStyleGOT()) {
2920 // ELF / PIC requires GOT in the EBX register before function calls via PLT
2923 RegsToPass.push_back(std::make_pair(unsigned(X86::EBX),
2924 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), getPointerTy())));
2926 // If we are tail calling and generating PIC/GOT style code load the
2927 // address of the callee into ECX. The value in ecx is used as target of
2928 // the tail jump. This is done to circumvent the ebx/callee-saved problem
2929 // for tail calls on PIC/GOT architectures. Normally we would just put the
2930 // address of GOT into ebx and then call target@PLT. But for tail calls
2931 // ebx would be restored (since ebx is callee saved) before jumping to the
2934 // Note: The actual moving to ECX is done further down.
2935 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2936 if (G && !G->getGlobal()->hasHiddenVisibility() &&
2937 !G->getGlobal()->hasProtectedVisibility())
2938 Callee = LowerGlobalAddress(Callee, DAG);
2939 else if (isa<ExternalSymbolSDNode>(Callee))
2940 Callee = LowerExternalSymbol(Callee, DAG);
2944 if (Is64Bit && isVarArg && !IsWin64 && !IsMustTail) {
2945 // From AMD64 ABI document:
2946 // For calls that may call functions that use varargs or stdargs
2947 // (prototype-less calls or calls to functions containing ellipsis (...) in
2948 // the declaration) %al is used as hidden argument to specify the number
2949 // of SSE registers used. The contents of %al do not need to match exactly
2950 // the number of registers, but must be an ubound on the number of SSE
2951 // registers used and is in the range 0 - 8 inclusive.
2953 // Count the number of XMM registers allocated.
2954 static const MCPhysReg XMMArgRegs[] = {
2955 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2956 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2958 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
2959 assert((Subtarget->hasSSE1() || !NumXMMRegs)
2960 && "SSE registers cannot be used when SSE is disabled");
2962 RegsToPass.push_back(std::make_pair(unsigned(X86::AL),
2963 DAG.getConstant(NumXMMRegs, MVT::i8)));
2966 if (Is64Bit && isVarArg && IsMustTail) {
2967 const auto &Forwards = X86Info->getForwardedMustTailRegParms();
2968 for (const auto &F : Forwards) {
2969 SDValue Val = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
2970 RegsToPass.push_back(std::make_pair(unsigned(F.PReg), Val));
2974 // For tail calls lower the arguments to the 'real' stack slots. Sibcalls
2975 // don't need this because the eligibility check rejects calls that require
2976 // shuffling arguments passed in memory.
2977 if (!IsSibcall && isTailCall) {
2978 // Force all the incoming stack arguments to be loaded from the stack
2979 // before any new outgoing arguments are stored to the stack, because the
2980 // outgoing stack slots may alias the incoming argument stack slots, and
2981 // the alias isn't otherwise explicit. This is slightly more conservative
2982 // than necessary, because it means that each store effectively depends
2983 // on every argument instead of just those arguments it would clobber.
2984 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2986 SmallVector<SDValue, 8> MemOpChains2;
2989 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2990 CCValAssign &VA = ArgLocs[i];
2993 assert(VA.isMemLoc());
2994 SDValue Arg = OutVals[i];
2995 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2996 // Skip inalloca arguments. They don't require any work.
2997 if (Flags.isInAlloca())
2999 // Create frame index.
3000 int32_t Offset = VA.getLocMemOffset()+FPDiff;
3001 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
3002 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
3003 FIN = DAG.getFrameIndex(FI, getPointerTy());
3005 if (Flags.isByVal()) {
3006 // Copy relative to framepointer.
3007 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
3008 if (!StackPtr.getNode())
3009 StackPtr = DAG.getCopyFromReg(Chain, dl,
3010 RegInfo->getStackRegister(),
3012 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
3014 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
3018 // Store relative to framepointer.
3019 MemOpChains2.push_back(
3020 DAG.getStore(ArgChain, dl, Arg, FIN,
3021 MachinePointerInfo::getFixedStack(FI),
3026 if (!MemOpChains2.empty())
3027 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains2);
3029 // Store the return address to the appropriate stack slot.
3030 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx,
3031 getPointerTy(), RegInfo->getSlotSize(),
3035 // Build a sequence of copy-to-reg nodes chained together with token chain
3036 // and flag operands which copy the outgoing args into registers.
3038 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3039 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3040 RegsToPass[i].second, InFlag);
3041 InFlag = Chain.getValue(1);
3044 if (DAG.getTarget().getCodeModel() == CodeModel::Large) {
3045 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
3046 // In the 64-bit large code model, we have to make all calls
3047 // through a register, since the call instruction's 32-bit
3048 // pc-relative offset may not be large enough to hold the whole
3050 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
3051 // If the callee is a GlobalAddress node (quite common, every direct call
3052 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
3055 // We should use extra load for direct calls to dllimported functions in
3057 const GlobalValue *GV = G->getGlobal();
3058 if (!GV->hasDLLImportStorageClass()) {
3059 unsigned char OpFlags = 0;
3060 bool ExtraLoad = false;
3061 unsigned WrapperKind = ISD::DELETED_NODE;
3063 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
3064 // external symbols most go through the PLT in PIC mode. If the symbol
3065 // has hidden or protected visibility, or if it is static or local, then
3066 // we don't need to use the PLT - we can directly call it.
3067 if (Subtarget->isTargetELF() &&
3068 DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
3069 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
3070 OpFlags = X86II::MO_PLT;
3071 } else if (Subtarget->isPICStyleStubAny() &&
3072 (GV->isDeclaration() || GV->isWeakForLinker()) &&
3073 (!Subtarget->getTargetTriple().isMacOSX() ||
3074 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3075 // PC-relative references to external symbols should go through $stub,
3076 // unless we're building with the leopard linker or later, which
3077 // automatically synthesizes these stubs.
3078 OpFlags = X86II::MO_DARWIN_STUB;
3079 } else if (Subtarget->isPICStyleRIPRel() &&
3080 isa<Function>(GV) &&
3081 cast<Function>(GV)->getAttributes().
3082 hasAttribute(AttributeSet::FunctionIndex,
3083 Attribute::NonLazyBind)) {
3084 // If the function is marked as non-lazy, generate an indirect call
3085 // which loads from the GOT directly. This avoids runtime overhead
3086 // at the cost of eager binding (and one extra byte of encoding).
3087 OpFlags = X86II::MO_GOTPCREL;
3088 WrapperKind = X86ISD::WrapperRIP;
3092 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
3093 G->getOffset(), OpFlags);
3095 // Add a wrapper if needed.
3096 if (WrapperKind != ISD::DELETED_NODE)
3097 Callee = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Callee);
3098 // Add extra indirection if needed.
3100 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Callee,
3101 MachinePointerInfo::getGOT(),
3102 false, false, false, 0);
3104 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3105 unsigned char OpFlags = 0;
3107 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
3108 // external symbols should go through the PLT.
3109 if (Subtarget->isTargetELF() &&
3110 DAG.getTarget().getRelocationModel() == Reloc::PIC_) {
3111 OpFlags = X86II::MO_PLT;
3112 } else if (Subtarget->isPICStyleStubAny() &&
3113 (!Subtarget->getTargetTriple().isMacOSX() ||
3114 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3115 // PC-relative references to external symbols should go through $stub,
3116 // unless we're building with the leopard linker or later, which
3117 // automatically synthesizes these stubs.
3118 OpFlags = X86II::MO_DARWIN_STUB;
3121 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
3123 } else if (Subtarget->isTarget64BitILP32() && Callee->getValueType(0) == MVT::i32) {
3124 // Zero-extend the 32-bit Callee address into a 64-bit according to x32 ABI
3125 Callee = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, Callee);
3128 // Returns a chain & a flag for retval copy to use.
3129 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3130 SmallVector<SDValue, 8> Ops;
3132 if (!IsSibcall && isTailCall) {
3133 Chain = DAG.getCALLSEQ_END(Chain,
3134 DAG.getIntPtrConstant(NumBytesToPop, true),
3135 DAG.getIntPtrConstant(0, true), InFlag, dl);
3136 InFlag = Chain.getValue(1);
3139 Ops.push_back(Chain);
3140 Ops.push_back(Callee);
3143 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
3145 // Add argument registers to the end of the list so that they are known live
3147 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3148 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3149 RegsToPass[i].second.getValueType()));
3151 // Add a register mask operand representing the call-preserved registers.
3152 const TargetRegisterInfo *TRI = DAG.getSubtarget().getRegisterInfo();
3153 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
3154 assert(Mask && "Missing call preserved mask for calling convention");
3155 Ops.push_back(DAG.getRegisterMask(Mask));
3157 if (InFlag.getNode())
3158 Ops.push_back(InFlag);
3162 //// If this is the first return lowered for this function, add the regs
3163 //// to the liveout set for the function.
3164 // This isn't right, although it's probably harmless on x86; liveouts
3165 // should be computed from returns not tail calls. Consider a void
3166 // function making a tail call to a function returning int.
3167 return DAG.getNode(X86ISD::TC_RETURN, dl, NodeTys, Ops);
3170 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops);
3171 InFlag = Chain.getValue(1);
3173 // Create the CALLSEQ_END node.
3174 unsigned NumBytesForCalleeToPop;
3175 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
3176 DAG.getTarget().Options.GuaranteedTailCallOpt))
3177 NumBytesForCalleeToPop = NumBytes; // Callee pops everything
3178 else if (!Is64Bit && !IsTailCallConvention(CallConv) &&
3179 !Subtarget->getTargetTriple().isOSMSVCRT() &&
3180 SR == StackStructReturn)
3181 // If this is a call to a struct-return function, the callee
3182 // pops the hidden struct pointer, so we have to push it back.
3183 // This is common for Darwin/X86, Linux & Mingw32 targets.
3184 // For MSVC Win32 targets, the caller pops the hidden struct pointer.
3185 NumBytesForCalleeToPop = 4;
3187 NumBytesForCalleeToPop = 0; // Callee pops nothing.
3189 // Returns a flag for retval copy to use.
3191 Chain = DAG.getCALLSEQ_END(Chain,
3192 DAG.getIntPtrConstant(NumBytesToPop, true),
3193 DAG.getIntPtrConstant(NumBytesForCalleeToPop,
3196 InFlag = Chain.getValue(1);
3199 // Handle result values, copying them out of physregs into vregs that we
3201 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3202 Ins, dl, DAG, InVals);
3205 //===----------------------------------------------------------------------===//
3206 // Fast Calling Convention (tail call) implementation
3207 //===----------------------------------------------------------------------===//
3209 // Like std call, callee cleans arguments, convention except that ECX is
3210 // reserved for storing the tail called function address. Only 2 registers are
3211 // free for argument passing (inreg). Tail call optimization is performed
3213 // * tailcallopt is enabled
3214 // * caller/callee are fastcc
3215 // On X86_64 architecture with GOT-style position independent code only local
3216 // (within module) calls are supported at the moment.
3217 // To keep the stack aligned according to platform abi the function
3218 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
3219 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
3220 // If a tail called function callee has more arguments than the caller the
3221 // caller needs to make sure that there is room to move the RETADDR to. This is
3222 // achieved by reserving an area the size of the argument delta right after the
3223 // original RETADDR, but before the saved framepointer or the spilled registers
3224 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
3236 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
3237 /// for a 16 byte align requirement.
3239 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
3240 SelectionDAG& DAG) const {
3241 MachineFunction &MF = DAG.getMachineFunction();
3242 const TargetMachine &TM = MF.getTarget();
3243 const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
3244 TM.getSubtargetImpl()->getRegisterInfo());
3245 const TargetFrameLowering &TFI = *TM.getSubtargetImpl()->getFrameLowering();
3246 unsigned StackAlignment = TFI.getStackAlignment();
3247 uint64_t AlignMask = StackAlignment - 1;
3248 int64_t Offset = StackSize;
3249 unsigned SlotSize = RegInfo->getSlotSize();
3250 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
3251 // Number smaller than 12 so just add the difference.
3252 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
3254 // Mask out lower bits, add stackalignment once plus the 12 bytes.
3255 Offset = ((~AlignMask) & Offset) + StackAlignment +
3256 (StackAlignment-SlotSize);
3261 /// MatchingStackOffset - Return true if the given stack call argument is
3262 /// already available in the same position (relatively) of the caller's
3263 /// incoming argument stack.
3265 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
3266 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
3267 const X86InstrInfo *TII) {
3268 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
3270 if (Arg.getOpcode() == ISD::CopyFromReg) {
3271 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
3272 if (!TargetRegisterInfo::isVirtualRegister(VR))
3274 MachineInstr *Def = MRI->getVRegDef(VR);
3277 if (!Flags.isByVal()) {
3278 if (!TII->isLoadFromStackSlot(Def, FI))
3281 unsigned Opcode = Def->getOpcode();
3282 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
3283 Def->getOperand(1).isFI()) {
3284 FI = Def->getOperand(1).getIndex();
3285 Bytes = Flags.getByValSize();
3289 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
3290 if (Flags.isByVal())
3291 // ByVal argument is passed in as a pointer but it's now being
3292 // dereferenced. e.g.
3293 // define @foo(%struct.X* %A) {
3294 // tail call @bar(%struct.X* byval %A)
3297 SDValue Ptr = Ld->getBasePtr();
3298 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
3301 FI = FINode->getIndex();
3302 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
3303 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
3304 FI = FINode->getIndex();
3305 Bytes = Flags.getByValSize();
3309 assert(FI != INT_MAX);
3310 if (!MFI->isFixedObjectIndex(FI))
3312 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
3315 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
3316 /// for tail call optimization. Targets which want to do tail call
3317 /// optimization should implement this function.
3319 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
3320 CallingConv::ID CalleeCC,
3322 bool isCalleeStructRet,
3323 bool isCallerStructRet,
3325 const SmallVectorImpl<ISD::OutputArg> &Outs,
3326 const SmallVectorImpl<SDValue> &OutVals,
3327 const SmallVectorImpl<ISD::InputArg> &Ins,
3328 SelectionDAG &DAG) const {
3329 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
3332 // If -tailcallopt is specified, make fastcc functions tail-callable.
3333 const MachineFunction &MF = DAG.getMachineFunction();
3334 const Function *CallerF = MF.getFunction();
3336 // If the function return type is x86_fp80 and the callee return type is not,
3337 // then the FP_EXTEND of the call result is not a nop. It's not safe to
3338 // perform a tailcall optimization here.
3339 if (CallerF->getReturnType()->isX86_FP80Ty() && !RetTy->isX86_FP80Ty())
3342 CallingConv::ID CallerCC = CallerF->getCallingConv();
3343 bool CCMatch = CallerCC == CalleeCC;
3344 bool IsCalleeWin64 = Subtarget->isCallingConvWin64(CalleeCC);
3345 bool IsCallerWin64 = Subtarget->isCallingConvWin64(CallerCC);
3347 if (DAG.getTarget().Options.GuaranteedTailCallOpt) {
3348 if (IsTailCallConvention(CalleeCC) && CCMatch)
3353 // Look for obvious safe cases to perform tail call optimization that do not
3354 // require ABI changes. This is what gcc calls sibcall.
3356 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
3357 // emit a special epilogue.
3358 const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
3359 DAG.getSubtarget().getRegisterInfo());
3360 if (RegInfo->needsStackRealignment(MF))
3363 // Also avoid sibcall optimization if either caller or callee uses struct
3364 // return semantics.
3365 if (isCalleeStructRet || isCallerStructRet)
3368 // An stdcall/thiscall caller is expected to clean up its arguments; the
3369 // callee isn't going to do that.
3370 // FIXME: this is more restrictive than needed. We could produce a tailcall
3371 // when the stack adjustment matches. For example, with a thiscall that takes
3372 // only one argument.
3373 if (!CCMatch && (CallerCC == CallingConv::X86_StdCall ||
3374 CallerCC == CallingConv::X86_ThisCall))
3377 // Do not sibcall optimize vararg calls unless all arguments are passed via
3379 if (isVarArg && !Outs.empty()) {
3381 // Optimizing for varargs on Win64 is unlikely to be safe without
3382 // additional testing.
3383 if (IsCalleeWin64 || IsCallerWin64)
3386 SmallVector<CCValAssign, 16> ArgLocs;
3387 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3390 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3391 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
3392 if (!ArgLocs[i].isRegLoc())
3396 // If the call result is in ST0 / ST1, it needs to be popped off the x87
3397 // stack. Therefore, if it's not used by the call it is not safe to optimize
3398 // this into a sibcall.
3399 bool Unused = false;
3400 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
3407 SmallVector<CCValAssign, 16> RVLocs;
3408 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(), RVLocs,
3410 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
3411 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3412 CCValAssign &VA = RVLocs[i];
3413 if (VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1)
3418 // If the calling conventions do not match, then we'd better make sure the
3419 // results are returned in the same way as what the caller expects.
3421 SmallVector<CCValAssign, 16> RVLocs1;
3422 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
3424 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
3426 SmallVector<CCValAssign, 16> RVLocs2;
3427 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
3429 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
3431 if (RVLocs1.size() != RVLocs2.size())
3433 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
3434 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
3436 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
3438 if (RVLocs1[i].isRegLoc()) {
3439 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
3442 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
3448 // If the callee takes no arguments then go on to check the results of the
3450 if (!Outs.empty()) {
3451 // Check if stack adjustment is needed. For now, do not do this if any
3452 // argument is passed on the stack.
3453 SmallVector<CCValAssign, 16> ArgLocs;
3454 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3457 // Allocate shadow area for Win64
3459 CCInfo.AllocateStack(32, 8);
3461 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3462 if (CCInfo.getNextStackOffset()) {
3463 MachineFunction &MF = DAG.getMachineFunction();
3464 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
3467 // Check if the arguments are already laid out in the right way as
3468 // the caller's fixed stack objects.
3469 MachineFrameInfo *MFI = MF.getFrameInfo();
3470 const MachineRegisterInfo *MRI = &MF.getRegInfo();
3471 const X86InstrInfo *TII =
3472 static_cast<const X86InstrInfo *>(DAG.getSubtarget().getInstrInfo());
3473 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3474 CCValAssign &VA = ArgLocs[i];
3475 SDValue Arg = OutVals[i];
3476 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3477 if (VA.getLocInfo() == CCValAssign::Indirect)
3479 if (!VA.isRegLoc()) {
3480 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
3487 // If the tailcall address may be in a register, then make sure it's
3488 // possible to register allocate for it. In 32-bit, the call address can
3489 // only target EAX, EDX, or ECX since the tail call must be scheduled after
3490 // callee-saved registers are restored. These happen to be the same
3491 // registers used to pass 'inreg' arguments so watch out for those.
3492 if (!Subtarget->is64Bit() &&
3493 ((!isa<GlobalAddressSDNode>(Callee) &&
3494 !isa<ExternalSymbolSDNode>(Callee)) ||
3495 DAG.getTarget().getRelocationModel() == Reloc::PIC_)) {
3496 unsigned NumInRegs = 0;
3497 // In PIC we need an extra register to formulate the address computation
3499 unsigned MaxInRegs =
3500 (DAG.getTarget().getRelocationModel() == Reloc::PIC_) ? 2 : 3;
3502 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3503 CCValAssign &VA = ArgLocs[i];
3506 unsigned Reg = VA.getLocReg();
3509 case X86::EAX: case X86::EDX: case X86::ECX:
3510 if (++NumInRegs == MaxInRegs)
3522 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
3523 const TargetLibraryInfo *libInfo) const {
3524 return X86::createFastISel(funcInfo, libInfo);
3527 //===----------------------------------------------------------------------===//
3528 // Other Lowering Hooks
3529 //===----------------------------------------------------------------------===//
3531 static bool MayFoldLoad(SDValue Op) {
3532 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
3535 static bool MayFoldIntoStore(SDValue Op) {
3536 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
3539 static bool isTargetShuffle(unsigned Opcode) {
3541 default: return false;
3542 case X86ISD::BLENDI:
3543 case X86ISD::PSHUFB:
3544 case X86ISD::PSHUFD:
3545 case X86ISD::PSHUFHW:
3546 case X86ISD::PSHUFLW:
3548 case X86ISD::PALIGNR:
3549 case X86ISD::MOVLHPS:
3550 case X86ISD::MOVLHPD:
3551 case X86ISD::MOVHLPS:
3552 case X86ISD::MOVLPS:
3553 case X86ISD::MOVLPD:
3554 case X86ISD::MOVSHDUP:
3555 case X86ISD::MOVSLDUP:
3556 case X86ISD::MOVDDUP:
3559 case X86ISD::UNPCKL:
3560 case X86ISD::UNPCKH:
3561 case X86ISD::VPERMILPI:
3562 case X86ISD::VPERM2X128:
3563 case X86ISD::VPERMI:
3568 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3569 SDValue V1, SelectionDAG &DAG) {
3571 default: llvm_unreachable("Unknown x86 shuffle node");
3572 case X86ISD::MOVSHDUP:
3573 case X86ISD::MOVSLDUP:
3574 case X86ISD::MOVDDUP:
3575 return DAG.getNode(Opc, dl, VT, V1);
3579 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3580 SDValue V1, unsigned TargetMask,
3581 SelectionDAG &DAG) {
3583 default: llvm_unreachable("Unknown x86 shuffle node");
3584 case X86ISD::PSHUFD:
3585 case X86ISD::PSHUFHW:
3586 case X86ISD::PSHUFLW:
3587 case X86ISD::VPERMILPI:
3588 case X86ISD::VPERMI:
3589 return DAG.getNode(Opc, dl, VT, V1, DAG.getConstant(TargetMask, MVT::i8));
3593 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3594 SDValue V1, SDValue V2, unsigned TargetMask,
3595 SelectionDAG &DAG) {
3597 default: llvm_unreachable("Unknown x86 shuffle node");
3598 case X86ISD::PALIGNR:
3599 case X86ISD::VALIGN:
3601 case X86ISD::VPERM2X128:
3602 return DAG.getNode(Opc, dl, VT, V1, V2,
3603 DAG.getConstant(TargetMask, MVT::i8));
3607 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3608 SDValue V1, SDValue V2, SelectionDAG &DAG) {
3610 default: llvm_unreachable("Unknown x86 shuffle node");
3611 case X86ISD::MOVLHPS:
3612 case X86ISD::MOVLHPD:
3613 case X86ISD::MOVHLPS:
3614 case X86ISD::MOVLPS:
3615 case X86ISD::MOVLPD:
3618 case X86ISD::UNPCKL:
3619 case X86ISD::UNPCKH:
3620 return DAG.getNode(Opc, dl, VT, V1, V2);
3624 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
3625 MachineFunction &MF = DAG.getMachineFunction();
3626 const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
3627 DAG.getSubtarget().getRegisterInfo());
3628 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
3629 int ReturnAddrIndex = FuncInfo->getRAIndex();
3631 if (ReturnAddrIndex == 0) {
3632 // Set up a frame object for the return address.
3633 unsigned SlotSize = RegInfo->getSlotSize();
3634 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize,
3637 FuncInfo->setRAIndex(ReturnAddrIndex);
3640 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
3643 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
3644 bool hasSymbolicDisplacement) {
3645 // Offset should fit into 32 bit immediate field.
3646 if (!isInt<32>(Offset))
3649 // If we don't have a symbolic displacement - we don't have any extra
3651 if (!hasSymbolicDisplacement)
3654 // FIXME: Some tweaks might be needed for medium code model.
3655 if (M != CodeModel::Small && M != CodeModel::Kernel)
3658 // For small code model we assume that latest object is 16MB before end of 31
3659 // bits boundary. We may also accept pretty large negative constants knowing
3660 // that all objects are in the positive half of address space.
3661 if (M == CodeModel::Small && Offset < 16*1024*1024)
3664 // For kernel code model we know that all object resist in the negative half
3665 // of 32bits address space. We may not accept negative offsets, since they may
3666 // be just off and we may accept pretty large positive ones.
3667 if (M == CodeModel::Kernel && Offset > 0)
3673 /// isCalleePop - Determines whether the callee is required to pop its
3674 /// own arguments. Callee pop is necessary to support tail calls.
3675 bool X86::isCalleePop(CallingConv::ID CallingConv,
3676 bool is64Bit, bool IsVarArg, bool TailCallOpt) {
3677 switch (CallingConv) {
3680 case CallingConv::X86_StdCall:
3681 case CallingConv::X86_FastCall:
3682 case CallingConv::X86_ThisCall:
3684 case CallingConv::Fast:
3685 case CallingConv::GHC:
3686 case CallingConv::HiPE:
3693 /// \brief Return true if the condition is an unsigned comparison operation.
3694 static bool isX86CCUnsigned(unsigned X86CC) {
3696 default: llvm_unreachable("Invalid integer condition!");
3697 case X86::COND_E: return true;
3698 case X86::COND_G: return false;
3699 case X86::COND_GE: return false;
3700 case X86::COND_L: return false;
3701 case X86::COND_LE: return false;
3702 case X86::COND_NE: return true;
3703 case X86::COND_B: return true;
3704 case X86::COND_A: return true;
3705 case X86::COND_BE: return true;
3706 case X86::COND_AE: return true;
3708 llvm_unreachable("covered switch fell through?!");
3711 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
3712 /// specific condition code, returning the condition code and the LHS/RHS of the
3713 /// comparison to make.
3714 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
3715 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
3717 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
3718 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
3719 // X > -1 -> X == 0, jump !sign.
3720 RHS = DAG.getConstant(0, RHS.getValueType());
3721 return X86::COND_NS;
3723 if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
3724 // X < 0 -> X == 0, jump on sign.
3727 if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
3729 RHS = DAG.getConstant(0, RHS.getValueType());
3730 return X86::COND_LE;
3734 switch (SetCCOpcode) {
3735 default: llvm_unreachable("Invalid integer condition!");
3736 case ISD::SETEQ: return X86::COND_E;
3737 case ISD::SETGT: return X86::COND_G;
3738 case ISD::SETGE: return X86::COND_GE;
3739 case ISD::SETLT: return X86::COND_L;
3740 case ISD::SETLE: return X86::COND_LE;
3741 case ISD::SETNE: return X86::COND_NE;
3742 case ISD::SETULT: return X86::COND_B;
3743 case ISD::SETUGT: return X86::COND_A;
3744 case ISD::SETULE: return X86::COND_BE;
3745 case ISD::SETUGE: return X86::COND_AE;
3749 // First determine if it is required or is profitable to flip the operands.
3751 // If LHS is a foldable load, but RHS is not, flip the condition.
3752 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
3753 !ISD::isNON_EXTLoad(RHS.getNode())) {
3754 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
3755 std::swap(LHS, RHS);
3758 switch (SetCCOpcode) {
3764 std::swap(LHS, RHS);
3768 // On a floating point condition, the flags are set as follows:
3770 // 0 | 0 | 0 | X > Y
3771 // 0 | 0 | 1 | X < Y
3772 // 1 | 0 | 0 | X == Y
3773 // 1 | 1 | 1 | unordered
3774 switch (SetCCOpcode) {
3775 default: llvm_unreachable("Condcode should be pre-legalized away");
3777 case ISD::SETEQ: return X86::COND_E;
3778 case ISD::SETOLT: // flipped
3780 case ISD::SETGT: return X86::COND_A;
3781 case ISD::SETOLE: // flipped
3783 case ISD::SETGE: return X86::COND_AE;
3784 case ISD::SETUGT: // flipped
3786 case ISD::SETLT: return X86::COND_B;
3787 case ISD::SETUGE: // flipped
3789 case ISD::SETLE: return X86::COND_BE;
3791 case ISD::SETNE: return X86::COND_NE;
3792 case ISD::SETUO: return X86::COND_P;
3793 case ISD::SETO: return X86::COND_NP;
3795 case ISD::SETUNE: return X86::COND_INVALID;
3799 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
3800 /// code. Current x86 isa includes the following FP cmov instructions:
3801 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
3802 static bool hasFPCMov(unsigned X86CC) {
3818 /// isFPImmLegal - Returns true if the target can instruction select the
3819 /// specified FP immediate natively. If false, the legalizer will
3820 /// materialize the FP immediate as a load from a constant pool.
3821 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3822 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
3823 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
3829 /// \brief Returns true if it is beneficial to convert a load of a constant
3830 /// to just the constant itself.
3831 bool X86TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
3833 assert(Ty->isIntegerTy());
3835 unsigned BitSize = Ty->getPrimitiveSizeInBits();
3836 if (BitSize == 0 || BitSize > 64)
3841 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
3842 /// the specified range (L, H].
3843 static bool isUndefOrInRange(int Val, int Low, int Hi) {
3844 return (Val < 0) || (Val >= Low && Val < Hi);
3847 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
3848 /// specified value.
3849 static bool isUndefOrEqual(int Val, int CmpVal) {
3850 return (Val < 0 || Val == CmpVal);
3853 /// isSequentialOrUndefInRange - Return true if every element in Mask, beginning
3854 /// from position Pos and ending in Pos+Size, falls within the specified
3855 /// sequential range (L, L+Pos]. or is undef.
3856 static bool isSequentialOrUndefInRange(ArrayRef<int> Mask,
3857 unsigned Pos, unsigned Size, int Low) {
3858 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
3859 if (!isUndefOrEqual(Mask[i], Low))
3864 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
3865 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
3866 /// the second operand.
3867 static bool isPSHUFDMask(ArrayRef<int> Mask, MVT VT) {
3868 if (VT == MVT::v4f32 || VT == MVT::v4i32 )
3869 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
3870 if (VT == MVT::v2f64 || VT == MVT::v2i64)
3871 return (Mask[0] < 2 && Mask[1] < 2);
3875 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
3876 /// is suitable for input to PSHUFHW.
3877 static bool isPSHUFHWMask(ArrayRef<int> Mask, MVT VT, bool HasInt256) {
3878 if (VT != MVT::v8i16 && (!HasInt256 || VT != MVT::v16i16))
3881 // Lower quadword copied in order or undef.
3882 if (!isSequentialOrUndefInRange(Mask, 0, 4, 0))
3885 // Upper quadword shuffled.
3886 for (unsigned i = 4; i != 8; ++i)
3887 if (!isUndefOrInRange(Mask[i], 4, 8))
3890 if (VT == MVT::v16i16) {
3891 // Lower quadword copied in order or undef.
3892 if (!isSequentialOrUndefInRange(Mask, 8, 4, 8))
3895 // Upper quadword shuffled.
3896 for (unsigned i = 12; i != 16; ++i)
3897 if (!isUndefOrInRange(Mask[i], 12, 16))
3904 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
3905 /// is suitable for input to PSHUFLW.
3906 static bool isPSHUFLWMask(ArrayRef<int> Mask, MVT VT, bool HasInt256) {
3907 if (VT != MVT::v8i16 && (!HasInt256 || VT != MVT::v16i16))
3910 // Upper quadword copied in order.
3911 if (!isSequentialOrUndefInRange(Mask, 4, 4, 4))
3914 // Lower quadword shuffled.
3915 for (unsigned i = 0; i != 4; ++i)
3916 if (!isUndefOrInRange(Mask[i], 0, 4))
3919 if (VT == MVT::v16i16) {
3920 // Upper quadword copied in order.
3921 if (!isSequentialOrUndefInRange(Mask, 12, 4, 12))
3924 // Lower quadword shuffled.
3925 for (unsigned i = 8; i != 12; ++i)
3926 if (!isUndefOrInRange(Mask[i], 8, 12))
3933 /// \brief Return true if the mask specifies a shuffle of elements that is
3934 /// suitable for input to intralane (palignr) or interlane (valign) vector
3936 static bool isAlignrMask(ArrayRef<int> Mask, MVT VT, bool InterLane) {
3937 unsigned NumElts = VT.getVectorNumElements();
3938 unsigned NumLanes = InterLane ? 1: VT.getSizeInBits()/128;
3939 unsigned NumLaneElts = NumElts/NumLanes;
3941 // Do not handle 64-bit element shuffles with palignr.
3942 if (NumLaneElts == 2)
3945 for (unsigned l = 0; l != NumElts; l+=NumLaneElts) {
3947 for (i = 0; i != NumLaneElts; ++i) {
3952 // Lane is all undef, go to next lane
3953 if (i == NumLaneElts)
3956 int Start = Mask[i+l];
3958 // Make sure its in this lane in one of the sources
3959 if (!isUndefOrInRange(Start, l, l+NumLaneElts) &&
3960 !isUndefOrInRange(Start, l+NumElts, l+NumElts+NumLaneElts))
3963 // If not lane 0, then we must match lane 0
3964 if (l != 0 && Mask[i] >= 0 && !isUndefOrEqual(Start, Mask[i]+l))
3967 // Correct second source to be contiguous with first source
3968 if (Start >= (int)NumElts)
3969 Start -= NumElts - NumLaneElts;
3971 // Make sure we're shifting in the right direction.
3972 if (Start <= (int)(i+l))
3977 // Check the rest of the elements to see if they are consecutive.
3978 for (++i; i != NumLaneElts; ++i) {
3979 int Idx = Mask[i+l];
3981 // Make sure its in this lane
3982 if (!isUndefOrInRange(Idx, l, l+NumLaneElts) &&
3983 !isUndefOrInRange(Idx, l+NumElts, l+NumElts+NumLaneElts))
3986 // If not lane 0, then we must match lane 0
3987 if (l != 0 && Mask[i] >= 0 && !isUndefOrEqual(Idx, Mask[i]+l))
3990 if (Idx >= (int)NumElts)
3991 Idx -= NumElts - NumLaneElts;
3993 if (!isUndefOrEqual(Idx, Start+i))
4002 /// \brief Return true if the node specifies a shuffle of elements that is
4003 /// suitable for input to PALIGNR.
4004 static bool isPALIGNRMask(ArrayRef<int> Mask, MVT VT,
4005 const X86Subtarget *Subtarget) {
4006 if ((VT.is128BitVector() && !Subtarget->hasSSSE3()) ||
4007 (VT.is256BitVector() && !Subtarget->hasInt256()) ||
4008 VT.is512BitVector())
4009 // FIXME: Add AVX512BW.
4012 return isAlignrMask(Mask, VT, false);
4015 /// \brief Return true if the node specifies a shuffle of elements that is
4016 /// suitable for input to VALIGN.
4017 static bool isVALIGNMask(ArrayRef<int> Mask, MVT VT,
4018 const X86Subtarget *Subtarget) {
4019 // FIXME: Add AVX512VL.
4020 if (!VT.is512BitVector() || !Subtarget->hasAVX512())
4022 return isAlignrMask(Mask, VT, true);
4025 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
4026 /// the two vector operands have swapped position.
4027 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask,
4028 unsigned NumElems) {
4029 for (unsigned i = 0; i != NumElems; ++i) {
4033 else if (idx < (int)NumElems)
4034 Mask[i] = idx + NumElems;
4036 Mask[i] = idx - NumElems;
4040 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
4041 /// specifies a shuffle of elements that is suitable for input to 128/256-bit
4042 /// SHUFPS and SHUFPD. If Commuted is true, then it checks for sources to be
4043 /// reverse of what x86 shuffles want.
4044 static bool isSHUFPMask(ArrayRef<int> Mask, MVT VT, bool Commuted = false) {
4046 unsigned NumElems = VT.getVectorNumElements();
4047 unsigned NumLanes = VT.getSizeInBits()/128;
4048 unsigned NumLaneElems = NumElems/NumLanes;
4050 if (NumLaneElems != 2 && NumLaneElems != 4)
4053 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4054 bool symetricMaskRequired =
4055 (VT.getSizeInBits() >= 256) && (EltSize == 32);
4057 // VSHUFPSY divides the resulting vector into 4 chunks.
4058 // The sources are also splitted into 4 chunks, and each destination
4059 // chunk must come from a different source chunk.
4061 // SRC1 => X7 X6 X5 X4 X3 X2 X1 X0
4062 // SRC2 => Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y9
4064 // DST => Y7..Y4, Y7..Y4, X7..X4, X7..X4,
4065 // Y3..Y0, Y3..Y0, X3..X0, X3..X0
4067 // VSHUFPDY divides the resulting vector into 4 chunks.
4068 // The sources are also splitted into 4 chunks, and each destination
4069 // chunk must come from a different source chunk.
4071 // SRC1 => X3 X2 X1 X0
4072 // SRC2 => Y3 Y2 Y1 Y0
4074 // DST => Y3..Y2, X3..X2, Y1..Y0, X1..X0
4076 SmallVector<int, 4> MaskVal(NumLaneElems, -1);
4077 unsigned HalfLaneElems = NumLaneElems/2;
4078 for (unsigned l = 0; l != NumElems; l += NumLaneElems) {
4079 for (unsigned i = 0; i != NumLaneElems; ++i) {
4080 int Idx = Mask[i+l];
4081 unsigned RngStart = l + ((Commuted == (i<HalfLaneElems)) ? NumElems : 0);
4082 if (!isUndefOrInRange(Idx, RngStart, RngStart+NumLaneElems))
4084 // For VSHUFPSY, the mask of the second half must be the same as the
4085 // first but with the appropriate offsets. This works in the same way as
4086 // VPERMILPS works with masks.
4087 if (!symetricMaskRequired || Idx < 0)
4089 if (MaskVal[i] < 0) {
4090 MaskVal[i] = Idx - l;
4093 if ((signed)(Idx - l) != MaskVal[i])
4101 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
4102 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
4103 static bool isMOVHLPSMask(ArrayRef<int> Mask, MVT VT) {
4104 if (!VT.is128BitVector())
4107 unsigned NumElems = VT.getVectorNumElements();
4112 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
4113 return isUndefOrEqual(Mask[0], 6) &&
4114 isUndefOrEqual(Mask[1], 7) &&
4115 isUndefOrEqual(Mask[2], 2) &&
4116 isUndefOrEqual(Mask[3], 3);
4119 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
4120 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
4122 static bool isMOVHLPS_v_undef_Mask(ArrayRef<int> Mask, MVT VT) {
4123 if (!VT.is128BitVector())
4126 unsigned NumElems = VT.getVectorNumElements();
4131 return isUndefOrEqual(Mask[0], 2) &&
4132 isUndefOrEqual(Mask[1], 3) &&
4133 isUndefOrEqual(Mask[2], 2) &&
4134 isUndefOrEqual(Mask[3], 3);
4137 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
4138 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
4139 static bool isMOVLPMask(ArrayRef<int> Mask, MVT VT) {
4140 if (!VT.is128BitVector())
4143 unsigned NumElems = VT.getVectorNumElements();
4145 if (NumElems != 2 && NumElems != 4)
4148 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
4149 if (!isUndefOrEqual(Mask[i], i + NumElems))
4152 for (unsigned i = NumElems/2, e = NumElems; i != e; ++i)
4153 if (!isUndefOrEqual(Mask[i], i))
4159 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
4160 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
4161 static bool isMOVLHPSMask(ArrayRef<int> Mask, MVT VT) {
4162 if (!VT.is128BitVector())
4165 unsigned NumElems = VT.getVectorNumElements();
4167 if (NumElems != 2 && NumElems != 4)
4170 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
4171 if (!isUndefOrEqual(Mask[i], i))
4174 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
4175 if (!isUndefOrEqual(Mask[i + e], i + NumElems))
4181 /// isINSERTPSMask - Return true if the specified VECTOR_SHUFFLE operand
4182 /// specifies a shuffle of elements that is suitable for input to INSERTPS.
4183 /// i. e: If all but one element come from the same vector.
4184 static bool isINSERTPSMask(ArrayRef<int> Mask, MVT VT) {
4185 // TODO: Deal with AVX's VINSERTPS
4186 if (!VT.is128BitVector() || (VT != MVT::v4f32 && VT != MVT::v4i32))
4189 unsigned CorrectPosV1 = 0;
4190 unsigned CorrectPosV2 = 0;
4191 for (int i = 0, e = (int)VT.getVectorNumElements(); i != e; ++i) {
4192 if (Mask[i] == -1) {
4200 else if (Mask[i] == i + 4)
4204 if (CorrectPosV1 == 3 || CorrectPosV2 == 3)
4205 // We have 3 elements (undefs count as elements from any vector) from one
4206 // vector, and one from another.
4213 // Some special combinations that can be optimized.
4216 SDValue Compact8x32ShuffleNode(ShuffleVectorSDNode *SVOp,
4217 SelectionDAG &DAG) {
4218 MVT VT = SVOp->getSimpleValueType(0);
4221 if (VT != MVT::v8i32 && VT != MVT::v8f32)
4224 ArrayRef<int> Mask = SVOp->getMask();
4226 // These are the special masks that may be optimized.
4227 static const int MaskToOptimizeEven[] = {0, 8, 2, 10, 4, 12, 6, 14};
4228 static const int MaskToOptimizeOdd[] = {1, 9, 3, 11, 5, 13, 7, 15};
4229 bool MatchEvenMask = true;
4230 bool MatchOddMask = true;
4231 for (int i=0; i<8; ++i) {
4232 if (!isUndefOrEqual(Mask[i], MaskToOptimizeEven[i]))
4233 MatchEvenMask = false;
4234 if (!isUndefOrEqual(Mask[i], MaskToOptimizeOdd[i]))
4235 MatchOddMask = false;
4238 if (!MatchEvenMask && !MatchOddMask)
4241 SDValue UndefNode = DAG.getNode(ISD::UNDEF, dl, VT);
4243 SDValue Op0 = SVOp->getOperand(0);
4244 SDValue Op1 = SVOp->getOperand(1);
4246 if (MatchEvenMask) {
4247 // Shift the second operand right to 32 bits.
4248 static const int ShiftRightMask[] = {-1, 0, -1, 2, -1, 4, -1, 6 };
4249 Op1 = DAG.getVectorShuffle(VT, dl, Op1, UndefNode, ShiftRightMask);
4251 // Shift the first operand left to 32 bits.
4252 static const int ShiftLeftMask[] = {1, -1, 3, -1, 5, -1, 7, -1 };
4253 Op0 = DAG.getVectorShuffle(VT, dl, Op0, UndefNode, ShiftLeftMask);
4255 static const int BlendMask[] = {0, 9, 2, 11, 4, 13, 6, 15};
4256 return DAG.getVectorShuffle(VT, dl, Op0, Op1, BlendMask);
4259 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
4260 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
4261 static bool isUNPCKLMask(ArrayRef<int> Mask, MVT VT,
4262 bool HasInt256, bool V2IsSplat = false) {
4264 assert(VT.getSizeInBits() >= 128 &&
4265 "Unsupported vector type for unpckl");
4267 unsigned NumElts = VT.getVectorNumElements();
4268 if (VT.is256BitVector() && NumElts != 4 && NumElts != 8 &&
4269 (!HasInt256 || (NumElts != 16 && NumElts != 32)))
4272 assert((!VT.is512BitVector() || VT.getScalarType().getSizeInBits() >= 32) &&
4273 "Unsupported vector type for unpckh");
4275 // AVX defines UNPCK* to operate independently on 128-bit lanes.
4276 unsigned NumLanes = VT.getSizeInBits()/128;
4277 unsigned NumLaneElts = NumElts/NumLanes;
4279 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
4280 for (unsigned i = 0, j = l; i != NumLaneElts; i += 2, ++j) {
4281 int BitI = Mask[l+i];
4282 int BitI1 = Mask[l+i+1];
4283 if (!isUndefOrEqual(BitI, j))
4286 if (!isUndefOrEqual(BitI1, NumElts))
4289 if (!isUndefOrEqual(BitI1, j + NumElts))
4298 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
4299 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
4300 static bool isUNPCKHMask(ArrayRef<int> Mask, MVT VT,
4301 bool HasInt256, bool V2IsSplat = false) {
4302 assert(VT.getSizeInBits() >= 128 &&
4303 "Unsupported vector type for unpckh");
4305 unsigned NumElts = VT.getVectorNumElements();
4306 if (VT.is256BitVector() && NumElts != 4 && NumElts != 8 &&
4307 (!HasInt256 || (NumElts != 16 && NumElts != 32)))
4310 assert((!VT.is512BitVector() || VT.getScalarType().getSizeInBits() >= 32) &&
4311 "Unsupported vector type for unpckh");
4313 // AVX defines UNPCK* to operate independently on 128-bit lanes.
4314 unsigned NumLanes = VT.getSizeInBits()/128;
4315 unsigned NumLaneElts = NumElts/NumLanes;
4317 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
4318 for (unsigned i = 0, j = l+NumLaneElts/2; i != NumLaneElts; i += 2, ++j) {
4319 int BitI = Mask[l+i];
4320 int BitI1 = Mask[l+i+1];
4321 if (!isUndefOrEqual(BitI, j))
4324 if (isUndefOrEqual(BitI1, NumElts))
4327 if (!isUndefOrEqual(BitI1, j+NumElts))
4335 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
4336 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
4338 static bool isUNPCKL_v_undef_Mask(ArrayRef<int> Mask, MVT VT, bool HasInt256) {
4339 unsigned NumElts = VT.getVectorNumElements();
4340 bool Is256BitVec = VT.is256BitVector();
4342 if (VT.is512BitVector())
4344 assert((VT.is128BitVector() || VT.is256BitVector()) &&
4345 "Unsupported vector type for unpckh");
4347 if (Is256BitVec && NumElts != 4 && NumElts != 8 &&
4348 (!HasInt256 || (NumElts != 16 && NumElts != 32)))
4351 // For 256-bit i64/f64, use MOVDDUPY instead, so reject the matching pattern
4352 // FIXME: Need a better way to get rid of this, there's no latency difference
4353 // between UNPCKLPD and MOVDDUP, the later should always be checked first and
4354 // the former later. We should also remove the "_undef" special mask.
4355 if (NumElts == 4 && Is256BitVec)
4358 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
4359 // independently on 128-bit lanes.
4360 unsigned NumLanes = VT.getSizeInBits()/128;
4361 unsigned NumLaneElts = NumElts/NumLanes;
4363 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
4364 for (unsigned i = 0, j = l; i != NumLaneElts; i += 2, ++j) {
4365 int BitI = Mask[l+i];
4366 int BitI1 = Mask[l+i+1];
4368 if (!isUndefOrEqual(BitI, j))
4370 if (!isUndefOrEqual(BitI1, j))
4378 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
4379 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
4381 static bool isUNPCKH_v_undef_Mask(ArrayRef<int> Mask, MVT VT, bool HasInt256) {
4382 unsigned NumElts = VT.getVectorNumElements();
4384 if (VT.is512BitVector())
4387 assert((VT.is128BitVector() || VT.is256BitVector()) &&
4388 "Unsupported vector type for unpckh");
4390 if (VT.is256BitVector() && NumElts != 4 && NumElts != 8 &&
4391 (!HasInt256 || (NumElts != 16 && NumElts != 32)))
4394 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
4395 // independently on 128-bit lanes.
4396 unsigned NumLanes = VT.getSizeInBits()/128;
4397 unsigned NumLaneElts = NumElts/NumLanes;
4399 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
4400 for (unsigned i = 0, j = l+NumLaneElts/2; i != NumLaneElts; i += 2, ++j) {
4401 int BitI = Mask[l+i];
4402 int BitI1 = Mask[l+i+1];
4403 if (!isUndefOrEqual(BitI, j))
4405 if (!isUndefOrEqual(BitI1, j))
4412 // Match for INSERTI64x4 INSERTF64x4 instructions (src0[0], src1[0]) or
4413 // (src1[0], src0[1]), manipulation with 256-bit sub-vectors
4414 static bool isINSERT64x4Mask(ArrayRef<int> Mask, MVT VT, unsigned int *Imm) {
4415 if (!VT.is512BitVector())
4418 unsigned NumElts = VT.getVectorNumElements();
4419 unsigned HalfSize = NumElts/2;
4420 if (isSequentialOrUndefInRange(Mask, 0, HalfSize, 0)) {
4421 if (isSequentialOrUndefInRange(Mask, HalfSize, HalfSize, NumElts)) {
4426 if (isSequentialOrUndefInRange(Mask, 0, HalfSize, NumElts)) {
4427 if (isSequentialOrUndefInRange(Mask, HalfSize, HalfSize, HalfSize)) {
4435 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
4436 /// specifies a shuffle of elements that is suitable for input to MOVSS,
4437 /// MOVSD, and MOVD, i.e. setting the lowest element.
4438 static bool isMOVLMask(ArrayRef<int> Mask, EVT VT) {
4439 if (VT.getVectorElementType().getSizeInBits() < 32)
4441 if (!VT.is128BitVector())
4444 unsigned NumElts = VT.getVectorNumElements();
4446 if (!isUndefOrEqual(Mask[0], NumElts))
4449 for (unsigned i = 1; i != NumElts; ++i)
4450 if (!isUndefOrEqual(Mask[i], i))
4456 /// isVPERM2X128Mask - Match 256-bit shuffles where the elements are considered
4457 /// as permutations between 128-bit chunks or halves. As an example: this
4459 /// vector_shuffle <4, 5, 6, 7, 12, 13, 14, 15>
4460 /// The first half comes from the second half of V1 and the second half from the
4461 /// the second half of V2.
4462 static bool isVPERM2X128Mask(ArrayRef<int> Mask, MVT VT, bool HasFp256) {
4463 if (!HasFp256 || !VT.is256BitVector())
4466 // The shuffle result is divided into half A and half B. In total the two
4467 // sources have 4 halves, namely: C, D, E, F. The final values of A and
4468 // B must come from C, D, E or F.
4469 unsigned HalfSize = VT.getVectorNumElements()/2;
4470 bool MatchA = false, MatchB = false;
4472 // Check if A comes from one of C, D, E, F.
4473 for (unsigned Half = 0; Half != 4; ++Half) {
4474 if (isSequentialOrUndefInRange(Mask, 0, HalfSize, Half*HalfSize)) {
4480 // Check if B comes from one of C, D, E, F.
4481 for (unsigned Half = 0; Half != 4; ++Half) {
4482 if (isSequentialOrUndefInRange(Mask, HalfSize, HalfSize, Half*HalfSize)) {
4488 return MatchA && MatchB;
4491 /// getShuffleVPERM2X128Immediate - Return the appropriate immediate to shuffle
4492 /// the specified VECTOR_MASK mask with VPERM2F128/VPERM2I128 instructions.
4493 static unsigned getShuffleVPERM2X128Immediate(ShuffleVectorSDNode *SVOp) {
4494 MVT VT = SVOp->getSimpleValueType(0);
4496 unsigned HalfSize = VT.getVectorNumElements()/2;
4498 unsigned FstHalf = 0, SndHalf = 0;
4499 for (unsigned i = 0; i < HalfSize; ++i) {
4500 if (SVOp->getMaskElt(i) > 0) {
4501 FstHalf = SVOp->getMaskElt(i)/HalfSize;
4505 for (unsigned i = HalfSize; i < HalfSize*2; ++i) {
4506 if (SVOp->getMaskElt(i) > 0) {
4507 SndHalf = SVOp->getMaskElt(i)/HalfSize;
4512 return (FstHalf | (SndHalf << 4));
4515 // Symetric in-lane mask. Each lane has 4 elements (for imm8)
4516 static bool isPermImmMask(ArrayRef<int> Mask, MVT VT, unsigned& Imm8) {
4517 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4521 unsigned NumElts = VT.getVectorNumElements();
4523 if (VT.is128BitVector() || (VT.is256BitVector() && EltSize == 64)) {
4524 for (unsigned i = 0; i != NumElts; ++i) {
4527 Imm8 |= Mask[i] << (i*2);
4532 unsigned LaneSize = 4;
4533 SmallVector<int, 4> MaskVal(LaneSize, -1);
4535 for (unsigned l = 0; l != NumElts; l += LaneSize) {
4536 for (unsigned i = 0; i != LaneSize; ++i) {
4537 if (!isUndefOrInRange(Mask[i+l], l, l+LaneSize))
4541 if (MaskVal[i] < 0) {
4542 MaskVal[i] = Mask[i+l] - l;
4543 Imm8 |= MaskVal[i] << (i*2);
4546 if (Mask[i+l] != (signed)(MaskVal[i]+l))
4553 /// isVPERMILPMask - Return true if the specified VECTOR_SHUFFLE operand
4554 /// specifies a shuffle of elements that is suitable for input to VPERMILPD*.
4555 /// Note that VPERMIL mask matching is different depending whether theunderlying
4556 /// type is 32 or 64. In the VPERMILPS the high half of the mask should point
4557 /// to the same elements of the low, but to the higher half of the source.
4558 /// In VPERMILPD the two lanes could be shuffled independently of each other
4559 /// with the same restriction that lanes can't be crossed. Also handles PSHUFDY.
4560 static bool isVPERMILPMask(ArrayRef<int> Mask, MVT VT) {
4561 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4562 if (VT.getSizeInBits() < 256 || EltSize < 32)
4564 bool symetricMaskRequired = (EltSize == 32);
4565 unsigned NumElts = VT.getVectorNumElements();
4567 unsigned NumLanes = VT.getSizeInBits()/128;
4568 unsigned LaneSize = NumElts/NumLanes;
4569 // 2 or 4 elements in one lane
4571 SmallVector<int, 4> ExpectedMaskVal(LaneSize, -1);
4572 for (unsigned l = 0; l != NumElts; l += LaneSize) {
4573 for (unsigned i = 0; i != LaneSize; ++i) {
4574 if (!isUndefOrInRange(Mask[i+l], l, l+LaneSize))
4576 if (symetricMaskRequired) {
4577 if (ExpectedMaskVal[i] < 0 && Mask[i+l] >= 0) {
4578 ExpectedMaskVal[i] = Mask[i+l] - l;
4581 if (!isUndefOrEqual(Mask[i+l], ExpectedMaskVal[i]+l))
4589 /// isCommutedMOVLMask - Returns true if the shuffle mask is except the reverse
4590 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
4591 /// element of vector 2 and the other elements to come from vector 1 in order.
4592 static bool isCommutedMOVLMask(ArrayRef<int> Mask, MVT VT,
4593 bool V2IsSplat = false, bool V2IsUndef = false) {
4594 if (!VT.is128BitVector())
4597 unsigned NumOps = VT.getVectorNumElements();
4598 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
4601 if (!isUndefOrEqual(Mask[0], 0))
4604 for (unsigned i = 1; i != NumOps; ++i)
4605 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
4606 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
4607 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
4613 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
4614 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
4615 /// Masks to match: <1, 1, 3, 3> or <1, 1, 3, 3, 5, 5, 7, 7>
4616 static bool isMOVSHDUPMask(ArrayRef<int> Mask, MVT VT,
4617 const X86Subtarget *Subtarget) {
4618 if (!Subtarget->hasSSE3())
4621 unsigned NumElems = VT.getVectorNumElements();
4623 if ((VT.is128BitVector() && NumElems != 4) ||
4624 (VT.is256BitVector() && NumElems != 8) ||
4625 (VT.is512BitVector() && NumElems != 16))
4628 // "i+1" is the value the indexed mask element must have
4629 for (unsigned i = 0; i != NumElems; i += 2)
4630 if (!isUndefOrEqual(Mask[i], i+1) ||
4631 !isUndefOrEqual(Mask[i+1], i+1))
4637 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
4638 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
4639 /// Masks to match: <0, 0, 2, 2> or <0, 0, 2, 2, 4, 4, 6, 6>
4640 static bool isMOVSLDUPMask(ArrayRef<int> Mask, MVT VT,
4641 const X86Subtarget *Subtarget) {
4642 if (!Subtarget->hasSSE3())
4645 unsigned NumElems = VT.getVectorNumElements();
4647 if ((VT.is128BitVector() && NumElems != 4) ||
4648 (VT.is256BitVector() && NumElems != 8) ||
4649 (VT.is512BitVector() && NumElems != 16))
4652 // "i" is the value the indexed mask element must have
4653 for (unsigned i = 0; i != NumElems; i += 2)
4654 if (!isUndefOrEqual(Mask[i], i) ||
4655 !isUndefOrEqual(Mask[i+1], i))
4661 /// isMOVDDUPYMask - Return true if the specified VECTOR_SHUFFLE operand
4662 /// specifies a shuffle of elements that is suitable for input to 256-bit
4663 /// version of MOVDDUP.
4664 static bool isMOVDDUPYMask(ArrayRef<int> Mask, MVT VT, bool HasFp256) {
4665 if (!HasFp256 || !VT.is256BitVector())
4668 unsigned NumElts = VT.getVectorNumElements();
4672 for (unsigned i = 0; i != NumElts/2; ++i)
4673 if (!isUndefOrEqual(Mask[i], 0))
4675 for (unsigned i = NumElts/2; i != NumElts; ++i)
4676 if (!isUndefOrEqual(Mask[i], NumElts/2))
4681 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
4682 /// specifies a shuffle of elements that is suitable for input to 128-bit
4683 /// version of MOVDDUP.
4684 static bool isMOVDDUPMask(ArrayRef<int> Mask, MVT VT) {
4685 if (!VT.is128BitVector())
4688 unsigned e = VT.getVectorNumElements() / 2;
4689 for (unsigned i = 0; i != e; ++i)
4690 if (!isUndefOrEqual(Mask[i], i))
4692 for (unsigned i = 0; i != e; ++i)
4693 if (!isUndefOrEqual(Mask[e+i], i))
4698 /// isVEXTRACTIndex - Return true if the specified
4699 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
4700 /// suitable for instruction that extract 128 or 256 bit vectors
4701 static bool isVEXTRACTIndex(SDNode *N, unsigned vecWidth) {
4702 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4703 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4706 // The index should be aligned on a vecWidth-bit boundary.
4708 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4710 MVT VT = N->getSimpleValueType(0);
4711 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4712 bool Result = (Index * ElSize) % vecWidth == 0;
4717 /// isVINSERTIndex - Return true if the specified INSERT_SUBVECTOR
4718 /// operand specifies a subvector insert that is suitable for input to
4719 /// insertion of 128 or 256-bit subvectors
4720 static bool isVINSERTIndex(SDNode *N, unsigned vecWidth) {
4721 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4722 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4724 // The index should be aligned on a vecWidth-bit boundary.
4726 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4728 MVT VT = N->getSimpleValueType(0);
4729 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4730 bool Result = (Index * ElSize) % vecWidth == 0;
4735 bool X86::isVINSERT128Index(SDNode *N) {
4736 return isVINSERTIndex(N, 128);
4739 bool X86::isVINSERT256Index(SDNode *N) {
4740 return isVINSERTIndex(N, 256);
4743 bool X86::isVEXTRACT128Index(SDNode *N) {
4744 return isVEXTRACTIndex(N, 128);
4747 bool X86::isVEXTRACT256Index(SDNode *N) {
4748 return isVEXTRACTIndex(N, 256);
4751 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
4752 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
4753 /// Handles 128-bit and 256-bit.
4754 static unsigned getShuffleSHUFImmediate(ShuffleVectorSDNode *N) {
4755 MVT VT = N->getSimpleValueType(0);
4757 assert((VT.getSizeInBits() >= 128) &&
4758 "Unsupported vector type for PSHUF/SHUFP");
4760 // Handle 128 and 256-bit vector lengths. AVX defines PSHUF/SHUFP to operate
4761 // independently on 128-bit lanes.
4762 unsigned NumElts = VT.getVectorNumElements();
4763 unsigned NumLanes = VT.getSizeInBits()/128;
4764 unsigned NumLaneElts = NumElts/NumLanes;
4766 assert((NumLaneElts == 2 || NumLaneElts == 4 || NumLaneElts == 8) &&
4767 "Only supports 2, 4 or 8 elements per lane");
4769 unsigned Shift = (NumLaneElts >= 4) ? 1 : 0;
4771 for (unsigned i = 0; i != NumElts; ++i) {
4772 int Elt = N->getMaskElt(i);
4773 if (Elt < 0) continue;
4774 Elt &= NumLaneElts - 1;
4775 unsigned ShAmt = (i << Shift) % 8;
4776 Mask |= Elt << ShAmt;
4782 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
4783 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
4784 static unsigned getShufflePSHUFHWImmediate(ShuffleVectorSDNode *N) {
4785 MVT VT = N->getSimpleValueType(0);
4787 assert((VT == MVT::v8i16 || VT == MVT::v16i16) &&
4788 "Unsupported vector type for PSHUFHW");
4790 unsigned NumElts = VT.getVectorNumElements();
4793 for (unsigned l = 0; l != NumElts; l += 8) {
4794 // 8 nodes per lane, but we only care about the last 4.
4795 for (unsigned i = 0; i < 4; ++i) {
4796 int Elt = N->getMaskElt(l+i+4);
4797 if (Elt < 0) continue;
4798 Elt &= 0x3; // only 2-bits.
4799 Mask |= Elt << (i * 2);
4806 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
4807 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
4808 static unsigned getShufflePSHUFLWImmediate(ShuffleVectorSDNode *N) {
4809 MVT VT = N->getSimpleValueType(0);
4811 assert((VT == MVT::v8i16 || VT == MVT::v16i16) &&
4812 "Unsupported vector type for PSHUFHW");
4814 unsigned NumElts = VT.getVectorNumElements();
4817 for (unsigned l = 0; l != NumElts; l += 8) {
4818 // 8 nodes per lane, but we only care about the first 4.
4819 for (unsigned i = 0; i < 4; ++i) {
4820 int Elt = N->getMaskElt(l+i);
4821 if (Elt < 0) continue;
4822 Elt &= 0x3; // only 2-bits
4823 Mask |= Elt << (i * 2);
4830 /// \brief Return the appropriate immediate to shuffle the specified
4831 /// VECTOR_SHUFFLE mask with the PALIGNR (if InterLane is false) or with
4832 /// VALIGN (if Interlane is true) instructions.
4833 static unsigned getShuffleAlignrImmediate(ShuffleVectorSDNode *SVOp,
4835 MVT VT = SVOp->getSimpleValueType(0);
4836 unsigned EltSize = InterLane ? 1 :
4837 VT.getVectorElementType().getSizeInBits() >> 3;
4839 unsigned NumElts = VT.getVectorNumElements();
4840 unsigned NumLanes = VT.is512BitVector() ? 1 : VT.getSizeInBits()/128;
4841 unsigned NumLaneElts = NumElts/NumLanes;
4845 for (i = 0; i != NumElts; ++i) {
4846 Val = SVOp->getMaskElt(i);
4850 if (Val >= (int)NumElts)
4851 Val -= NumElts - NumLaneElts;
4853 assert(Val - i > 0 && "PALIGNR imm should be positive");
4854 return (Val - i) * EltSize;
4857 /// \brief Return the appropriate immediate to shuffle the specified
4858 /// VECTOR_SHUFFLE mask with the PALIGNR instruction.
4859 static unsigned getShufflePALIGNRImmediate(ShuffleVectorSDNode *SVOp) {
4860 return getShuffleAlignrImmediate(SVOp, false);
4863 /// \brief Return the appropriate immediate to shuffle the specified
4864 /// VECTOR_SHUFFLE mask with the VALIGN instruction.
4865 static unsigned getShuffleVALIGNImmediate(ShuffleVectorSDNode *SVOp) {
4866 return getShuffleAlignrImmediate(SVOp, true);
4870 static unsigned getExtractVEXTRACTImmediate(SDNode *N, unsigned vecWidth) {
4871 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4872 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4873 llvm_unreachable("Illegal extract subvector for VEXTRACT");
4876 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4878 MVT VecVT = N->getOperand(0).getSimpleValueType();
4879 MVT ElVT = VecVT.getVectorElementType();
4881 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4882 return Index / NumElemsPerChunk;
4885 static unsigned getInsertVINSERTImmediate(SDNode *N, unsigned vecWidth) {
4886 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4887 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4888 llvm_unreachable("Illegal insert subvector for VINSERT");
4891 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4893 MVT VecVT = N->getSimpleValueType(0);
4894 MVT ElVT = VecVT.getVectorElementType();
4896 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4897 return Index / NumElemsPerChunk;
4900 /// getExtractVEXTRACT128Immediate - Return the appropriate immediate
4901 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF128
4902 /// and VINSERTI128 instructions.
4903 unsigned X86::getExtractVEXTRACT128Immediate(SDNode *N) {
4904 return getExtractVEXTRACTImmediate(N, 128);
4907 /// getExtractVEXTRACT256Immediate - Return the appropriate immediate
4908 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF64x4
4909 /// and VINSERTI64x4 instructions.
4910 unsigned X86::getExtractVEXTRACT256Immediate(SDNode *N) {
4911 return getExtractVEXTRACTImmediate(N, 256);
4914 /// getInsertVINSERT128Immediate - Return the appropriate immediate
4915 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF128
4916 /// and VINSERTI128 instructions.
4917 unsigned X86::getInsertVINSERT128Immediate(SDNode *N) {
4918 return getInsertVINSERTImmediate(N, 128);
4921 /// getInsertVINSERT256Immediate - Return the appropriate immediate
4922 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF46x4
4923 /// and VINSERTI64x4 instructions.
4924 unsigned X86::getInsertVINSERT256Immediate(SDNode *N) {
4925 return getInsertVINSERTImmediate(N, 256);
4928 /// isZero - Returns true if Elt is a constant integer zero
4929 static bool isZero(SDValue V) {
4930 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
4931 return C && C->isNullValue();
4934 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
4936 bool X86::isZeroNode(SDValue Elt) {
4939 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Elt))
4940 return CFP->getValueAPF().isPosZero();
4944 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
4945 /// match movhlps. The lower half elements should come from upper half of
4946 /// V1 (and in order), and the upper half elements should come from the upper
4947 /// half of V2 (and in order).
4948 static bool ShouldXformToMOVHLPS(ArrayRef<int> Mask, MVT VT) {
4949 if (!VT.is128BitVector())
4951 if (VT.getVectorNumElements() != 4)
4953 for (unsigned i = 0, e = 2; i != e; ++i)
4954 if (!isUndefOrEqual(Mask[i], i+2))
4956 for (unsigned i = 2; i != 4; ++i)
4957 if (!isUndefOrEqual(Mask[i], i+4))
4962 /// isScalarLoadToVector - Returns true if the node is a scalar load that
4963 /// is promoted to a vector. It also returns the LoadSDNode by reference if
4965 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = nullptr) {
4966 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
4968 N = N->getOperand(0).getNode();
4969 if (!ISD::isNON_EXTLoad(N))
4972 *LD = cast<LoadSDNode>(N);
4976 // Test whether the given value is a vector value which will be legalized
4978 static bool WillBeConstantPoolLoad(SDNode *N) {
4979 if (N->getOpcode() != ISD::BUILD_VECTOR)
4982 // Check for any non-constant elements.
4983 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
4984 switch (N->getOperand(i).getNode()->getOpcode()) {
4986 case ISD::ConstantFP:
4993 // Vectors of all-zeros and all-ones are materialized with special
4994 // instructions rather than being loaded.
4995 return !ISD::isBuildVectorAllZeros(N) &&
4996 !ISD::isBuildVectorAllOnes(N);
4999 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
5000 /// match movlp{s|d}. The lower half elements should come from lower half of
5001 /// V1 (and in order), and the upper half elements should come from the upper
5002 /// half of V2 (and in order). And since V1 will become the source of the
5003 /// MOVLP, it must be either a vector load or a scalar load to vector.
5004 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
5005 ArrayRef<int> Mask, MVT VT) {
5006 if (!VT.is128BitVector())
5009 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
5011 // Is V2 is a vector load, don't do this transformation. We will try to use
5012 // load folding shufps op.
5013 if (ISD::isNON_EXTLoad(V2) || WillBeConstantPoolLoad(V2))
5016 unsigned NumElems = VT.getVectorNumElements();
5018 if (NumElems != 2 && NumElems != 4)
5020 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
5021 if (!isUndefOrEqual(Mask[i], i))
5023 for (unsigned i = NumElems/2, e = NumElems; i != e; ++i)
5024 if (!isUndefOrEqual(Mask[i], i+NumElems))
5029 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
5030 /// to an zero vector.
5031 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
5032 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
5033 SDValue V1 = N->getOperand(0);
5034 SDValue V2 = N->getOperand(1);
5035 unsigned NumElems = N->getValueType(0).getVectorNumElements();
5036 for (unsigned i = 0; i != NumElems; ++i) {
5037 int Idx = N->getMaskElt(i);
5038 if (Idx >= (int)NumElems) {
5039 unsigned Opc = V2.getOpcode();
5040 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
5042 if (Opc != ISD::BUILD_VECTOR ||
5043 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
5045 } else if (Idx >= 0) {
5046 unsigned Opc = V1.getOpcode();
5047 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
5049 if (Opc != ISD::BUILD_VECTOR ||
5050 !X86::isZeroNode(V1.getOperand(Idx)))
5057 /// getZeroVector - Returns a vector of specified type with all zero elements.
5059 static SDValue getZeroVector(EVT VT, const X86Subtarget *Subtarget,
5060 SelectionDAG &DAG, SDLoc dl) {
5061 assert(VT.isVector() && "Expected a vector type");
5063 // Always build SSE zero vectors as <4 x i32> bitcasted
5064 // to their dest type. This ensures they get CSE'd.
5066 if (VT.is128BitVector()) { // SSE
5067 if (Subtarget->hasSSE2()) { // SSE2
5068 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
5069 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
5071 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
5072 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
5074 } else if (VT.is256BitVector()) { // AVX
5075 if (Subtarget->hasInt256()) { // AVX2
5076 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
5077 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
5078 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
5080 // 256-bit logic and arithmetic instructions in AVX are all
5081 // floating-point, no support for integer ops. Emit fp zeroed vectors.
5082 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
5083 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
5084 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops);
5086 } else if (VT.is512BitVector()) { // AVX-512
5087 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
5088 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
5089 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
5090 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
5091 } else if (VT.getScalarType() == MVT::i1) {
5092 assert(VT.getVectorNumElements() <= 16 && "Unexpected vector type");
5093 SDValue Cst = DAG.getTargetConstant(0, MVT::i1);
5094 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5095 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5097 llvm_unreachable("Unexpected vector type");
5099 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
5102 /// getOnesVector - Returns a vector of specified type with all bits set.
5103 /// Always build ones vectors as <4 x i32> or <8 x i32>. For 256-bit types with
5104 /// no AVX2 supprt, use two <4 x i32> inserted in a <8 x i32> appropriately.
5105 /// Then bitcast to their original type, ensuring they get CSE'd.
5106 static SDValue getOnesVector(MVT VT, bool HasInt256, SelectionDAG &DAG,
5108 assert(VT.isVector() && "Expected a vector type");
5110 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
5112 if (VT.is256BitVector()) {
5113 if (HasInt256) { // AVX2
5114 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
5115 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
5117 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
5118 Vec = Concat128BitVectors(Vec, Vec, MVT::v8i32, 8, DAG, dl);
5120 } else if (VT.is128BitVector()) {
5121 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
5123 llvm_unreachable("Unexpected vector type");
5125 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
5128 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
5129 /// that point to V2 points to its first element.
5130 static void NormalizeMask(SmallVectorImpl<int> &Mask, unsigned NumElems) {
5131 for (unsigned i = 0; i != NumElems; ++i) {
5132 if (Mask[i] > (int)NumElems) {
5138 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
5139 /// operation of specified width.
5140 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1,
5142 unsigned NumElems = VT.getVectorNumElements();
5143 SmallVector<int, 8> Mask;
5144 Mask.push_back(NumElems);
5145 for (unsigned i = 1; i != NumElems; ++i)
5147 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
5150 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
5151 static SDValue getUnpackl(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
5153 unsigned NumElems = VT.getVectorNumElements();
5154 SmallVector<int, 8> Mask;
5155 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
5157 Mask.push_back(i + NumElems);
5159 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
5162 /// getUnpackh - Returns a vector_shuffle node for an unpackh operation.
5163 static SDValue getUnpackh(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
5165 unsigned NumElems = VT.getVectorNumElements();
5166 SmallVector<int, 8> Mask;
5167 for (unsigned i = 0, Half = NumElems/2; i != Half; ++i) {
5168 Mask.push_back(i + Half);
5169 Mask.push_back(i + NumElems + Half);
5171 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
5174 // PromoteSplati8i16 - All i16 and i8 vector types can't be used directly by
5175 // a generic shuffle instruction because the target has no such instructions.
5176 // Generate shuffles which repeat i16 and i8 several times until they can be
5177 // represented by v4f32 and then be manipulated by target suported shuffles.
5178 static SDValue PromoteSplati8i16(SDValue V, SelectionDAG &DAG, int &EltNo) {
5179 MVT VT = V.getSimpleValueType();
5180 int NumElems = VT.getVectorNumElements();
5183 while (NumElems > 4) {
5184 if (EltNo < NumElems/2) {
5185 V = getUnpackl(DAG, dl, VT, V, V);
5187 V = getUnpackh(DAG, dl, VT, V, V);
5188 EltNo -= NumElems/2;
5195 /// getLegalSplat - Generate a legal splat with supported x86 shuffles
5196 static SDValue getLegalSplat(SelectionDAG &DAG, SDValue V, int EltNo) {
5197 MVT VT = V.getSimpleValueType();
5200 if (VT.is128BitVector()) {
5201 V = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, V);
5202 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
5203 V = DAG.getVectorShuffle(MVT::v4f32, dl, V, DAG.getUNDEF(MVT::v4f32),
5205 } else if (VT.is256BitVector()) {
5206 // To use VPERMILPS to splat scalars, the second half of indicies must
5207 // refer to the higher part, which is a duplication of the lower one,
5208 // because VPERMILPS can only handle in-lane permutations.
5209 int SplatMask[8] = { EltNo, EltNo, EltNo, EltNo,
5210 EltNo+4, EltNo+4, EltNo+4, EltNo+4 };
5212 V = DAG.getNode(ISD::BITCAST, dl, MVT::v8f32, V);
5213 V = DAG.getVectorShuffle(MVT::v8f32, dl, V, DAG.getUNDEF(MVT::v8f32),
5216 llvm_unreachable("Vector size not supported");
5218 return DAG.getNode(ISD::BITCAST, dl, VT, V);
5221 /// PromoteSplat - Splat is promoted to target supported vector shuffles.
5222 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG) {
5223 MVT SrcVT = SV->getSimpleValueType(0);
5224 SDValue V1 = SV->getOperand(0);
5227 int EltNo = SV->getSplatIndex();
5228 int NumElems = SrcVT.getVectorNumElements();
5229 bool Is256BitVec = SrcVT.is256BitVector();
5231 assert(((SrcVT.is128BitVector() && NumElems > 4) || Is256BitVec) &&
5232 "Unknown how to promote splat for type");
5234 // Extract the 128-bit part containing the splat element and update
5235 // the splat element index when it refers to the higher register.
5237 V1 = Extract128BitVector(V1, EltNo, DAG, dl);
5238 if (EltNo >= NumElems/2)
5239 EltNo -= NumElems/2;
5242 // All i16 and i8 vector types can't be used directly by a generic shuffle
5243 // instruction because the target has no such instruction. Generate shuffles
5244 // which repeat i16 and i8 several times until they fit in i32, and then can
5245 // be manipulated by target suported shuffles.
5246 MVT EltVT = SrcVT.getVectorElementType();
5247 if (EltVT == MVT::i8 || EltVT == MVT::i16)
5248 V1 = PromoteSplati8i16(V1, DAG, EltNo);
5250 // Recreate the 256-bit vector and place the same 128-bit vector
5251 // into the low and high part. This is necessary because we want
5252 // to use VPERM* to shuffle the vectors
5254 V1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, SrcVT, V1, V1);
5257 return getLegalSplat(DAG, V1, EltNo);
5260 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
5261 /// vector of zero or undef vector. This produces a shuffle where the low
5262 /// element of V2 is swizzled into the zero/undef vector, landing at element
5263 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
5264 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
5266 const X86Subtarget *Subtarget,
5267 SelectionDAG &DAG) {
5268 MVT VT = V2.getSimpleValueType();
5270 ? getZeroVector(VT, Subtarget, DAG, SDLoc(V2)) : DAG.getUNDEF(VT);
5271 unsigned NumElems = VT.getVectorNumElements();
5272 SmallVector<int, 16> MaskVec;
5273 for (unsigned i = 0; i != NumElems; ++i)
5274 // If this is the insertion idx, put the low elt of V2 here.
5275 MaskVec.push_back(i == Idx ? NumElems : i);
5276 return DAG.getVectorShuffle(VT, SDLoc(V2), V1, V2, &MaskVec[0]);
5279 /// getTargetShuffleMask - Calculates the shuffle mask corresponding to the
5280 /// target specific opcode. Returns true if the Mask could be calculated. Sets
5281 /// IsUnary to true if only uses one source. Note that this will set IsUnary for
5282 /// shuffles which use a single input multiple times, and in those cases it will
5283 /// adjust the mask to only have indices within that single input.
5284 static bool getTargetShuffleMask(SDNode *N, MVT VT,
5285 SmallVectorImpl<int> &Mask, bool &IsUnary) {
5286 unsigned NumElems = VT.getVectorNumElements();
5290 bool IsFakeUnary = false;
5291 switch(N->getOpcode()) {
5292 case X86ISD::BLENDI:
5293 ImmN = N->getOperand(N->getNumOperands()-1);
5294 DecodeBLENDMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
5297 ImmN = N->getOperand(N->getNumOperands()-1);
5298 DecodeSHUFPMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
5299 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
5301 case X86ISD::UNPCKH:
5302 DecodeUNPCKHMask(VT, Mask);
5303 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
5305 case X86ISD::UNPCKL:
5306 DecodeUNPCKLMask(VT, Mask);
5307 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
5309 case X86ISD::MOVHLPS:
5310 DecodeMOVHLPSMask(NumElems, Mask);
5311 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
5313 case X86ISD::MOVLHPS:
5314 DecodeMOVLHPSMask(NumElems, Mask);
5315 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
5317 case X86ISD::PALIGNR:
5318 ImmN = N->getOperand(N->getNumOperands()-1);
5319 DecodePALIGNRMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
5321 case X86ISD::PSHUFD:
5322 case X86ISD::VPERMILPI:
5323 ImmN = N->getOperand(N->getNumOperands()-1);
5324 DecodePSHUFMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
5327 case X86ISD::PSHUFHW:
5328 ImmN = N->getOperand(N->getNumOperands()-1);
5329 DecodePSHUFHWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
5332 case X86ISD::PSHUFLW:
5333 ImmN = N->getOperand(N->getNumOperands()-1);
5334 DecodePSHUFLWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
5337 case X86ISD::PSHUFB: {
5339 SDValue MaskNode = N->getOperand(1);
5340 while (MaskNode->getOpcode() == ISD::BITCAST)
5341 MaskNode = MaskNode->getOperand(0);
5343 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
5344 // If we have a build-vector, then things are easy.
5345 EVT VT = MaskNode.getValueType();
5346 assert(VT.isVector() &&
5347 "Can't produce a non-vector with a build_vector!");
5348 if (!VT.isInteger())
5351 int NumBytesPerElement = VT.getVectorElementType().getSizeInBits() / 8;
5353 SmallVector<uint64_t, 32> RawMask;
5354 for (int i = 0, e = MaskNode->getNumOperands(); i < e; ++i) {
5355 SDValue Op = MaskNode->getOperand(i);
5356 if (Op->getOpcode() == ISD::UNDEF) {
5357 RawMask.push_back((uint64_t)SM_SentinelUndef);
5360 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
5363 APInt MaskElement = CN->getAPIntValue();
5365 // We now have to decode the element which could be any integer size and
5366 // extract each byte of it.
5367 for (int j = 0; j < NumBytesPerElement; ++j) {
5368 // Note that this is x86 and so always little endian: the low byte is
5369 // the first byte of the mask.
5370 RawMask.push_back(MaskElement.getLoBits(8).getZExtValue());
5371 MaskElement = MaskElement.lshr(8);
5374 DecodePSHUFBMask(RawMask, Mask);
5378 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
5382 SDValue Ptr = MaskLoad->getBasePtr();
5383 if (Ptr->getOpcode() == X86ISD::Wrapper)
5384 Ptr = Ptr->getOperand(0);
5386 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
5387 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
5390 if (auto *C = dyn_cast<Constant>(MaskCP->getConstVal())) {
5391 // FIXME: Support AVX-512 here.
5392 Type *Ty = C->getType();
5393 if (!Ty->isVectorTy() || (Ty->getVectorNumElements() != 16 &&
5394 Ty->getVectorNumElements() != 32))
5397 DecodePSHUFBMask(C, Mask);
5403 case X86ISD::VPERMI:
5404 ImmN = N->getOperand(N->getNumOperands()-1);
5405 DecodeVPERMMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
5409 case X86ISD::MOVSD: {
5410 // The index 0 always comes from the first element of the second source,
5411 // this is why MOVSS and MOVSD are used in the first place. The other
5412 // elements come from the other positions of the first source vector
5413 Mask.push_back(NumElems);
5414 for (unsigned i = 1; i != NumElems; ++i) {
5419 case X86ISD::VPERM2X128:
5420 ImmN = N->getOperand(N->getNumOperands()-1);
5421 DecodeVPERM2X128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
5422 if (Mask.empty()) return false;
5424 case X86ISD::MOVSLDUP:
5425 DecodeMOVSLDUPMask(VT, Mask);
5427 case X86ISD::MOVSHDUP:
5428 DecodeMOVSHDUPMask(VT, Mask);
5430 case X86ISD::MOVDDUP:
5431 case X86ISD::MOVLHPD:
5432 case X86ISD::MOVLPD:
5433 case X86ISD::MOVLPS:
5434 // Not yet implemented
5436 default: llvm_unreachable("unknown target shuffle node");
5439 // If we have a fake unary shuffle, the shuffle mask is spread across two
5440 // inputs that are actually the same node. Re-map the mask to always point
5441 // into the first input.
5444 if (M >= (int)Mask.size())
5450 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
5451 /// element of the result of the vector shuffle.
5452 static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG,
5455 return SDValue(); // Limit search depth.
5457 SDValue V = SDValue(N, 0);
5458 EVT VT = V.getValueType();
5459 unsigned Opcode = V.getOpcode();
5461 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
5462 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
5463 int Elt = SV->getMaskElt(Index);
5466 return DAG.getUNDEF(VT.getVectorElementType());
5468 unsigned NumElems = VT.getVectorNumElements();
5469 SDValue NewV = (Elt < (int)NumElems) ? SV->getOperand(0)
5470 : SV->getOperand(1);
5471 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG, Depth+1);
5474 // Recurse into target specific vector shuffles to find scalars.
5475 if (isTargetShuffle(Opcode)) {
5476 MVT ShufVT = V.getSimpleValueType();
5477 unsigned NumElems = ShufVT.getVectorNumElements();
5478 SmallVector<int, 16> ShuffleMask;
5481 if (!getTargetShuffleMask(N, ShufVT, ShuffleMask, IsUnary))
5484 int Elt = ShuffleMask[Index];
5486 return DAG.getUNDEF(ShufVT.getVectorElementType());
5488 SDValue NewV = (Elt < (int)NumElems) ? N->getOperand(0)
5490 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG,
5494 // Actual nodes that may contain scalar elements
5495 if (Opcode == ISD::BITCAST) {
5496 V = V.getOperand(0);
5497 EVT SrcVT = V.getValueType();
5498 unsigned NumElems = VT.getVectorNumElements();
5500 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
5504 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
5505 return (Index == 0) ? V.getOperand(0)
5506 : DAG.getUNDEF(VT.getVectorElementType());
5508 if (V.getOpcode() == ISD::BUILD_VECTOR)
5509 return V.getOperand(Index);
5514 /// getNumOfConsecutiveZeros - Return the number of elements of a vector
5515 /// shuffle operation which come from a consecutively from a zero. The
5516 /// search can start in two different directions, from left or right.
5517 /// We count undefs as zeros until PreferredNum is reached.
5518 static unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp,
5519 unsigned NumElems, bool ZerosFromLeft,
5521 unsigned PreferredNum = -1U) {
5522 unsigned NumZeros = 0;
5523 for (unsigned i = 0; i != NumElems; ++i) {
5524 unsigned Index = ZerosFromLeft ? i : NumElems - i - 1;
5525 SDValue Elt = getShuffleScalarElt(SVOp, Index, DAG, 0);
5529 if (X86::isZeroNode(Elt))
5531 else if (Elt.getOpcode() == ISD::UNDEF) // Undef as zero up to PreferredNum.
5532 NumZeros = std::min(NumZeros + 1, PreferredNum);
5540 /// isShuffleMaskConsecutive - Check if the shuffle mask indicies [MaskI, MaskE)
5541 /// correspond consecutively to elements from one of the vector operands,
5542 /// starting from its index OpIdx. Also tell OpNum which source vector operand.
5544 bool isShuffleMaskConsecutive(ShuffleVectorSDNode *SVOp,
5545 unsigned MaskI, unsigned MaskE, unsigned OpIdx,
5546 unsigned NumElems, unsigned &OpNum) {
5547 bool SeenV1 = false;
5548 bool SeenV2 = false;
5550 for (unsigned i = MaskI; i != MaskE; ++i, ++OpIdx) {
5551 int Idx = SVOp->getMaskElt(i);
5552 // Ignore undef indicies
5556 if (Idx < (int)NumElems)
5561 // Only accept consecutive elements from the same vector
5562 if ((Idx % NumElems != OpIdx) || (SeenV1 && SeenV2))
5566 OpNum = SeenV1 ? 0 : 1;
5570 /// isVectorShiftRight - Returns true if the shuffle can be implemented as a
5571 /// logical left shift of a vector.
5572 static bool isVectorShiftRight(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
5573 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
5575 SVOp->getSimpleValueType(0).getVectorNumElements();
5576 unsigned NumZeros = getNumOfConsecutiveZeros(
5577 SVOp, NumElems, false /* check zeros from right */, DAG,
5578 SVOp->getMaskElt(0));
5584 // Considering the elements in the mask that are not consecutive zeros,
5585 // check if they consecutively come from only one of the source vectors.
5587 // V1 = {X, A, B, C} 0
5589 // vector_shuffle V1, V2 <1, 2, 3, X>
5591 if (!isShuffleMaskConsecutive(SVOp,
5592 0, // Mask Start Index
5593 NumElems-NumZeros, // Mask End Index(exclusive)
5594 NumZeros, // Where to start looking in the src vector
5595 NumElems, // Number of elements in vector
5596 OpSrc)) // Which source operand ?
5601 ShVal = SVOp->getOperand(OpSrc);
5605 /// isVectorShiftLeft - Returns true if the shuffle can be implemented as a
5606 /// logical left shift of a vector.
5607 static bool isVectorShiftLeft(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
5608 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
5610 SVOp->getSimpleValueType(0).getVectorNumElements();
5611 unsigned NumZeros = getNumOfConsecutiveZeros(
5612 SVOp, NumElems, true /* check zeros from left */, DAG,
5613 NumElems - SVOp->getMaskElt(NumElems - 1) - 1);
5619 // Considering the elements in the mask that are not consecutive zeros,
5620 // check if they consecutively come from only one of the source vectors.
5622 // 0 { A, B, X, X } = V2
5624 // vector_shuffle V1, V2 <X, X, 4, 5>
5626 if (!isShuffleMaskConsecutive(SVOp,
5627 NumZeros, // Mask Start Index
5628 NumElems, // Mask End Index(exclusive)
5629 0, // Where to start looking in the src vector
5630 NumElems, // Number of elements in vector
5631 OpSrc)) // Which source operand ?
5636 ShVal = SVOp->getOperand(OpSrc);
5640 /// isVectorShift - Returns true if the shuffle can be implemented as a
5641 /// logical left or right shift of a vector.
5642 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
5643 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
5644 // Although the logic below support any bitwidth size, there are no
5645 // shift instructions which handle more than 128-bit vectors.
5646 if (!SVOp->getSimpleValueType(0).is128BitVector())
5649 if (isVectorShiftLeft(SVOp, DAG, isLeft, ShVal, ShAmt) ||
5650 isVectorShiftRight(SVOp, DAG, isLeft, ShVal, ShAmt))
5656 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
5658 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
5659 unsigned NumNonZero, unsigned NumZero,
5661 const X86Subtarget* Subtarget,
5662 const TargetLowering &TLI) {
5669 for (unsigned i = 0; i < 16; ++i) {
5670 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
5671 if (ThisIsNonZero && First) {
5673 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
5675 V = DAG.getUNDEF(MVT::v8i16);
5680 SDValue ThisElt, LastElt;
5681 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
5682 if (LastIsNonZero) {
5683 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
5684 MVT::i16, Op.getOperand(i-1));
5686 if (ThisIsNonZero) {
5687 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
5688 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
5689 ThisElt, DAG.getConstant(8, MVT::i8));
5691 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
5695 if (ThisElt.getNode())
5696 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
5697 DAG.getIntPtrConstant(i/2));
5701 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V);
5704 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
5706 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
5707 unsigned NumNonZero, unsigned NumZero,
5709 const X86Subtarget* Subtarget,
5710 const TargetLowering &TLI) {
5717 for (unsigned i = 0; i < 8; ++i) {
5718 bool isNonZero = (NonZeros & (1 << i)) != 0;
5722 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
5724 V = DAG.getUNDEF(MVT::v8i16);
5727 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
5728 MVT::v8i16, V, Op.getOperand(i),
5729 DAG.getIntPtrConstant(i));
5736 /// LowerBuildVectorv4x32 - Custom lower build_vector of v4i32 or v4f32.
5737 static SDValue LowerBuildVectorv4x32(SDValue Op, unsigned NumElems,
5738 unsigned NonZeros, unsigned NumNonZero,
5739 unsigned NumZero, SelectionDAG &DAG,
5740 const X86Subtarget *Subtarget,
5741 const TargetLowering &TLI) {
5742 // We know there's at least one non-zero element
5743 unsigned FirstNonZeroIdx = 0;
5744 SDValue FirstNonZero = Op->getOperand(FirstNonZeroIdx);
5745 while (FirstNonZero.getOpcode() == ISD::UNDEF ||
5746 X86::isZeroNode(FirstNonZero)) {
5748 FirstNonZero = Op->getOperand(FirstNonZeroIdx);
5751 if (FirstNonZero.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5752 !isa<ConstantSDNode>(FirstNonZero.getOperand(1)))
5755 SDValue V = FirstNonZero.getOperand(0);
5756 MVT VVT = V.getSimpleValueType();
5757 if (!Subtarget->hasSSE41() || (VVT != MVT::v4f32 && VVT != MVT::v4i32))
5760 unsigned FirstNonZeroDst =
5761 cast<ConstantSDNode>(FirstNonZero.getOperand(1))->getZExtValue();
5762 unsigned CorrectIdx = FirstNonZeroDst == FirstNonZeroIdx;
5763 unsigned IncorrectIdx = CorrectIdx ? -1U : FirstNonZeroIdx;
5764 unsigned IncorrectDst = CorrectIdx ? -1U : FirstNonZeroDst;
5766 for (unsigned Idx = FirstNonZeroIdx + 1; Idx < NumElems; ++Idx) {
5767 SDValue Elem = Op.getOperand(Idx);
5768 if (Elem.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elem))
5771 // TODO: What else can be here? Deal with it.
5772 if (Elem.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
5775 // TODO: Some optimizations are still possible here
5776 // ex: Getting one element from a vector, and the rest from another.
5777 if (Elem.getOperand(0) != V)
5780 unsigned Dst = cast<ConstantSDNode>(Elem.getOperand(1))->getZExtValue();
5783 else if (IncorrectIdx == -1U) {
5787 // There was already one element with an incorrect index.
5788 // We can't optimize this case to an insertps.
5792 if (NumNonZero == CorrectIdx || NumNonZero == CorrectIdx + 1) {
5794 EVT VT = Op.getSimpleValueType();
5795 unsigned ElementMoveMask = 0;
5796 if (IncorrectIdx == -1U)
5797 ElementMoveMask = FirstNonZeroIdx << 6 | FirstNonZeroIdx << 4;
5799 ElementMoveMask = IncorrectDst << 6 | IncorrectIdx << 4;
5801 SDValue InsertpsMask =
5802 DAG.getIntPtrConstant(ElementMoveMask | (~NonZeros & 0xf));
5803 return DAG.getNode(X86ISD::INSERTPS, dl, VT, V, V, InsertpsMask);
5809 /// getVShift - Return a vector logical shift node.
5811 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
5812 unsigned NumBits, SelectionDAG &DAG,
5813 const TargetLowering &TLI, SDLoc dl) {
5814 assert(VT.is128BitVector() && "Unknown type for VShift");
5815 EVT ShVT = MVT::v2i64;
5816 unsigned Opc = isLeft ? X86ISD::VSHLDQ : X86ISD::VSRLDQ;
5817 SrcOp = DAG.getNode(ISD::BITCAST, dl, ShVT, SrcOp);
5818 return DAG.getNode(ISD::BITCAST, dl, VT,
5819 DAG.getNode(Opc, dl, ShVT, SrcOp,
5820 DAG.getConstant(NumBits,
5821 TLI.getScalarShiftAmountTy(SrcOp.getValueType()))));
5825 LowerAsSplatVectorLoad(SDValue SrcOp, MVT VT, SDLoc dl, SelectionDAG &DAG) {
5827 // Check if the scalar load can be widened into a vector load. And if
5828 // the address is "base + cst" see if the cst can be "absorbed" into
5829 // the shuffle mask.
5830 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
5831 SDValue Ptr = LD->getBasePtr();
5832 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
5834 EVT PVT = LD->getValueType(0);
5835 if (PVT != MVT::i32 && PVT != MVT::f32)
5840 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
5841 FI = FINode->getIndex();
5843 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
5844 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
5845 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
5846 Offset = Ptr.getConstantOperandVal(1);
5847 Ptr = Ptr.getOperand(0);
5852 // FIXME: 256-bit vector instructions don't require a strict alignment,
5853 // improve this code to support it better.
5854 unsigned RequiredAlign = VT.getSizeInBits()/8;
5855 SDValue Chain = LD->getChain();
5856 // Make sure the stack object alignment is at least 16 or 32.
5857 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5858 if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
5859 if (MFI->isFixedObjectIndex(FI)) {
5860 // Can't change the alignment. FIXME: It's possible to compute
5861 // the exact stack offset and reference FI + adjust offset instead.
5862 // If someone *really* cares about this. That's the way to implement it.
5865 MFI->setObjectAlignment(FI, RequiredAlign);
5869 // (Offset % 16 or 32) must be multiple of 4. Then address is then
5870 // Ptr + (Offset & ~15).
5873 if ((Offset % RequiredAlign) & 3)
5875 int64_t StartOffset = Offset & ~(RequiredAlign-1);
5877 Ptr = DAG.getNode(ISD::ADD, SDLoc(Ptr), Ptr.getValueType(),
5878 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
5880 int EltNo = (Offset - StartOffset) >> 2;
5881 unsigned NumElems = VT.getVectorNumElements();
5883 EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
5884 SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
5885 LD->getPointerInfo().getWithOffset(StartOffset),
5886 false, false, false, 0);
5888 SmallVector<int, 8> Mask;
5889 for (unsigned i = 0; i != NumElems; ++i)
5890 Mask.push_back(EltNo);
5892 return DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), &Mask[0]);
5898 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
5899 /// vector of type 'VT', see if the elements can be replaced by a single large
5900 /// load which has the same value as a build_vector whose operands are 'elts'.
5902 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
5904 /// FIXME: we'd also like to handle the case where the last elements are zero
5905 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
5906 /// There's even a handy isZeroNode for that purpose.
5907 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
5908 SDLoc &DL, SelectionDAG &DAG,
5909 bool isAfterLegalize) {
5910 EVT EltVT = VT.getVectorElementType();
5911 unsigned NumElems = Elts.size();
5913 LoadSDNode *LDBase = nullptr;
5914 unsigned LastLoadedElt = -1U;
5916 // For each element in the initializer, see if we've found a load or an undef.
5917 // If we don't find an initial load element, or later load elements are
5918 // non-consecutive, bail out.
5919 for (unsigned i = 0; i < NumElems; ++i) {
5920 SDValue Elt = Elts[i];
5922 if (!Elt.getNode() ||
5923 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
5926 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
5928 LDBase = cast<LoadSDNode>(Elt.getNode());
5932 if (Elt.getOpcode() == ISD::UNDEF)
5935 LoadSDNode *LD = cast<LoadSDNode>(Elt);
5936 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
5941 // If we have found an entire vector of loads and undefs, then return a large
5942 // load of the entire vector width starting at the base pointer. If we found
5943 // consecutive loads for the low half, generate a vzext_load node.
5944 if (LastLoadedElt == NumElems - 1) {
5946 if (isAfterLegalize &&
5947 !DAG.getTargetLoweringInfo().isOperationLegal(ISD::LOAD, VT))
5950 SDValue NewLd = SDValue();
5952 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
5953 NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
5954 LDBase->getPointerInfo(),
5955 LDBase->isVolatile(), LDBase->isNonTemporal(),
5956 LDBase->isInvariant(), 0);
5957 NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
5958 LDBase->getPointerInfo(),
5959 LDBase->isVolatile(), LDBase->isNonTemporal(),
5960 LDBase->isInvariant(), LDBase->getAlignment());
5962 if (LDBase->hasAnyUseOfValue(1)) {
5963 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5965 SDValue(NewLd.getNode(), 1));
5966 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5967 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5968 SDValue(NewLd.getNode(), 1));
5973 if (NumElems == 4 && LastLoadedElt == 1 &&
5974 DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
5975 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
5976 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
5978 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys, Ops, MVT::i64,
5979 LDBase->getPointerInfo(),
5980 LDBase->getAlignment(),
5981 false/*isVolatile*/, true/*ReadMem*/,
5984 // Make sure the newly-created LOAD is in the same position as LDBase in
5985 // terms of dependency. We create a TokenFactor for LDBase and ResNode, and
5986 // update uses of LDBase's output chain to use the TokenFactor.
5987 if (LDBase->hasAnyUseOfValue(1)) {
5988 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5989 SDValue(LDBase, 1), SDValue(ResNode.getNode(), 1));
5990 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5991 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5992 SDValue(ResNode.getNode(), 1));
5995 return DAG.getNode(ISD::BITCAST, DL, VT, ResNode);
6000 /// LowerVectorBroadcast - Attempt to use the vbroadcast instruction
6001 /// to generate a splat value for the following cases:
6002 /// 1. A splat BUILD_VECTOR which uses a single scalar load, or a constant.
6003 /// 2. A splat shuffle which uses a scalar_to_vector node which comes from
6004 /// a scalar load, or a constant.
6005 /// The VBROADCAST node is returned when a pattern is found,
6006 /// or SDValue() otherwise.
6007 static SDValue LowerVectorBroadcast(SDValue Op, const X86Subtarget* Subtarget,
6008 SelectionDAG &DAG) {
6009 // VBROADCAST requires AVX.
6010 // TODO: Splats could be generated for non-AVX CPUs using SSE
6011 // instructions, but there's less potential gain for only 128-bit vectors.
6012 if (!Subtarget->hasAVX())
6015 MVT VT = Op.getSimpleValueType();
6018 assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&
6019 "Unsupported vector type for broadcast.");
6024 switch (Op.getOpcode()) {
6026 // Unknown pattern found.
6029 case ISD::BUILD_VECTOR: {
6030 auto *BVOp = cast<BuildVectorSDNode>(Op.getNode());
6031 BitVector UndefElements;
6032 SDValue Splat = BVOp->getSplatValue(&UndefElements);
6034 // We need a splat of a single value to use broadcast, and it doesn't
6035 // make any sense if the value is only in one element of the vector.
6036 if (!Splat || (VT.getVectorNumElements() - UndefElements.count()) <= 1)
6040 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
6041 Ld.getOpcode() == ISD::ConstantFP);
6043 // Make sure that all of the users of a non-constant load are from the
6044 // BUILD_VECTOR node.
6045 if (!ConstSplatVal && !BVOp->isOnlyUserOf(Ld.getNode()))
6050 case ISD::VECTOR_SHUFFLE: {
6051 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
6053 // Shuffles must have a splat mask where the first element is
6055 if ((!SVOp->isSplat()) || SVOp->getMaskElt(0) != 0)
6058 SDValue Sc = Op.getOperand(0);
6059 if (Sc.getOpcode() != ISD::SCALAR_TO_VECTOR &&
6060 Sc.getOpcode() != ISD::BUILD_VECTOR) {
6062 if (!Subtarget->hasInt256())
6065 // Use the register form of the broadcast instruction available on AVX2.
6066 if (VT.getSizeInBits() >= 256)
6067 Sc = Extract128BitVector(Sc, 0, DAG, dl);
6068 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Sc);
6071 Ld = Sc.getOperand(0);
6072 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
6073 Ld.getOpcode() == ISD::ConstantFP);
6075 // The scalar_to_vector node and the suspected
6076 // load node must have exactly one user.
6077 // Constants may have multiple users.
6079 // AVX-512 has register version of the broadcast
6080 bool hasRegVer = Subtarget->hasAVX512() && VT.is512BitVector() &&
6081 Ld.getValueType().getSizeInBits() >= 32;
6082 if (!ConstSplatVal && ((!Sc.hasOneUse() || !Ld.hasOneUse()) &&
6089 unsigned ScalarSize = Ld.getValueType().getSizeInBits();
6090 bool IsGE256 = (VT.getSizeInBits() >= 256);
6092 // When optimizing for size, generate up to 5 extra bytes for a broadcast
6093 // instruction to save 8 or more bytes of constant pool data.
6094 // TODO: If multiple splats are generated to load the same constant,
6095 // it may be detrimental to overall size. There needs to be a way to detect
6096 // that condition to know if this is truly a size win.
6097 const Function *F = DAG.getMachineFunction().getFunction();
6098 bool OptForSize = F->getAttributes().
6099 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize);
6101 // Handle broadcasting a single constant scalar from the constant pool
6103 // On Sandybridge (no AVX2), it is still better to load a constant vector
6104 // from the constant pool and not to broadcast it from a scalar.
6105 // But override that restriction when optimizing for size.
6106 // TODO: Check if splatting is recommended for other AVX-capable CPUs.
6107 if (ConstSplatVal && (Subtarget->hasAVX2() || OptForSize)) {
6108 EVT CVT = Ld.getValueType();
6109 assert(!CVT.isVector() && "Must not broadcast a vector type");
6111 // Splat f32, i32, v4f64, v4i64 in all cases with AVX2.
6112 // For size optimization, also splat v2f64 and v2i64, and for size opt
6113 // with AVX2, also splat i8 and i16.
6114 // With pattern matching, the VBROADCAST node may become a VMOVDDUP.
6115 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
6116 (OptForSize && (ScalarSize == 64 || Subtarget->hasAVX2()))) {
6117 const Constant *C = nullptr;
6118 if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld))
6119 C = CI->getConstantIntValue();
6120 else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld))
6121 C = CF->getConstantFPValue();
6123 assert(C && "Invalid constant type");
6125 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6126 SDValue CP = DAG.getConstantPool(C, TLI.getPointerTy());
6127 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
6128 Ld = DAG.getLoad(CVT, dl, DAG.getEntryNode(), CP,
6129 MachinePointerInfo::getConstantPool(),
6130 false, false, false, Alignment);
6132 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
6136 bool IsLoad = ISD::isNormalLoad(Ld.getNode());
6138 // Handle AVX2 in-register broadcasts.
6139 if (!IsLoad && Subtarget->hasInt256() &&
6140 (ScalarSize == 32 || (IsGE256 && ScalarSize == 64)))
6141 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
6143 // The scalar source must be a normal load.
6147 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64))
6148 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
6150 // The integer check is needed for the 64-bit into 128-bit so it doesn't match
6151 // double since there is no vbroadcastsd xmm
6152 if (Subtarget->hasInt256() && Ld.getValueType().isInteger()) {
6153 if (ScalarSize == 8 || ScalarSize == 16 || ScalarSize == 64)
6154 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
6157 // Unsupported broadcast.
6161 /// \brief For an EXTRACT_VECTOR_ELT with a constant index return the real
6162 /// underlying vector and index.
6164 /// Modifies \p ExtractedFromVec to the real vector and returns the real
6166 static int getUnderlyingExtractedFromVec(SDValue &ExtractedFromVec,
6168 int Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue();
6169 if (!isa<ShuffleVectorSDNode>(ExtractedFromVec))
6172 // For 256-bit vectors, LowerEXTRACT_VECTOR_ELT_SSE4 may have already
6174 // (extract_vector_elt (v8f32 %vreg1), Constant<6>)
6176 // (extract_vector_elt (vector_shuffle<2,u,u,u>
6177 // (extract_subvector (v8f32 %vreg0), Constant<4>),
6180 // In this case the vector is the extract_subvector expression and the index
6181 // is 2, as specified by the shuffle.
6182 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(ExtractedFromVec);
6183 SDValue ShuffleVec = SVOp->getOperand(0);
6184 MVT ShuffleVecVT = ShuffleVec.getSimpleValueType();
6185 assert(ShuffleVecVT.getVectorElementType() ==
6186 ExtractedFromVec.getSimpleValueType().getVectorElementType());
6188 int ShuffleIdx = SVOp->getMaskElt(Idx);
6189 if (isUndefOrInRange(ShuffleIdx, 0, ShuffleVecVT.getVectorNumElements())) {
6190 ExtractedFromVec = ShuffleVec;
6196 static SDValue buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) {
6197 MVT VT = Op.getSimpleValueType();
6199 // Skip if insert_vec_elt is not supported.
6200 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6201 if (!TLI.isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT))
6205 unsigned NumElems = Op.getNumOperands();
6209 SmallVector<unsigned, 4> InsertIndices;
6210 SmallVector<int, 8> Mask(NumElems, -1);
6212 for (unsigned i = 0; i != NumElems; ++i) {
6213 unsigned Opc = Op.getOperand(i).getOpcode();
6215 if (Opc == ISD::UNDEF)
6218 if (Opc != ISD::EXTRACT_VECTOR_ELT) {
6219 // Quit if more than 1 elements need inserting.
6220 if (InsertIndices.size() > 1)
6223 InsertIndices.push_back(i);
6227 SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0);
6228 SDValue ExtIdx = Op.getOperand(i).getOperand(1);
6229 // Quit if non-constant index.
6230 if (!isa<ConstantSDNode>(ExtIdx))
6232 int Idx = getUnderlyingExtractedFromVec(ExtractedFromVec, ExtIdx);
6234 // Quit if extracted from vector of different type.
6235 if (ExtractedFromVec.getValueType() != VT)
6238 if (!VecIn1.getNode())
6239 VecIn1 = ExtractedFromVec;
6240 else if (VecIn1 != ExtractedFromVec) {
6241 if (!VecIn2.getNode())
6242 VecIn2 = ExtractedFromVec;
6243 else if (VecIn2 != ExtractedFromVec)
6244 // Quit if more than 2 vectors to shuffle
6248 if (ExtractedFromVec == VecIn1)
6250 else if (ExtractedFromVec == VecIn2)
6251 Mask[i] = Idx + NumElems;
6254 if (!VecIn1.getNode())
6257 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
6258 SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, &Mask[0]);
6259 for (unsigned i = 0, e = InsertIndices.size(); i != e; ++i) {
6260 unsigned Idx = InsertIndices[i];
6261 NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx),
6262 DAG.getIntPtrConstant(Idx));
6268 // Lower BUILD_VECTOR operation for v8i1 and v16i1 types.
6270 X86TargetLowering::LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const {
6272 MVT VT = Op.getSimpleValueType();
6273 assert((VT.getVectorElementType() == MVT::i1) && (VT.getSizeInBits() <= 16) &&
6274 "Unexpected type in LowerBUILD_VECTORvXi1!");
6277 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
6278 SDValue Cst = DAG.getTargetConstant(0, MVT::i1);
6279 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
6280 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
6283 if (ISD::isBuildVectorAllOnes(Op.getNode())) {
6284 SDValue Cst = DAG.getTargetConstant(1, MVT::i1);
6285 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
6286 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
6289 bool AllContants = true;
6290 uint64_t Immediate = 0;
6291 int NonConstIdx = -1;
6292 bool IsSplat = true;
6293 unsigned NumNonConsts = 0;
6294 unsigned NumConsts = 0;
6295 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
6296 SDValue In = Op.getOperand(idx);
6297 if (In.getOpcode() == ISD::UNDEF)
6299 if (!isa<ConstantSDNode>(In)) {
6300 AllContants = false;
6306 if (cast<ConstantSDNode>(In)->getZExtValue())
6307 Immediate |= (1ULL << idx);
6309 if (In != Op.getOperand(0))
6314 SDValue FullMask = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1,
6315 DAG.getConstant(Immediate, MVT::i16));
6316 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, FullMask,
6317 DAG.getIntPtrConstant(0));
6320 if (NumNonConsts == 1 && NonConstIdx != 0) {
6323 SDValue VecAsImm = DAG.getConstant(Immediate,
6324 MVT::getIntegerVT(VT.getSizeInBits()));
6325 DstVec = DAG.getNode(ISD::BITCAST, dl, VT, VecAsImm);
6328 DstVec = DAG.getUNDEF(VT);
6329 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
6330 Op.getOperand(NonConstIdx),
6331 DAG.getIntPtrConstant(NonConstIdx));
6333 if (!IsSplat && (NonConstIdx != 0))
6334 llvm_unreachable("Unsupported BUILD_VECTOR operation");
6335 MVT SelectVT = (VT == MVT::v16i1)? MVT::i16 : MVT::i8;
6338 Select = DAG.getNode(ISD::SELECT, dl, SelectVT, Op.getOperand(0),
6339 DAG.getConstant(-1, SelectVT),
6340 DAG.getConstant(0, SelectVT));
6342 Select = DAG.getNode(ISD::SELECT, dl, SelectVT, Op.getOperand(0),
6343 DAG.getConstant((Immediate | 1), SelectVT),
6344 DAG.getConstant(Immediate, SelectVT));
6345 return DAG.getNode(ISD::BITCAST, dl, VT, Select);
6348 /// \brief Return true if \p N implements a horizontal binop and return the
6349 /// operands for the horizontal binop into V0 and V1.
6351 /// This is a helper function of PerformBUILD_VECTORCombine.
6352 /// This function checks that the build_vector \p N in input implements a
6353 /// horizontal operation. Parameter \p Opcode defines the kind of horizontal
6354 /// operation to match.
6355 /// For example, if \p Opcode is equal to ISD::ADD, then this function
6356 /// checks if \p N implements a horizontal arithmetic add; if instead \p Opcode
6357 /// is equal to ISD::SUB, then this function checks if this is a horizontal
6360 /// This function only analyzes elements of \p N whose indices are
6361 /// in range [BaseIdx, LastIdx).
6362 static bool isHorizontalBinOp(const BuildVectorSDNode *N, unsigned Opcode,
6364 unsigned BaseIdx, unsigned LastIdx,
6365 SDValue &V0, SDValue &V1) {
6366 EVT VT = N->getValueType(0);
6368 assert(BaseIdx * 2 <= LastIdx && "Invalid Indices in input!");
6369 assert(VT.isVector() && VT.getVectorNumElements() >= LastIdx &&
6370 "Invalid Vector in input!");
6372 bool IsCommutable = (Opcode == ISD::ADD || Opcode == ISD::FADD);
6373 bool CanFold = true;
6374 unsigned ExpectedVExtractIdx = BaseIdx;
6375 unsigned NumElts = LastIdx - BaseIdx;
6376 V0 = DAG.getUNDEF(VT);
6377 V1 = DAG.getUNDEF(VT);
6379 // Check if N implements a horizontal binop.
6380 for (unsigned i = 0, e = NumElts; i != e && CanFold; ++i) {
6381 SDValue Op = N->getOperand(i + BaseIdx);
6384 if (Op->getOpcode() == ISD::UNDEF) {
6385 // Update the expected vector extract index.
6386 if (i * 2 == NumElts)
6387 ExpectedVExtractIdx = BaseIdx;
6388 ExpectedVExtractIdx += 2;
6392 CanFold = Op->getOpcode() == Opcode && Op->hasOneUse();
6397 SDValue Op0 = Op.getOperand(0);
6398 SDValue Op1 = Op.getOperand(1);
6400 // Try to match the following pattern:
6401 // (BINOP (extract_vector_elt A, I), (extract_vector_elt A, I+1))
6402 CanFold = (Op0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
6403 Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
6404 Op0.getOperand(0) == Op1.getOperand(0) &&
6405 isa<ConstantSDNode>(Op0.getOperand(1)) &&
6406 isa<ConstantSDNode>(Op1.getOperand(1)));
6410 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
6411 unsigned I1 = cast<ConstantSDNode>(Op1.getOperand(1))->getZExtValue();
6413 if (i * 2 < NumElts) {
6414 if (V0.getOpcode() == ISD::UNDEF)
6415 V0 = Op0.getOperand(0);
6417 if (V1.getOpcode() == ISD::UNDEF)
6418 V1 = Op0.getOperand(0);
6419 if (i * 2 == NumElts)
6420 ExpectedVExtractIdx = BaseIdx;
6423 SDValue Expected = (i * 2 < NumElts) ? V0 : V1;
6424 if (I0 == ExpectedVExtractIdx)
6425 CanFold = I1 == I0 + 1 && Op0.getOperand(0) == Expected;
6426 else if (IsCommutable && I1 == ExpectedVExtractIdx) {
6427 // Try to match the following dag sequence:
6428 // (BINOP (extract_vector_elt A, I+1), (extract_vector_elt A, I))
6429 CanFold = I0 == I1 + 1 && Op1.getOperand(0) == Expected;
6433 ExpectedVExtractIdx += 2;
6439 /// \brief Emit a sequence of two 128-bit horizontal add/sub followed by
6440 /// a concat_vector.
6442 /// This is a helper function of PerformBUILD_VECTORCombine.
6443 /// This function expects two 256-bit vectors called V0 and V1.
6444 /// At first, each vector is split into two separate 128-bit vectors.
6445 /// Then, the resulting 128-bit vectors are used to implement two
6446 /// horizontal binary operations.
6448 /// The kind of horizontal binary operation is defined by \p X86Opcode.
6450 /// \p Mode specifies how the 128-bit parts of V0 and V1 are passed in input to
6451 /// the two new horizontal binop.
6452 /// When Mode is set, the first horizontal binop dag node would take as input
6453 /// the lower 128-bit of V0 and the upper 128-bit of V0. The second
6454 /// horizontal binop dag node would take as input the lower 128-bit of V1
6455 /// and the upper 128-bit of V1.
6457 /// HADD V0_LO, V0_HI
6458 /// HADD V1_LO, V1_HI
6460 /// Otherwise, the first horizontal binop dag node takes as input the lower
6461 /// 128-bit of V0 and the lower 128-bit of V1, and the second horizontal binop
6462 /// dag node takes the the upper 128-bit of V0 and the upper 128-bit of V1.
6464 /// HADD V0_LO, V1_LO
6465 /// HADD V0_HI, V1_HI
6467 /// If \p isUndefLO is set, then the algorithm propagates UNDEF to the lower
6468 /// 128-bits of the result. If \p isUndefHI is set, then UNDEF is propagated to
6469 /// the upper 128-bits of the result.
6470 static SDValue ExpandHorizontalBinOp(const SDValue &V0, const SDValue &V1,
6471 SDLoc DL, SelectionDAG &DAG,
6472 unsigned X86Opcode, bool Mode,
6473 bool isUndefLO, bool isUndefHI) {
6474 EVT VT = V0.getValueType();
6475 assert(VT.is256BitVector() && VT == V1.getValueType() &&
6476 "Invalid nodes in input!");
6478 unsigned NumElts = VT.getVectorNumElements();
6479 SDValue V0_LO = Extract128BitVector(V0, 0, DAG, DL);
6480 SDValue V0_HI = Extract128BitVector(V0, NumElts/2, DAG, DL);
6481 SDValue V1_LO = Extract128BitVector(V1, 0, DAG, DL);
6482 SDValue V1_HI = Extract128BitVector(V1, NumElts/2, DAG, DL);
6483 EVT NewVT = V0_LO.getValueType();
6485 SDValue LO = DAG.getUNDEF(NewVT);
6486 SDValue HI = DAG.getUNDEF(NewVT);
6489 // Don't emit a horizontal binop if the result is expected to be UNDEF.
6490 if (!isUndefLO && V0->getOpcode() != ISD::UNDEF)
6491 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V0_HI);
6492 if (!isUndefHI && V1->getOpcode() != ISD::UNDEF)
6493 HI = DAG.getNode(X86Opcode, DL, NewVT, V1_LO, V1_HI);
6495 // Don't emit a horizontal binop if the result is expected to be UNDEF.
6496 if (!isUndefLO && (V0_LO->getOpcode() != ISD::UNDEF ||
6497 V1_LO->getOpcode() != ISD::UNDEF))
6498 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V1_LO);
6500 if (!isUndefHI && (V0_HI->getOpcode() != ISD::UNDEF ||
6501 V1_HI->getOpcode() != ISD::UNDEF))
6502 HI = DAG.getNode(X86Opcode, DL, NewVT, V0_HI, V1_HI);
6505 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LO, HI);
6508 /// \brief Try to fold a build_vector that performs an 'addsub' into the
6509 /// sequence of 'vadd + vsub + blendi'.
6510 static SDValue matchAddSub(const BuildVectorSDNode *BV, SelectionDAG &DAG,
6511 const X86Subtarget *Subtarget) {
6513 EVT VT = BV->getValueType(0);
6514 unsigned NumElts = VT.getVectorNumElements();
6515 SDValue InVec0 = DAG.getUNDEF(VT);
6516 SDValue InVec1 = DAG.getUNDEF(VT);
6518 assert((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v4f32 ||
6519 VT == MVT::v2f64) && "build_vector with an invalid type found!");
6521 // Odd-numbered elements in the input build vector are obtained from
6522 // adding two integer/float elements.
6523 // Even-numbered elements in the input build vector are obtained from
6524 // subtracting two integer/float elements.
6525 unsigned ExpectedOpcode = ISD::FSUB;
6526 unsigned NextExpectedOpcode = ISD::FADD;
6527 bool AddFound = false;
6528 bool SubFound = false;
6530 for (unsigned i = 0, e = NumElts; i != e; i++) {
6531 SDValue Op = BV->getOperand(i);
6533 // Skip 'undef' values.
6534 unsigned Opcode = Op.getOpcode();
6535 if (Opcode == ISD::UNDEF) {
6536 std::swap(ExpectedOpcode, NextExpectedOpcode);
6540 // Early exit if we found an unexpected opcode.
6541 if (Opcode != ExpectedOpcode)
6544 SDValue Op0 = Op.getOperand(0);
6545 SDValue Op1 = Op.getOperand(1);
6547 // Try to match the following pattern:
6548 // (BINOP (extract_vector_elt A, i), (extract_vector_elt B, i))
6549 // Early exit if we cannot match that sequence.
6550 if (Op0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
6551 Op1.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
6552 !isa<ConstantSDNode>(Op0.getOperand(1)) ||
6553 !isa<ConstantSDNode>(Op1.getOperand(1)) ||
6554 Op0.getOperand(1) != Op1.getOperand(1))
6557 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
6561 // We found a valid add/sub node. Update the information accordingly.
6567 // Update InVec0 and InVec1.
6568 if (InVec0.getOpcode() == ISD::UNDEF)
6569 InVec0 = Op0.getOperand(0);
6570 if (InVec1.getOpcode() == ISD::UNDEF)
6571 InVec1 = Op1.getOperand(0);
6573 // Make sure that operands in input to each add/sub node always
6574 // come from a same pair of vectors.
6575 if (InVec0 != Op0.getOperand(0)) {
6576 if (ExpectedOpcode == ISD::FSUB)
6579 // FADD is commutable. Try to commute the operands
6580 // and then test again.
6581 std::swap(Op0, Op1);
6582 if (InVec0 != Op0.getOperand(0))
6586 if (InVec1 != Op1.getOperand(0))
6589 // Update the pair of expected opcodes.
6590 std::swap(ExpectedOpcode, NextExpectedOpcode);
6593 // Don't try to fold this build_vector into an ADDSUB if the inputs are undef.
6594 if (AddFound && SubFound && InVec0.getOpcode() != ISD::UNDEF &&
6595 InVec1.getOpcode() != ISD::UNDEF)
6596 return DAG.getNode(X86ISD::ADDSUB, DL, VT, InVec0, InVec1);
6601 static SDValue PerformBUILD_VECTORCombine(SDNode *N, SelectionDAG &DAG,
6602 const X86Subtarget *Subtarget) {
6604 EVT VT = N->getValueType(0);
6605 unsigned NumElts = VT.getVectorNumElements();
6606 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(N);
6607 SDValue InVec0, InVec1;
6609 // Try to match an ADDSUB.
6610 if ((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
6611 (Subtarget->hasAVX() && (VT == MVT::v8f32 || VT == MVT::v4f64))) {
6612 SDValue Value = matchAddSub(BV, DAG, Subtarget);
6613 if (Value.getNode())
6617 // Try to match horizontal ADD/SUB.
6618 unsigned NumUndefsLO = 0;
6619 unsigned NumUndefsHI = 0;
6620 unsigned Half = NumElts/2;
6622 // Count the number of UNDEF operands in the build_vector in input.
6623 for (unsigned i = 0, e = Half; i != e; ++i)
6624 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
6627 for (unsigned i = Half, e = NumElts; i != e; ++i)
6628 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
6631 // Early exit if this is either a build_vector of all UNDEFs or all the
6632 // operands but one are UNDEF.
6633 if (NumUndefsLO + NumUndefsHI + 1 >= NumElts)
6636 if ((VT == MVT::v4f32 || VT == MVT::v2f64) && Subtarget->hasSSE3()) {
6637 // Try to match an SSE3 float HADD/HSUB.
6638 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
6639 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
6641 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
6642 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
6643 } else if ((VT == MVT::v4i32 || VT == MVT::v8i16) && Subtarget->hasSSSE3()) {
6644 // Try to match an SSSE3 integer HADD/HSUB.
6645 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
6646 return DAG.getNode(X86ISD::HADD, DL, VT, InVec0, InVec1);
6648 if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
6649 return DAG.getNode(X86ISD::HSUB, DL, VT, InVec0, InVec1);
6652 if (!Subtarget->hasAVX())
6655 if ((VT == MVT::v8f32 || VT == MVT::v4f64)) {
6656 // Try to match an AVX horizontal add/sub of packed single/double
6657 // precision floating point values from 256-bit vectors.
6658 SDValue InVec2, InVec3;
6659 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, Half, InVec0, InVec1) &&
6660 isHorizontalBinOp(BV, ISD::FADD, DAG, Half, NumElts, InVec2, InVec3) &&
6661 ((InVec0.getOpcode() == ISD::UNDEF ||
6662 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6663 ((InVec1.getOpcode() == ISD::UNDEF ||
6664 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6665 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
6667 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, Half, InVec0, InVec1) &&
6668 isHorizontalBinOp(BV, ISD::FSUB, DAG, Half, NumElts, InVec2, InVec3) &&
6669 ((InVec0.getOpcode() == ISD::UNDEF ||
6670 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6671 ((InVec1.getOpcode() == ISD::UNDEF ||
6672 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6673 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
6674 } else if (VT == MVT::v8i32 || VT == MVT::v16i16) {
6675 // Try to match an AVX2 horizontal add/sub of signed integers.
6676 SDValue InVec2, InVec3;
6678 bool CanFold = true;
6680 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, Half, InVec0, InVec1) &&
6681 isHorizontalBinOp(BV, ISD::ADD, DAG, Half, NumElts, InVec2, InVec3) &&
6682 ((InVec0.getOpcode() == ISD::UNDEF ||
6683 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6684 ((InVec1.getOpcode() == ISD::UNDEF ||
6685 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6686 X86Opcode = X86ISD::HADD;
6687 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, Half, InVec0, InVec1) &&
6688 isHorizontalBinOp(BV, ISD::SUB, DAG, Half, NumElts, InVec2, InVec3) &&
6689 ((InVec0.getOpcode() == ISD::UNDEF ||
6690 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6691 ((InVec1.getOpcode() == ISD::UNDEF ||
6692 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6693 X86Opcode = X86ISD::HSUB;
6698 // Fold this build_vector into a single horizontal add/sub.
6699 // Do this only if the target has AVX2.
6700 if (Subtarget->hasAVX2())
6701 return DAG.getNode(X86Opcode, DL, VT, InVec0, InVec1);
6703 // Do not try to expand this build_vector into a pair of horizontal
6704 // add/sub if we can emit a pair of scalar add/sub.
6705 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
6708 // Convert this build_vector into a pair of horizontal binop followed by
6710 bool isUndefLO = NumUndefsLO == Half;
6711 bool isUndefHI = NumUndefsHI == Half;
6712 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, false,
6713 isUndefLO, isUndefHI);
6717 if ((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v8i32 ||
6718 VT == MVT::v16i16) && Subtarget->hasAVX()) {
6720 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
6721 X86Opcode = X86ISD::HADD;
6722 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
6723 X86Opcode = X86ISD::HSUB;
6724 else if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
6725 X86Opcode = X86ISD::FHADD;
6726 else if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
6727 X86Opcode = X86ISD::FHSUB;
6731 // Don't try to expand this build_vector into a pair of horizontal add/sub
6732 // if we can simply emit a pair of scalar add/sub.
6733 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
6736 // Convert this build_vector into two horizontal add/sub followed by
6738 bool isUndefLO = NumUndefsLO == Half;
6739 bool isUndefHI = NumUndefsHI == Half;
6740 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, true,
6741 isUndefLO, isUndefHI);
6748 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
6751 MVT VT = Op.getSimpleValueType();
6752 MVT ExtVT = VT.getVectorElementType();
6753 unsigned NumElems = Op.getNumOperands();
6755 // Generate vectors for predicate vectors.
6756 if (VT.getScalarType() == MVT::i1 && Subtarget->hasAVX512())
6757 return LowerBUILD_VECTORvXi1(Op, DAG);
6759 // Vectors containing all zeros can be matched by pxor and xorps later
6760 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
6761 // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
6762 // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
6763 if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32)
6766 return getZeroVector(VT, Subtarget, DAG, dl);
6769 // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
6770 // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use
6771 // vpcmpeqd on 256-bit vectors.
6772 if (Subtarget->hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) {
6773 if (VT == MVT::v4i32 || (VT == MVT::v8i32 && Subtarget->hasInt256()))
6776 if (!VT.is512BitVector())
6777 return getOnesVector(VT, Subtarget->hasInt256(), DAG, dl);
6780 SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG);
6781 if (Broadcast.getNode())
6784 unsigned EVTBits = ExtVT.getSizeInBits();
6786 unsigned NumZero = 0;
6787 unsigned NumNonZero = 0;
6788 unsigned NonZeros = 0;
6789 bool IsAllConstants = true;
6790 SmallSet<SDValue, 8> Values;
6791 for (unsigned i = 0; i < NumElems; ++i) {
6792 SDValue Elt = Op.getOperand(i);
6793 if (Elt.getOpcode() == ISD::UNDEF)
6796 if (Elt.getOpcode() != ISD::Constant &&
6797 Elt.getOpcode() != ISD::ConstantFP)
6798 IsAllConstants = false;
6799 if (X86::isZeroNode(Elt))
6802 NonZeros |= (1 << i);
6807 // All undef vector. Return an UNDEF. All zero vectors were handled above.
6808 if (NumNonZero == 0)
6809 return DAG.getUNDEF(VT);
6811 // Special case for single non-zero, non-undef, element.
6812 if (NumNonZero == 1) {
6813 unsigned Idx = countTrailingZeros(NonZeros);
6814 SDValue Item = Op.getOperand(Idx);
6816 // If this is an insertion of an i64 value on x86-32, and if the top bits of
6817 // the value are obviously zero, truncate the value to i32 and do the
6818 // insertion that way. Only do this if the value is non-constant or if the
6819 // value is a constant being inserted into element 0. It is cheaper to do
6820 // a constant pool load than it is to do a movd + shuffle.
6821 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
6822 (!IsAllConstants || Idx == 0)) {
6823 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
6825 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
6826 EVT VecVT = MVT::v4i32;
6827 unsigned VecElts = 4;
6829 // Truncate the value (which may itself be a constant) to i32, and
6830 // convert it to a vector with movd (S2V+shuffle to zero extend).
6831 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
6832 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
6834 // If using the new shuffle lowering, just directly insert this.
6835 if (ExperimentalVectorShuffleLowering)
6837 ISD::BITCAST, dl, VT,
6838 getShuffleVectorZeroOrUndef(Item, Idx * 2, true, Subtarget, DAG));
6840 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6842 // Now we have our 32-bit value zero extended in the low element of
6843 // a vector. If Idx != 0, swizzle it into place.
6845 SmallVector<int, 4> Mask;
6846 Mask.push_back(Idx);
6847 for (unsigned i = 1; i != VecElts; ++i)
6849 Item = DAG.getVectorShuffle(VecVT, dl, Item, DAG.getUNDEF(VecVT),
6852 return DAG.getNode(ISD::BITCAST, dl, VT, Item);
6856 // If we have a constant or non-constant insertion into the low element of
6857 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
6858 // the rest of the elements. This will be matched as movd/movq/movss/movsd
6859 // depending on what the source datatype is.
6862 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6864 if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
6865 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
6866 if (VT.is256BitVector() || VT.is512BitVector()) {
6867 SDValue ZeroVec = getZeroVector(VT, Subtarget, DAG, dl);
6868 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, ZeroVec,
6869 Item, DAG.getIntPtrConstant(0));
6871 assert(VT.is128BitVector() && "Expected an SSE value type!");
6872 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6873 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
6874 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6877 if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
6878 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
6879 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
6880 if (VT.is256BitVector()) {
6881 SDValue ZeroVec = getZeroVector(MVT::v8i32, Subtarget, DAG, dl);
6882 Item = Insert128BitVector(ZeroVec, Item, 0, DAG, dl);
6884 assert(VT.is128BitVector() && "Expected an SSE value type!");
6885 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6887 return DAG.getNode(ISD::BITCAST, dl, VT, Item);
6891 // Is it a vector logical left shift?
6892 if (NumElems == 2 && Idx == 1 &&
6893 X86::isZeroNode(Op.getOperand(0)) &&
6894 !X86::isZeroNode(Op.getOperand(1))) {
6895 unsigned NumBits = VT.getSizeInBits();
6896 return getVShift(true, VT,
6897 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6898 VT, Op.getOperand(1)),
6899 NumBits/2, DAG, *this, dl);
6902 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
6905 // Otherwise, if this is a vector with i32 or f32 elements, and the element
6906 // is a non-constant being inserted into an element other than the low one,
6907 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
6908 // movd/movss) to move this into the low element, then shuffle it into
6910 if (EVTBits == 32) {
6911 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6913 // If using the new shuffle lowering, just directly insert this.
6914 if (ExperimentalVectorShuffleLowering)
6915 return getShuffleVectorZeroOrUndef(Item, Idx, NumZero > 0, Subtarget, DAG);
6917 // Turn it into a shuffle of zero and zero-extended scalar to vector.
6918 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0, Subtarget, DAG);
6919 SmallVector<int, 8> MaskVec;
6920 for (unsigned i = 0; i != NumElems; ++i)
6921 MaskVec.push_back(i == Idx ? 0 : 1);
6922 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
6926 // Splat is obviously ok. Let legalizer expand it to a shuffle.
6927 if (Values.size() == 1) {
6928 if (EVTBits == 32) {
6929 // Instead of a shuffle like this:
6930 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
6931 // Check if it's possible to issue this instead.
6932 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
6933 unsigned Idx = countTrailingZeros(NonZeros);
6934 SDValue Item = Op.getOperand(Idx);
6935 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
6936 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
6941 // A vector full of immediates; various special cases are already
6942 // handled, so this is best done with a single constant-pool load.
6946 // For AVX-length vectors, build the individual 128-bit pieces and use
6947 // shuffles to put them in place.
6948 if (VT.is256BitVector() || VT.is512BitVector()) {
6949 SmallVector<SDValue, 64> V;
6950 for (unsigned i = 0; i != NumElems; ++i)
6951 V.push_back(Op.getOperand(i));
6953 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
6955 // Build both the lower and upper subvector.
6956 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
6957 makeArrayRef(&V[0], NumElems/2));
6958 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
6959 makeArrayRef(&V[NumElems / 2], NumElems/2));
6961 // Recreate the wider vector with the lower and upper part.
6962 if (VT.is256BitVector())
6963 return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
6964 return Concat256BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
6967 // Let legalizer expand 2-wide build_vectors.
6968 if (EVTBits == 64) {
6969 if (NumNonZero == 1) {
6970 // One half is zero or undef.
6971 unsigned Idx = countTrailingZeros(NonZeros);
6972 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
6973 Op.getOperand(Idx));
6974 return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG);
6979 // If element VT is < 32 bits, convert it to inserts into a zero vector.
6980 if (EVTBits == 8 && NumElems == 16) {
6981 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
6983 if (V.getNode()) return V;
6986 if (EVTBits == 16 && NumElems == 8) {
6987 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
6989 if (V.getNode()) return V;
6992 // If element VT is == 32 bits and has 4 elems, try to generate an INSERTPS
6993 if (EVTBits == 32 && NumElems == 4) {
6994 SDValue V = LowerBuildVectorv4x32(Op, NumElems, NonZeros, NumNonZero,
6995 NumZero, DAG, Subtarget, *this);
7000 // If element VT is == 32 bits, turn it into a number of shuffles.
7001 SmallVector<SDValue, 8> V(NumElems);
7002 if (NumElems == 4 && NumZero > 0) {
7003 for (unsigned i = 0; i < 4; ++i) {
7004 bool isZero = !(NonZeros & (1 << i));
7006 V[i] = getZeroVector(VT, Subtarget, DAG, dl);
7008 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
7011 for (unsigned i = 0; i < 2; ++i) {
7012 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
7015 V[i] = V[i*2]; // Must be a zero vector.
7018 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
7021 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
7024 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
7029 bool Reverse1 = (NonZeros & 0x3) == 2;
7030 bool Reverse2 = ((NonZeros & (0x3 << 2)) >> 2) == 2;
7034 static_cast<int>(Reverse2 ? NumElems+1 : NumElems),
7035 static_cast<int>(Reverse2 ? NumElems : NumElems+1)
7037 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
7040 if (Values.size() > 1 && VT.is128BitVector()) {
7041 // Check for a build vector of consecutive loads.
7042 for (unsigned i = 0; i < NumElems; ++i)
7043 V[i] = Op.getOperand(i);
7045 // Check for elements which are consecutive loads.
7046 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false);
7050 // Check for a build vector from mostly shuffle plus few inserting.
7051 SDValue Sh = buildFromShuffleMostly(Op, DAG);
7055 // For SSE 4.1, use insertps to put the high elements into the low element.
7056 if (getSubtarget()->hasSSE41()) {
7058 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
7059 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
7061 Result = DAG.getUNDEF(VT);
7063 for (unsigned i = 1; i < NumElems; ++i) {
7064 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
7065 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
7066 Op.getOperand(i), DAG.getIntPtrConstant(i));
7071 // Otherwise, expand into a number of unpckl*, start by extending each of
7072 // our (non-undef) elements to the full vector width with the element in the
7073 // bottom slot of the vector (which generates no code for SSE).
7074 for (unsigned i = 0; i < NumElems; ++i) {
7075 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
7076 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
7078 V[i] = DAG.getUNDEF(VT);
7081 // Next, we iteratively mix elements, e.g. for v4f32:
7082 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
7083 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
7084 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
7085 unsigned EltStride = NumElems >> 1;
7086 while (EltStride != 0) {
7087 for (unsigned i = 0; i < EltStride; ++i) {
7088 // If V[i+EltStride] is undef and this is the first round of mixing,
7089 // then it is safe to just drop this shuffle: V[i] is already in the
7090 // right place, the one element (since it's the first round) being
7091 // inserted as undef can be dropped. This isn't safe for successive
7092 // rounds because they will permute elements within both vectors.
7093 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
7094 EltStride == NumElems/2)
7097 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
7106 // LowerAVXCONCAT_VECTORS - 256-bit AVX can use the vinsertf128 instruction
7107 // to create 256-bit vectors from two other 128-bit ones.
7108 static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
7110 MVT ResVT = Op.getSimpleValueType();
7112 assert((ResVT.is256BitVector() ||
7113 ResVT.is512BitVector()) && "Value type must be 256-/512-bit wide");
7115 SDValue V1 = Op.getOperand(0);
7116 SDValue V2 = Op.getOperand(1);
7117 unsigned NumElems = ResVT.getVectorNumElements();
7118 if(ResVT.is256BitVector())
7119 return Concat128BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
7121 if (Op.getNumOperands() == 4) {
7122 MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(),
7123 ResVT.getVectorNumElements()/2);
7124 SDValue V3 = Op.getOperand(2);
7125 SDValue V4 = Op.getOperand(3);
7126 return Concat256BitVectors(Concat128BitVectors(V1, V2, HalfVT, NumElems/2, DAG, dl),
7127 Concat128BitVectors(V3, V4, HalfVT, NumElems/2, DAG, dl), ResVT, NumElems, DAG, dl);
7129 return Concat256BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
7132 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
7133 MVT LLVM_ATTRIBUTE_UNUSED VT = Op.getSimpleValueType();
7134 assert((VT.is256BitVector() && Op.getNumOperands() == 2) ||
7135 (VT.is512BitVector() && (Op.getNumOperands() == 2 ||
7136 Op.getNumOperands() == 4)));
7138 // AVX can use the vinsertf128 instruction to create 256-bit vectors
7139 // from two other 128-bit ones.
7141 // 512-bit vector may contain 2 256-bit vectors or 4 128-bit vectors
7142 return LowerAVXCONCAT_VECTORS(Op, DAG);
7146 //===----------------------------------------------------------------------===//
7147 // Vector shuffle lowering
7149 // This is an experimental code path for lowering vector shuffles on x86. It is
7150 // designed to handle arbitrary vector shuffles and blends, gracefully
7151 // degrading performance as necessary. It works hard to recognize idiomatic
7152 // shuffles and lower them to optimal instruction patterns without leaving
7153 // a framework that allows reasonably efficient handling of all vector shuffle
7155 //===----------------------------------------------------------------------===//
7157 /// \brief Tiny helper function to identify a no-op mask.
7159 /// This is a somewhat boring predicate function. It checks whether the mask
7160 /// array input, which is assumed to be a single-input shuffle mask of the kind
7161 /// used by the X86 shuffle instructions (not a fully general
7162 /// ShuffleVectorSDNode mask) requires any shuffles to occur. Both undef and an
7163 /// in-place shuffle are 'no-op's.
7164 static bool isNoopShuffleMask(ArrayRef<int> Mask) {
7165 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7166 if (Mask[i] != -1 && Mask[i] != i)
7171 /// \brief Helper function to classify a mask as a single-input mask.
7173 /// This isn't a generic single-input test because in the vector shuffle
7174 /// lowering we canonicalize single inputs to be the first input operand. This
7175 /// means we can more quickly test for a single input by only checking whether
7176 /// an input from the second operand exists. We also assume that the size of
7177 /// mask corresponds to the size of the input vectors which isn't true in the
7178 /// fully general case.
7179 static bool isSingleInputShuffleMask(ArrayRef<int> Mask) {
7181 if (M >= (int)Mask.size())
7186 /// \brief Test whether there are elements crossing 128-bit lanes in this
7189 /// X86 divides up its shuffles into in-lane and cross-lane shuffle operations
7190 /// and we routinely test for these.
7191 static bool is128BitLaneCrossingShuffleMask(MVT VT, ArrayRef<int> Mask) {
7192 int LaneSize = 128 / VT.getScalarSizeInBits();
7193 int Size = Mask.size();
7194 for (int i = 0; i < Size; ++i)
7195 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
7200 /// \brief Test whether a shuffle mask is equivalent within each 128-bit lane.
7202 /// This checks a shuffle mask to see if it is performing the same
7203 /// 128-bit lane-relative shuffle in each 128-bit lane. This trivially implies
7204 /// that it is also not lane-crossing. It may however involve a blend from the
7205 /// same lane of a second vector.
7207 /// The specific repeated shuffle mask is populated in \p RepeatedMask, as it is
7208 /// non-trivial to compute in the face of undef lanes. The representation is
7209 /// *not* suitable for use with existing 128-bit shuffles as it will contain
7210 /// entries from both V1 and V2 inputs to the wider mask.
7212 is128BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask,
7213 SmallVectorImpl<int> &RepeatedMask) {
7214 int LaneSize = 128 / VT.getScalarSizeInBits();
7215 RepeatedMask.resize(LaneSize, -1);
7216 int Size = Mask.size();
7217 for (int i = 0; i < Size; ++i) {
7220 if ((Mask[i] % Size) / LaneSize != i / LaneSize)
7221 // This entry crosses lanes, so there is no way to model this shuffle.
7224 // Ok, handle the in-lane shuffles by detecting if and when they repeat.
7225 if (RepeatedMask[i % LaneSize] == -1)
7226 // This is the first non-undef entry in this slot of a 128-bit lane.
7227 RepeatedMask[i % LaneSize] =
7228 Mask[i] < Size ? Mask[i] % LaneSize : Mask[i] % LaneSize + Size;
7229 else if (RepeatedMask[i % LaneSize] + (i / LaneSize) * LaneSize != Mask[i])
7230 // Found a mismatch with the repeated mask.
7236 // Hide this symbol with an anonymous namespace instead of 'static' so that MSVC
7237 // 2013 will allow us to use it as a non-type template parameter.
7240 /// \brief Implementation of the \c isShuffleEquivalent variadic functor.
7242 /// See its documentation for details.
7243 bool isShuffleEquivalentImpl(ArrayRef<int> Mask, ArrayRef<const int *> Args) {
7244 if (Mask.size() != Args.size())
7246 for (int i = 0, e = Mask.size(); i < e; ++i) {
7247 assert(*Args[i] >= 0 && "Arguments must be positive integers!");
7248 if (Mask[i] != -1 && Mask[i] != *Args[i])
7256 /// \brief Checks whether a shuffle mask is equivalent to an explicit list of
7259 /// This is a fast way to test a shuffle mask against a fixed pattern:
7261 /// if (isShuffleEquivalent(Mask, 3, 2, 1, 0)) { ... }
7263 /// It returns true if the mask is exactly as wide as the argument list, and
7264 /// each element of the mask is either -1 (signifying undef) or the value given
7265 /// in the argument.
7266 static const VariadicFunction1<
7267 bool, ArrayRef<int>, int, isShuffleEquivalentImpl> isShuffleEquivalent = {};
7269 /// \brief Get a 4-lane 8-bit shuffle immediate for a mask.
7271 /// This helper function produces an 8-bit shuffle immediate corresponding to
7272 /// the ubiquitous shuffle encoding scheme used in x86 instructions for
7273 /// shuffling 4 lanes. It can be used with most of the PSHUF instructions for
7276 /// NB: We rely heavily on "undef" masks preserving the input lane.
7277 static SDValue getV4X86ShuffleImm8ForMask(ArrayRef<int> Mask,
7278 SelectionDAG &DAG) {
7279 assert(Mask.size() == 4 && "Only 4-lane shuffle masks");
7280 assert(Mask[0] >= -1 && Mask[0] < 4 && "Out of bound mask element!");
7281 assert(Mask[1] >= -1 && Mask[1] < 4 && "Out of bound mask element!");
7282 assert(Mask[2] >= -1 && Mask[2] < 4 && "Out of bound mask element!");
7283 assert(Mask[3] >= -1 && Mask[3] < 4 && "Out of bound mask element!");
7286 Imm |= (Mask[0] == -1 ? 0 : Mask[0]) << 0;
7287 Imm |= (Mask[1] == -1 ? 1 : Mask[1]) << 2;
7288 Imm |= (Mask[2] == -1 ? 2 : Mask[2]) << 4;
7289 Imm |= (Mask[3] == -1 ? 3 : Mask[3]) << 6;
7290 return DAG.getConstant(Imm, MVT::i8);
7293 /// \brief Try to emit a blend instruction for a shuffle.
7295 /// This doesn't do any checks for the availability of instructions for blending
7296 /// these values. It relies on the availability of the X86ISD::BLENDI pattern to
7297 /// be matched in the backend with the type given. What it does check for is
7298 /// that the shuffle mask is in fact a blend.
7299 static SDValue lowerVectorShuffleAsBlend(SDLoc DL, MVT VT, SDValue V1,
7300 SDValue V2, ArrayRef<int> Mask,
7301 const X86Subtarget *Subtarget,
7302 SelectionDAG &DAG) {
7304 unsigned BlendMask = 0;
7305 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
7306 if (Mask[i] >= Size) {
7307 if (Mask[i] != i + Size)
7308 return SDValue(); // Shuffled V2 input!
7309 BlendMask |= 1u << i;
7312 if (Mask[i] >= 0 && Mask[i] != i)
7313 return SDValue(); // Shuffled V1 input!
7315 switch (VT.SimpleTy) {
7320 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V2,
7321 DAG.getConstant(BlendMask, MVT::i8));
7325 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
7329 // If we have AVX2 it is faster to use VPBLENDD when the shuffle fits into
7330 // that instruction.
7331 if (Subtarget->hasAVX2()) {
7332 // Scale the blend by the number of 32-bit dwords per element.
7333 int Scale = VT.getScalarSizeInBits() / 32;
7335 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7336 if (Mask[i] >= Size)
7337 for (int j = 0; j < Scale; ++j)
7338 BlendMask |= 1u << (i * Scale + j);
7340 MVT BlendVT = VT.getSizeInBits() > 128 ? MVT::v8i32 : MVT::v4i32;
7341 V1 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V1);
7342 V2 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V2);
7343 return DAG.getNode(ISD::BITCAST, DL, VT,
7344 DAG.getNode(X86ISD::BLENDI, DL, BlendVT, V1, V2,
7345 DAG.getConstant(BlendMask, MVT::i8)));
7349 // For integer shuffles we need to expand the mask and cast the inputs to
7350 // v8i16s prior to blending.
7351 int Scale = 8 / VT.getVectorNumElements();
7353 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7354 if (Mask[i] >= Size)
7355 for (int j = 0; j < Scale; ++j)
7356 BlendMask |= 1u << (i * Scale + j);
7358 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1);
7359 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V2);
7360 return DAG.getNode(ISD::BITCAST, DL, VT,
7361 DAG.getNode(X86ISD::BLENDI, DL, MVT::v8i16, V1, V2,
7362 DAG.getConstant(BlendMask, MVT::i8)));
7366 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
7367 SmallVector<int, 8> RepeatedMask;
7368 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
7369 // We can lower these with PBLENDW which is mirrored across 128-bit lanes.
7370 assert(RepeatedMask.size() == 8 && "Repeated mask size doesn't match!");
7372 for (int i = 0; i < 8; ++i)
7373 if (RepeatedMask[i] >= 16)
7374 BlendMask |= 1u << i;
7375 return DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2,
7376 DAG.getConstant(BlendMask, MVT::i8));
7381 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
7382 SDValue PBLENDVMask[32];
7383 // Scale the blend by the number of bytes per element.
7384 int Scale = VT.getScalarSizeInBits() / 8;
7385 assert(Mask.size() * Scale == 32 && "Not a 256-bit vector!");
7386 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7387 for (int j = 0; j < Scale; ++j)
7388 PBLENDVMask[Scale * i + j] =
7389 Mask[i] < 0 ? DAG.getUNDEF(MVT::i8)
7390 : DAG.getConstant(Mask[i] < Size ? 0 : 0x80, MVT::i8);
7392 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, V1);
7393 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, V2);
7395 ISD::BITCAST, DL, VT,
7396 DAG.getNode(ISD::VSELECT, DL, MVT::v32i8,
7397 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PBLENDVMask),
7402 llvm_unreachable("Not a supported integer vector type!");
7406 /// \brief Generic routine to lower a shuffle and blend as a decomposed set of
7407 /// unblended shuffles followed by an unshuffled blend.
7409 /// This matches the extremely common pattern for handling combined
7410 /// shuffle+blend operations on newer X86 ISAs where we have very fast blend
7412 static SDValue lowerVectorShuffleAsDecomposedShuffleBlend(SDLoc DL, MVT VT,
7416 SelectionDAG &DAG) {
7417 // Shuffle the input elements into the desired positions in V1 and V2 and
7418 // blend them together.
7419 SmallVector<int, 32> V1Mask(Mask.size(), -1);
7420 SmallVector<int, 32> V2Mask(Mask.size(), -1);
7421 SmallVector<int, 32> BlendMask(Mask.size(), -1);
7422 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7423 if (Mask[i] >= 0 && Mask[i] < Size) {
7424 V1Mask[i] = Mask[i];
7426 } else if (Mask[i] >= Size) {
7427 V2Mask[i] = Mask[i] - Size;
7428 BlendMask[i] = i + Size;
7431 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
7432 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
7433 return DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
7436 /// \brief Try to lower a vector shuffle as a byte rotation.
7438 /// We have a generic PALIGNR instruction in x86 that will do an arbitrary
7439 /// byte-rotation of a the concatentation of two vectors. This routine will
7440 /// try to generically lower a vector shuffle through such an instruction. It
7441 /// does not check for the availability of PALIGNR-based lowerings, only the
7442 /// applicability of this strategy to the given mask. This matches shuffle
7443 /// vectors that look like:
7445 /// v8i16 [11, 12, 13, 14, 15, 0, 1, 2]
7447 /// Essentially it concatenates V1 and V2, shifts right by some number of
7448 /// elements, and takes the low elements as the result. Note that while this is
7449 /// specified as a *right shift* because x86 is little-endian, it is a *left
7450 /// rotate* of the vector lanes.
7452 /// Note that this only handles 128-bit vector widths currently.
7453 static SDValue lowerVectorShuffleAsByteRotate(SDLoc DL, MVT VT, SDValue V1,
7456 SelectionDAG &DAG) {
7457 assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!");
7459 // We need to detect various ways of spelling a rotation:
7460 // [11, 12, 13, 14, 15, 0, 1, 2]
7461 // [-1, 12, 13, 14, -1, -1, 1, -1]
7462 // [-1, -1, -1, -1, -1, -1, 1, 2]
7463 // [ 3, 4, 5, 6, 7, 8, 9, 10]
7464 // [-1, 4, 5, 6, -1, -1, 9, -1]
7465 // [-1, 4, 5, 6, -1, -1, -1, -1]
7468 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
7471 assert(Mask[i] >= 0 && "Only -1 is a valid negative mask element!");
7473 // Based on the mod-Size value of this mask element determine where
7474 // a rotated vector would have started.
7475 int StartIdx = i - (Mask[i] % Size);
7477 // The identity rotation isn't interesting, stop.
7480 // If we found the tail of a vector the rotation must be the missing
7481 // front. If we found the head of a vector, it must be how much of the head.
7482 int CandidateRotation = StartIdx < 0 ? -StartIdx : Size - StartIdx;
7485 Rotation = CandidateRotation;
7486 else if (Rotation != CandidateRotation)
7487 // The rotations don't match, so we can't match this mask.
7490 // Compute which value this mask is pointing at.
7491 SDValue MaskV = Mask[i] < Size ? V1 : V2;
7493 // Compute which of the two target values this index should be assigned to.
7494 // This reflects whether the high elements are remaining or the low elements
7496 SDValue &TargetV = StartIdx < 0 ? Hi : Lo;
7498 // Either set up this value if we've not encountered it before, or check
7499 // that it remains consistent.
7502 else if (TargetV != MaskV)
7503 // This may be a rotation, but it pulls from the inputs in some
7504 // unsupported interleaving.
7508 // Check that we successfully analyzed the mask, and normalize the results.
7509 assert(Rotation != 0 && "Failed to locate a viable rotation!");
7510 assert((Lo || Hi) && "Failed to find a rotated input vector!");
7516 // Cast the inputs to v16i8 to match PALIGNR.
7517 Lo = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Lo);
7518 Hi = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Hi);
7520 assert(VT.getSizeInBits() == 128 &&
7521 "Rotate-based lowering only supports 128-bit lowering!");
7522 assert(Mask.size() <= 16 &&
7523 "Can shuffle at most 16 bytes in a 128-bit vector!");
7524 // The actual rotate instruction rotates bytes, so we need to scale the
7525 // rotation based on how many bytes are in the vector.
7526 int Scale = 16 / Mask.size();
7528 return DAG.getNode(ISD::BITCAST, DL, VT,
7529 DAG.getNode(X86ISD::PALIGNR, DL, MVT::v16i8, Hi, Lo,
7530 DAG.getConstant(Rotation * Scale, MVT::i8)));
7533 /// \brief Compute whether each element of a shuffle is zeroable.
7535 /// A "zeroable" vector shuffle element is one which can be lowered to zero.
7536 /// Either it is an undef element in the shuffle mask, the element of the input
7537 /// referenced is undef, or the element of the input referenced is known to be
7538 /// zero. Many x86 shuffles can zero lanes cheaply and we often want to handle
7539 /// as many lanes with this technique as possible to simplify the remaining
7541 static SmallBitVector computeZeroableShuffleElements(ArrayRef<int> Mask,
7542 SDValue V1, SDValue V2) {
7543 SmallBitVector Zeroable(Mask.size(), false);
7545 bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
7546 bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
7548 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
7550 // Handle the easy cases.
7551 if (M < 0 || (M >= 0 && M < Size && V1IsZero) || (M >= Size && V2IsZero)) {
7556 // If this is an index into a build_vector node, dig out the input value and
7558 SDValue V = M < Size ? V1 : V2;
7559 if (V.getOpcode() != ISD::BUILD_VECTOR)
7562 SDValue Input = V.getOperand(M % Size);
7563 // The UNDEF opcode check really should be dead code here, but not quite
7564 // worth asserting on (it isn't invalid, just unexpected).
7565 if (Input.getOpcode() == ISD::UNDEF || X86::isZeroNode(Input))
7572 /// \brief Lower a vector shuffle as a zero or any extension.
7574 /// Given a specific number of elements, element bit width, and extension
7575 /// stride, produce either a zero or any extension based on the available
7576 /// features of the subtarget.
7577 static SDValue lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7578 SDLoc DL, MVT VT, int NumElements, int Scale, bool AnyExt, SDValue InputV,
7579 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7580 assert(Scale > 1 && "Need a scale to extend.");
7581 int EltBits = VT.getSizeInBits() / NumElements;
7582 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
7583 "Only 8, 16, and 32 bit elements can be extended.");
7584 assert(Scale * EltBits <= 64 && "Cannot zero extend past 64 bits.");
7586 // Found a valid zext mask! Try various lowering strategies based on the
7587 // input type and available ISA extensions.
7588 if (Subtarget->hasSSE41()) {
7589 MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements);
7590 MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits * Scale),
7591 NumElements / Scale);
7592 InputV = DAG.getNode(ISD::BITCAST, DL, InputVT, InputV);
7593 return DAG.getNode(ISD::BITCAST, DL, VT,
7594 DAG.getNode(X86ISD::VZEXT, DL, ExtVT, InputV));
7597 // For any extends we can cheat for larger element sizes and use shuffle
7598 // instructions that can fold with a load and/or copy.
7599 if (AnyExt && EltBits == 32) {
7600 int PSHUFDMask[4] = {0, -1, 1, -1};
7602 ISD::BITCAST, DL, VT,
7603 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7604 DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, InputV),
7605 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
7607 if (AnyExt && EltBits == 16 && Scale > 2) {
7608 int PSHUFDMask[4] = {0, -1, 0, -1};
7609 InputV = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7610 DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, InputV),
7611 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG));
7612 int PSHUFHWMask[4] = {1, -1, -1, -1};
7614 ISD::BITCAST, DL, VT,
7615 DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16,
7616 DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, InputV),
7617 getV4X86ShuffleImm8ForMask(PSHUFHWMask, DAG)));
7620 // If this would require more than 2 unpack instructions to expand, use
7621 // pshufb when available. We can only use more than 2 unpack instructions
7622 // when zero extending i8 elements which also makes it easier to use pshufb.
7623 if (Scale > 4 && EltBits == 8 && Subtarget->hasSSSE3()) {
7624 assert(NumElements == 16 && "Unexpected byte vector width!");
7625 SDValue PSHUFBMask[16];
7626 for (int i = 0; i < 16; ++i)
7628 DAG.getConstant((i % Scale == 0) ? i / Scale : 0x80, MVT::i8);
7629 InputV = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, InputV);
7630 return DAG.getNode(ISD::BITCAST, DL, VT,
7631 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, InputV,
7632 DAG.getNode(ISD::BUILD_VECTOR, DL,
7633 MVT::v16i8, PSHUFBMask)));
7636 // Otherwise emit a sequence of unpacks.
7638 MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements);
7639 SDValue Ext = AnyExt ? DAG.getUNDEF(InputVT)
7640 : getZeroVector(InputVT, Subtarget, DAG, DL);
7641 InputV = DAG.getNode(ISD::BITCAST, DL, InputVT, InputV);
7642 InputV = DAG.getNode(X86ISD::UNPCKL, DL, InputVT, InputV, Ext);
7646 } while (Scale > 1);
7647 return DAG.getNode(ISD::BITCAST, DL, VT, InputV);
7650 /// \brief Try to lower a vector shuffle as a zero extension on any micrarch.
7652 /// This routine will try to do everything in its power to cleverly lower
7653 /// a shuffle which happens to match the pattern of a zero extend. It doesn't
7654 /// check for the profitability of this lowering, it tries to aggressively
7655 /// match this pattern. It will use all of the micro-architectural details it
7656 /// can to emit an efficient lowering. It handles both blends with all-zero
7657 /// inputs to explicitly zero-extend and undef-lanes (sometimes undef due to
7658 /// masking out later).
7660 /// The reason we have dedicated lowering for zext-style shuffles is that they
7661 /// are both incredibly common and often quite performance sensitive.
7662 static SDValue lowerVectorShuffleAsZeroOrAnyExtend(
7663 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7664 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7665 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7667 int Bits = VT.getSizeInBits();
7668 int NumElements = Mask.size();
7670 // Define a helper function to check a particular ext-scale and lower to it if
7672 auto Lower = [&](int Scale) -> SDValue {
7675 for (int i = 0; i < NumElements; ++i) {
7677 continue; // Valid anywhere but doesn't tell us anything.
7678 if (i % Scale != 0) {
7679 // Each of the extend elements needs to be zeroable.
7683 // We no lorger are in the anyext case.
7688 // Each of the base elements needs to be consecutive indices into the
7689 // same input vector.
7690 SDValue V = Mask[i] < NumElements ? V1 : V2;
7693 else if (InputV != V)
7694 return SDValue(); // Flip-flopping inputs.
7696 if (Mask[i] % NumElements != i / Scale)
7697 return SDValue(); // Non-consecutive strided elemenst.
7700 // If we fail to find an input, we have a zero-shuffle which should always
7701 // have already been handled.
7702 // FIXME: Maybe handle this here in case during blending we end up with one?
7706 return lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7707 DL, VT, NumElements, Scale, AnyExt, InputV, Subtarget, DAG);
7710 // The widest scale possible for extending is to a 64-bit integer.
7711 assert(Bits % 64 == 0 &&
7712 "The number of bits in a vector must be divisible by 64 on x86!");
7713 int NumExtElements = Bits / 64;
7715 // Each iteration, try extending the elements half as much, but into twice as
7717 for (; NumExtElements < NumElements; NumExtElements *= 2) {
7718 assert(NumElements % NumExtElements == 0 &&
7719 "The input vector size must be divisble by the extended size.");
7720 if (SDValue V = Lower(NumElements / NumExtElements))
7724 // No viable ext lowering found.
7728 /// \brief Try to lower insertion of a single element into a zero vector.
7730 /// This is a common pattern that we have especially efficient patterns to lower
7731 /// across all subtarget feature sets.
7732 static SDValue lowerVectorShuffleAsElementInsertion(
7733 MVT VT, SDLoc DL, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7734 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7735 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7737 int V2Index = std::find_if(Mask.begin(), Mask.end(),
7738 [&Mask](int M) { return M >= (int)Mask.size(); }) -
7740 if (Mask.size() == 2) {
7741 if (!Zeroable[V2Index ^ 1]) {
7742 // For 2-wide masks we may be able to just invert the inputs. We use an xor
7743 // with 2 to flip from {2,3} to {0,1} and vice versa.
7744 int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2),
7745 Mask[1] < 0 ? -1 : (Mask[1] ^ 2)};
7746 if (Zeroable[V2Index])
7747 return lowerVectorShuffleAsElementInsertion(VT, DL, V2, V1, InverseMask,
7753 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7754 if (i != V2Index && !Zeroable[i])
7755 return SDValue(); // Not inserting into a zero vector.
7758 // Step over any bitcasts on either input so we can scan the actual
7759 // BUILD_VECTOR nodes.
7760 while (V1.getOpcode() == ISD::BITCAST)
7761 V1 = V1.getOperand(0);
7762 while (V2.getOpcode() == ISD::BITCAST)
7763 V2 = V2.getOperand(0);
7765 // Check for a single input from a SCALAR_TO_VECTOR node.
7766 // FIXME: All of this should be canonicalized into INSERT_VECTOR_ELT and
7767 // all the smarts here sunk into that routine. However, the current
7768 // lowering of BUILD_VECTOR makes that nearly impossible until the old
7769 // vector shuffle lowering is dead.
7770 if (!((V2.getOpcode() == ISD::SCALAR_TO_VECTOR &&
7771 Mask[V2Index] == (int)Mask.size()) ||
7772 V2.getOpcode() == ISD::BUILD_VECTOR))
7775 SDValue V2S = V2.getOperand(Mask[V2Index] - Mask.size());
7777 // First, we need to zext the scalar if it is smaller than an i32.
7779 MVT EltVT = VT.getVectorElementType();
7780 V2S = DAG.getNode(ISD::BITCAST, DL, EltVT, V2S);
7781 if (EltVT == MVT::i8 || EltVT == MVT::i16) {
7782 // Zero-extend directly to i32.
7784 V2S = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, V2S);
7787 V2 = DAG.getNode(X86ISD::VZEXT_MOVL, DL, ExtVT,
7788 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtVT, V2S));
7790 V2 = DAG.getNode(ISD::BITCAST, DL, VT, V2);
7793 // If we have 4 or fewer lanes we can cheaply shuffle the element into
7794 // the desired position. Otherwise it is more efficient to do a vector
7795 // shift left. We know that we can do a vector shift left because all
7796 // the inputs are zero.
7797 if (VT.isFloatingPoint() || VT.getVectorNumElements() <= 4) {
7798 SmallVector<int, 4> V2Shuffle(Mask.size(), 1);
7799 V2Shuffle[V2Index] = 0;
7800 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Shuffle);
7802 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, V2);
7804 X86ISD::VSHLDQ, DL, MVT::v2i64, V2,
7806 V2Index * EltVT.getSizeInBits(),
7807 DAG.getTargetLoweringInfo().getScalarShiftAmountTy(MVT::v2i64)));
7808 V2 = DAG.getNode(ISD::BITCAST, DL, VT, V2);
7814 /// \brief Handle lowering of 2-lane 64-bit floating point shuffles.
7816 /// This is the basis function for the 2-lane 64-bit shuffles as we have full
7817 /// support for floating point shuffles but not integer shuffles. These
7818 /// instructions will incur a domain crossing penalty on some chips though so
7819 /// it is better to avoid lowering through this for integer vectors where
7821 static SDValue lowerV2F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7822 const X86Subtarget *Subtarget,
7823 SelectionDAG &DAG) {
7825 assert(Op.getSimpleValueType() == MVT::v2f64 && "Bad shuffle type!");
7826 assert(V1.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
7827 assert(V2.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
7828 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7829 ArrayRef<int> Mask = SVOp->getMask();
7830 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
7832 if (isSingleInputShuffleMask(Mask)) {
7833 // Straight shuffle of a single input vector. Simulate this by using the
7834 // single input as both of the "inputs" to this instruction..
7835 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1);
7837 if (Subtarget->hasAVX()) {
7838 // If we have AVX, we can use VPERMILPS which will allow folding a load
7839 // into the shuffle.
7840 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v2f64, V1,
7841 DAG.getConstant(SHUFPDMask, MVT::i8));
7844 return DAG.getNode(X86ISD::SHUFP, SDLoc(Op), MVT::v2f64, V1, V1,
7845 DAG.getConstant(SHUFPDMask, MVT::i8));
7847 assert(Mask[0] >= 0 && Mask[0] < 2 && "Non-canonicalized blend!");
7848 assert(Mask[1] >= 2 && "Non-canonicalized blend!");
7850 // Use dedicated unpack instructions for masks that match their pattern.
7851 if (isShuffleEquivalent(Mask, 0, 2))
7852 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2f64, V1, V2);
7853 if (isShuffleEquivalent(Mask, 1, 3))
7854 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2f64, V1, V2);
7856 // If we have a single input, insert that into V1 if we can do so cheaply.
7857 if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1)
7858 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7859 MVT::v2f64, DL, V1, V2, Mask, Subtarget, DAG))
7862 if (Subtarget->hasSSE41())
7863 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2f64, V1, V2, Mask,
7867 unsigned SHUFPDMask = (Mask[0] == 1) | (((Mask[1] - 2) == 1) << 1);
7868 return DAG.getNode(X86ISD::SHUFP, SDLoc(Op), MVT::v2f64, V1, V2,
7869 DAG.getConstant(SHUFPDMask, MVT::i8));
7872 /// \brief Handle lowering of 2-lane 64-bit integer shuffles.
7874 /// Tries to lower a 2-lane 64-bit shuffle using shuffle operations provided by
7875 /// the integer unit to minimize domain crossing penalties. However, for blends
7876 /// it falls back to the floating point shuffle operation with appropriate bit
7878 static SDValue lowerV2I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7879 const X86Subtarget *Subtarget,
7880 SelectionDAG &DAG) {
7882 assert(Op.getSimpleValueType() == MVT::v2i64 && "Bad shuffle type!");
7883 assert(V1.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
7884 assert(V2.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
7885 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7886 ArrayRef<int> Mask = SVOp->getMask();
7887 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
7889 if (isSingleInputShuffleMask(Mask)) {
7890 // Straight shuffle of a single input vector. For everything from SSE2
7891 // onward this has a single fast instruction with no scary immediates.
7892 // We have to map the mask as it is actually a v4i32 shuffle instruction.
7893 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V1);
7894 int WidenedMask[4] = {
7895 std::max(Mask[0], 0) * 2, std::max(Mask[0], 0) * 2 + 1,
7896 std::max(Mask[1], 0) * 2, std::max(Mask[1], 0) * 2 + 1};
7898 ISD::BITCAST, DL, MVT::v2i64,
7899 DAG.getNode(X86ISD::PSHUFD, SDLoc(Op), MVT::v4i32, V1,
7900 getV4X86ShuffleImm8ForMask(WidenedMask, DAG)));
7903 // Use dedicated unpack instructions for masks that match their pattern.
7904 if (isShuffleEquivalent(Mask, 0, 2))
7905 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, V1, V2);
7906 if (isShuffleEquivalent(Mask, 1, 3))
7907 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2i64, V1, V2);
7909 // If we have a single input from V2 insert that into V1 if we can do so
7911 if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1)
7912 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7913 MVT::v2i64, DL, V1, V2, Mask, Subtarget, DAG))
7916 if (Subtarget->hasSSE41())
7917 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2i64, V1, V2, Mask,
7921 // Try to use rotation instructions if available.
7922 if (Subtarget->hasSSSE3())
7923 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
7924 DL, MVT::v2i64, V1, V2, Mask, DAG))
7927 // We implement this with SHUFPD which is pretty lame because it will likely
7928 // incur 2 cycles of stall for integer vectors on Nehalem and older chips.
7929 // However, all the alternatives are still more cycles and newer chips don't
7930 // have this problem. It would be really nice if x86 had better shuffles here.
7931 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, V1);
7932 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, V2);
7933 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7934 DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask));
7937 /// \brief Lower a vector shuffle using the SHUFPS instruction.
7939 /// This is a helper routine dedicated to lowering vector shuffles using SHUFPS.
7940 /// It makes no assumptions about whether this is the *best* lowering, it simply
7942 static SDValue lowerVectorShuffleWithSHUPFS(SDLoc DL, MVT VT,
7943 ArrayRef<int> Mask, SDValue V1,
7944 SDValue V2, SelectionDAG &DAG) {
7945 SDValue LowV = V1, HighV = V2;
7946 int NewMask[4] = {Mask[0], Mask[1], Mask[2], Mask[3]};
7949 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
7951 if (NumV2Elements == 1) {
7953 std::find_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }) -
7956 // Compute the index adjacent to V2Index and in the same half by toggling
7958 int V2AdjIndex = V2Index ^ 1;
7960 if (Mask[V2AdjIndex] == -1) {
7961 // Handles all the cases where we have a single V2 element and an undef.
7962 // This will only ever happen in the high lanes because we commute the
7963 // vector otherwise.
7965 std::swap(LowV, HighV);
7966 NewMask[V2Index] -= 4;
7968 // Handle the case where the V2 element ends up adjacent to a V1 element.
7969 // To make this work, blend them together as the first step.
7970 int V1Index = V2AdjIndex;
7971 int BlendMask[4] = {Mask[V2Index] - 4, 0, Mask[V1Index], 0};
7972 V2 = DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
7973 getV4X86ShuffleImm8ForMask(BlendMask, DAG));
7975 // Now proceed to reconstruct the final blend as we have the necessary
7976 // high or low half formed.
7983 NewMask[V1Index] = 2; // We put the V1 element in V2[2].
7984 NewMask[V2Index] = 0; // We shifted the V2 element into V2[0].
7986 } else if (NumV2Elements == 2) {
7987 if (Mask[0] < 4 && Mask[1] < 4) {
7988 // Handle the easy case where we have V1 in the low lanes and V2 in the
7989 // high lanes. We never see this reversed because we sort the shuffle.
7993 // We have a mixture of V1 and V2 in both low and high lanes. Rather than
7994 // trying to place elements directly, just blend them and set up the final
7995 // shuffle to place them.
7997 // The first two blend mask elements are for V1, the second two are for
7999 int BlendMask[4] = {Mask[0] < 4 ? Mask[0] : Mask[1],
8000 Mask[2] < 4 ? Mask[2] : Mask[3],
8001 (Mask[0] >= 4 ? Mask[0] : Mask[1]) - 4,
8002 (Mask[2] >= 4 ? Mask[2] : Mask[3]) - 4};
8003 V1 = DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
8004 getV4X86ShuffleImm8ForMask(BlendMask, DAG));
8006 // Now we do a normal shuffle of V1 by giving V1 as both operands to
8009 NewMask[0] = Mask[0] < 4 ? 0 : 2;
8010 NewMask[1] = Mask[0] < 4 ? 2 : 0;
8011 NewMask[2] = Mask[2] < 4 ? 1 : 3;
8012 NewMask[3] = Mask[2] < 4 ? 3 : 1;
8015 return DAG.getNode(X86ISD::SHUFP, DL, VT, LowV, HighV,
8016 getV4X86ShuffleImm8ForMask(NewMask, DAG));
8019 /// \brief Lower 4-lane 32-bit floating point shuffles.
8021 /// Uses instructions exclusively from the floating point unit to minimize
8022 /// domain crossing penalties, as these are sufficient to implement all v4f32
8024 static SDValue lowerV4F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8025 const X86Subtarget *Subtarget,
8026 SelectionDAG &DAG) {
8028 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
8029 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8030 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8031 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8032 ArrayRef<int> Mask = SVOp->getMask();
8033 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8036 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8038 if (NumV2Elements == 0) {
8039 if (Subtarget->hasAVX()) {
8040 // If we have AVX, we can use VPERMILPS which will allow folding a load
8041 // into the shuffle.
8042 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f32, V1,
8043 getV4X86ShuffleImm8ForMask(Mask, DAG));
8046 // Otherwise, use a straight shuffle of a single input vector. We pass the
8047 // input vector to both operands to simulate this with a SHUFPS.
8048 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1,
8049 getV4X86ShuffleImm8ForMask(Mask, DAG));
8052 // Use dedicated unpack instructions for masks that match their pattern.
8053 if (isShuffleEquivalent(Mask, 0, 4, 1, 5))
8054 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V1, V2);
8055 if (isShuffleEquivalent(Mask, 2, 6, 3, 7))
8056 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V1, V2);
8058 // There are special ways we can lower some single-element blends. However, we
8059 // have custom ways we can lower more complex single-element blends below that
8060 // we defer to if both this and BLENDPS fail to match, so restrict this to
8061 // when the V2 input is targeting element 0 of the mask -- that is the fast
8063 if (NumV2Elements == 1 && Mask[0] >= 4)
8064 if (SDValue V = lowerVectorShuffleAsElementInsertion(MVT::v4f32, DL, V1, V2,
8065 Mask, Subtarget, DAG))
8068 if (Subtarget->hasSSE41())
8069 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f32, V1, V2, Mask,
8073 // Check for whether we can use INSERTPS to perform the blend. We only use
8074 // INSERTPS when the V1 elements are already in the correct locations
8075 // because otherwise we can just always use two SHUFPS instructions which
8076 // are much smaller to encode than a SHUFPS and an INSERTPS.
8077 if (NumV2Elements == 1 && Subtarget->hasSSE41()) {
8079 std::find_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }) -
8082 // When using INSERTPS we can zero any lane of the destination. Collect
8083 // the zero inputs into a mask and drop them from the lanes of V1 which
8084 // actually need to be present as inputs to the INSERTPS.
8085 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
8087 // Synthesize a shuffle mask for the non-zero and non-v2 inputs.
8088 bool InsertNeedsShuffle = false;
8090 for (int i = 0; i < 4; ++i)
8094 } else if (Mask[i] != i) {
8095 InsertNeedsShuffle = true;
8100 // We don't want to use INSERTPS or other insertion techniques if it will
8101 // require shuffling anyways.
8102 if (!InsertNeedsShuffle) {
8103 // If all of V1 is zeroable, replace it with undef.
8104 if ((ZMask | 1 << V2Index) == 0xF)
8105 V1 = DAG.getUNDEF(MVT::v4f32);
8107 unsigned InsertPSMask = (Mask[V2Index] - 4) << 6 | V2Index << 4 | ZMask;
8108 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
8110 // Insert the V2 element into the desired position.
8111 return DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
8112 DAG.getConstant(InsertPSMask, MVT::i8));
8116 // Otherwise fall back to a SHUFPS lowering strategy.
8117 return lowerVectorShuffleWithSHUPFS(DL, MVT::v4f32, Mask, V1, V2, DAG);
8120 /// \brief Lower 4-lane i32 vector shuffles.
8122 /// We try to handle these with integer-domain shuffles where we can, but for
8123 /// blends we use the floating point domain blend instructions.
8124 static SDValue lowerV4I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8125 const X86Subtarget *Subtarget,
8126 SelectionDAG &DAG) {
8128 assert(Op.getSimpleValueType() == MVT::v4i32 && "Bad shuffle type!");
8129 assert(V1.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8130 assert(V2.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8131 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8132 ArrayRef<int> Mask = SVOp->getMask();
8133 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8136 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8138 if (NumV2Elements == 0) {
8139 // Straight shuffle of a single input vector. For everything from SSE2
8140 // onward this has a single fast instruction with no scary immediates.
8141 // We coerce the shuffle pattern to be compatible with UNPCK instructions
8142 // but we aren't actually going to use the UNPCK instruction because doing
8143 // so prevents folding a load into this instruction or making a copy.
8144 const int UnpackLoMask[] = {0, 0, 1, 1};
8145 const int UnpackHiMask[] = {2, 2, 3, 3};
8146 if (isShuffleEquivalent(Mask, 0, 0, 1, 1))
8147 Mask = UnpackLoMask;
8148 else if (isShuffleEquivalent(Mask, 2, 2, 3, 3))
8149 Mask = UnpackHiMask;
8151 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
8152 getV4X86ShuffleImm8ForMask(Mask, DAG));
8155 // Whenever we can lower this as a zext, that instruction is strictly faster
8156 // than any alternative.
8157 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v4i32, V1, V2,
8158 Mask, Subtarget, DAG))
8161 // Use dedicated unpack instructions for masks that match their pattern.
8162 if (isShuffleEquivalent(Mask, 0, 4, 1, 5))
8163 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V1, V2);
8164 if (isShuffleEquivalent(Mask, 2, 6, 3, 7))
8165 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V1, V2);
8167 // There are special ways we can lower some single-element blends.
8168 if (NumV2Elements == 1)
8169 if (SDValue V = lowerVectorShuffleAsElementInsertion(MVT::v4i32, DL, V1, V2,
8170 Mask, Subtarget, DAG))
8173 if (Subtarget->hasSSE41())
8174 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i32, V1, V2, Mask,
8178 // Try to use rotation instructions if available.
8179 if (Subtarget->hasSSSE3())
8180 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8181 DL, MVT::v4i32, V1, V2, Mask, DAG))
8184 // We implement this with SHUFPS because it can blend from two vectors.
8185 // Because we're going to eventually use SHUFPS, we use SHUFPS even to build
8186 // up the inputs, bypassing domain shift penalties that we would encur if we
8187 // directly used PSHUFD on Nehalem and older. For newer chips, this isn't
8189 return DAG.getNode(ISD::BITCAST, DL, MVT::v4i32,
8190 DAG.getVectorShuffle(
8192 DAG.getNode(ISD::BITCAST, DL, MVT::v4f32, V1),
8193 DAG.getNode(ISD::BITCAST, DL, MVT::v4f32, V2), Mask));
8196 /// \brief Lowering of single-input v8i16 shuffles is the cornerstone of SSE2
8197 /// shuffle lowering, and the most complex part.
8199 /// The lowering strategy is to try to form pairs of input lanes which are
8200 /// targeted at the same half of the final vector, and then use a dword shuffle
8201 /// to place them onto the right half, and finally unpack the paired lanes into
8202 /// their final position.
8204 /// The exact breakdown of how to form these dword pairs and align them on the
8205 /// correct sides is really tricky. See the comments within the function for
8206 /// more of the details.
8207 static SDValue lowerV8I16SingleInputVectorShuffle(
8208 SDLoc DL, SDValue V, MutableArrayRef<int> Mask,
8209 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
8210 assert(V.getSimpleValueType() == MVT::v8i16 && "Bad input type!");
8211 MutableArrayRef<int> LoMask = Mask.slice(0, 4);
8212 MutableArrayRef<int> HiMask = Mask.slice(4, 4);
8214 SmallVector<int, 4> LoInputs;
8215 std::copy_if(LoMask.begin(), LoMask.end(), std::back_inserter(LoInputs),
8216 [](int M) { return M >= 0; });
8217 std::sort(LoInputs.begin(), LoInputs.end());
8218 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), LoInputs.end());
8219 SmallVector<int, 4> HiInputs;
8220 std::copy_if(HiMask.begin(), HiMask.end(), std::back_inserter(HiInputs),
8221 [](int M) { return M >= 0; });
8222 std::sort(HiInputs.begin(), HiInputs.end());
8223 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), HiInputs.end());
8225 std::lower_bound(LoInputs.begin(), LoInputs.end(), 4) - LoInputs.begin();
8226 int NumHToL = LoInputs.size() - NumLToL;
8228 std::lower_bound(HiInputs.begin(), HiInputs.end(), 4) - HiInputs.begin();
8229 int NumHToH = HiInputs.size() - NumLToH;
8230 MutableArrayRef<int> LToLInputs(LoInputs.data(), NumLToL);
8231 MutableArrayRef<int> LToHInputs(HiInputs.data(), NumLToH);
8232 MutableArrayRef<int> HToLInputs(LoInputs.data() + NumLToL, NumHToL);
8233 MutableArrayRef<int> HToHInputs(HiInputs.data() + NumLToH, NumHToH);
8235 // Use dedicated unpack instructions for masks that match their pattern.
8236 if (isShuffleEquivalent(Mask, 0, 0, 1, 1, 2, 2, 3, 3))
8237 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V, V);
8238 if (isShuffleEquivalent(Mask, 4, 4, 5, 5, 6, 6, 7, 7))
8239 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V, V);
8241 // Try to use rotation instructions if available.
8242 if (Subtarget->hasSSSE3())
8243 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8244 DL, MVT::v8i16, V, V, Mask, DAG))
8247 // Simplify the 1-into-3 and 3-into-1 cases with a single pshufd. For all
8248 // such inputs we can swap two of the dwords across the half mark and end up
8249 // with <=2 inputs to each half in each half. Once there, we can fall through
8250 // to the generic code below. For example:
8252 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8253 // Mask: [0, 1, 2, 7, 4, 5, 6, 3] -----------------> [0, 1, 4, 7, 2, 3, 6, 5]
8255 // However in some very rare cases we have a 1-into-3 or 3-into-1 on one half
8256 // and an existing 2-into-2 on the other half. In this case we may have to
8257 // pre-shuffle the 2-into-2 half to avoid turning it into a 3-into-1 or
8258 // 1-into-3 which could cause us to cycle endlessly fixing each side in turn.
8259 // Fortunately, we don't have to handle anything but a 2-into-2 pattern
8260 // because any other situation (including a 3-into-1 or 1-into-3 in the other
8261 // half than the one we target for fixing) will be fixed when we re-enter this
8262 // path. We will also combine away any sequence of PSHUFD instructions that
8263 // result into a single instruction. Here is an example of the tricky case:
8265 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8266 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -THIS-IS-BAD!!!!-> [5, 7, 1, 0, 4, 7, 5, 3]
8268 // This now has a 1-into-3 in the high half! Instead, we do two shuffles:
8270 // Input: [a, b, c, d, e, f, g, h] PSHUFHW[0,2,1,3]-> [a, b, c, d, e, g, f, h]
8271 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -----------------> [3, 7, 1, 0, 2, 7, 3, 6]
8273 // Input: [a, b, c, d, e, g, f, h] -PSHUFD[0,2,1,3]-> [a, b, e, g, c, d, f, h]
8274 // Mask: [3, 7, 1, 0, 2, 7, 3, 6] -----------------> [5, 7, 1, 0, 4, 7, 5, 6]
8276 // The result is fine to be handled by the generic logic.
8277 auto balanceSides = [&](ArrayRef<int> AToAInputs, ArrayRef<int> BToAInputs,
8278 ArrayRef<int> BToBInputs, ArrayRef<int> AToBInputs,
8279 int AOffset, int BOffset) {
8280 assert((AToAInputs.size() == 3 || AToAInputs.size() == 1) &&
8281 "Must call this with A having 3 or 1 inputs from the A half.");
8282 assert((BToAInputs.size() == 1 || BToAInputs.size() == 3) &&
8283 "Must call this with B having 1 or 3 inputs from the B half.");
8284 assert(AToAInputs.size() + BToAInputs.size() == 4 &&
8285 "Must call this with either 3:1 or 1:3 inputs (summing to 4).");
8287 // Compute the index of dword with only one word among the three inputs in
8288 // a half by taking the sum of the half with three inputs and subtracting
8289 // the sum of the actual three inputs. The difference is the remaining
8292 int &TripleDWord = AToAInputs.size() == 3 ? ADWord : BDWord;
8293 int &OneInputDWord = AToAInputs.size() == 3 ? BDWord : ADWord;
8294 int TripleInputOffset = AToAInputs.size() == 3 ? AOffset : BOffset;
8295 ArrayRef<int> TripleInputs = AToAInputs.size() == 3 ? AToAInputs : BToAInputs;
8296 int OneInput = AToAInputs.size() == 3 ? BToAInputs[0] : AToAInputs[0];
8297 int TripleInputSum = 0 + 1 + 2 + 3 + (4 * TripleInputOffset);
8298 int TripleNonInputIdx =
8299 TripleInputSum - std::accumulate(TripleInputs.begin(), TripleInputs.end(), 0);
8300 TripleDWord = TripleNonInputIdx / 2;
8302 // We use xor with one to compute the adjacent DWord to whichever one the
8304 OneInputDWord = (OneInput / 2) ^ 1;
8306 // Check for one tricky case: We're fixing a 3<-1 or a 1<-3 shuffle for AToA
8307 // and BToA inputs. If there is also such a problem with the BToB and AToB
8308 // inputs, we don't try to fix it necessarily -- we'll recurse and see it in
8309 // the next pass. However, if we have a 2<-2 in the BToB and AToB inputs, it
8310 // is essential that we don't *create* a 3<-1 as then we might oscillate.
8311 if (BToBInputs.size() == 2 && AToBInputs.size() == 2) {
8312 // Compute how many inputs will be flipped by swapping these DWords. We
8314 // to balance this to ensure we don't form a 3-1 shuffle in the other
8316 int NumFlippedAToBInputs =
8317 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord) +
8318 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord + 1);
8319 int NumFlippedBToBInputs =
8320 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord) +
8321 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord + 1);
8322 if ((NumFlippedAToBInputs == 1 &&
8323 (NumFlippedBToBInputs == 0 || NumFlippedBToBInputs == 2)) ||
8324 (NumFlippedBToBInputs == 1 &&
8325 (NumFlippedAToBInputs == 0 || NumFlippedAToBInputs == 2))) {
8326 // We choose whether to fix the A half or B half based on whether that
8327 // half has zero flipped inputs. At zero, we may not be able to fix it
8328 // with that half. We also bias towards fixing the B half because that
8329 // will more commonly be the high half, and we have to bias one way.
8330 auto FixFlippedInputs = [&V, &DL, &Mask, &DAG](int PinnedIdx, int DWord,
8331 ArrayRef<int> Inputs) {
8332 int FixIdx = PinnedIdx ^ 1; // The adjacent slot to the pinned slot.
8333 bool IsFixIdxInput = std::find(Inputs.begin(), Inputs.end(),
8334 PinnedIdx ^ 1) != Inputs.end();
8335 // Determine whether the free index is in the flipped dword or the
8336 // unflipped dword based on where the pinned index is. We use this bit
8337 // in an xor to conditionally select the adjacent dword.
8338 int FixFreeIdx = 2 * (DWord ^ (PinnedIdx / 2 == DWord));
8339 bool IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8340 FixFreeIdx) != Inputs.end();
8341 if (IsFixIdxInput == IsFixFreeIdxInput)
8343 IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8344 FixFreeIdx) != Inputs.end();
8345 assert(IsFixIdxInput != IsFixFreeIdxInput &&
8346 "We need to be changing the number of flipped inputs!");
8347 int PSHUFHalfMask[] = {0, 1, 2, 3};
8348 std::swap(PSHUFHalfMask[FixFreeIdx % 4], PSHUFHalfMask[FixIdx % 4]);
8349 V = DAG.getNode(FixIdx < 4 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW, DL,
8351 getV4X86ShuffleImm8ForMask(PSHUFHalfMask, DAG));
8354 if (M != -1 && M == FixIdx)
8356 else if (M != -1 && M == FixFreeIdx)
8359 if (NumFlippedBToBInputs != 0) {
8361 BToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
8362 FixFlippedInputs(BPinnedIdx, BDWord, BToBInputs);
8364 assert(NumFlippedAToBInputs != 0 && "Impossible given predicates!");
8366 AToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
8367 FixFlippedInputs(APinnedIdx, ADWord, AToBInputs);
8372 int PSHUFDMask[] = {0, 1, 2, 3};
8373 PSHUFDMask[ADWord] = BDWord;
8374 PSHUFDMask[BDWord] = ADWord;
8375 V = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
8376 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
8377 DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V),
8378 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
8380 // Adjust the mask to match the new locations of A and B.
8382 if (M != -1 && M/2 == ADWord)
8383 M = 2 * BDWord + M % 2;
8384 else if (M != -1 && M/2 == BDWord)
8385 M = 2 * ADWord + M % 2;
8387 // Recurse back into this routine to re-compute state now that this isn't
8388 // a 3 and 1 problem.
8389 return DAG.getVectorShuffle(MVT::v8i16, DL, V, DAG.getUNDEF(MVT::v8i16),
8392 if ((NumLToL == 3 && NumHToL == 1) || (NumLToL == 1 && NumHToL == 3))
8393 return balanceSides(LToLInputs, HToLInputs, HToHInputs, LToHInputs, 0, 4);
8394 else if ((NumHToH == 3 && NumLToH == 1) || (NumHToH == 1 && NumLToH == 3))
8395 return balanceSides(HToHInputs, LToHInputs, LToLInputs, HToLInputs, 4, 0);
8397 // At this point there are at most two inputs to the low and high halves from
8398 // each half. That means the inputs can always be grouped into dwords and
8399 // those dwords can then be moved to the correct half with a dword shuffle.
8400 // We use at most one low and one high word shuffle to collect these paired
8401 // inputs into dwords, and finally a dword shuffle to place them.
8402 int PSHUFLMask[4] = {-1, -1, -1, -1};
8403 int PSHUFHMask[4] = {-1, -1, -1, -1};
8404 int PSHUFDMask[4] = {-1, -1, -1, -1};
8406 // First fix the masks for all the inputs that are staying in their
8407 // original halves. This will then dictate the targets of the cross-half
8409 auto fixInPlaceInputs =
8410 [&PSHUFDMask](ArrayRef<int> InPlaceInputs, ArrayRef<int> IncomingInputs,
8411 MutableArrayRef<int> SourceHalfMask,
8412 MutableArrayRef<int> HalfMask, int HalfOffset) {
8413 if (InPlaceInputs.empty())
8415 if (InPlaceInputs.size() == 1) {
8416 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8417 InPlaceInputs[0] - HalfOffset;
8418 PSHUFDMask[InPlaceInputs[0] / 2] = InPlaceInputs[0] / 2;
8421 if (IncomingInputs.empty()) {
8422 // Just fix all of the in place inputs.
8423 for (int Input : InPlaceInputs) {
8424 SourceHalfMask[Input - HalfOffset] = Input - HalfOffset;
8425 PSHUFDMask[Input / 2] = Input / 2;
8430 assert(InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!");
8431 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8432 InPlaceInputs[0] - HalfOffset;
8433 // Put the second input next to the first so that they are packed into
8434 // a dword. We find the adjacent index by toggling the low bit.
8435 int AdjIndex = InPlaceInputs[0] ^ 1;
8436 SourceHalfMask[AdjIndex - HalfOffset] = InPlaceInputs[1] - HalfOffset;
8437 std::replace(HalfMask.begin(), HalfMask.end(), InPlaceInputs[1], AdjIndex);
8438 PSHUFDMask[AdjIndex / 2] = AdjIndex / 2;
8440 fixInPlaceInputs(LToLInputs, HToLInputs, PSHUFLMask, LoMask, 0);
8441 fixInPlaceInputs(HToHInputs, LToHInputs, PSHUFHMask, HiMask, 4);
8443 // Now gather the cross-half inputs and place them into a free dword of
8444 // their target half.
8445 // FIXME: This operation could almost certainly be simplified dramatically to
8446 // look more like the 3-1 fixing operation.
8447 auto moveInputsToRightHalf = [&PSHUFDMask](
8448 MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs,
8449 MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask,
8450 MutableArrayRef<int> FinalSourceHalfMask, int SourceOffset,
8452 auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) {
8453 return SourceHalfMask[Word] != -1 && SourceHalfMask[Word] != Word;
8455 auto isDWordClobbered = [&isWordClobbered](ArrayRef<int> SourceHalfMask,
8457 int LowWord = Word & ~1;
8458 int HighWord = Word | 1;
8459 return isWordClobbered(SourceHalfMask, LowWord) ||
8460 isWordClobbered(SourceHalfMask, HighWord);
8463 if (IncomingInputs.empty())
8466 if (ExistingInputs.empty()) {
8467 // Map any dwords with inputs from them into the right half.
8468 for (int Input : IncomingInputs) {
8469 // If the source half mask maps over the inputs, turn those into
8470 // swaps and use the swapped lane.
8471 if (isWordClobbered(SourceHalfMask, Input - SourceOffset)) {
8472 if (SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == -1) {
8473 SourceHalfMask[SourceHalfMask[Input - SourceOffset]] =
8474 Input - SourceOffset;
8475 // We have to swap the uses in our half mask in one sweep.
8476 for (int &M : HalfMask)
8477 if (M == SourceHalfMask[Input - SourceOffset] + SourceOffset)
8479 else if (M == Input)
8480 M = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8482 assert(SourceHalfMask[SourceHalfMask[Input - SourceOffset]] ==
8483 Input - SourceOffset &&
8484 "Previous placement doesn't match!");
8486 // Note that this correctly re-maps both when we do a swap and when
8487 // we observe the other side of the swap above. We rely on that to
8488 // avoid swapping the members of the input list directly.
8489 Input = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8492 // Map the input's dword into the correct half.
8493 if (PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == -1)
8494 PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] = Input / 2;
8496 assert(PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] ==
8498 "Previous placement doesn't match!");
8501 // And just directly shift any other-half mask elements to be same-half
8502 // as we will have mirrored the dword containing the element into the
8503 // same position within that half.
8504 for (int &M : HalfMask)
8505 if (M >= SourceOffset && M < SourceOffset + 4) {
8506 M = M - SourceOffset + DestOffset;
8507 assert(M >= 0 && "This should never wrap below zero!");
8512 // Ensure we have the input in a viable dword of its current half. This
8513 // is particularly tricky because the original position may be clobbered
8514 // by inputs being moved and *staying* in that half.
8515 if (IncomingInputs.size() == 1) {
8516 if (isWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8517 int InputFixed = std::find(std::begin(SourceHalfMask),
8518 std::end(SourceHalfMask), -1) -
8519 std::begin(SourceHalfMask) + SourceOffset;
8520 SourceHalfMask[InputFixed - SourceOffset] =
8521 IncomingInputs[0] - SourceOffset;
8522 std::replace(HalfMask.begin(), HalfMask.end(), IncomingInputs[0],
8524 IncomingInputs[0] = InputFixed;
8526 } else if (IncomingInputs.size() == 2) {
8527 if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 ||
8528 isDWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8529 // We have two non-adjacent or clobbered inputs we need to extract from
8530 // the source half. To do this, we need to map them into some adjacent
8531 // dword slot in the source mask.
8532 int InputsFixed[2] = {IncomingInputs[0] - SourceOffset,
8533 IncomingInputs[1] - SourceOffset};
8535 // If there is a free slot in the source half mask adjacent to one of
8536 // the inputs, place the other input in it. We use (Index XOR 1) to
8537 // compute an adjacent index.
8538 if (!isWordClobbered(SourceHalfMask, InputsFixed[0]) &&
8539 SourceHalfMask[InputsFixed[0] ^ 1] == -1) {
8540 SourceHalfMask[InputsFixed[0]] = InputsFixed[0];
8541 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
8542 InputsFixed[1] = InputsFixed[0] ^ 1;
8543 } else if (!isWordClobbered(SourceHalfMask, InputsFixed[1]) &&
8544 SourceHalfMask[InputsFixed[1] ^ 1] == -1) {
8545 SourceHalfMask[InputsFixed[1]] = InputsFixed[1];
8546 SourceHalfMask[InputsFixed[1] ^ 1] = InputsFixed[0];
8547 InputsFixed[0] = InputsFixed[1] ^ 1;
8548 } else if (SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] == -1 &&
8549 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] == -1) {
8550 // The two inputs are in the same DWord but it is clobbered and the
8551 // adjacent DWord isn't used at all. Move both inputs to the free
8553 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] = InputsFixed[0];
8554 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] = InputsFixed[1];
8555 InputsFixed[0] = 2 * ((InputsFixed[0] / 2) ^ 1);
8556 InputsFixed[1] = 2 * ((InputsFixed[0] / 2) ^ 1) + 1;
8558 // The only way we hit this point is if there is no clobbering
8559 // (because there are no off-half inputs to this half) and there is no
8560 // free slot adjacent to one of the inputs. In this case, we have to
8561 // swap an input with a non-input.
8562 for (int i = 0; i < 4; ++i)
8563 assert((SourceHalfMask[i] == -1 || SourceHalfMask[i] == i) &&
8564 "We can't handle any clobbers here!");
8565 assert(InputsFixed[1] != (InputsFixed[0] ^ 1) &&
8566 "Cannot have adjacent inputs here!");
8568 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
8569 SourceHalfMask[InputsFixed[1]] = InputsFixed[0] ^ 1;
8571 // We also have to update the final source mask in this case because
8572 // it may need to undo the above swap.
8573 for (int &M : FinalSourceHalfMask)
8574 if (M == (InputsFixed[0] ^ 1) + SourceOffset)
8575 M = InputsFixed[1] + SourceOffset;
8576 else if (M == InputsFixed[1] + SourceOffset)
8577 M = (InputsFixed[0] ^ 1) + SourceOffset;
8579 InputsFixed[1] = InputsFixed[0] ^ 1;
8582 // Point everything at the fixed inputs.
8583 for (int &M : HalfMask)
8584 if (M == IncomingInputs[0])
8585 M = InputsFixed[0] + SourceOffset;
8586 else if (M == IncomingInputs[1])
8587 M = InputsFixed[1] + SourceOffset;
8589 IncomingInputs[0] = InputsFixed[0] + SourceOffset;
8590 IncomingInputs[1] = InputsFixed[1] + SourceOffset;
8593 llvm_unreachable("Unhandled input size!");
8596 // Now hoist the DWord down to the right half.
8597 int FreeDWord = (PSHUFDMask[DestOffset / 2] == -1 ? 0 : 1) + DestOffset / 2;
8598 assert(PSHUFDMask[FreeDWord] == -1 && "DWord not free");
8599 PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2;
8600 for (int &M : HalfMask)
8601 for (int Input : IncomingInputs)
8603 M = FreeDWord * 2 + Input % 2;
8605 moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask, HiMask,
8606 /*SourceOffset*/ 4, /*DestOffset*/ 0);
8607 moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask, LoMask,
8608 /*SourceOffset*/ 0, /*DestOffset*/ 4);
8610 // Now enact all the shuffles we've computed to move the inputs into their
8612 if (!isNoopShuffleMask(PSHUFLMask))
8613 V = DAG.getNode(X86ISD::PSHUFLW, DL, MVT::v8i16, V,
8614 getV4X86ShuffleImm8ForMask(PSHUFLMask, DAG));
8615 if (!isNoopShuffleMask(PSHUFHMask))
8616 V = DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16, V,
8617 getV4X86ShuffleImm8ForMask(PSHUFHMask, DAG));
8618 if (!isNoopShuffleMask(PSHUFDMask))
8619 V = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
8620 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
8621 DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V),
8622 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
8624 // At this point, each half should contain all its inputs, and we can then
8625 // just shuffle them into their final position.
8626 assert(std::count_if(LoMask.begin(), LoMask.end(),
8627 [](int M) { return M >= 4; }) == 0 &&
8628 "Failed to lift all the high half inputs to the low mask!");
8629 assert(std::count_if(HiMask.begin(), HiMask.end(),
8630 [](int M) { return M >= 0 && M < 4; }) == 0 &&
8631 "Failed to lift all the low half inputs to the high mask!");
8633 // Do a half shuffle for the low mask.
8634 if (!isNoopShuffleMask(LoMask))
8635 V = DAG.getNode(X86ISD::PSHUFLW, DL, MVT::v8i16, V,
8636 getV4X86ShuffleImm8ForMask(LoMask, DAG));
8638 // Do a half shuffle with the high mask after shifting its values down.
8639 for (int &M : HiMask)
8642 if (!isNoopShuffleMask(HiMask))
8643 V = DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16, V,
8644 getV4X86ShuffleImm8ForMask(HiMask, DAG));
8649 /// \brief Detect whether the mask pattern should be lowered through
8652 /// This essentially tests whether viewing the mask as an interleaving of two
8653 /// sub-sequences reduces the cross-input traffic of a blend operation. If so,
8654 /// lowering it through interleaving is a significantly better strategy.
8655 static bool shouldLowerAsInterleaving(ArrayRef<int> Mask) {
8656 int NumEvenInputs[2] = {0, 0};
8657 int NumOddInputs[2] = {0, 0};
8658 int NumLoInputs[2] = {0, 0};
8659 int NumHiInputs[2] = {0, 0};
8660 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
8664 int InputIdx = Mask[i] >= Size;
8667 ++NumLoInputs[InputIdx];
8669 ++NumHiInputs[InputIdx];
8672 ++NumEvenInputs[InputIdx];
8674 ++NumOddInputs[InputIdx];
8677 // The minimum number of cross-input results for both the interleaved and
8678 // split cases. If interleaving results in fewer cross-input results, return
8680 int InterleavedCrosses = std::min(NumEvenInputs[1] + NumOddInputs[0],
8681 NumEvenInputs[0] + NumOddInputs[1]);
8682 int SplitCrosses = std::min(NumLoInputs[1] + NumHiInputs[0],
8683 NumLoInputs[0] + NumHiInputs[1]);
8684 return InterleavedCrosses < SplitCrosses;
8687 /// \brief Blend two v8i16 vectors using a naive unpack strategy.
8689 /// This strategy only works when the inputs from each vector fit into a single
8690 /// half of that vector, and generally there are not so many inputs as to leave
8691 /// the in-place shuffles required highly constrained (and thus expensive). It
8692 /// shifts all the inputs into a single side of both input vectors and then
8693 /// uses an unpack to interleave these inputs in a single vector. At that
8694 /// point, we will fall back on the generic single input shuffle lowering.
8695 static SDValue lowerV8I16BasicBlendVectorShuffle(SDLoc DL, SDValue V1,
8697 MutableArrayRef<int> Mask,
8698 const X86Subtarget *Subtarget,
8699 SelectionDAG &DAG) {
8700 assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad input type!");
8701 assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad input type!");
8702 SmallVector<int, 3> LoV1Inputs, HiV1Inputs, LoV2Inputs, HiV2Inputs;
8703 for (int i = 0; i < 8; ++i)
8704 if (Mask[i] >= 0 && Mask[i] < 4)
8705 LoV1Inputs.push_back(i);
8706 else if (Mask[i] >= 4 && Mask[i] < 8)
8707 HiV1Inputs.push_back(i);
8708 else if (Mask[i] >= 8 && Mask[i] < 12)
8709 LoV2Inputs.push_back(i);
8710 else if (Mask[i] >= 12)
8711 HiV2Inputs.push_back(i);
8713 int NumV1Inputs = LoV1Inputs.size() + HiV1Inputs.size();
8714 int NumV2Inputs = LoV2Inputs.size() + HiV2Inputs.size();
8717 assert(NumV1Inputs > 0 && NumV1Inputs <= 3 && "At most 3 inputs supported");
8718 assert(NumV2Inputs > 0 && NumV2Inputs <= 3 && "At most 3 inputs supported");
8719 assert(NumV1Inputs + NumV2Inputs <= 4 && "At most 4 combined inputs");
8721 bool MergeFromLo = LoV1Inputs.size() + LoV2Inputs.size() >=
8722 HiV1Inputs.size() + HiV2Inputs.size();
8724 auto moveInputsToHalf = [&](SDValue V, ArrayRef<int> LoInputs,
8725 ArrayRef<int> HiInputs, bool MoveToLo,
8727 ArrayRef<int> GoodInputs = MoveToLo ? LoInputs : HiInputs;
8728 ArrayRef<int> BadInputs = MoveToLo ? HiInputs : LoInputs;
8729 if (BadInputs.empty())
8732 int MoveMask[] = {-1, -1, -1, -1, -1, -1, -1, -1};
8733 int MoveOffset = MoveToLo ? 0 : 4;
8735 if (GoodInputs.empty()) {
8736 for (int BadInput : BadInputs) {
8737 MoveMask[Mask[BadInput] % 4 + MoveOffset] = Mask[BadInput] - MaskOffset;
8738 Mask[BadInput] = Mask[BadInput] % 4 + MoveOffset + MaskOffset;
8741 if (GoodInputs.size() == 2) {
8742 // If the low inputs are spread across two dwords, pack them into
8744 MoveMask[MoveOffset] = Mask[GoodInputs[0]] - MaskOffset;
8745 MoveMask[MoveOffset + 1] = Mask[GoodInputs[1]] - MaskOffset;
8746 Mask[GoodInputs[0]] = MoveOffset + MaskOffset;
8747 Mask[GoodInputs[1]] = MoveOffset + 1 + MaskOffset;
8749 // Otherwise pin the good inputs.
8750 for (int GoodInput : GoodInputs)
8751 MoveMask[Mask[GoodInput] - MaskOffset] = Mask[GoodInput] - MaskOffset;
8754 if (BadInputs.size() == 2) {
8755 // If we have two bad inputs then there may be either one or two good
8756 // inputs fixed in place. Find a fixed input, and then find the *other*
8757 // two adjacent indices by using modular arithmetic.
8759 std::find_if(std::begin(MoveMask) + MoveOffset, std::end(MoveMask),
8760 [](int M) { return M >= 0; }) -
8761 std::begin(MoveMask);
8763 ((((GoodMaskIdx - MoveOffset) & ~1) + 2) % 4) + MoveOffset;
8764 assert(MoveMask[MoveMaskIdx] == -1 && "Expected empty slot");
8765 assert(MoveMask[MoveMaskIdx + 1] == -1 && "Expected empty slot");
8766 MoveMask[MoveMaskIdx] = Mask[BadInputs[0]] - MaskOffset;
8767 MoveMask[MoveMaskIdx + 1] = Mask[BadInputs[1]] - MaskOffset;
8768 Mask[BadInputs[0]] = MoveMaskIdx + MaskOffset;
8769 Mask[BadInputs[1]] = MoveMaskIdx + 1 + MaskOffset;
8771 assert(BadInputs.size() == 1 && "All sizes handled");
8772 int MoveMaskIdx = std::find(std::begin(MoveMask) + MoveOffset,
8773 std::end(MoveMask), -1) -
8774 std::begin(MoveMask);
8775 MoveMask[MoveMaskIdx] = Mask[BadInputs[0]] - MaskOffset;
8776 Mask[BadInputs[0]] = MoveMaskIdx + MaskOffset;
8780 return DAG.getVectorShuffle(MVT::v8i16, DL, V, DAG.getUNDEF(MVT::v8i16),
8783 V1 = moveInputsToHalf(V1, LoV1Inputs, HiV1Inputs, MergeFromLo,
8785 V2 = moveInputsToHalf(V2, LoV2Inputs, HiV2Inputs, MergeFromLo,
8788 // FIXME: Select an interleaving of the merge of V1 and V2 that minimizes
8789 // cross-half traffic in the final shuffle.
8791 // Munge the mask to be a single-input mask after the unpack merges the
8795 M = 2 * (M % 4) + (M / 8);
8797 return DAG.getVectorShuffle(
8798 MVT::v8i16, DL, DAG.getNode(MergeFromLo ? X86ISD::UNPCKL : X86ISD::UNPCKH,
8799 DL, MVT::v8i16, V1, V2),
8800 DAG.getUNDEF(MVT::v8i16), Mask);
8803 /// \brief Generic lowering of 8-lane i16 shuffles.
8805 /// This handles both single-input shuffles and combined shuffle/blends with
8806 /// two inputs. The single input shuffles are immediately delegated to
8807 /// a dedicated lowering routine.
8809 /// The blends are lowered in one of three fundamental ways. If there are few
8810 /// enough inputs, it delegates to a basic UNPCK-based strategy. If the shuffle
8811 /// of the input is significantly cheaper when lowered as an interleaving of
8812 /// the two inputs, try to interleave them. Otherwise, blend the low and high
8813 /// halves of the inputs separately (making them have relatively few inputs)
8814 /// and then concatenate them.
8815 static SDValue lowerV8I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8816 const X86Subtarget *Subtarget,
8817 SelectionDAG &DAG) {
8819 assert(Op.getSimpleValueType() == MVT::v8i16 && "Bad shuffle type!");
8820 assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
8821 assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
8822 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8823 ArrayRef<int> OrigMask = SVOp->getMask();
8824 int MaskStorage[8] = {OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
8825 OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7]};
8826 MutableArrayRef<int> Mask(MaskStorage);
8828 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
8830 // Whenever we can lower this as a zext, that instruction is strictly faster
8831 // than any alternative.
8832 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
8833 DL, MVT::v8i16, V1, V2, OrigMask, Subtarget, DAG))
8836 auto isV1 = [](int M) { return M >= 0 && M < 8; };
8837 auto isV2 = [](int M) { return M >= 8; };
8839 int NumV1Inputs = std::count_if(Mask.begin(), Mask.end(), isV1);
8840 int NumV2Inputs = std::count_if(Mask.begin(), Mask.end(), isV2);
8842 if (NumV2Inputs == 0)
8843 return lowerV8I16SingleInputVectorShuffle(DL, V1, Mask, Subtarget, DAG);
8845 assert(NumV1Inputs > 0 && "All single-input shuffles should be canonicalized "
8846 "to be V1-input shuffles.");
8848 // There are special ways we can lower some single-element blends.
8849 if (NumV2Inputs == 1)
8850 if (SDValue V = lowerVectorShuffleAsElementInsertion(MVT::v8i16, DL, V1, V2,
8851 Mask, Subtarget, DAG))
8854 if (Subtarget->hasSSE41())
8855 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i16, V1, V2, Mask,
8859 // Try to use rotation instructions if available.
8860 if (Subtarget->hasSSSE3())
8861 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(DL, MVT::v8i16, V1, V2, Mask, DAG))
8864 if (NumV1Inputs + NumV2Inputs <= 4)
8865 return lowerV8I16BasicBlendVectorShuffle(DL, V1, V2, Mask, Subtarget, DAG);
8867 // Check whether an interleaving lowering is likely to be more efficient.
8868 // This isn't perfect but it is a strong heuristic that tends to work well on
8869 // the kinds of shuffles that show up in practice.
8871 // FIXME: Handle 1x, 2x, and 4x interleaving.
8872 if (shouldLowerAsInterleaving(Mask)) {
8873 // FIXME: Figure out whether we should pack these into the low or high
8876 int EMask[8], OMask[8];
8877 for (int i = 0; i < 4; ++i) {
8878 EMask[i] = Mask[2*i];
8879 OMask[i] = Mask[2*i + 1];
8884 SDValue Evens = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, EMask);
8885 SDValue Odds = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, OMask);
8887 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, Evens, Odds);
8890 int LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
8891 int HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
8893 for (int i = 0; i < 4; ++i) {
8894 LoBlendMask[i] = Mask[i];
8895 HiBlendMask[i] = Mask[i + 4];
8898 SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, LoBlendMask);
8899 SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, HiBlendMask);
8900 LoV = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, LoV);
8901 HiV = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, HiV);
8903 return DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
8904 DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, LoV, HiV));
8907 /// \brief Check whether a compaction lowering can be done by dropping even
8908 /// elements and compute how many times even elements must be dropped.
8910 /// This handles shuffles which take every Nth element where N is a power of
8911 /// two. Example shuffle masks:
8913 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 0, 2, 4, 6, 8, 10, 12, 14
8914 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30
8915 /// N = 2: 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12
8916 /// N = 2: 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28
8917 /// N = 3: 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8
8918 /// N = 3: 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24
8920 /// Any of these lanes can of course be undef.
8922 /// This routine only supports N <= 3.
8923 /// FIXME: Evaluate whether either AVX or AVX-512 have any opportunities here
8926 /// \returns N above, or the number of times even elements must be dropped if
8927 /// there is such a number. Otherwise returns zero.
8928 static int canLowerByDroppingEvenElements(ArrayRef<int> Mask) {
8929 // Figure out whether we're looping over two inputs or just one.
8930 bool IsSingleInput = isSingleInputShuffleMask(Mask);
8932 // The modulus for the shuffle vector entries is based on whether this is
8933 // a single input or not.
8934 int ShuffleModulus = Mask.size() * (IsSingleInput ? 1 : 2);
8935 assert(isPowerOf2_32((uint32_t)ShuffleModulus) &&
8936 "We should only be called with masks with a power-of-2 size!");
8938 uint64_t ModMask = (uint64_t)ShuffleModulus - 1;
8940 // We track whether the input is viable for all power-of-2 strides 2^1, 2^2,
8941 // and 2^3 simultaneously. This is because we may have ambiguity with
8942 // partially undef inputs.
8943 bool ViableForN[3] = {true, true, true};
8945 for (int i = 0, e = Mask.size(); i < e; ++i) {
8946 // Ignore undef lanes, we'll optimistically collapse them to the pattern we
8951 bool IsAnyViable = false;
8952 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
8953 if (ViableForN[j]) {
8956 // The shuffle mask must be equal to (i * 2^N) % M.
8957 if ((uint64_t)Mask[i] == (((uint64_t)i << N) & ModMask))
8960 ViableForN[j] = false;
8962 // Early exit if we exhaust the possible powers of two.
8967 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
8971 // Return 0 as there is no viable power of two.
8975 /// \brief Generic lowering of v16i8 shuffles.
8977 /// This is a hybrid strategy to lower v16i8 vectors. It first attempts to
8978 /// detect any complexity reducing interleaving. If that doesn't help, it uses
8979 /// UNPCK to spread the i8 elements across two i16-element vectors, and uses
8980 /// the existing lowering for v8i16 blends on each half, finally PACK-ing them
8982 static SDValue lowerV16I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8983 const X86Subtarget *Subtarget,
8984 SelectionDAG &DAG) {
8986 assert(Op.getSimpleValueType() == MVT::v16i8 && "Bad shuffle type!");
8987 assert(V1.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
8988 assert(V2.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
8989 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8990 ArrayRef<int> OrigMask = SVOp->getMask();
8991 assert(OrigMask.size() == 16 && "Unexpected mask size for v16 shuffle!");
8993 // Try to use rotation instructions if available.
8994 if (Subtarget->hasSSSE3())
8995 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(DL, MVT::v16i8, V1, V2,
8999 // Try to use a zext lowering.
9000 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
9001 DL, MVT::v16i8, V1, V2, OrigMask, Subtarget, DAG))
9004 int MaskStorage[16] = {
9005 OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
9006 OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7],
9007 OrigMask[8], OrigMask[9], OrigMask[10], OrigMask[11],
9008 OrigMask[12], OrigMask[13], OrigMask[14], OrigMask[15]};
9009 MutableArrayRef<int> Mask(MaskStorage);
9010 MutableArrayRef<int> LoMask = Mask.slice(0, 8);
9011 MutableArrayRef<int> HiMask = Mask.slice(8, 8);
9014 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 16; });
9016 // For single-input shuffles, there are some nicer lowering tricks we can use.
9017 if (NumV2Elements == 0) {
9018 // Check whether we can widen this to an i16 shuffle by duplicating bytes.
9019 // Notably, this handles splat and partial-splat shuffles more efficiently.
9020 // However, it only makes sense if the pre-duplication shuffle simplifies
9021 // things significantly. Currently, this means we need to be able to
9022 // express the pre-duplication shuffle as an i16 shuffle.
9024 // FIXME: We should check for other patterns which can be widened into an
9025 // i16 shuffle as well.
9026 auto canWidenViaDuplication = [](ArrayRef<int> Mask) {
9027 for (int i = 0; i < 16; i += 2)
9028 if (Mask[i] != -1 && Mask[i + 1] != -1 && Mask[i] != Mask[i + 1])
9033 auto tryToWidenViaDuplication = [&]() -> SDValue {
9034 if (!canWidenViaDuplication(Mask))
9036 SmallVector<int, 4> LoInputs;
9037 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(LoInputs),
9038 [](int M) { return M >= 0 && M < 8; });
9039 std::sort(LoInputs.begin(), LoInputs.end());
9040 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()),
9042 SmallVector<int, 4> HiInputs;
9043 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(HiInputs),
9044 [](int M) { return M >= 8; });
9045 std::sort(HiInputs.begin(), HiInputs.end());
9046 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()),
9049 bool TargetLo = LoInputs.size() >= HiInputs.size();
9050 ArrayRef<int> InPlaceInputs = TargetLo ? LoInputs : HiInputs;
9051 ArrayRef<int> MovingInputs = TargetLo ? HiInputs : LoInputs;
9053 int PreDupI16Shuffle[] = {-1, -1, -1, -1, -1, -1, -1, -1};
9054 SmallDenseMap<int, int, 8> LaneMap;
9055 for (int I : InPlaceInputs) {
9056 PreDupI16Shuffle[I/2] = I/2;
9059 int j = TargetLo ? 0 : 4, je = j + 4;
9060 for (int i = 0, ie = MovingInputs.size(); i < ie; ++i) {
9061 // Check if j is already a shuffle of this input. This happens when
9062 // there are two adjacent bytes after we move the low one.
9063 if (PreDupI16Shuffle[j] != MovingInputs[i] / 2) {
9064 // If we haven't yet mapped the input, search for a slot into which
9066 while (j < je && PreDupI16Shuffle[j] != -1)
9070 // We can't place the inputs into a single half with a simple i16 shuffle, so bail.
9073 // Map this input with the i16 shuffle.
9074 PreDupI16Shuffle[j] = MovingInputs[i] / 2;
9077 // Update the lane map based on the mapping we ended up with.
9078 LaneMap[MovingInputs[i]] = 2 * j + MovingInputs[i] % 2;
9081 ISD::BITCAST, DL, MVT::v16i8,
9082 DAG.getVectorShuffle(MVT::v8i16, DL,
9083 DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1),
9084 DAG.getUNDEF(MVT::v8i16), PreDupI16Shuffle));
9086 // Unpack the bytes to form the i16s that will be shuffled into place.
9087 V1 = DAG.getNode(TargetLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
9088 MVT::v16i8, V1, V1);
9090 int PostDupI16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9091 for (int i = 0; i < 16; i += 2) {
9093 PostDupI16Shuffle[i / 2] = LaneMap[Mask[i]] - (TargetLo ? 0 : 8);
9094 assert(PostDupI16Shuffle[i / 2] < 8 && "Invalid v8 shuffle mask!");
9097 ISD::BITCAST, DL, MVT::v16i8,
9098 DAG.getVectorShuffle(MVT::v8i16, DL,
9099 DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1),
9100 DAG.getUNDEF(MVT::v8i16), PostDupI16Shuffle));
9102 if (SDValue V = tryToWidenViaDuplication())
9106 // Check whether an interleaving lowering is likely to be more efficient.
9107 // This isn't perfect but it is a strong heuristic that tends to work well on
9108 // the kinds of shuffles that show up in practice.
9110 // FIXME: We need to handle other interleaving widths (i16, i32, ...).
9111 if (shouldLowerAsInterleaving(Mask)) {
9112 // FIXME: Figure out whether we should pack these into the low or high
9115 int EMask[16], OMask[16];
9116 for (int i = 0; i < 8; ++i) {
9117 EMask[i] = Mask[2*i];
9118 OMask[i] = Mask[2*i + 1];
9123 SDValue Evens = DAG.getVectorShuffle(MVT::v16i8, DL, V1, V2, EMask);
9124 SDValue Odds = DAG.getVectorShuffle(MVT::v16i8, DL, V1, V2, OMask);
9126 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, Evens, Odds);
9129 // Check for SSSE3 which lets us lower all v16i8 shuffles much more directly
9130 // with PSHUFB. It is important to do this before we attempt to generate any
9131 // blends but after all of the single-input lowerings. If the single input
9132 // lowerings can find an instruction sequence that is faster than a PSHUFB, we
9133 // want to preserve that and we can DAG combine any longer sequences into
9134 // a PSHUFB in the end. But once we start blending from multiple inputs,
9135 // the complexity of DAG combining bad patterns back into PSHUFB is too high,
9136 // and there are *very* few patterns that would actually be faster than the
9137 // PSHUFB approach because of its ability to zero lanes.
9139 // FIXME: The only exceptions to the above are blends which are exact
9140 // interleavings with direct instructions supporting them. We currently don't
9141 // handle those well here.
9142 if (Subtarget->hasSSSE3()) {
9145 for (int i = 0; i < 16; ++i)
9146 if (Mask[i] == -1) {
9147 V1Mask[i] = V2Mask[i] = DAG.getUNDEF(MVT::i8);
9149 V1Mask[i] = DAG.getConstant(Mask[i] < 16 ? Mask[i] : 0x80, MVT::i8);
9151 DAG.getConstant(Mask[i] < 16 ? 0x80 : Mask[i] - 16, MVT::i8);
9153 V1 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, V1,
9154 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V1Mask));
9155 if (isSingleInputShuffleMask(Mask))
9156 return V1; // Single inputs are easy.
9158 // Otherwise, blend the two.
9159 V2 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, V2,
9160 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V2Mask));
9161 return DAG.getNode(ISD::OR, DL, MVT::v16i8, V1, V2);
9164 // There are special ways we can lower some single-element blends.
9165 if (NumV2Elements == 1)
9166 if (SDValue V = lowerVectorShuffleAsElementInsertion(MVT::v16i8, DL, V1, V2,
9167 Mask, Subtarget, DAG))
9170 // Check whether a compaction lowering can be done. This handles shuffles
9171 // which take every Nth element for some even N. See the helper function for
9174 // We special case these as they can be particularly efficiently handled with
9175 // the PACKUSB instruction on x86 and they show up in common patterns of
9176 // rearranging bytes to truncate wide elements.
9177 if (int NumEvenDrops = canLowerByDroppingEvenElements(Mask)) {
9178 // NumEvenDrops is the power of two stride of the elements. Another way of
9179 // thinking about it is that we need to drop the even elements this many
9180 // times to get the original input.
9181 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9183 // First we need to zero all the dropped bytes.
9184 assert(NumEvenDrops <= 3 &&
9185 "No support for dropping even elements more than 3 times.");
9186 // We use the mask type to pick which bytes are preserved based on how many
9187 // elements are dropped.
9188 MVT MaskVTs[] = { MVT::v8i16, MVT::v4i32, MVT::v2i64 };
9189 SDValue ByteClearMask =
9190 DAG.getNode(ISD::BITCAST, DL, MVT::v16i8,
9191 DAG.getConstant(0xFF, MaskVTs[NumEvenDrops - 1]));
9192 V1 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V1, ByteClearMask);
9194 V2 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V2, ByteClearMask);
9196 // Now pack things back together.
9197 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1);
9198 V2 = IsSingleInput ? V1 : DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V2);
9199 SDValue Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, V1, V2);
9200 for (int i = 1; i < NumEvenDrops; ++i) {
9201 Result = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, Result);
9202 Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, Result, Result);
9208 int V1LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9209 int V1HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9210 int V2LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9211 int V2HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9213 auto buildBlendMasks = [](MutableArrayRef<int> HalfMask,
9214 MutableArrayRef<int> V1HalfBlendMask,
9215 MutableArrayRef<int> V2HalfBlendMask) {
9216 for (int i = 0; i < 8; ++i)
9217 if (HalfMask[i] >= 0 && HalfMask[i] < 16) {
9218 V1HalfBlendMask[i] = HalfMask[i];
9220 } else if (HalfMask[i] >= 16) {
9221 V2HalfBlendMask[i] = HalfMask[i] - 16;
9222 HalfMask[i] = i + 8;
9225 buildBlendMasks(LoMask, V1LoBlendMask, V2LoBlendMask);
9226 buildBlendMasks(HiMask, V1HiBlendMask, V2HiBlendMask);
9228 SDValue Zero = getZeroVector(MVT::v8i16, Subtarget, DAG, DL);
9230 auto buildLoAndHiV8s = [&](SDValue V, MutableArrayRef<int> LoBlendMask,
9231 MutableArrayRef<int> HiBlendMask) {
9233 // Check if any of the odd lanes in the v16i8 are used. If not, we can mask
9234 // them out and avoid using UNPCK{L,H} to extract the elements of V as
9236 if (std::none_of(LoBlendMask.begin(), LoBlendMask.end(),
9237 [](int M) { return M >= 0 && M % 2 == 1; }) &&
9238 std::none_of(HiBlendMask.begin(), HiBlendMask.end(),
9239 [](int M) { return M >= 0 && M % 2 == 1; })) {
9240 // Use a mask to drop the high bytes.
9241 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V);
9242 V1 = DAG.getNode(ISD::AND, DL, MVT::v8i16, V1,
9243 DAG.getConstant(0x00FF, MVT::v8i16));
9245 // This will be a single vector shuffle instead of a blend so nuke V2.
9246 V2 = DAG.getUNDEF(MVT::v8i16);
9248 // Squash the masks to point directly into V1.
9249 for (int &M : LoBlendMask)
9252 for (int &M : HiBlendMask)
9256 // Otherwise just unpack the low half of V into V1 and the high half into
9257 // V2 so that we can blend them as i16s.
9258 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
9259 DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V, Zero));
9260 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
9261 DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V, Zero));
9264 SDValue BlendedLo = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, LoBlendMask);
9265 SDValue BlendedHi = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, HiBlendMask);
9266 return std::make_pair(BlendedLo, BlendedHi);
9268 SDValue V1Lo, V1Hi, V2Lo, V2Hi;
9269 std::tie(V1Lo, V1Hi) = buildLoAndHiV8s(V1, V1LoBlendMask, V1HiBlendMask);
9270 std::tie(V2Lo, V2Hi) = buildLoAndHiV8s(V2, V2LoBlendMask, V2HiBlendMask);
9272 SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, V1Lo, V2Lo, LoMask);
9273 SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, V1Hi, V2Hi, HiMask);
9275 return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, LoV, HiV);
9278 /// \brief Dispatching routine to lower various 128-bit x86 vector shuffles.
9280 /// This routine breaks down the specific type of 128-bit shuffle and
9281 /// dispatches to the lowering routines accordingly.
9282 static SDValue lower128BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9283 MVT VT, const X86Subtarget *Subtarget,
9284 SelectionDAG &DAG) {
9285 switch (VT.SimpleTy) {
9287 return lowerV2I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9289 return lowerV2F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9291 return lowerV4I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9293 return lowerV4F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9295 return lowerV8I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
9297 return lowerV16I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
9300 llvm_unreachable("Unimplemented!");
9304 /// \brief Generic routine to split a 256-bit vector shuffle into 128-bit
9307 /// There is a severely limited set of shuffles available in AVX1 for 256-bit
9308 /// vectors resulting in routinely needing to split the shuffle into two 128-bit
9309 /// shuffles. This can be done generically for any 256-bit vector shuffle and so
9310 /// we encode the logic here for specific shuffle lowering routines to bail to
9311 /// when they exhaust the features avaible to more directly handle the shuffle.
9312 static SDValue splitAndLower256BitVectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9313 SDValue V2, ArrayRef<int> Mask,
9314 SelectionDAG &DAG) {
9315 assert(VT.getSizeInBits() == 256 && "Only for 256-bit vector shuffles!");
9316 assert(V1.getSimpleValueType() == VT && "Bad operand type!");
9317 assert(V2.getSimpleValueType() == VT && "Bad operand type!");
9319 ArrayRef<int> LoMask = Mask.slice(0, Mask.size()/2);
9320 ArrayRef<int> HiMask = Mask.slice(Mask.size()/2);
9322 int NumElements = VT.getVectorNumElements();
9323 int SplitNumElements = NumElements / 2;
9324 MVT ScalarVT = VT.getScalarType();
9325 MVT SplitVT = MVT::getVectorVT(ScalarVT, NumElements / 2);
9327 SDValue LoV1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SplitVT, V1,
9328 DAG.getIntPtrConstant(0));
9329 SDValue HiV1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SplitVT, V1,
9330 DAG.getIntPtrConstant(SplitNumElements));
9331 SDValue LoV2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SplitVT, V2,
9332 DAG.getIntPtrConstant(0));
9333 SDValue HiV2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SplitVT, V2,
9334 DAG.getIntPtrConstant(SplitNumElements));
9336 // Now create two 4-way blends of these half-width vectors.
9337 auto HalfBlend = [&](ArrayRef<int> HalfMask) {
9338 SmallVector<int, 16> V1BlendMask, V2BlendMask, BlendMask;
9339 for (int i = 0; i < SplitNumElements; ++i) {
9340 int M = HalfMask[i];
9341 if (M >= NumElements) {
9342 V2BlendMask.push_back(M - NumElements);
9343 V1BlendMask.push_back(-1);
9344 BlendMask.push_back(SplitNumElements + i);
9345 } else if (M >= 0) {
9346 V2BlendMask.push_back(-1);
9347 V1BlendMask.push_back(M);
9348 BlendMask.push_back(i);
9350 V2BlendMask.push_back(-1);
9351 V1BlendMask.push_back(-1);
9352 BlendMask.push_back(-1);
9355 SDValue V1Blend = DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9356 SDValue V2Blend = DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9357 return DAG.getVectorShuffle(SplitVT, DL, V1Blend, V2Blend, BlendMask);
9359 SDValue Lo = HalfBlend(LoMask);
9360 SDValue Hi = HalfBlend(HiMask);
9361 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
9364 /// \brief Lower a vector shuffle crossing multiple 128-bit lanes as
9365 /// a permutation and blend of those lanes.
9367 /// This essentially blends the out-of-lane inputs to each lane into the lane
9368 /// from a permuted copy of the vector. This lowering strategy results in four
9369 /// instructions in the worst case for a single-input cross lane shuffle which
9370 /// is lower than any other fully general cross-lane shuffle strategy I'm aware
9371 /// of. Special cases for each particular shuffle pattern should be handled
9372 /// prior to trying this lowering.
9373 static SDValue lowerVectorShuffleAsLanePermuteAndBlend(SDLoc DL, MVT VT,
9374 SDValue V1, SDValue V2,
9376 SelectionDAG &DAG) {
9377 // FIXME: This should probably be generalized for 512-bit vectors as well.
9378 assert(VT.getSizeInBits() == 256 && "Only for 256-bit vector shuffles!");
9379 int LaneSize = Mask.size() / 2;
9381 // If there are only inputs from one 128-bit lane, splitting will in fact be
9382 // less expensive. The flags track wether the given lane contains an element
9383 // that crosses to another lane.
9384 bool LaneCrossing[2] = {false, false};
9385 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9386 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
9387 LaneCrossing[(Mask[i] % Size) / LaneSize] = true;
9388 if (!LaneCrossing[0] || !LaneCrossing[1])
9389 return splitAndLower256BitVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9391 if (isSingleInputShuffleMask(Mask)) {
9392 SmallVector<int, 32> FlippedBlendMask;
9393 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9394 FlippedBlendMask.push_back(
9395 Mask[i] < 0 ? -1 : ((Mask[i] / LaneSize == i / LaneSize)
9397 : Mask[i] % LaneSize +
9398 (i / LaneSize) * LaneSize + Size));
9400 // Flip the vector, and blend the results which should now be in-lane. The
9401 // VPERM2X128 mask uses the low 2 bits for the low source and bits 4 and
9402 // 5 for the high source. The value 3 selects the high half of source 2 and
9403 // the value 2 selects the low half of source 2. We only use source 2 to
9404 // allow folding it into a memory operand.
9405 unsigned PERMMask = 3 | 2 << 4;
9406 SDValue Flipped = DAG.getNode(X86ISD::VPERM2X128, DL, VT, DAG.getUNDEF(VT),
9407 V1, DAG.getConstant(PERMMask, MVT::i8));
9408 return DAG.getVectorShuffle(VT, DL, V1, Flipped, FlippedBlendMask);
9411 // This now reduces to two single-input shuffles of V1 and V2 which at worst
9412 // will be handled by the above logic and a blend of the results, much like
9413 // other patterns in AVX.
9414 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
9417 /// \brief Handle lowering of 4-lane 64-bit floating point shuffles.
9419 /// Also ends up handling lowering of 4-lane 64-bit integer shuffles when AVX2
9420 /// isn't available.
9421 static SDValue lowerV4F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9422 const X86Subtarget *Subtarget,
9423 SelectionDAG &DAG) {
9425 assert(V1.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
9426 assert(V2.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
9427 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9428 ArrayRef<int> Mask = SVOp->getMask();
9429 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
9431 if (isSingleInputShuffleMask(Mask)) {
9432 if (!is128BitLaneCrossingShuffleMask(MVT::v4f64, Mask)) {
9433 // Non-half-crossing single input shuffles can be lowerid with an
9434 // interleaved permutation.
9435 unsigned VPERMILPMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1) |
9436 ((Mask[2] == 3) << 2) | ((Mask[3] == 3) << 3);
9437 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f64, V1,
9438 DAG.getConstant(VPERMILPMask, MVT::i8));
9441 // With AVX2 we have direct support for this permutation.
9442 if (Subtarget->hasAVX2())
9443 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4f64, V1,
9444 getV4X86ShuffleImm8ForMask(Mask, DAG));
9446 // Otherwise, fall back.
9447 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v4f64, V1, V2, Mask,
9451 // X86 has dedicated unpack instructions that can handle specific blend
9452 // operations: UNPCKH and UNPCKL.
9453 if (isShuffleEquivalent(Mask, 0, 4, 2, 6))
9454 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V1, V2);
9455 if (isShuffleEquivalent(Mask, 1, 5, 3, 7))
9456 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V1, V2);
9458 // If we have a single input to the zero element, insert that into V1 if we
9459 // can do so cheaply.
9461 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
9462 if (NumV2Elements == 1 && Mask[0] >= 4)
9463 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
9464 MVT::v4f64, DL, V1, V2, Mask, Subtarget, DAG))
9467 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f64, V1, V2, Mask,
9471 // Check if the blend happens to exactly fit that of SHUFPD.
9472 if (Mask[0] < 4 && (Mask[1] == -1 || Mask[1] >= 4) &&
9473 Mask[2] < 4 && (Mask[3] == -1 || Mask[3] >= 4)) {
9474 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 5) << 1) |
9475 ((Mask[2] == 3) << 2) | ((Mask[3] == 7) << 3);
9476 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f64, V1, V2,
9477 DAG.getConstant(SHUFPDMask, MVT::i8));
9479 if ((Mask[0] == -1 || Mask[0] >= 4) && Mask[1] < 4 &&
9480 (Mask[2] == -1 || Mask[2] >= 4) && Mask[3] < 4) {
9481 unsigned SHUFPDMask = (Mask[0] == 5) | ((Mask[1] == 1) << 1) |
9482 ((Mask[2] == 7) << 2) | ((Mask[3] == 3) << 3);
9483 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f64, V2, V1,
9484 DAG.getConstant(SHUFPDMask, MVT::i8));
9487 // Otherwise fall back on generic blend lowering.
9488 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4f64, V1, V2,
9492 /// \brief Handle lowering of 4-lane 64-bit integer shuffles.
9494 /// This routine is only called when we have AVX2 and thus a reasonable
9495 /// instruction set for v4i64 shuffling..
9496 static SDValue lowerV4I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9497 const X86Subtarget *Subtarget,
9498 SelectionDAG &DAG) {
9500 assert(V1.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
9501 assert(V2.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
9502 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9503 ArrayRef<int> Mask = SVOp->getMask();
9504 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
9505 assert(Subtarget->hasAVX2() && "We can only lower v4i64 with AVX2!");
9507 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i64, V1, V2, Mask,
9511 // When the shuffle is mirrored between the 128-bit lanes of the unit, we can
9512 // use lower latency instructions that will operate on both 128-bit lanes.
9513 SmallVector<int, 2> RepeatedMask;
9514 if (is128BitLaneRepeatedShuffleMask(MVT::v4i64, Mask, RepeatedMask)) {
9515 if (isSingleInputShuffleMask(Mask)) {
9516 int PSHUFDMask[] = {-1, -1, -1, -1};
9517 for (int i = 0; i < 2; ++i)
9518 if (RepeatedMask[i] >= 0) {
9519 PSHUFDMask[2 * i] = 2 * RepeatedMask[i];
9520 PSHUFDMask[2 * i + 1] = 2 * RepeatedMask[i] + 1;
9523 ISD::BITCAST, DL, MVT::v4i64,
9524 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32,
9525 DAG.getNode(ISD::BITCAST, DL, MVT::v8i32, V1),
9526 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
9529 // Use dedicated unpack instructions for masks that match their pattern.
9530 if (isShuffleEquivalent(Mask, 0, 4, 2, 6))
9531 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V1, V2);
9532 if (isShuffleEquivalent(Mask, 1, 5, 3, 7))
9533 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V1, V2);
9536 // AVX2 provides a direct instruction for permuting a single input across
9538 if (isSingleInputShuffleMask(Mask))
9539 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4i64, V1,
9540 getV4X86ShuffleImm8ForMask(Mask, DAG));
9542 // Otherwise fall back on generic blend lowering.
9543 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i64, V1, V2,
9547 /// \brief Handle lowering of 8-lane 32-bit floating point shuffles.
9549 /// Also ends up handling lowering of 8-lane 32-bit integer shuffles when AVX2
9550 /// isn't available.
9551 static SDValue lowerV8F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9552 const X86Subtarget *Subtarget,
9553 SelectionDAG &DAG) {
9555 assert(V1.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
9556 assert(V2.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
9557 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9558 ArrayRef<int> Mask = SVOp->getMask();
9559 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9561 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8f32, V1, V2, Mask,
9565 // If the shuffle mask is repeated in each 128-bit lane, we have many more
9566 // options to efficiently lower the shuffle.
9567 SmallVector<int, 2> RepeatedMask;
9568 if (is128BitLaneRepeatedShuffleMask(MVT::v8f32, Mask, RepeatedMask)) {
9569 if (isSingleInputShuffleMask(Mask))
9570 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f32, V1,
9571 getV4X86ShuffleImm8ForMask(RepeatedMask, DAG));
9573 // Use dedicated unpack instructions for masks that match their pattern.
9574 if (isShuffleEquivalent(Mask, 0, 8, 1, 9, 4, 12, 5, 13))
9575 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V1, V2);
9576 if (isShuffleEquivalent(Mask, 2, 10, 3, 11, 6, 14, 7, 15))
9577 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V1, V2);
9579 // Otherwise, fall back to a SHUFPS sequence. Here it is important that we
9580 // have already handled any direct blends.
9581 int SHUFPSMask[] = {Mask[0], Mask[1], Mask[2], Mask[3]};
9582 for (int &M : SHUFPSMask)
9585 return lowerVectorShuffleWithSHUPFS(DL, MVT::v8f32, SHUFPSMask, V1, V2, DAG);
9588 // If we have a single input shuffle with different shuffle patterns in the
9589 // two 128-bit lanes use the variable mask to VPERMILPS.
9590 if (isSingleInputShuffleMask(Mask)) {
9591 SDValue VPermMask[8];
9592 for (int i = 0; i < 8; ++i)
9593 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
9594 : DAG.getConstant(Mask[i], MVT::i32);
9595 if (!is128BitLaneCrossingShuffleMask(MVT::v8f32, Mask))
9597 X86ISD::VPERMILPV, DL, MVT::v8f32, V1,
9598 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask));
9600 if (Subtarget->hasAVX2())
9601 return DAG.getNode(X86ISD::VPERMV, DL, MVT::v8f32,
9602 DAG.getNode(ISD::BITCAST, DL, MVT::v8f32,
9603 DAG.getNode(ISD::BUILD_VECTOR, DL,
9604 MVT::v8i32, VPermMask)),
9607 // Otherwise, fall back.
9608 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v8f32, V1, V2, Mask,
9612 // Otherwise fall back on generic blend lowering.
9613 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8f32, V1, V2,
9617 /// \brief Handle lowering of 8-lane 32-bit integer shuffles.
9619 /// This routine is only called when we have AVX2 and thus a reasonable
9620 /// instruction set for v8i32 shuffling..
9621 static SDValue lowerV8I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9622 const X86Subtarget *Subtarget,
9623 SelectionDAG &DAG) {
9625 assert(V1.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
9626 assert(V2.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
9627 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9628 ArrayRef<int> Mask = SVOp->getMask();
9629 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9630 assert(Subtarget->hasAVX2() && "We can only lower v8i32 with AVX2!");
9632 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i32, V1, V2, Mask,
9636 // If the shuffle mask is repeated in each 128-bit lane we can use more
9637 // efficient instructions that mirror the shuffles across the two 128-bit
9639 SmallVector<int, 4> RepeatedMask;
9640 if (is128BitLaneRepeatedShuffleMask(MVT::v8i32, Mask, RepeatedMask)) {
9641 assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!");
9642 if (isSingleInputShuffleMask(Mask))
9643 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, V1,
9644 getV4X86ShuffleImm8ForMask(RepeatedMask, DAG));
9646 // Use dedicated unpack instructions for masks that match their pattern.
9647 if (isShuffleEquivalent(Mask, 0, 8, 1, 9, 4, 12, 5, 13))
9648 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V1, V2);
9649 if (isShuffleEquivalent(Mask, 2, 10, 3, 11, 6, 14, 7, 15))
9650 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V1, V2);
9653 // If the shuffle patterns aren't repeated but it is a single input, directly
9654 // generate a cross-lane VPERMD instruction.
9655 if (isSingleInputShuffleMask(Mask)) {
9656 SDValue VPermMask[8];
9657 for (int i = 0; i < 8; ++i)
9658 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
9659 : DAG.getConstant(Mask[i], MVT::i32);
9661 X86ISD::VPERMV, DL, MVT::v8i32,
9662 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1);
9665 // Otherwise fall back on generic blend lowering.
9666 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i32, V1, V2,
9670 /// \brief Handle lowering of 16-lane 16-bit integer shuffles.
9672 /// This routine is only called when we have AVX2 and thus a reasonable
9673 /// instruction set for v16i16 shuffling..
9674 static SDValue lowerV16I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9675 const X86Subtarget *Subtarget,
9676 SelectionDAG &DAG) {
9678 assert(V1.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
9679 assert(V2.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
9680 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9681 ArrayRef<int> Mask = SVOp->getMask();
9682 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
9683 assert(Subtarget->hasAVX2() && "We can only lower v16i16 with AVX2!");
9685 // There are no generalized cross-lane shuffle operations available on i16
9687 if (is128BitLaneCrossingShuffleMask(MVT::v16i16, Mask))
9688 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v16i16, V1, V2,
9691 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i16, V1, V2, Mask,
9695 // Use dedicated unpack instructions for masks that match their pattern.
9696 if (isShuffleEquivalent(Mask,
9697 // First 128-bit lane:
9698 0, 16, 1, 17, 2, 18, 3, 19,
9699 // Second 128-bit lane:
9700 8, 24, 9, 25, 10, 26, 11, 27))
9701 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i16, V1, V2);
9702 if (isShuffleEquivalent(Mask,
9703 // First 128-bit lane:
9704 4, 20, 5, 21, 6, 22, 7, 23,
9705 // Second 128-bit lane:
9706 12, 28, 13, 29, 14, 30, 15, 31))
9707 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i16, V1, V2);
9709 if (isSingleInputShuffleMask(Mask)) {
9710 SDValue PSHUFBMask[32];
9711 for (int i = 0; i < 16; ++i) {
9712 if (Mask[i] == -1) {
9713 PSHUFBMask[2 * i] = PSHUFBMask[2 * i + 1] = DAG.getUNDEF(MVT::i8);
9717 int M = i < 8 ? Mask[i] : Mask[i] - 8;
9718 assert(M >= 0 && M < 8 && "Invalid single-input mask!");
9719 PSHUFBMask[2 * i] = DAG.getConstant(2 * M, MVT::i8);
9720 PSHUFBMask[2 * i + 1] = DAG.getConstant(2 * M + 1, MVT::i8);
9723 ISD::BITCAST, DL, MVT::v16i16,
9725 X86ISD::PSHUFB, DL, MVT::v32i8,
9726 DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, V1),
9727 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask)));
9730 // Otherwise fall back on generic blend lowering.
9731 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v16i16, V1, V2,
9735 /// \brief Handle lowering of 32-lane 8-bit integer shuffles.
9737 /// This routine is only called when we have AVX2 and thus a reasonable
9738 /// instruction set for v32i8 shuffling..
9739 static SDValue lowerV32I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9740 const X86Subtarget *Subtarget,
9741 SelectionDAG &DAG) {
9743 assert(V1.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
9744 assert(V2.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
9745 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9746 ArrayRef<int> Mask = SVOp->getMask();
9747 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
9748 assert(Subtarget->hasAVX2() && "We can only lower v32i8 with AVX2!");
9750 // There are no generalized cross-lane shuffle operations available on i8
9752 if (is128BitLaneCrossingShuffleMask(MVT::v32i8, Mask))
9753 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v32i8, V1, V2,
9756 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v32i8, V1, V2, Mask,
9760 // Use dedicated unpack instructions for masks that match their pattern.
9761 // Note that these are repeated 128-bit lane unpacks, not unpacks across all
9763 if (isShuffleEquivalent(
9765 // First 128-bit lane:
9766 0, 32, 1, 33, 2, 34, 3, 35, 4, 36, 5, 37, 6, 38, 7, 39,
9767 // Second 128-bit lane:
9768 16, 48, 17, 49, 18, 50, 19, 51, 20, 52, 21, 53, 22, 54, 23, 55))
9769 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v32i8, V1, V2);
9770 if (isShuffleEquivalent(
9772 // First 128-bit lane:
9773 8, 40, 9, 41, 10, 42, 11, 43, 12, 44, 13, 45, 14, 46, 15, 47,
9774 // Second 128-bit lane:
9775 24, 56, 25, 57, 26, 58, 27, 59, 28, 60, 29, 61, 30, 62, 31, 63))
9776 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v32i8, V1, V2);
9778 if (isSingleInputShuffleMask(Mask)) {
9779 SDValue PSHUFBMask[32];
9780 for (int i = 0; i < 32; ++i)
9783 ? DAG.getUNDEF(MVT::i8)
9784 : DAG.getConstant(Mask[i] < 16 ? Mask[i] : Mask[i] - 16, MVT::i8);
9787 X86ISD::PSHUFB, DL, MVT::v32i8, V1,
9788 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask));
9791 // Otherwise fall back on generic blend lowering.
9792 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v32i8, V1, V2,
9796 /// \brief High-level routine to lower various 256-bit x86 vector shuffles.
9798 /// This routine either breaks down the specific type of a 256-bit x86 vector
9799 /// shuffle or splits it into two 128-bit shuffles and fuses the results back
9800 /// together based on the available instructions.
9801 static SDValue lower256BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9802 MVT VT, const X86Subtarget *Subtarget,
9803 SelectionDAG &DAG) {
9805 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9806 ArrayRef<int> Mask = SVOp->getMask();
9808 // There is a really nice hard cut-over between AVX1 and AVX2 that means we can
9809 // check for those subtargets here and avoid much of the subtarget querying in
9810 // the per-vector-type lowering routines. With AVX1 we have essentially *zero*
9811 // ability to manipulate a 256-bit vector with integer types. Since we'll use
9812 // floating point types there eventually, just immediately cast everything to
9813 // a float and operate entirely in that domain.
9814 if (VT.isInteger() && !Subtarget->hasAVX2()) {
9815 int ElementBits = VT.getScalarSizeInBits();
9816 if (ElementBits < 32)
9817 // No floating point type available, decompose into 128-bit vectors.
9818 return splitAndLower256BitVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9820 MVT FpVT = MVT::getVectorVT(MVT::getFloatingPointVT(ElementBits),
9821 VT.getVectorNumElements());
9822 V1 = DAG.getNode(ISD::BITCAST, DL, FpVT, V1);
9823 V2 = DAG.getNode(ISD::BITCAST, DL, FpVT, V2);
9824 return DAG.getNode(ISD::BITCAST, DL, VT,
9825 DAG.getVectorShuffle(FpVT, DL, V1, V2, Mask));
9828 switch (VT.SimpleTy) {
9830 return lowerV4F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9832 return lowerV4I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9834 return lowerV8F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9836 return lowerV8I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9838 return lowerV16I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
9840 return lowerV32I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
9843 llvm_unreachable("Not a valid 256-bit x86 vector type!");
9847 /// \brief Tiny helper function to test whether a shuffle mask could be
9848 /// simplified by widening the elements being shuffled.
9849 static bool canWidenShuffleElements(ArrayRef<int> Mask) {
9850 for (int i = 0, Size = Mask.size(); i < Size; i += 2)
9851 if ((Mask[i] != -1 && Mask[i] % 2 != 0) ||
9852 (Mask[i + 1] != -1 && (Mask[i + 1] % 2 != 1 ||
9853 (Mask[i] != -1 && Mask[i] + 1 != Mask[i + 1]))))
9859 /// \brief Top-level lowering for x86 vector shuffles.
9861 /// This handles decomposition, canonicalization, and lowering of all x86
9862 /// vector shuffles. Most of the specific lowering strategies are encapsulated
9863 /// above in helper routines. The canonicalization attempts to widen shuffles
9864 /// to involve fewer lanes of wider elements, consolidate symmetric patterns
9865 /// s.t. only one of the two inputs needs to be tested, etc.
9866 static SDValue lowerVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
9867 SelectionDAG &DAG) {
9868 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9869 ArrayRef<int> Mask = SVOp->getMask();
9870 SDValue V1 = Op.getOperand(0);
9871 SDValue V2 = Op.getOperand(1);
9872 MVT VT = Op.getSimpleValueType();
9873 int NumElements = VT.getVectorNumElements();
9876 assert(VT.getSizeInBits() != 64 && "Can't lower MMX shuffles");
9878 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
9879 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
9880 if (V1IsUndef && V2IsUndef)
9881 return DAG.getUNDEF(VT);
9883 // When we create a shuffle node we put the UNDEF node to second operand,
9884 // but in some cases the first operand may be transformed to UNDEF.
9885 // In this case we should just commute the node.
9887 return DAG.getCommutedVectorShuffle(*SVOp);
9889 // Check for non-undef masks pointing at an undef vector and make the masks
9890 // undef as well. This makes it easier to match the shuffle based solely on
9894 if (M >= NumElements) {
9895 SmallVector<int, 8> NewMask(Mask.begin(), Mask.end());
9896 for (int &M : NewMask)
9897 if (M >= NumElements)
9899 return DAG.getVectorShuffle(VT, dl, V1, V2, NewMask);
9902 // For integer vector shuffles, try to collapse them into a shuffle of fewer
9903 // lanes but wider integers. We cap this to not form integers larger than i64
9904 // but it might be interesting to form i128 integers to handle flipping the
9905 // low and high halves of AVX 256-bit vectors.
9906 if (VT.isInteger() && VT.getScalarSizeInBits() < 64 &&
9907 canWidenShuffleElements(Mask)) {
9908 SmallVector<int, 8> NewMask;
9909 for (int i = 0, Size = Mask.size(); i < Size; i += 2)
9910 NewMask.push_back(Mask[i] != -1
9912 : (Mask[i + 1] != -1 ? Mask[i + 1] / 2 : -1));
9914 MVT::getVectorVT(MVT::getIntegerVT(VT.getScalarSizeInBits() * 2),
9915 VT.getVectorNumElements() / 2);
9916 V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, V1);
9917 V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, V2);
9918 return DAG.getNode(ISD::BITCAST, dl, VT,
9919 DAG.getVectorShuffle(NewVT, dl, V1, V2, NewMask));
9922 int NumV1Elements = 0, NumUndefElements = 0, NumV2Elements = 0;
9923 for (int M : SVOp->getMask())
9926 else if (M < NumElements)
9931 // Commute the shuffle as needed such that more elements come from V1 than
9932 // V2. This allows us to match the shuffle pattern strictly on how many
9933 // elements come from V1 without handling the symmetric cases.
9934 if (NumV2Elements > NumV1Elements)
9935 return DAG.getCommutedVectorShuffle(*SVOp);
9937 // When the number of V1 and V2 elements are the same, try to minimize the
9938 // number of uses of V2 in the low half of the vector. When that is tied,
9939 // ensure that the sum of indices for V1 is equal to or lower than the sum
9941 if (NumV1Elements == NumV2Elements) {
9942 int LowV1Elements = 0, LowV2Elements = 0;
9943 for (int M : SVOp->getMask().slice(0, NumElements / 2))
9944 if (M >= NumElements)
9948 if (LowV2Elements > LowV1Elements)
9949 return DAG.getCommutedVectorShuffle(*SVOp);
9951 int SumV1Indices = 0, SumV2Indices = 0;
9952 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
9953 if (SVOp->getMask()[i] >= NumElements)
9955 else if (SVOp->getMask()[i] >= 0)
9957 if (SumV2Indices < SumV1Indices)
9958 return DAG.getCommutedVectorShuffle(*SVOp);
9961 // For each vector width, delegate to a specialized lowering routine.
9962 if (VT.getSizeInBits() == 128)
9963 return lower128BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
9965 if (VT.getSizeInBits() == 256)
9966 return lower256BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
9968 llvm_unreachable("Unimplemented!");
9972 //===----------------------------------------------------------------------===//
9973 // Legacy vector shuffle lowering
9975 // This code is the legacy code handling vector shuffles until the above
9976 // replaces its functionality and performance.
9977 //===----------------------------------------------------------------------===//
9979 static bool isBlendMask(ArrayRef<int> MaskVals, MVT VT, bool hasSSE41,
9980 bool hasInt256, unsigned *MaskOut = nullptr) {
9981 MVT EltVT = VT.getVectorElementType();
9983 // There is no blend with immediate in AVX-512.
9984 if (VT.is512BitVector())
9987 if (!hasSSE41 || EltVT == MVT::i8)
9989 if (!hasInt256 && VT == MVT::v16i16)
9992 unsigned MaskValue = 0;
9993 unsigned NumElems = VT.getVectorNumElements();
9994 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
9995 unsigned NumLanes = (NumElems - 1) / 8 + 1;
9996 unsigned NumElemsInLane = NumElems / NumLanes;
9998 // Blend for v16i16 should be symetric for the both lanes.
9999 for (unsigned i = 0; i < NumElemsInLane; ++i) {
10001 int SndLaneEltIdx = (NumLanes == 2) ? MaskVals[i + NumElemsInLane] : -1;
10002 int EltIdx = MaskVals[i];
10004 if ((EltIdx < 0 || EltIdx == (int)i) &&
10005 (SndLaneEltIdx < 0 || SndLaneEltIdx == (int)(i + NumElemsInLane)))
10008 if (((unsigned)EltIdx == (i + NumElems)) &&
10009 (SndLaneEltIdx < 0 ||
10010 (unsigned)SndLaneEltIdx == i + NumElems + NumElemsInLane))
10011 MaskValue |= (1 << i);
10017 *MaskOut = MaskValue;
10021 // Try to lower a shuffle node into a simple blend instruction.
10022 // This function assumes isBlendMask returns true for this
10023 // SuffleVectorSDNode
10024 static SDValue LowerVECTOR_SHUFFLEtoBlend(ShuffleVectorSDNode *SVOp,
10025 unsigned MaskValue,
10026 const X86Subtarget *Subtarget,
10027 SelectionDAG &DAG) {
10028 MVT VT = SVOp->getSimpleValueType(0);
10029 MVT EltVT = VT.getVectorElementType();
10030 assert(isBlendMask(SVOp->getMask(), VT, Subtarget->hasSSE41(),
10031 Subtarget->hasInt256() && "Trying to lower a "
10032 "VECTOR_SHUFFLE to a Blend but "
10033 "with the wrong mask"));
10034 SDValue V1 = SVOp->getOperand(0);
10035 SDValue V2 = SVOp->getOperand(1);
10037 unsigned NumElems = VT.getVectorNumElements();
10039 // Convert i32 vectors to floating point if it is not AVX2.
10040 // AVX2 introduced VPBLENDD instruction for 128 and 256-bit vectors.
10042 if (EltVT == MVT::i64 || (EltVT == MVT::i32 && !Subtarget->hasInt256())) {
10043 BlendVT = MVT::getVectorVT(MVT::getFloatingPointVT(EltVT.getSizeInBits()),
10045 V1 = DAG.getNode(ISD::BITCAST, dl, VT, V1);
10046 V2 = DAG.getNode(ISD::BITCAST, dl, VT, V2);
10049 SDValue Ret = DAG.getNode(X86ISD::BLENDI, dl, BlendVT, V1, V2,
10050 DAG.getConstant(MaskValue, MVT::i32));
10051 return DAG.getNode(ISD::BITCAST, dl, VT, Ret);
10054 /// In vector type \p VT, return true if the element at index \p InputIdx
10055 /// falls on a different 128-bit lane than \p OutputIdx.
10056 static bool ShuffleCrosses128bitLane(MVT VT, unsigned InputIdx,
10057 unsigned OutputIdx) {
10058 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
10059 return InputIdx * EltSize / 128 != OutputIdx * EltSize / 128;
10062 /// Generate a PSHUFB if possible. Selects elements from \p V1 according to
10063 /// \p MaskVals. MaskVals[OutputIdx] = InputIdx specifies that we want to
10064 /// shuffle the element at InputIdx in V1 to OutputIdx in the result. If \p
10065 /// MaskVals refers to elements outside of \p V1 or is undef (-1), insert a
10067 static SDValue getPSHUFB(ArrayRef<int> MaskVals, SDValue V1, SDLoc &dl,
10068 SelectionDAG &DAG) {
10069 MVT VT = V1.getSimpleValueType();
10070 assert(VT.is128BitVector() || VT.is256BitVector());
10072 MVT EltVT = VT.getVectorElementType();
10073 unsigned EltSizeInBytes = EltVT.getSizeInBits() / 8;
10074 unsigned NumElts = VT.getVectorNumElements();
10076 SmallVector<SDValue, 32> PshufbMask;
10077 for (unsigned OutputIdx = 0; OutputIdx < NumElts; ++OutputIdx) {
10078 int InputIdx = MaskVals[OutputIdx];
10079 unsigned InputByteIdx;
10081 if (InputIdx < 0 || NumElts <= (unsigned)InputIdx)
10082 InputByteIdx = 0x80;
10084 // Cross lane is not allowed.
10085 if (ShuffleCrosses128bitLane(VT, InputIdx, OutputIdx))
10087 InputByteIdx = InputIdx * EltSizeInBytes;
10088 // Index is an byte offset within the 128-bit lane.
10089 InputByteIdx &= 0xf;
10092 for (unsigned j = 0; j < EltSizeInBytes; ++j) {
10093 PshufbMask.push_back(DAG.getConstant(InputByteIdx, MVT::i8));
10094 if (InputByteIdx != 0x80)
10099 MVT ShufVT = MVT::getVectorVT(MVT::i8, PshufbMask.size());
10101 V1 = DAG.getNode(ISD::BITCAST, dl, ShufVT, V1);
10102 return DAG.getNode(X86ISD::PSHUFB, dl, ShufVT, V1,
10103 DAG.getNode(ISD::BUILD_VECTOR, dl, ShufVT, PshufbMask));
10106 // v8i16 shuffles - Prefer shuffles in the following order:
10107 // 1. [all] pshuflw, pshufhw, optional move
10108 // 2. [ssse3] 1 x pshufb
10109 // 3. [ssse3] 2 x pshufb + 1 x por
10110 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
10112 LowerVECTOR_SHUFFLEv8i16(SDValue Op, const X86Subtarget *Subtarget,
10113 SelectionDAG &DAG) {
10114 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10115 SDValue V1 = SVOp->getOperand(0);
10116 SDValue V2 = SVOp->getOperand(1);
10118 SmallVector<int, 8> MaskVals;
10120 // Determine if more than 1 of the words in each of the low and high quadwords
10121 // of the result come from the same quadword of one of the two inputs. Undef
10122 // mask values count as coming from any quadword, for better codegen.
10124 // Lo/HiQuad[i] = j indicates how many words from the ith quad of the input
10125 // feeds this quad. For i, 0 and 1 refer to V1, 2 and 3 refer to V2.
10126 unsigned LoQuad[] = { 0, 0, 0, 0 };
10127 unsigned HiQuad[] = { 0, 0, 0, 0 };
10128 // Indices of quads used.
10129 std::bitset<4> InputQuads;
10130 for (unsigned i = 0; i < 8; ++i) {
10131 unsigned *Quad = i < 4 ? LoQuad : HiQuad;
10132 int EltIdx = SVOp->getMaskElt(i);
10133 MaskVals.push_back(EltIdx);
10141 ++Quad[EltIdx / 4];
10142 InputQuads.set(EltIdx / 4);
10145 int BestLoQuad = -1;
10146 unsigned MaxQuad = 1;
10147 for (unsigned i = 0; i < 4; ++i) {
10148 if (LoQuad[i] > MaxQuad) {
10150 MaxQuad = LoQuad[i];
10154 int BestHiQuad = -1;
10156 for (unsigned i = 0; i < 4; ++i) {
10157 if (HiQuad[i] > MaxQuad) {
10159 MaxQuad = HiQuad[i];
10163 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
10164 // of the two input vectors, shuffle them into one input vector so only a
10165 // single pshufb instruction is necessary. If there are more than 2 input
10166 // quads, disable the next transformation since it does not help SSSE3.
10167 bool V1Used = InputQuads[0] || InputQuads[1];
10168 bool V2Used = InputQuads[2] || InputQuads[3];
10169 if (Subtarget->hasSSSE3()) {
10170 if (InputQuads.count() == 2 && V1Used && V2Used) {
10171 BestLoQuad = InputQuads[0] ? 0 : 1;
10172 BestHiQuad = InputQuads[2] ? 2 : 3;
10174 if (InputQuads.count() > 2) {
10180 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
10181 // the shuffle mask. If a quad is scored as -1, that means that it contains
10182 // words from all 4 input quadwords.
10184 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
10186 BestLoQuad < 0 ? 0 : BestLoQuad,
10187 BestHiQuad < 0 ? 1 : BestHiQuad
10189 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
10190 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V1),
10191 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V2), &MaskV[0]);
10192 NewV = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, NewV);
10194 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
10195 // source words for the shuffle, to aid later transformations.
10196 bool AllWordsInNewV = true;
10197 bool InOrder[2] = { true, true };
10198 for (unsigned i = 0; i != 8; ++i) {
10199 int idx = MaskVals[i];
10201 InOrder[i/4] = false;
10202 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
10204 AllWordsInNewV = false;
10208 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
10209 if (AllWordsInNewV) {
10210 for (int i = 0; i != 8; ++i) {
10211 int idx = MaskVals[i];
10214 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
10215 if ((idx != i) && idx < 4)
10217 if ((idx != i) && idx > 3)
10226 // If we've eliminated the use of V2, and the new mask is a pshuflw or
10227 // pshufhw, that's as cheap as it gets. Return the new shuffle.
10228 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
10229 unsigned Opc = pshufhw ? X86ISD::PSHUFHW : X86ISD::PSHUFLW;
10230 unsigned TargetMask = 0;
10231 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
10232 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
10233 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(NewV.getNode());
10234 TargetMask = pshufhw ? getShufflePSHUFHWImmediate(SVOp):
10235 getShufflePSHUFLWImmediate(SVOp);
10236 V1 = NewV.getOperand(0);
10237 return getTargetShuffleNode(Opc, dl, MVT::v8i16, V1, TargetMask, DAG);
10241 // Promote splats to a larger type which usually leads to more efficient code.
10242 // FIXME: Is this true if pshufb is available?
10243 if (SVOp->isSplat())
10244 return PromoteSplat(SVOp, DAG);
10246 // If we have SSSE3, and all words of the result are from 1 input vector,
10247 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
10248 // is present, fall back to case 4.
10249 if (Subtarget->hasSSSE3()) {
10250 SmallVector<SDValue,16> pshufbMask;
10252 // If we have elements from both input vectors, set the high bit of the
10253 // shuffle mask element to zero out elements that come from V2 in the V1
10254 // mask, and elements that come from V1 in the V2 mask, so that the two
10255 // results can be OR'd together.
10256 bool TwoInputs = V1Used && V2Used;
10257 V1 = getPSHUFB(MaskVals, V1, dl, DAG);
10259 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
10261 // Calculate the shuffle mask for the second input, shuffle it, and
10262 // OR it with the first shuffled input.
10263 CommuteVectorShuffleMask(MaskVals, 8);
10264 V2 = getPSHUFB(MaskVals, V2, dl, DAG);
10265 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
10266 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
10269 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
10270 // and update MaskVals with new element order.
10271 std::bitset<8> InOrder;
10272 if (BestLoQuad >= 0) {
10273 int MaskV[] = { -1, -1, -1, -1, 4, 5, 6, 7 };
10274 for (int i = 0; i != 4; ++i) {
10275 int idx = MaskVals[i];
10278 } else if ((idx / 4) == BestLoQuad) {
10279 MaskV[i] = idx & 3;
10283 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
10286 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSE2()) {
10287 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(NewV.getNode());
10288 NewV = getTargetShuffleNode(X86ISD::PSHUFLW, dl, MVT::v8i16,
10289 NewV.getOperand(0),
10290 getShufflePSHUFLWImmediate(SVOp), DAG);
10294 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
10295 // and update MaskVals with the new element order.
10296 if (BestHiQuad >= 0) {
10297 int MaskV[] = { 0, 1, 2, 3, -1, -1, -1, -1 };
10298 for (unsigned i = 4; i != 8; ++i) {
10299 int idx = MaskVals[i];
10302 } else if ((idx / 4) == BestHiQuad) {
10303 MaskV[i] = (idx & 3) + 4;
10307 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
10310 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSE2()) {
10311 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(NewV.getNode());
10312 NewV = getTargetShuffleNode(X86ISD::PSHUFHW, dl, MVT::v8i16,
10313 NewV.getOperand(0),
10314 getShufflePSHUFHWImmediate(SVOp), DAG);
10318 // In case BestHi & BestLo were both -1, which means each quadword has a word
10319 // from each of the four input quadwords, calculate the InOrder bitvector now
10320 // before falling through to the insert/extract cleanup.
10321 if (BestLoQuad == -1 && BestHiQuad == -1) {
10323 for (int i = 0; i != 8; ++i)
10324 if (MaskVals[i] < 0 || MaskVals[i] == i)
10328 // The other elements are put in the right place using pextrw and pinsrw.
10329 for (unsigned i = 0; i != 8; ++i) {
10332 int EltIdx = MaskVals[i];
10335 SDValue ExtOp = (EltIdx < 8) ?
10336 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
10337 DAG.getIntPtrConstant(EltIdx)) :
10338 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
10339 DAG.getIntPtrConstant(EltIdx - 8));
10340 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
10341 DAG.getIntPtrConstant(i));
10346 /// \brief v16i16 shuffles
10348 /// FIXME: We only support generation of a single pshufb currently. We can
10349 /// generalize the other applicable cases from LowerVECTOR_SHUFFLEv8i16 as
10350 /// well (e.g 2 x pshufb + 1 x por).
10352 LowerVECTOR_SHUFFLEv16i16(SDValue Op, SelectionDAG &DAG) {
10353 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10354 SDValue V1 = SVOp->getOperand(0);
10355 SDValue V2 = SVOp->getOperand(1);
10358 if (V2.getOpcode() != ISD::UNDEF)
10361 SmallVector<int, 16> MaskVals(SVOp->getMask().begin(), SVOp->getMask().end());
10362 return getPSHUFB(MaskVals, V1, dl, DAG);
10365 // v16i8 shuffles - Prefer shuffles in the following order:
10366 // 1. [ssse3] 1 x pshufb
10367 // 2. [ssse3] 2 x pshufb + 1 x por
10368 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
10369 static SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
10370 const X86Subtarget* Subtarget,
10371 SelectionDAG &DAG) {
10372 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
10373 SDValue V1 = SVOp->getOperand(0);
10374 SDValue V2 = SVOp->getOperand(1);
10376 ArrayRef<int> MaskVals = SVOp->getMask();
10378 // Promote splats to a larger type which usually leads to more efficient code.
10379 // FIXME: Is this true if pshufb is available?
10380 if (SVOp->isSplat())
10381 return PromoteSplat(SVOp, DAG);
10383 // If we have SSSE3, case 1 is generated when all result bytes come from
10384 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
10385 // present, fall back to case 3.
10387 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
10388 if (Subtarget->hasSSSE3()) {
10389 SmallVector<SDValue,16> pshufbMask;
10391 // If all result elements are from one input vector, then only translate
10392 // undef mask values to 0x80 (zero out result) in the pshufb mask.
10394 // Otherwise, we have elements from both input vectors, and must zero out
10395 // elements that come from V2 in the first mask, and V1 in the second mask
10396 // so that we can OR them together.
10397 for (unsigned i = 0; i != 16; ++i) {
10398 int EltIdx = MaskVals[i];
10399 if (EltIdx < 0 || EltIdx >= 16)
10401 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
10403 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
10404 DAG.getNode(ISD::BUILD_VECTOR, dl,
10405 MVT::v16i8, pshufbMask));
10407 // As PSHUFB will zero elements with negative indices, it's safe to ignore
10408 // the 2nd operand if it's undefined or zero.
10409 if (V2.getOpcode() == ISD::UNDEF ||
10410 ISD::isBuildVectorAllZeros(V2.getNode()))
10413 // Calculate the shuffle mask for the second input, shuffle it, and
10414 // OR it with the first shuffled input.
10415 pshufbMask.clear();
10416 for (unsigned i = 0; i != 16; ++i) {
10417 int EltIdx = MaskVals[i];
10418 EltIdx = (EltIdx < 16) ? 0x80 : EltIdx - 16;
10419 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
10421 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
10422 DAG.getNode(ISD::BUILD_VECTOR, dl,
10423 MVT::v16i8, pshufbMask));
10424 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
10427 // No SSSE3 - Calculate in place words and then fix all out of place words
10428 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
10429 // the 16 different words that comprise the two doublequadword input vectors.
10430 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
10431 V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V2);
10433 for (int i = 0; i != 8; ++i) {
10434 int Elt0 = MaskVals[i*2];
10435 int Elt1 = MaskVals[i*2+1];
10437 // This word of the result is all undef, skip it.
10438 if (Elt0 < 0 && Elt1 < 0)
10441 // This word of the result is already in the correct place, skip it.
10442 if ((Elt0 == i*2) && (Elt1 == i*2+1))
10445 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
10446 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
10449 // If Elt0 and Elt1 are defined, are consecutive, and can be load
10450 // using a single extract together, load it and store it.
10451 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
10452 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
10453 DAG.getIntPtrConstant(Elt1 / 2));
10454 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
10455 DAG.getIntPtrConstant(i));
10459 // If Elt1 is defined, extract it from the appropriate source. If the
10460 // source byte is not also odd, shift the extracted word left 8 bits
10461 // otherwise clear the bottom 8 bits if we need to do an or.
10463 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
10464 DAG.getIntPtrConstant(Elt1 / 2));
10465 if ((Elt1 & 1) == 0)
10466 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
10468 TLI.getShiftAmountTy(InsElt.getValueType())));
10469 else if (Elt0 >= 0)
10470 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
10471 DAG.getConstant(0xFF00, MVT::i16));
10473 // If Elt0 is defined, extract it from the appropriate source. If the
10474 // source byte is not also even, shift the extracted word right 8 bits. If
10475 // Elt1 was also defined, OR the extracted values together before
10476 // inserting them in the result.
10478 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
10479 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
10480 if ((Elt0 & 1) != 0)
10481 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
10483 TLI.getShiftAmountTy(InsElt0.getValueType())));
10484 else if (Elt1 >= 0)
10485 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
10486 DAG.getConstant(0x00FF, MVT::i16));
10487 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
10490 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
10491 DAG.getIntPtrConstant(i));
10493 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, NewV);
10496 // v32i8 shuffles - Translate to VPSHUFB if possible.
10498 SDValue LowerVECTOR_SHUFFLEv32i8(ShuffleVectorSDNode *SVOp,
10499 const X86Subtarget *Subtarget,
10500 SelectionDAG &DAG) {
10501 MVT VT = SVOp->getSimpleValueType(0);
10502 SDValue V1 = SVOp->getOperand(0);
10503 SDValue V2 = SVOp->getOperand(1);
10505 SmallVector<int, 32> MaskVals(SVOp->getMask().begin(), SVOp->getMask().end());
10507 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
10508 bool V1IsAllZero = ISD::isBuildVectorAllZeros(V1.getNode());
10509 bool V2IsAllZero = ISD::isBuildVectorAllZeros(V2.getNode());
10511 // VPSHUFB may be generated if
10512 // (1) one of input vector is undefined or zeroinitializer.
10513 // The mask value 0x80 puts 0 in the corresponding slot of the vector.
10514 // And (2) the mask indexes don't cross the 128-bit lane.
10515 if (VT != MVT::v32i8 || !Subtarget->hasInt256() ||
10516 (!V2IsUndef && !V2IsAllZero && !V1IsAllZero))
10519 if (V1IsAllZero && !V2IsAllZero) {
10520 CommuteVectorShuffleMask(MaskVals, 32);
10523 return getPSHUFB(MaskVals, V1, dl, DAG);
10526 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
10527 /// ones, or rewriting v4i32 / v4f32 as 2 wide ones if possible. This can be
10528 /// done when every pair / quad of shuffle mask elements point to elements in
10529 /// the right sequence. e.g.
10530 /// vector_shuffle X, Y, <2, 3, | 10, 11, | 0, 1, | 14, 15>
10532 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
10533 SelectionDAG &DAG) {
10534 MVT VT = SVOp->getSimpleValueType(0);
10536 unsigned NumElems = VT.getVectorNumElements();
10539 switch (VT.SimpleTy) {
10540 default: llvm_unreachable("Unexpected!");
10543 return SDValue(SVOp, 0);
10544 case MVT::v4f32: NewVT = MVT::v2f64; Scale = 2; break;
10545 case MVT::v4i32: NewVT = MVT::v2i64; Scale = 2; break;
10546 case MVT::v8i16: NewVT = MVT::v4i32; Scale = 2; break;
10547 case MVT::v16i8: NewVT = MVT::v4i32; Scale = 4; break;
10548 case MVT::v16i16: NewVT = MVT::v8i32; Scale = 2; break;
10549 case MVT::v32i8: NewVT = MVT::v8i32; Scale = 4; break;
10552 SmallVector<int, 8> MaskVec;
10553 for (unsigned i = 0; i != NumElems; i += Scale) {
10555 for (unsigned j = 0; j != Scale; ++j) {
10556 int EltIdx = SVOp->getMaskElt(i+j);
10560 StartIdx = (EltIdx / Scale);
10561 if (EltIdx != (int)(StartIdx*Scale + j))
10564 MaskVec.push_back(StartIdx);
10567 SDValue V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, SVOp->getOperand(0));
10568 SDValue V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, SVOp->getOperand(1));
10569 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
10572 /// getVZextMovL - Return a zero-extending vector move low node.
10574 static SDValue getVZextMovL(MVT VT, MVT OpVT,
10575 SDValue SrcOp, SelectionDAG &DAG,
10576 const X86Subtarget *Subtarget, SDLoc dl) {
10577 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
10578 LoadSDNode *LD = nullptr;
10579 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
10580 LD = dyn_cast<LoadSDNode>(SrcOp);
10582 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
10584 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
10585 if ((ExtVT != MVT::i64 || Subtarget->is64Bit()) &&
10586 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
10587 SrcOp.getOperand(0).getOpcode() == ISD::BITCAST &&
10588 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
10590 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
10591 return DAG.getNode(ISD::BITCAST, dl, VT,
10592 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
10593 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
10595 SrcOp.getOperand(0)
10601 return DAG.getNode(ISD::BITCAST, dl, VT,
10602 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
10603 DAG.getNode(ISD::BITCAST, dl,
10607 /// LowerVECTOR_SHUFFLE_256 - Handle all 256-bit wide vectors shuffles
10608 /// which could not be matched by any known target speficic shuffle
10610 LowerVECTOR_SHUFFLE_256(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
10612 SDValue NewOp = Compact8x32ShuffleNode(SVOp, DAG);
10613 if (NewOp.getNode())
10616 MVT VT = SVOp->getSimpleValueType(0);
10618 unsigned NumElems = VT.getVectorNumElements();
10619 unsigned NumLaneElems = NumElems / 2;
10622 MVT EltVT = VT.getVectorElementType();
10623 MVT NVT = MVT::getVectorVT(EltVT, NumLaneElems);
10626 SmallVector<int, 16> Mask;
10627 for (unsigned l = 0; l < 2; ++l) {
10628 // Build a shuffle mask for the output, discovering on the fly which
10629 // input vectors to use as shuffle operands (recorded in InputUsed).
10630 // If building a suitable shuffle vector proves too hard, then bail
10631 // out with UseBuildVector set.
10632 bool UseBuildVector = false;
10633 int InputUsed[2] = { -1, -1 }; // Not yet discovered.
10634 unsigned LaneStart = l * NumLaneElems;
10635 for (unsigned i = 0; i != NumLaneElems; ++i) {
10636 // The mask element. This indexes into the input.
10637 int Idx = SVOp->getMaskElt(i+LaneStart);
10639 // the mask element does not index into any input vector.
10640 Mask.push_back(-1);
10644 // The input vector this mask element indexes into.
10645 int Input = Idx / NumLaneElems;
10647 // Turn the index into an offset from the start of the input vector.
10648 Idx -= Input * NumLaneElems;
10650 // Find or create a shuffle vector operand to hold this input.
10652 for (OpNo = 0; OpNo < array_lengthof(InputUsed); ++OpNo) {
10653 if (InputUsed[OpNo] == Input)
10654 // This input vector is already an operand.
10656 if (InputUsed[OpNo] < 0) {
10657 // Create a new operand for this input vector.
10658 InputUsed[OpNo] = Input;
10663 if (OpNo >= array_lengthof(InputUsed)) {
10664 // More than two input vectors used! Give up on trying to create a
10665 // shuffle vector. Insert all elements into a BUILD_VECTOR instead.
10666 UseBuildVector = true;
10670 // Add the mask index for the new shuffle vector.
10671 Mask.push_back(Idx + OpNo * NumLaneElems);
10674 if (UseBuildVector) {
10675 SmallVector<SDValue, 16> SVOps;
10676 for (unsigned i = 0; i != NumLaneElems; ++i) {
10677 // The mask element. This indexes into the input.
10678 int Idx = SVOp->getMaskElt(i+LaneStart);
10680 SVOps.push_back(DAG.getUNDEF(EltVT));
10684 // The input vector this mask element indexes into.
10685 int Input = Idx / NumElems;
10687 // Turn the index into an offset from the start of the input vector.
10688 Idx -= Input * NumElems;
10690 // Extract the vector element by hand.
10691 SVOps.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
10692 SVOp->getOperand(Input),
10693 DAG.getIntPtrConstant(Idx)));
10696 // Construct the output using a BUILD_VECTOR.
10697 Output[l] = DAG.getNode(ISD::BUILD_VECTOR, dl, NVT, SVOps);
10698 } else if (InputUsed[0] < 0) {
10699 // No input vectors were used! The result is undefined.
10700 Output[l] = DAG.getUNDEF(NVT);
10702 SDValue Op0 = Extract128BitVector(SVOp->getOperand(InputUsed[0] / 2),
10703 (InputUsed[0] % 2) * NumLaneElems,
10705 // If only one input was used, use an undefined vector for the other.
10706 SDValue Op1 = (InputUsed[1] < 0) ? DAG.getUNDEF(NVT) :
10707 Extract128BitVector(SVOp->getOperand(InputUsed[1] / 2),
10708 (InputUsed[1] % 2) * NumLaneElems, DAG, dl);
10709 // At least one input vector was used. Create a new shuffle vector.
10710 Output[l] = DAG.getVectorShuffle(NVT, dl, Op0, Op1, &Mask[0]);
10716 // Concatenate the result back
10717 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Output[0], Output[1]);
10720 /// LowerVECTOR_SHUFFLE_128v4 - Handle all 128-bit wide vectors with
10721 /// 4 elements, and match them with several different shuffle types.
10723 LowerVECTOR_SHUFFLE_128v4(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
10724 SDValue V1 = SVOp->getOperand(0);
10725 SDValue V2 = SVOp->getOperand(1);
10727 MVT VT = SVOp->getSimpleValueType(0);
10729 assert(VT.is128BitVector() && "Unsupported vector size");
10731 std::pair<int, int> Locs[4];
10732 int Mask1[] = { -1, -1, -1, -1 };
10733 SmallVector<int, 8> PermMask(SVOp->getMask().begin(), SVOp->getMask().end());
10735 unsigned NumHi = 0;
10736 unsigned NumLo = 0;
10737 for (unsigned i = 0; i != 4; ++i) {
10738 int Idx = PermMask[i];
10740 Locs[i] = std::make_pair(-1, -1);
10742 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
10744 Locs[i] = std::make_pair(0, NumLo);
10745 Mask1[NumLo] = Idx;
10748 Locs[i] = std::make_pair(1, NumHi);
10750 Mask1[2+NumHi] = Idx;
10756 if (NumLo <= 2 && NumHi <= 2) {
10757 // If no more than two elements come from either vector. This can be
10758 // implemented with two shuffles. First shuffle gather the elements.
10759 // The second shuffle, which takes the first shuffle as both of its
10760 // vector operands, put the elements into the right order.
10761 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
10763 int Mask2[] = { -1, -1, -1, -1 };
10765 for (unsigned i = 0; i != 4; ++i)
10766 if (Locs[i].first != -1) {
10767 unsigned Idx = (i < 2) ? 0 : 4;
10768 Idx += Locs[i].first * 2 + Locs[i].second;
10772 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
10775 if (NumLo == 3 || NumHi == 3) {
10776 // Otherwise, we must have three elements from one vector, call it X, and
10777 // one element from the other, call it Y. First, use a shufps to build an
10778 // intermediate vector with the one element from Y and the element from X
10779 // that will be in the same half in the final destination (the indexes don't
10780 // matter). Then, use a shufps to build the final vector, taking the half
10781 // containing the element from Y from the intermediate, and the other half
10784 // Normalize it so the 3 elements come from V1.
10785 CommuteVectorShuffleMask(PermMask, 4);
10789 // Find the element from V2.
10791 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
10792 int Val = PermMask[HiIndex];
10799 Mask1[0] = PermMask[HiIndex];
10801 Mask1[2] = PermMask[HiIndex^1];
10803 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
10805 if (HiIndex >= 2) {
10806 Mask1[0] = PermMask[0];
10807 Mask1[1] = PermMask[1];
10808 Mask1[2] = HiIndex & 1 ? 6 : 4;
10809 Mask1[3] = HiIndex & 1 ? 4 : 6;
10810 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
10813 Mask1[0] = HiIndex & 1 ? 2 : 0;
10814 Mask1[1] = HiIndex & 1 ? 0 : 2;
10815 Mask1[2] = PermMask[2];
10816 Mask1[3] = PermMask[3];
10821 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
10824 // Break it into (shuffle shuffle_hi, shuffle_lo).
10825 int LoMask[] = { -1, -1, -1, -1 };
10826 int HiMask[] = { -1, -1, -1, -1 };
10828 int *MaskPtr = LoMask;
10829 unsigned MaskIdx = 0;
10830 unsigned LoIdx = 0;
10831 unsigned HiIdx = 2;
10832 for (unsigned i = 0; i != 4; ++i) {
10839 int Idx = PermMask[i];
10841 Locs[i] = std::make_pair(-1, -1);
10842 } else if (Idx < 4) {
10843 Locs[i] = std::make_pair(MaskIdx, LoIdx);
10844 MaskPtr[LoIdx] = Idx;
10847 Locs[i] = std::make_pair(MaskIdx, HiIdx);
10848 MaskPtr[HiIdx] = Idx;
10853 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
10854 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
10855 int MaskOps[] = { -1, -1, -1, -1 };
10856 for (unsigned i = 0; i != 4; ++i)
10857 if (Locs[i].first != -1)
10858 MaskOps[i] = Locs[i].first * 4 + Locs[i].second;
10859 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
10862 static bool MayFoldVectorLoad(SDValue V) {
10863 while (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
10864 V = V.getOperand(0);
10866 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
10867 V = V.getOperand(0);
10868 if (V.hasOneUse() && V.getOpcode() == ISD::BUILD_VECTOR &&
10869 V.getNumOperands() == 2 && V.getOperand(1).getOpcode() == ISD::UNDEF)
10870 // BUILD_VECTOR (load), undef
10871 V = V.getOperand(0);
10873 return MayFoldLoad(V);
10877 SDValue getMOVDDup(SDValue &Op, SDLoc &dl, SDValue V1, SelectionDAG &DAG) {
10878 MVT VT = Op.getSimpleValueType();
10880 // Canonizalize to v2f64.
10881 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, V1);
10882 return DAG.getNode(ISD::BITCAST, dl, VT,
10883 getTargetShuffleNode(X86ISD::MOVDDUP, dl, MVT::v2f64,
10888 SDValue getMOVLowToHigh(SDValue &Op, SDLoc &dl, SelectionDAG &DAG,
10890 SDValue V1 = Op.getOperand(0);
10891 SDValue V2 = Op.getOperand(1);
10892 MVT VT = Op.getSimpleValueType();
10894 assert(VT != MVT::v2i64 && "unsupported shuffle type");
10896 if (HasSSE2 && VT == MVT::v2f64)
10897 return getTargetShuffleNode(X86ISD::MOVLHPD, dl, VT, V1, V2, DAG);
10899 // v4f32 or v4i32: canonizalized to v4f32 (which is legal for SSE1)
10900 return DAG.getNode(ISD::BITCAST, dl, VT,
10901 getTargetShuffleNode(X86ISD::MOVLHPS, dl, MVT::v4f32,
10902 DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, V1),
10903 DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, V2), DAG));
10907 SDValue getMOVHighToLow(SDValue &Op, SDLoc &dl, SelectionDAG &DAG) {
10908 SDValue V1 = Op.getOperand(0);
10909 SDValue V2 = Op.getOperand(1);
10910 MVT VT = Op.getSimpleValueType();
10912 assert((VT == MVT::v4i32 || VT == MVT::v4f32) &&
10913 "unsupported shuffle type");
10915 if (V2.getOpcode() == ISD::UNDEF)
10919 return getTargetShuffleNode(X86ISD::MOVHLPS, dl, VT, V1, V2, DAG);
10923 SDValue getMOVLP(SDValue &Op, SDLoc &dl, SelectionDAG &DAG, bool HasSSE2) {
10924 SDValue V1 = Op.getOperand(0);
10925 SDValue V2 = Op.getOperand(1);
10926 MVT VT = Op.getSimpleValueType();
10927 unsigned NumElems = VT.getVectorNumElements();
10929 // Use MOVLPS and MOVLPD in case V1 or V2 are loads. During isel, the second
10930 // operand of these instructions is only memory, so check if there's a
10931 // potencial load folding here, otherwise use SHUFPS or MOVSD to match the
10933 bool CanFoldLoad = false;
10935 // Trivial case, when V2 comes from a load.
10936 if (MayFoldVectorLoad(V2))
10937 CanFoldLoad = true;
10939 // When V1 is a load, it can be folded later into a store in isel, example:
10940 // (store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)), addr:$src1)
10942 // (MOVLPSmr addr:$src1, VR128:$src2)
10943 // So, recognize this potential and also use MOVLPS or MOVLPD
10944 else if (MayFoldVectorLoad(V1) && MayFoldIntoStore(Op))
10945 CanFoldLoad = true;
10947 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10949 if (HasSSE2 && NumElems == 2)
10950 return getTargetShuffleNode(X86ISD::MOVLPD, dl, VT, V1, V2, DAG);
10953 // If we don't care about the second element, proceed to use movss.
10954 if (SVOp->getMaskElt(1) != -1)
10955 return getTargetShuffleNode(X86ISD::MOVLPS, dl, VT, V1, V2, DAG);
10958 // movl and movlp will both match v2i64, but v2i64 is never matched by
10959 // movl earlier because we make it strict to avoid messing with the movlp load
10960 // folding logic (see the code above getMOVLP call). Match it here then,
10961 // this is horrible, but will stay like this until we move all shuffle
10962 // matching to x86 specific nodes. Note that for the 1st condition all
10963 // types are matched with movsd.
10965 // FIXME: isMOVLMask should be checked and matched before getMOVLP,
10966 // as to remove this logic from here, as much as possible
10967 if (NumElems == 2 || !isMOVLMask(SVOp->getMask(), VT))
10968 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
10969 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
10972 assert(VT != MVT::v4i32 && "unsupported shuffle type");
10974 // Invert the operand order and use SHUFPS to match it.
10975 return getTargetShuffleNode(X86ISD::SHUFP, dl, VT, V2, V1,
10976 getShuffleSHUFImmediate(SVOp), DAG);
10979 static SDValue NarrowVectorLoadToElement(LoadSDNode *Load, unsigned Index,
10980 SelectionDAG &DAG) {
10982 MVT VT = Load->getSimpleValueType(0);
10983 MVT EVT = VT.getVectorElementType();
10984 SDValue Addr = Load->getOperand(1);
10985 SDValue NewAddr = DAG.getNode(
10986 ISD::ADD, dl, Addr.getSimpleValueType(), Addr,
10987 DAG.getConstant(Index * EVT.getStoreSize(), Addr.getSimpleValueType()));
10990 DAG.getLoad(EVT, dl, Load->getChain(), NewAddr,
10991 DAG.getMachineFunction().getMachineMemOperand(
10992 Load->getMemOperand(), 0, EVT.getStoreSize()));
10996 // It is only safe to call this function if isINSERTPSMask is true for
10997 // this shufflevector mask.
10998 static SDValue getINSERTPS(ShuffleVectorSDNode *SVOp, SDLoc &dl,
10999 SelectionDAG &DAG) {
11000 // Generate an insertps instruction when inserting an f32 from memory onto a
11001 // v4f32 or when copying a member from one v4f32 to another.
11002 // We also use it for transferring i32 from one register to another,
11003 // since it simply copies the same bits.
11004 // If we're transferring an i32 from memory to a specific element in a
11005 // register, we output a generic DAG that will match the PINSRD
11007 MVT VT = SVOp->getSimpleValueType(0);
11008 MVT EVT = VT.getVectorElementType();
11009 SDValue V1 = SVOp->getOperand(0);
11010 SDValue V2 = SVOp->getOperand(1);
11011 auto Mask = SVOp->getMask();
11012 assert((VT == MVT::v4f32 || VT == MVT::v4i32) &&
11013 "unsupported vector type for insertps/pinsrd");
11015 auto FromV1Predicate = [](const int &i) { return i < 4 && i > -1; };
11016 auto FromV2Predicate = [](const int &i) { return i >= 4; };
11017 int FromV1 = std::count_if(Mask.begin(), Mask.end(), FromV1Predicate);
11021 unsigned DestIndex;
11025 DestIndex = std::find_if(Mask.begin(), Mask.end(), FromV1Predicate) -
11028 // If we have 1 element from each vector, we have to check if we're
11029 // changing V1's element's place. If so, we're done. Otherwise, we
11030 // should assume we're changing V2's element's place and behave
11032 int FromV2 = std::count_if(Mask.begin(), Mask.end(), FromV2Predicate);
11033 assert(DestIndex <= INT32_MAX && "truncated destination index");
11034 if (FromV1 == FromV2 &&
11035 static_cast<int>(DestIndex) == Mask[DestIndex] % 4) {
11039 std::find_if(Mask.begin(), Mask.end(), FromV2Predicate) - Mask.begin();
11042 assert(std::count_if(Mask.begin(), Mask.end(), FromV2Predicate) == 1 &&
11043 "More than one element from V1 and from V2, or no elements from one "
11044 "of the vectors. This case should not have returned true from "
11049 std::find_if(Mask.begin(), Mask.end(), FromV2Predicate) - Mask.begin();
11052 // Get an index into the source vector in the range [0,4) (the mask is
11053 // in the range [0,8) because it can address V1 and V2)
11054 unsigned SrcIndex = Mask[DestIndex] % 4;
11055 if (MayFoldLoad(From)) {
11056 // Trivial case, when From comes from a load and is only used by the
11057 // shuffle. Make it use insertps from the vector that we need from that
11060 NarrowVectorLoadToElement(cast<LoadSDNode>(From), SrcIndex, DAG);
11061 if (!NewLoad.getNode())
11064 if (EVT == MVT::f32) {
11065 // Create this as a scalar to vector to match the instruction pattern.
11066 SDValue LoadScalarToVector =
11067 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, NewLoad);
11068 SDValue InsertpsMask = DAG.getIntPtrConstant(DestIndex << 4);
11069 return DAG.getNode(X86ISD::INSERTPS, dl, VT, To, LoadScalarToVector,
11071 } else { // EVT == MVT::i32
11072 // If we're getting an i32 from memory, use an INSERT_VECTOR_ELT
11073 // instruction, to match the PINSRD instruction, which loads an i32 to a
11074 // certain vector element.
11075 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, To, NewLoad,
11076 DAG.getConstant(DestIndex, MVT::i32));
11080 // Vector-element-to-vector
11081 SDValue InsertpsMask = DAG.getIntPtrConstant(DestIndex << 4 | SrcIndex << 6);
11082 return DAG.getNode(X86ISD::INSERTPS, dl, VT, To, From, InsertpsMask);
11085 // Reduce a vector shuffle to zext.
11086 static SDValue LowerVectorIntExtend(SDValue Op, const X86Subtarget *Subtarget,
11087 SelectionDAG &DAG) {
11088 // PMOVZX is only available from SSE41.
11089 if (!Subtarget->hasSSE41())
11092 MVT VT = Op.getSimpleValueType();
11094 // Only AVX2 support 256-bit vector integer extending.
11095 if (!Subtarget->hasInt256() && VT.is256BitVector())
11098 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11100 SDValue V1 = Op.getOperand(0);
11101 SDValue V2 = Op.getOperand(1);
11102 unsigned NumElems = VT.getVectorNumElements();
11104 // Extending is an unary operation and the element type of the source vector
11105 // won't be equal to or larger than i64.
11106 if (V2.getOpcode() != ISD::UNDEF || !VT.isInteger() ||
11107 VT.getVectorElementType() == MVT::i64)
11110 // Find the expansion ratio, e.g. expanding from i8 to i32 has a ratio of 4.
11111 unsigned Shift = 1; // Start from 2, i.e. 1 << 1.
11112 while ((1U << Shift) < NumElems) {
11113 if (SVOp->getMaskElt(1U << Shift) == 1)
11116 // The maximal ratio is 8, i.e. from i8 to i64.
11121 // Check the shuffle mask.
11122 unsigned Mask = (1U << Shift) - 1;
11123 for (unsigned i = 0; i != NumElems; ++i) {
11124 int EltIdx = SVOp->getMaskElt(i);
11125 if ((i & Mask) != 0 && EltIdx != -1)
11127 if ((i & Mask) == 0 && (unsigned)EltIdx != (i >> Shift))
11131 unsigned NBits = VT.getVectorElementType().getSizeInBits() << Shift;
11132 MVT NeVT = MVT::getIntegerVT(NBits);
11133 MVT NVT = MVT::getVectorVT(NeVT, NumElems >> Shift);
11135 if (!DAG.getTargetLoweringInfo().isTypeLegal(NVT))
11138 // Simplify the operand as it's prepared to be fed into shuffle.
11139 unsigned SignificantBits = NVT.getSizeInBits() >> Shift;
11140 if (V1.getOpcode() == ISD::BITCAST &&
11141 V1.getOperand(0).getOpcode() == ISD::SCALAR_TO_VECTOR &&
11142 V1.getOperand(0).getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
11143 V1.getOperand(0).getOperand(0)
11144 .getSimpleValueType().getSizeInBits() == SignificantBits) {
11145 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast x)
11146 SDValue V = V1.getOperand(0).getOperand(0).getOperand(0);
11147 ConstantSDNode *CIdx =
11148 dyn_cast<ConstantSDNode>(V1.getOperand(0).getOperand(0).getOperand(1));
11149 // If it's foldable, i.e. normal load with single use, we will let code
11150 // selection to fold it. Otherwise, we will short the conversion sequence.
11151 if (CIdx && CIdx->getZExtValue() == 0 &&
11152 (!ISD::isNormalLoad(V.getNode()) || !V.hasOneUse())) {
11153 MVT FullVT = V.getSimpleValueType();
11154 MVT V1VT = V1.getSimpleValueType();
11155 if (FullVT.getSizeInBits() > V1VT.getSizeInBits()) {
11156 // The "ext_vec_elt" node is wider than the result node.
11157 // In this case we should extract subvector from V.
11158 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast (extract_subvector x)).
11159 unsigned Ratio = FullVT.getSizeInBits() / V1VT.getSizeInBits();
11160 MVT SubVecVT = MVT::getVectorVT(FullVT.getVectorElementType(),
11161 FullVT.getVectorNumElements()/Ratio);
11162 V = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVecVT, V,
11163 DAG.getIntPtrConstant(0));
11165 V1 = DAG.getNode(ISD::BITCAST, DL, V1VT, V);
11169 return DAG.getNode(ISD::BITCAST, DL, VT,
11170 DAG.getNode(X86ISD::VZEXT, DL, NVT, V1));
11173 static SDValue NormalizeVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
11174 SelectionDAG &DAG) {
11175 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11176 MVT VT = Op.getSimpleValueType();
11178 SDValue V1 = Op.getOperand(0);
11179 SDValue V2 = Op.getOperand(1);
11181 if (isZeroShuffle(SVOp))
11182 return getZeroVector(VT, Subtarget, DAG, dl);
11184 // Handle splat operations
11185 if (SVOp->isSplat()) {
11186 // Use vbroadcast whenever the splat comes from a foldable load
11187 SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG);
11188 if (Broadcast.getNode())
11192 // Check integer expanding shuffles.
11193 SDValue NewOp = LowerVectorIntExtend(Op, Subtarget, DAG);
11194 if (NewOp.getNode())
11197 // If the shuffle can be profitably rewritten as a narrower shuffle, then
11199 if (VT == MVT::v8i16 || VT == MVT::v16i8 || VT == MVT::v16i16 ||
11200 VT == MVT::v32i8) {
11201 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG);
11202 if (NewOp.getNode())
11203 return DAG.getNode(ISD::BITCAST, dl, VT, NewOp);
11204 } else if (VT.is128BitVector() && Subtarget->hasSSE2()) {
11205 // FIXME: Figure out a cleaner way to do this.
11206 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
11207 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG);
11208 if (NewOp.getNode()) {
11209 MVT NewVT = NewOp.getSimpleValueType();
11210 if (isCommutedMOVLMask(cast<ShuffleVectorSDNode>(NewOp)->getMask(),
11211 NewVT, true, false))
11212 return getVZextMovL(VT, NewVT, NewOp.getOperand(0), DAG, Subtarget,
11215 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
11216 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG);
11217 if (NewOp.getNode()) {
11218 MVT NewVT = NewOp.getSimpleValueType();
11219 if (isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)->getMask(), NewVT))
11220 return getVZextMovL(VT, NewVT, NewOp.getOperand(1), DAG, Subtarget,
11229 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
11230 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11231 SDValue V1 = Op.getOperand(0);
11232 SDValue V2 = Op.getOperand(1);
11233 MVT VT = Op.getSimpleValueType();
11235 unsigned NumElems = VT.getVectorNumElements();
11236 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
11237 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
11238 bool V1IsSplat = false;
11239 bool V2IsSplat = false;
11240 bool HasSSE2 = Subtarget->hasSSE2();
11241 bool HasFp256 = Subtarget->hasFp256();
11242 bool HasInt256 = Subtarget->hasInt256();
11243 MachineFunction &MF = DAG.getMachineFunction();
11244 bool OptForSize = MF.getFunction()->getAttributes().
11245 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize);
11247 // Check if we should use the experimental vector shuffle lowering. If so,
11248 // delegate completely to that code path.
11249 if (ExperimentalVectorShuffleLowering)
11250 return lowerVectorShuffle(Op, Subtarget, DAG);
11252 assert(VT.getSizeInBits() != 64 && "Can't lower MMX shuffles");
11254 if (V1IsUndef && V2IsUndef)
11255 return DAG.getUNDEF(VT);
11257 // When we create a shuffle node we put the UNDEF node to second operand,
11258 // but in some cases the first operand may be transformed to UNDEF.
11259 // In this case we should just commute the node.
11261 return DAG.getCommutedVectorShuffle(*SVOp);
11263 // Vector shuffle lowering takes 3 steps:
11265 // 1) Normalize the input vectors. Here splats, zeroed vectors, profitable
11266 // narrowing and commutation of operands should be handled.
11267 // 2) Matching of shuffles with known shuffle masks to x86 target specific
11269 // 3) Rewriting of unmatched masks into new generic shuffle operations,
11270 // so the shuffle can be broken into other shuffles and the legalizer can
11271 // try the lowering again.
11273 // The general idea is that no vector_shuffle operation should be left to
11274 // be matched during isel, all of them must be converted to a target specific
11277 // Normalize the input vectors. Here splats, zeroed vectors, profitable
11278 // narrowing and commutation of operands should be handled. The actual code
11279 // doesn't include all of those, work in progress...
11280 SDValue NewOp = NormalizeVectorShuffle(Op, Subtarget, DAG);
11281 if (NewOp.getNode())
11284 SmallVector<int, 8> M(SVOp->getMask().begin(), SVOp->getMask().end());
11286 // NOTE: isPSHUFDMask can also match both masks below (unpckl_undef and
11287 // unpckh_undef). Only use pshufd if speed is more important than size.
11288 if (OptForSize && isUNPCKL_v_undef_Mask(M, VT, HasInt256))
11289 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V1, DAG);
11290 if (OptForSize && isUNPCKH_v_undef_Mask(M, VT, HasInt256))
11291 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V1, DAG);
11293 if (isMOVDDUPMask(M, VT) && Subtarget->hasSSE3() &&
11294 V2IsUndef && MayFoldVectorLoad(V1))
11295 return getMOVDDup(Op, dl, V1, DAG);
11297 if (isMOVHLPS_v_undef_Mask(M, VT))
11298 return getMOVHighToLow(Op, dl, DAG);
11300 // Use to match splats
11301 if (HasSSE2 && isUNPCKHMask(M, VT, HasInt256) && V2IsUndef &&
11302 (VT == MVT::v2f64 || VT == MVT::v2i64))
11303 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V1, DAG);
11305 if (isPSHUFDMask(M, VT)) {
11306 // The actual implementation will match the mask in the if above and then
11307 // during isel it can match several different instructions, not only pshufd
11308 // as its name says, sad but true, emulate the behavior for now...
11309 if (isMOVDDUPMask(M, VT) && ((VT == MVT::v4f32 || VT == MVT::v2i64)))
11310 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V1, DAG);
11312 unsigned TargetMask = getShuffleSHUFImmediate(SVOp);
11314 if (HasSSE2 && (VT == MVT::v4f32 || VT == MVT::v4i32))
11315 return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1, TargetMask, DAG);
11317 if (HasFp256 && (VT == MVT::v4f32 || VT == MVT::v2f64))
11318 return getTargetShuffleNode(X86ISD::VPERMILPI, dl, VT, V1, TargetMask,
11321 return getTargetShuffleNode(X86ISD::SHUFP, dl, VT, V1, V1,
11325 if (isPALIGNRMask(M, VT, Subtarget))
11326 return getTargetShuffleNode(X86ISD::PALIGNR, dl, VT, V1, V2,
11327 getShufflePALIGNRImmediate(SVOp),
11330 if (isVALIGNMask(M, VT, Subtarget))
11331 return getTargetShuffleNode(X86ISD::VALIGN, dl, VT, V1, V2,
11332 getShuffleVALIGNImmediate(SVOp),
11335 // Check if this can be converted into a logical shift.
11336 bool isLeft = false;
11337 unsigned ShAmt = 0;
11339 bool isShift = HasSSE2 && isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
11340 if (isShift && ShVal.hasOneUse()) {
11341 // If the shifted value has multiple uses, it may be cheaper to use
11342 // v_set0 + movlhps or movhlps, etc.
11343 MVT EltVT = VT.getVectorElementType();
11344 ShAmt *= EltVT.getSizeInBits();
11345 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
11348 if (isMOVLMask(M, VT)) {
11349 if (ISD::isBuildVectorAllZeros(V1.getNode()))
11350 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
11351 if (!isMOVLPMask(M, VT)) {
11352 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
11353 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
11355 if (VT == MVT::v4i32 || VT == MVT::v4f32)
11356 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
11360 // FIXME: fold these into legal mask.
11361 if (isMOVLHPSMask(M, VT) && !isUNPCKLMask(M, VT, HasInt256))
11362 return getMOVLowToHigh(Op, dl, DAG, HasSSE2);
11364 if (isMOVHLPSMask(M, VT))
11365 return getMOVHighToLow(Op, dl, DAG);
11367 if (V2IsUndef && isMOVSHDUPMask(M, VT, Subtarget))
11368 return getTargetShuffleNode(X86ISD::MOVSHDUP, dl, VT, V1, DAG);
11370 if (V2IsUndef && isMOVSLDUPMask(M, VT, Subtarget))
11371 return getTargetShuffleNode(X86ISD::MOVSLDUP, dl, VT, V1, DAG);
11373 if (isMOVLPMask(M, VT))
11374 return getMOVLP(Op, dl, DAG, HasSSE2);
11376 if (ShouldXformToMOVHLPS(M, VT) ||
11377 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), M, VT))
11378 return DAG.getCommutedVectorShuffle(*SVOp);
11381 // No better options. Use a vshldq / vsrldq.
11382 MVT EltVT = VT.getVectorElementType();
11383 ShAmt *= EltVT.getSizeInBits();
11384 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
11387 bool Commuted = false;
11388 // FIXME: This should also accept a bitcast of a splat? Be careful, not
11389 // 1,1,1,1 -> v8i16 though.
11390 BitVector UndefElements;
11391 if (auto *BVOp = dyn_cast<BuildVectorSDNode>(V1.getNode()))
11392 if (BVOp->getConstantSplatNode(&UndefElements) && UndefElements.none())
11394 if (auto *BVOp = dyn_cast<BuildVectorSDNode>(V2.getNode()))
11395 if (BVOp->getConstantSplatNode(&UndefElements) && UndefElements.none())
11398 // Canonicalize the splat or undef, if present, to be on the RHS.
11399 if (!V2IsUndef && V1IsSplat && !V2IsSplat) {
11400 CommuteVectorShuffleMask(M, NumElems);
11402 std::swap(V1IsSplat, V2IsSplat);
11406 if (isCommutedMOVLMask(M, VT, V2IsSplat, V2IsUndef)) {
11407 // Shuffling low element of v1 into undef, just return v1.
11410 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
11411 // the instruction selector will not match, so get a canonical MOVL with
11412 // swapped operands to undo the commute.
11413 return getMOVL(DAG, dl, VT, V2, V1);
11416 if (isUNPCKLMask(M, VT, HasInt256))
11417 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V2, DAG);
11419 if (isUNPCKHMask(M, VT, HasInt256))
11420 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V2, DAG);
11423 // Normalize mask so all entries that point to V2 points to its first
11424 // element then try to match unpck{h|l} again. If match, return a
11425 // new vector_shuffle with the corrected mask.p
11426 SmallVector<int, 8> NewMask(M.begin(), M.end());
11427 NormalizeMask(NewMask, NumElems);
11428 if (isUNPCKLMask(NewMask, VT, HasInt256, true))
11429 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V2, DAG);
11430 if (isUNPCKHMask(NewMask, VT, HasInt256, true))
11431 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V2, DAG);
11435 // Commute is back and try unpck* again.
11436 // FIXME: this seems wrong.
11437 CommuteVectorShuffleMask(M, NumElems);
11439 std::swap(V1IsSplat, V2IsSplat);
11441 if (isUNPCKLMask(M, VT, HasInt256))
11442 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V2, DAG);
11444 if (isUNPCKHMask(M, VT, HasInt256))
11445 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V2, DAG);
11448 // Normalize the node to match x86 shuffle ops if needed
11449 if (!V2IsUndef && (isSHUFPMask(M, VT, /* Commuted */ true)))
11450 return DAG.getCommutedVectorShuffle(*SVOp);
11452 // The checks below are all present in isShuffleMaskLegal, but they are
11453 // inlined here right now to enable us to directly emit target specific
11454 // nodes, and remove one by one until they don't return Op anymore.
11456 if (ShuffleVectorSDNode::isSplatMask(&M[0], VT) &&
11457 SVOp->getSplatIndex() == 0 && V2IsUndef) {
11458 if (VT == MVT::v2f64 || VT == MVT::v2i64)
11459 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V1, DAG);
11462 if (isPSHUFHWMask(M, VT, HasInt256))
11463 return getTargetShuffleNode(X86ISD::PSHUFHW, dl, VT, V1,
11464 getShufflePSHUFHWImmediate(SVOp),
11467 if (isPSHUFLWMask(M, VT, HasInt256))
11468 return getTargetShuffleNode(X86ISD::PSHUFLW, dl, VT, V1,
11469 getShufflePSHUFLWImmediate(SVOp),
11472 unsigned MaskValue;
11473 if (isBlendMask(M, VT, Subtarget->hasSSE41(), Subtarget->hasInt256(),
11475 return LowerVECTOR_SHUFFLEtoBlend(SVOp, MaskValue, Subtarget, DAG);
11477 if (isSHUFPMask(M, VT))
11478 return getTargetShuffleNode(X86ISD::SHUFP, dl, VT, V1, V2,
11479 getShuffleSHUFImmediate(SVOp), DAG);
11481 if (isUNPCKL_v_undef_Mask(M, VT, HasInt256))
11482 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V1, DAG);
11483 if (isUNPCKH_v_undef_Mask(M, VT, HasInt256))
11484 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V1, DAG);
11486 //===--------------------------------------------------------------------===//
11487 // Generate target specific nodes for 128 or 256-bit shuffles only
11488 // supported in the AVX instruction set.
11491 // Handle VMOVDDUPY permutations
11492 if (V2IsUndef && isMOVDDUPYMask(M, VT, HasFp256))
11493 return getTargetShuffleNode(X86ISD::MOVDDUP, dl, VT, V1, DAG);
11495 // Handle VPERMILPS/D* permutations
11496 if (isVPERMILPMask(M, VT)) {
11497 if ((HasInt256 && VT == MVT::v8i32) || VT == MVT::v16i32)
11498 return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1,
11499 getShuffleSHUFImmediate(SVOp), DAG);
11500 return getTargetShuffleNode(X86ISD::VPERMILPI, dl, VT, V1,
11501 getShuffleSHUFImmediate(SVOp), DAG);
11505 if (VT.is512BitVector() && isINSERT64x4Mask(M, VT, &Idx))
11506 return Insert256BitVector(V1, Extract256BitVector(V2, 0, DAG, dl),
11507 Idx*(NumElems/2), DAG, dl);
11509 // Handle VPERM2F128/VPERM2I128 permutations
11510 if (isVPERM2X128Mask(M, VT, HasFp256))
11511 return getTargetShuffleNode(X86ISD::VPERM2X128, dl, VT, V1,
11512 V2, getShuffleVPERM2X128Immediate(SVOp), DAG);
11514 if (Subtarget->hasSSE41() && isINSERTPSMask(M, VT))
11515 return getINSERTPS(SVOp, dl, DAG);
11518 if (V2IsUndef && HasInt256 && isPermImmMask(M, VT, Imm8))
11519 return getTargetShuffleNode(X86ISD::VPERMI, dl, VT, V1, Imm8, DAG);
11521 if ((V2IsUndef && HasInt256 && VT.is256BitVector() && NumElems == 8) ||
11522 VT.is512BitVector()) {
11523 MVT MaskEltVT = MVT::getIntegerVT(VT.getVectorElementType().getSizeInBits());
11524 MVT MaskVectorVT = MVT::getVectorVT(MaskEltVT, NumElems);
11525 SmallVector<SDValue, 16> permclMask;
11526 for (unsigned i = 0; i != NumElems; ++i) {
11527 permclMask.push_back(DAG.getConstant((M[i]>=0) ? M[i] : 0, MaskEltVT));
11530 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVectorVT, permclMask);
11532 // Bitcast is for VPERMPS since mask is v8i32 but node takes v8f32
11533 return DAG.getNode(X86ISD::VPERMV, dl, VT,
11534 DAG.getNode(ISD::BITCAST, dl, VT, Mask), V1);
11535 return DAG.getNode(X86ISD::VPERMV3, dl, VT, V1,
11536 DAG.getNode(ISD::BITCAST, dl, VT, Mask), V2);
11539 //===--------------------------------------------------------------------===//
11540 // Since no target specific shuffle was selected for this generic one,
11541 // lower it into other known shuffles. FIXME: this isn't true yet, but
11542 // this is the plan.
11545 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
11546 if (VT == MVT::v8i16) {
11547 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(Op, Subtarget, DAG);
11548 if (NewOp.getNode())
11552 if (VT == MVT::v16i16 && Subtarget->hasInt256()) {
11553 SDValue NewOp = LowerVECTOR_SHUFFLEv16i16(Op, DAG);
11554 if (NewOp.getNode())
11558 if (VT == MVT::v16i8) {
11559 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, Subtarget, DAG);
11560 if (NewOp.getNode())
11564 if (VT == MVT::v32i8) {
11565 SDValue NewOp = LowerVECTOR_SHUFFLEv32i8(SVOp, Subtarget, DAG);
11566 if (NewOp.getNode())
11570 // Handle all 128-bit wide vectors with 4 elements, and match them with
11571 // several different shuffle types.
11572 if (NumElems == 4 && VT.is128BitVector())
11573 return LowerVECTOR_SHUFFLE_128v4(SVOp, DAG);
11575 // Handle general 256-bit shuffles
11576 if (VT.is256BitVector())
11577 return LowerVECTOR_SHUFFLE_256(SVOp, DAG);
11582 // This function assumes its argument is a BUILD_VECTOR of constants or
11583 // undef SDNodes. i.e: ISD::isBuildVectorOfConstantSDNodes(BuildVector) is
11585 static bool BUILD_VECTORtoBlendMask(BuildVectorSDNode *BuildVector,
11586 unsigned &MaskValue) {
11588 unsigned NumElems = BuildVector->getNumOperands();
11589 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
11590 unsigned NumLanes = (NumElems - 1) / 8 + 1;
11591 unsigned NumElemsInLane = NumElems / NumLanes;
11593 // Blend for v16i16 should be symetric for the both lanes.
11594 for (unsigned i = 0; i < NumElemsInLane; ++i) {
11595 SDValue EltCond = BuildVector->getOperand(i);
11596 SDValue SndLaneEltCond =
11597 (NumLanes == 2) ? BuildVector->getOperand(i + NumElemsInLane) : EltCond;
11599 int Lane1Cond = -1, Lane2Cond = -1;
11600 if (isa<ConstantSDNode>(EltCond))
11601 Lane1Cond = !isZero(EltCond);
11602 if (isa<ConstantSDNode>(SndLaneEltCond))
11603 Lane2Cond = !isZero(SndLaneEltCond);
11605 if (Lane1Cond == Lane2Cond || Lane2Cond < 0)
11606 // Lane1Cond != 0, means we want the first argument.
11607 // Lane1Cond == 0, means we want the second argument.
11608 // The encoding of this argument is 0 for the first argument, 1
11609 // for the second. Therefore, invert the condition.
11610 MaskValue |= !Lane1Cond << i;
11611 else if (Lane1Cond < 0)
11612 MaskValue |= !Lane2Cond << i;
11619 // Try to lower a vselect node into a simple blend instruction.
11620 static SDValue LowerVSELECTtoBlend(SDValue Op, const X86Subtarget *Subtarget,
11621 SelectionDAG &DAG) {
11622 SDValue Cond = Op.getOperand(0);
11623 SDValue LHS = Op.getOperand(1);
11624 SDValue RHS = Op.getOperand(2);
11626 MVT VT = Op.getSimpleValueType();
11627 MVT EltVT = VT.getVectorElementType();
11628 unsigned NumElems = VT.getVectorNumElements();
11630 // There is no blend with immediate in AVX-512.
11631 if (VT.is512BitVector())
11634 if (!Subtarget->hasSSE41() || EltVT == MVT::i8)
11636 if (!Subtarget->hasInt256() && VT == MVT::v16i16)
11639 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
11642 // Check the mask for BLEND and build the value.
11643 unsigned MaskValue = 0;
11644 if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue))
11647 // Convert i32 vectors to floating point if it is not AVX2.
11648 // AVX2 introduced VPBLENDD instruction for 128 and 256-bit vectors.
11650 if (EltVT == MVT::i64 || (EltVT == MVT::i32 && !Subtarget->hasInt256())) {
11651 BlendVT = MVT::getVectorVT(MVT::getFloatingPointVT(EltVT.getSizeInBits()),
11653 LHS = DAG.getNode(ISD::BITCAST, dl, VT, LHS);
11654 RHS = DAG.getNode(ISD::BITCAST, dl, VT, RHS);
11657 SDValue Ret = DAG.getNode(X86ISD::BLENDI, dl, BlendVT, LHS, RHS,
11658 DAG.getConstant(MaskValue, MVT::i32));
11659 return DAG.getNode(ISD::BITCAST, dl, VT, Ret);
11662 SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const {
11663 // A vselect where all conditions and data are constants can be optimized into
11664 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
11665 if (ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(0).getNode()) &&
11666 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(1).getNode()) &&
11667 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(2).getNode()))
11670 SDValue BlendOp = LowerVSELECTtoBlend(Op, Subtarget, DAG);
11671 if (BlendOp.getNode())
11674 // Some types for vselect were previously set to Expand, not Legal or
11675 // Custom. Return an empty SDValue so we fall-through to Expand, after
11676 // the Custom lowering phase.
11677 MVT VT = Op.getSimpleValueType();
11678 switch (VT.SimpleTy) {
11683 if (Subtarget->hasBWI() && Subtarget->hasVLX())
11688 // We couldn't create a "Blend with immediate" node.
11689 // This node should still be legal, but we'll have to emit a blendv*
11694 static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
11695 MVT VT = Op.getSimpleValueType();
11698 if (!Op.getOperand(0).getSimpleValueType().is128BitVector())
11701 if (VT.getSizeInBits() == 8) {
11702 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
11703 Op.getOperand(0), Op.getOperand(1));
11704 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
11705 DAG.getValueType(VT));
11706 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11709 if (VT.getSizeInBits() == 16) {
11710 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11711 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
11713 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
11714 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11715 DAG.getNode(ISD::BITCAST, dl,
11718 Op.getOperand(1)));
11719 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
11720 Op.getOperand(0), Op.getOperand(1));
11721 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
11722 DAG.getValueType(VT));
11723 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11726 if (VT == MVT::f32) {
11727 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
11728 // the result back to FR32 register. It's only worth matching if the
11729 // result has a single use which is a store or a bitcast to i32. And in
11730 // the case of a store, it's not worth it if the index is a constant 0,
11731 // because a MOVSSmr can be used instead, which is smaller and faster.
11732 if (!Op.hasOneUse())
11734 SDNode *User = *Op.getNode()->use_begin();
11735 if ((User->getOpcode() != ISD::STORE ||
11736 (isa<ConstantSDNode>(Op.getOperand(1)) &&
11737 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
11738 (User->getOpcode() != ISD::BITCAST ||
11739 User->getValueType(0) != MVT::i32))
11741 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11742 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32,
11745 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, Extract);
11748 if (VT == MVT::i32 || VT == MVT::i64) {
11749 // ExtractPS/pextrq works with constant index.
11750 if (isa<ConstantSDNode>(Op.getOperand(1)))
11756 /// Extract one bit from mask vector, like v16i1 or v8i1.
11757 /// AVX-512 feature.
11759 X86TargetLowering::ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const {
11760 SDValue Vec = Op.getOperand(0);
11762 MVT VecVT = Vec.getSimpleValueType();
11763 SDValue Idx = Op.getOperand(1);
11764 MVT EltVT = Op.getSimpleValueType();
11766 assert((EltVT == MVT::i1) && "Unexpected operands in ExtractBitFromMaskVector");
11768 // variable index can't be handled in mask registers,
11769 // extend vector to VR512
11770 if (!isa<ConstantSDNode>(Idx)) {
11771 MVT ExtVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
11772 SDValue Ext = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVT, Vec);
11773 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
11774 ExtVT.getVectorElementType(), Ext, Idx);
11775 return DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
11778 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11779 const TargetRegisterClass* rc = getRegClassFor(VecVT);
11780 unsigned MaxSift = rc->getSize()*8 - 1;
11781 Vec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, Vec,
11782 DAG.getConstant(MaxSift - IdxVal, MVT::i8));
11783 Vec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, Vec,
11784 DAG.getConstant(MaxSift, MVT::i8));
11785 return DAG.getNode(X86ISD::VEXTRACT, dl, MVT::i1, Vec,
11786 DAG.getIntPtrConstant(0));
11790 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
11791 SelectionDAG &DAG) const {
11793 SDValue Vec = Op.getOperand(0);
11794 MVT VecVT = Vec.getSimpleValueType();
11795 SDValue Idx = Op.getOperand(1);
11797 if (Op.getSimpleValueType() == MVT::i1)
11798 return ExtractBitFromMaskVector(Op, DAG);
11800 if (!isa<ConstantSDNode>(Idx)) {
11801 if (VecVT.is512BitVector() ||
11802 (VecVT.is256BitVector() && Subtarget->hasInt256() &&
11803 VecVT.getVectorElementType().getSizeInBits() == 32)) {
11806 MVT::getIntegerVT(VecVT.getVectorElementType().getSizeInBits());
11807 MVT MaskVT = MVT::getVectorVT(MaskEltVT, VecVT.getSizeInBits() /
11808 MaskEltVT.getSizeInBits());
11810 Idx = DAG.getZExtOrTrunc(Idx, dl, MaskEltVT);
11811 SDValue Mask = DAG.getNode(X86ISD::VINSERT, dl, MaskVT,
11812 getZeroVector(MaskVT, Subtarget, DAG, dl),
11813 Idx, DAG.getConstant(0, getPointerTy()));
11814 SDValue Perm = DAG.getNode(X86ISD::VPERMV, dl, VecVT, Mask, Vec);
11815 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(),
11816 Perm, DAG.getConstant(0, getPointerTy()));
11821 // If this is a 256-bit vector result, first extract the 128-bit vector and
11822 // then extract the element from the 128-bit vector.
11823 if (VecVT.is256BitVector() || VecVT.is512BitVector()) {
11825 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11826 // Get the 128-bit vector.
11827 Vec = Extract128BitVector(Vec, IdxVal, DAG, dl);
11828 MVT EltVT = VecVT.getVectorElementType();
11830 unsigned ElemsPerChunk = 128 / EltVT.getSizeInBits();
11832 //if (IdxVal >= NumElems/2)
11833 // IdxVal -= NumElems/2;
11834 IdxVal -= (IdxVal/ElemsPerChunk)*ElemsPerChunk;
11835 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
11836 DAG.getConstant(IdxVal, MVT::i32));
11839 assert(VecVT.is128BitVector() && "Unexpected vector length");
11841 if (Subtarget->hasSSE41()) {
11842 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
11847 MVT VT = Op.getSimpleValueType();
11848 // TODO: handle v16i8.
11849 if (VT.getSizeInBits() == 16) {
11850 SDValue Vec = Op.getOperand(0);
11851 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11853 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
11854 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11855 DAG.getNode(ISD::BITCAST, dl,
11857 Op.getOperand(1)));
11858 // Transform it so it match pextrw which produces a 32-bit result.
11859 MVT EltVT = MVT::i32;
11860 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
11861 Op.getOperand(0), Op.getOperand(1));
11862 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
11863 DAG.getValueType(VT));
11864 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11867 if (VT.getSizeInBits() == 32) {
11868 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11872 // SHUFPS the element to the lowest double word, then movss.
11873 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
11874 MVT VVT = Op.getOperand(0).getSimpleValueType();
11875 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
11876 DAG.getUNDEF(VVT), Mask);
11877 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
11878 DAG.getIntPtrConstant(0));
11881 if (VT.getSizeInBits() == 64) {
11882 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
11883 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
11884 // to match extract_elt for f64.
11885 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11889 // UNPCKHPD the element to the lowest double word, then movsd.
11890 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
11891 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
11892 int Mask[2] = { 1, -1 };
11893 MVT VVT = Op.getOperand(0).getSimpleValueType();
11894 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
11895 DAG.getUNDEF(VVT), Mask);
11896 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
11897 DAG.getIntPtrConstant(0));
11903 /// Insert one bit to mask vector, like v16i1 or v8i1.
11904 /// AVX-512 feature.
11906 X86TargetLowering::InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const {
11908 SDValue Vec = Op.getOperand(0);
11909 SDValue Elt = Op.getOperand(1);
11910 SDValue Idx = Op.getOperand(2);
11911 MVT VecVT = Vec.getSimpleValueType();
11913 if (!isa<ConstantSDNode>(Idx)) {
11914 // Non constant index. Extend source and destination,
11915 // insert element and then truncate the result.
11916 MVT ExtVecVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
11917 MVT ExtEltVT = (VecVT == MVT::v8i1 ? MVT::i64 : MVT::i32);
11918 SDValue ExtOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ExtVecVT,
11919 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVecVT, Vec),
11920 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtEltVT, Elt), Idx);
11921 return DAG.getNode(ISD::TRUNCATE, dl, VecVT, ExtOp);
11924 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11925 SDValue EltInVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Elt);
11926 if (Vec.getOpcode() == ISD::UNDEF)
11927 return DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
11928 DAG.getConstant(IdxVal, MVT::i8));
11929 const TargetRegisterClass* rc = getRegClassFor(VecVT);
11930 unsigned MaxSift = rc->getSize()*8 - 1;
11931 EltInVec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
11932 DAG.getConstant(MaxSift, MVT::i8));
11933 EltInVec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, EltInVec,
11934 DAG.getConstant(MaxSift - IdxVal, MVT::i8));
11935 return DAG.getNode(ISD::OR, dl, VecVT, Vec, EltInVec);
11938 SDValue X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
11939 SelectionDAG &DAG) const {
11940 MVT VT = Op.getSimpleValueType();
11941 MVT EltVT = VT.getVectorElementType();
11943 if (EltVT == MVT::i1)
11944 return InsertBitToMaskVector(Op, DAG);
11947 SDValue N0 = Op.getOperand(0);
11948 SDValue N1 = Op.getOperand(1);
11949 SDValue N2 = Op.getOperand(2);
11950 if (!isa<ConstantSDNode>(N2))
11952 auto *N2C = cast<ConstantSDNode>(N2);
11953 unsigned IdxVal = N2C->getZExtValue();
11955 // If the vector is wider than 128 bits, extract the 128-bit subvector, insert
11956 // into that, and then insert the subvector back into the result.
11957 if (VT.is256BitVector() || VT.is512BitVector()) {
11958 // Get the desired 128-bit vector half.
11959 SDValue V = Extract128BitVector(N0, IdxVal, DAG, dl);
11961 // Insert the element into the desired half.
11962 unsigned NumEltsIn128 = 128 / EltVT.getSizeInBits();
11963 unsigned IdxIn128 = IdxVal - (IdxVal / NumEltsIn128) * NumEltsIn128;
11965 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1,
11966 DAG.getConstant(IdxIn128, MVT::i32));
11968 // Insert the changed part back to the 256-bit vector
11969 return Insert128BitVector(N0, V, IdxVal, DAG, dl);
11971 assert(VT.is128BitVector() && "Only 128-bit vector types should be left!");
11973 if (Subtarget->hasSSE41()) {
11974 if (EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) {
11976 if (VT == MVT::v8i16) {
11977 Opc = X86ISD::PINSRW;
11979 assert(VT == MVT::v16i8);
11980 Opc = X86ISD::PINSRB;
11983 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
11985 if (N1.getValueType() != MVT::i32)
11986 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11987 if (N2.getValueType() != MVT::i32)
11988 N2 = DAG.getIntPtrConstant(IdxVal);
11989 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
11992 if (EltVT == MVT::f32) {
11993 // Bits [7:6] of the constant are the source select. This will always be
11994 // zero here. The DAG Combiner may combine an extract_elt index into
11996 // bits. For example (insert (extract, 3), 2) could be matched by
11998 // the '3' into bits [7:6] of X86ISD::INSERTPS.
11999 // Bits [5:4] of the constant are the destination select. This is the
12000 // value of the incoming immediate.
12001 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
12002 // combine either bitwise AND or insert of float 0.0 to set these bits.
12003 N2 = DAG.getIntPtrConstant(IdxVal << 4);
12004 // Create this as a scalar to vector..
12005 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
12006 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
12009 if (EltVT == MVT::i32 || EltVT == MVT::i64) {
12010 // PINSR* works with constant index.
12015 if (EltVT == MVT::i8)
12018 if (EltVT.getSizeInBits() == 16) {
12019 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
12020 // as its second argument.
12021 if (N1.getValueType() != MVT::i32)
12022 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
12023 if (N2.getValueType() != MVT::i32)
12024 N2 = DAG.getIntPtrConstant(IdxVal);
12025 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
12030 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
12032 MVT OpVT = Op.getSimpleValueType();
12034 // If this is a 256-bit vector result, first insert into a 128-bit
12035 // vector and then insert into the 256-bit vector.
12036 if (!OpVT.is128BitVector()) {
12037 // Insert into a 128-bit vector.
12038 unsigned SizeFactor = OpVT.getSizeInBits()/128;
12039 MVT VT128 = MVT::getVectorVT(OpVT.getVectorElementType(),
12040 OpVT.getVectorNumElements() / SizeFactor);
12042 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
12044 // Insert the 128-bit vector.
12045 return Insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl);
12048 if (OpVT == MVT::v1i64 &&
12049 Op.getOperand(0).getValueType() == MVT::i64)
12050 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
12052 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
12053 assert(OpVT.is128BitVector() && "Expected an SSE type!");
12054 return DAG.getNode(ISD::BITCAST, dl, OpVT,
12055 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,AnyExt));
12058 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
12059 // a simple subregister reference or explicit instructions to grab
12060 // upper bits of a vector.
12061 static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
12062 SelectionDAG &DAG) {
12064 SDValue In = Op.getOperand(0);
12065 SDValue Idx = Op.getOperand(1);
12066 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
12067 MVT ResVT = Op.getSimpleValueType();
12068 MVT InVT = In.getSimpleValueType();
12070 if (Subtarget->hasFp256()) {
12071 if (ResVT.is128BitVector() &&
12072 (InVT.is256BitVector() || InVT.is512BitVector()) &&
12073 isa<ConstantSDNode>(Idx)) {
12074 return Extract128BitVector(In, IdxVal, DAG, dl);
12076 if (ResVT.is256BitVector() && InVT.is512BitVector() &&
12077 isa<ConstantSDNode>(Idx)) {
12078 return Extract256BitVector(In, IdxVal, DAG, dl);
12084 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
12085 // simple superregister reference or explicit instructions to insert
12086 // the upper bits of a vector.
12087 static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
12088 SelectionDAG &DAG) {
12089 if (Subtarget->hasFp256()) {
12090 SDLoc dl(Op.getNode());
12091 SDValue Vec = Op.getNode()->getOperand(0);
12092 SDValue SubVec = Op.getNode()->getOperand(1);
12093 SDValue Idx = Op.getNode()->getOperand(2);
12095 if ((Op.getNode()->getSimpleValueType(0).is256BitVector() ||
12096 Op.getNode()->getSimpleValueType(0).is512BitVector()) &&
12097 SubVec.getNode()->getSimpleValueType(0).is128BitVector() &&
12098 isa<ConstantSDNode>(Idx)) {
12099 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
12100 return Insert128BitVector(Vec, SubVec, IdxVal, DAG, dl);
12103 if (Op.getNode()->getSimpleValueType(0).is512BitVector() &&
12104 SubVec.getNode()->getSimpleValueType(0).is256BitVector() &&
12105 isa<ConstantSDNode>(Idx)) {
12106 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
12107 return Insert256BitVector(Vec, SubVec, IdxVal, DAG, dl);
12113 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
12114 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
12115 // one of the above mentioned nodes. It has to be wrapped because otherwise
12116 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
12117 // be used to form addressing mode. These wrapped nodes will be selected
12120 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
12121 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
12123 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
12124 // global base reg.
12125 unsigned char OpFlag = 0;
12126 unsigned WrapperKind = X86ISD::Wrapper;
12127 CodeModel::Model M = DAG.getTarget().getCodeModel();
12129 if (Subtarget->isPICStyleRIPRel() &&
12130 (M == CodeModel::Small || M == CodeModel::Kernel))
12131 WrapperKind = X86ISD::WrapperRIP;
12132 else if (Subtarget->isPICStyleGOT())
12133 OpFlag = X86II::MO_GOTOFF;
12134 else if (Subtarget->isPICStyleStubPIC())
12135 OpFlag = X86II::MO_PIC_BASE_OFFSET;
12137 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
12138 CP->getAlignment(),
12139 CP->getOffset(), OpFlag);
12141 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
12142 // With PIC, the address is actually $g + Offset.
12144 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
12145 DAG.getNode(X86ISD::GlobalBaseReg,
12146 SDLoc(), getPointerTy()),
12153 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
12154 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
12156 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
12157 // global base reg.
12158 unsigned char OpFlag = 0;
12159 unsigned WrapperKind = X86ISD::Wrapper;
12160 CodeModel::Model M = DAG.getTarget().getCodeModel();
12162 if (Subtarget->isPICStyleRIPRel() &&
12163 (M == CodeModel::Small || M == CodeModel::Kernel))
12164 WrapperKind = X86ISD::WrapperRIP;
12165 else if (Subtarget->isPICStyleGOT())
12166 OpFlag = X86II::MO_GOTOFF;
12167 else if (Subtarget->isPICStyleStubPIC())
12168 OpFlag = X86II::MO_PIC_BASE_OFFSET;
12170 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
12173 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
12175 // With PIC, the address is actually $g + Offset.
12177 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
12178 DAG.getNode(X86ISD::GlobalBaseReg,
12179 SDLoc(), getPointerTy()),
12186 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
12187 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
12189 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
12190 // global base reg.
12191 unsigned char OpFlag = 0;
12192 unsigned WrapperKind = X86ISD::Wrapper;
12193 CodeModel::Model M = DAG.getTarget().getCodeModel();
12195 if (Subtarget->isPICStyleRIPRel() &&
12196 (M == CodeModel::Small || M == CodeModel::Kernel)) {
12197 if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF())
12198 OpFlag = X86II::MO_GOTPCREL;
12199 WrapperKind = X86ISD::WrapperRIP;
12200 } else if (Subtarget->isPICStyleGOT()) {
12201 OpFlag = X86II::MO_GOT;
12202 } else if (Subtarget->isPICStyleStubPIC()) {
12203 OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE;
12204 } else if (Subtarget->isPICStyleStubNoDynamic()) {
12205 OpFlag = X86II::MO_DARWIN_NONLAZY;
12208 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
12211 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
12213 // With PIC, the address is actually $g + Offset.
12214 if (DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
12215 !Subtarget->is64Bit()) {
12216 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
12217 DAG.getNode(X86ISD::GlobalBaseReg,
12218 SDLoc(), getPointerTy()),
12222 // For symbols that require a load from a stub to get the address, emit the
12224 if (isGlobalStubReference(OpFlag))
12225 Result = DAG.getLoad(getPointerTy(), DL, DAG.getEntryNode(), Result,
12226 MachinePointerInfo::getGOT(), false, false, false, 0);
12232 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
12233 // Create the TargetBlockAddressAddress node.
12234 unsigned char OpFlags =
12235 Subtarget->ClassifyBlockAddressReference();
12236 CodeModel::Model M = DAG.getTarget().getCodeModel();
12237 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
12238 int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset();
12240 SDValue Result = DAG.getTargetBlockAddress(BA, getPointerTy(), Offset,
12243 if (Subtarget->isPICStyleRIPRel() &&
12244 (M == CodeModel::Small || M == CodeModel::Kernel))
12245 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
12247 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
12249 // With PIC, the address is actually $g + Offset.
12250 if (isGlobalRelativeToPICBase(OpFlags)) {
12251 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
12252 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
12260 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
12261 int64_t Offset, SelectionDAG &DAG) const {
12262 // Create the TargetGlobalAddress node, folding in the constant
12263 // offset if it is legal.
12264 unsigned char OpFlags =
12265 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget());
12266 CodeModel::Model M = DAG.getTarget().getCodeModel();
12268 if (OpFlags == X86II::MO_NO_FLAG &&
12269 X86::isOffsetSuitableForCodeModel(Offset, M)) {
12270 // A direct static reference to a global.
12271 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
12274 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
12277 if (Subtarget->isPICStyleRIPRel() &&
12278 (M == CodeModel::Small || M == CodeModel::Kernel))
12279 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
12281 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
12283 // With PIC, the address is actually $g + Offset.
12284 if (isGlobalRelativeToPICBase(OpFlags)) {
12285 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
12286 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
12290 // For globals that require a load from a stub to get the address, emit the
12292 if (isGlobalStubReference(OpFlags))
12293 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
12294 MachinePointerInfo::getGOT(), false, false, false, 0);
12296 // If there was a non-zero offset that we didn't fold, create an explicit
12297 // addition for it.
12299 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
12300 DAG.getConstant(Offset, getPointerTy()));
12306 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
12307 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
12308 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
12309 return LowerGlobalAddress(GV, SDLoc(Op), Offset, DAG);
12313 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
12314 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
12315 unsigned char OperandFlags, bool LocalDynamic = false) {
12316 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
12317 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
12319 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
12320 GA->getValueType(0),
12324 X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR
12328 SDValue Ops[] = { Chain, TGA, *InFlag };
12329 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
12331 SDValue Ops[] = { Chain, TGA };
12332 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
12335 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
12336 MFI->setAdjustsStack(true);
12338 SDValue Flag = Chain.getValue(1);
12339 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
12342 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
12344 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
12347 SDLoc dl(GA); // ? function entry point might be better
12348 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
12349 DAG.getNode(X86ISD::GlobalBaseReg,
12350 SDLoc(), PtrVT), InFlag);
12351 InFlag = Chain.getValue(1);
12353 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
12356 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
12358 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
12360 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT,
12361 X86::RAX, X86II::MO_TLSGD);
12364 static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA,
12370 // Get the start address of the TLS block for this module.
12371 X86MachineFunctionInfo* MFI = DAG.getMachineFunction()
12372 .getInfo<X86MachineFunctionInfo>();
12373 MFI->incNumLocalDynamicTLSAccesses();
12377 Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, X86::RAX,
12378 X86II::MO_TLSLD, /*LocalDynamic=*/true);
12381 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
12382 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), InFlag);
12383 InFlag = Chain.getValue(1);
12384 Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX,
12385 X86II::MO_TLSLDM, /*LocalDynamic=*/true);
12388 // Note: the CleanupLocalDynamicTLSPass will remove redundant computations
12392 unsigned char OperandFlags = X86II::MO_DTPOFF;
12393 unsigned WrapperKind = X86ISD::Wrapper;
12394 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
12395 GA->getValueType(0),
12396 GA->getOffset(), OperandFlags);
12397 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
12399 // Add x@dtpoff with the base.
12400 return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base);
12403 // Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model.
12404 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
12405 const EVT PtrVT, TLSModel::Model model,
12406 bool is64Bit, bool isPIC) {
12409 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
12410 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
12411 is64Bit ? 257 : 256));
12413 SDValue ThreadPointer =
12414 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), DAG.getIntPtrConstant(0),
12415 MachinePointerInfo(Ptr), false, false, false, 0);
12417 unsigned char OperandFlags = 0;
12418 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
12420 unsigned WrapperKind = X86ISD::Wrapper;
12421 if (model == TLSModel::LocalExec) {
12422 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
12423 } else if (model == TLSModel::InitialExec) {
12425 OperandFlags = X86II::MO_GOTTPOFF;
12426 WrapperKind = X86ISD::WrapperRIP;
12428 OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF;
12431 llvm_unreachable("Unexpected model");
12434 // emit "addl x@ntpoff,%eax" (local exec)
12435 // or "addl x@indntpoff,%eax" (initial exec)
12436 // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic)
12438 DAG.getTargetGlobalAddress(GA->getGlobal(), dl, GA->getValueType(0),
12439 GA->getOffset(), OperandFlags);
12440 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
12442 if (model == TLSModel::InitialExec) {
12443 if (isPIC && !is64Bit) {
12444 Offset = DAG.getNode(ISD::ADD, dl, PtrVT,
12445 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
12449 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
12450 MachinePointerInfo::getGOT(), false, false, false, 0);
12453 // The address of the thread local variable is the add of the thread
12454 // pointer with the offset of the variable.
12455 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
12459 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
12461 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
12462 const GlobalValue *GV = GA->getGlobal();
12464 if (Subtarget->isTargetELF()) {
12465 TLSModel::Model model = DAG.getTarget().getTLSModel(GV);
12468 case TLSModel::GeneralDynamic:
12469 if (Subtarget->is64Bit())
12470 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
12471 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
12472 case TLSModel::LocalDynamic:
12473 return LowerToTLSLocalDynamicModel(GA, DAG, getPointerTy(),
12474 Subtarget->is64Bit());
12475 case TLSModel::InitialExec:
12476 case TLSModel::LocalExec:
12477 return LowerToTLSExecModel(
12478 GA, DAG, getPointerTy(), model, Subtarget->is64Bit(),
12479 DAG.getTarget().getRelocationModel() == Reloc::PIC_);
12481 llvm_unreachable("Unknown TLS model.");
12484 if (Subtarget->isTargetDarwin()) {
12485 // Darwin only has one model of TLS. Lower to that.
12486 unsigned char OpFlag = 0;
12487 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
12488 X86ISD::WrapperRIP : X86ISD::Wrapper;
12490 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
12491 // global base reg.
12492 bool PIC32 = (DAG.getTarget().getRelocationModel() == Reloc::PIC_) &&
12493 !Subtarget->is64Bit();
12495 OpFlag = X86II::MO_TLVP_PIC_BASE;
12497 OpFlag = X86II::MO_TLVP;
12499 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
12500 GA->getValueType(0),
12501 GA->getOffset(), OpFlag);
12502 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
12504 // With PIC32, the address is actually $g + Offset.
12506 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
12507 DAG.getNode(X86ISD::GlobalBaseReg,
12508 SDLoc(), getPointerTy()),
12511 // Lowering the machine isd will make sure everything is in the right
12513 SDValue Chain = DAG.getEntryNode();
12514 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
12515 SDValue Args[] = { Chain, Offset };
12516 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args);
12518 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
12519 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
12520 MFI->setAdjustsStack(true);
12522 // And our return value (tls address) is in the standard call return value
12524 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
12525 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy(),
12526 Chain.getValue(1));
12529 if (Subtarget->isTargetKnownWindowsMSVC() ||
12530 Subtarget->isTargetWindowsGNU()) {
12531 // Just use the implicit TLS architecture
12532 // Need to generate someting similar to:
12533 // mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage
12535 // mov ecx, dword [rel _tls_index]: Load index (from C runtime)
12536 // mov rcx, qword [rdx+rcx*8]
12537 // mov eax, .tls$:tlsvar
12538 // [rax+rcx] contains the address
12539 // Windows 64bit: gs:0x58
12540 // Windows 32bit: fs:__tls_array
12543 SDValue Chain = DAG.getEntryNode();
12545 // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or
12546 // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly
12547 // use its literal value of 0x2C.
12548 Value *Ptr = Constant::getNullValue(Subtarget->is64Bit()
12549 ? Type::getInt8PtrTy(*DAG.getContext(),
12551 : Type::getInt32PtrTy(*DAG.getContext(),
12555 Subtarget->is64Bit()
12556 ? DAG.getIntPtrConstant(0x58)
12557 : (Subtarget->isTargetWindowsGNU()
12558 ? DAG.getIntPtrConstant(0x2C)
12559 : DAG.getExternalSymbol("_tls_array", getPointerTy()));
12561 SDValue ThreadPointer =
12562 DAG.getLoad(getPointerTy(), dl, Chain, TlsArray,
12563 MachinePointerInfo(Ptr), false, false, false, 0);
12565 // Load the _tls_index variable
12566 SDValue IDX = DAG.getExternalSymbol("_tls_index", getPointerTy());
12567 if (Subtarget->is64Bit())
12568 IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, getPointerTy(), Chain,
12569 IDX, MachinePointerInfo(), MVT::i32,
12570 false, false, false, 0);
12572 IDX = DAG.getLoad(getPointerTy(), dl, Chain, IDX, MachinePointerInfo(),
12573 false, false, false, 0);
12575 SDValue Scale = DAG.getConstant(Log2_64_Ceil(TD->getPointerSize()),
12577 IDX = DAG.getNode(ISD::SHL, dl, getPointerTy(), IDX, Scale);
12579 SDValue res = DAG.getNode(ISD::ADD, dl, getPointerTy(), ThreadPointer, IDX);
12580 res = DAG.getLoad(getPointerTy(), dl, Chain, res, MachinePointerInfo(),
12581 false, false, false, 0);
12583 // Get the offset of start of .tls section
12584 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
12585 GA->getValueType(0),
12586 GA->getOffset(), X86II::MO_SECREL);
12587 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), TGA);
12589 // The address of the thread local variable is the add of the thread
12590 // pointer with the offset of the variable.
12591 return DAG.getNode(ISD::ADD, dl, getPointerTy(), res, Offset);
12594 llvm_unreachable("TLS not implemented for this target.");
12597 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values
12598 /// and take a 2 x i32 value to shift plus a shift amount.
12599 static SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) {
12600 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
12601 MVT VT = Op.getSimpleValueType();
12602 unsigned VTBits = VT.getSizeInBits();
12604 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
12605 SDValue ShOpLo = Op.getOperand(0);
12606 SDValue ShOpHi = Op.getOperand(1);
12607 SDValue ShAmt = Op.getOperand(2);
12608 // X86ISD::SHLD and X86ISD::SHRD have defined overflow behavior but the
12609 // generic ISD nodes haven't. Insert an AND to be safe, it's optimized away
12611 SDValue SafeShAmt = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
12612 DAG.getConstant(VTBits - 1, MVT::i8));
12613 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
12614 DAG.getConstant(VTBits - 1, MVT::i8))
12615 : DAG.getConstant(0, VT);
12617 SDValue Tmp2, Tmp3;
12618 if (Op.getOpcode() == ISD::SHL_PARTS) {
12619 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
12620 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, SafeShAmt);
12622 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
12623 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, SafeShAmt);
12626 // If the shift amount is larger or equal than the width of a part we can't
12627 // rely on the results of shld/shrd. Insert a test and select the appropriate
12628 // values for large shift amounts.
12629 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
12630 DAG.getConstant(VTBits, MVT::i8));
12631 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
12632 AndNode, DAG.getConstant(0, MVT::i8));
12635 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
12636 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
12637 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
12639 if (Op.getOpcode() == ISD::SHL_PARTS) {
12640 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
12641 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
12643 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
12644 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
12647 SDValue Ops[2] = { Lo, Hi };
12648 return DAG.getMergeValues(Ops, dl);
12651 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
12652 SelectionDAG &DAG) const {
12653 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
12655 if (SrcVT.isVector())
12658 assert(SrcVT <= MVT::i64 && SrcVT >= MVT::i16 &&
12659 "Unknown SINT_TO_FP to lower!");
12661 // These are really Legal; return the operand so the caller accepts it as
12663 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
12665 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
12666 Subtarget->is64Bit()) {
12671 unsigned Size = SrcVT.getSizeInBits()/8;
12672 MachineFunction &MF = DAG.getMachineFunction();
12673 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
12674 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
12675 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
12677 MachinePointerInfo::getFixedStack(SSFI),
12679 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
12682 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
12684 SelectionDAG &DAG) const {
12688 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
12690 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
12692 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
12694 unsigned ByteSize = SrcVT.getSizeInBits()/8;
12696 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
12697 MachineMemOperand *MMO;
12699 int SSFI = FI->getIndex();
12701 DAG.getMachineFunction()
12702 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
12703 MachineMemOperand::MOLoad, ByteSize, ByteSize);
12705 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
12706 StackSlot = StackSlot.getOperand(1);
12708 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
12709 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
12711 Tys, Ops, SrcVT, MMO);
12714 Chain = Result.getValue(1);
12715 SDValue InFlag = Result.getValue(2);
12717 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
12718 // shouldn't be necessary except that RFP cannot be live across
12719 // multiple blocks. When stackifier is fixed, they can be uncoupled.
12720 MachineFunction &MF = DAG.getMachineFunction();
12721 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
12722 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
12723 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
12724 Tys = DAG.getVTList(MVT::Other);
12726 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
12728 MachineMemOperand *MMO =
12729 DAG.getMachineFunction()
12730 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
12731 MachineMemOperand::MOStore, SSFISize, SSFISize);
12733 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
12734 Ops, Op.getValueType(), MMO);
12735 Result = DAG.getLoad(Op.getValueType(), DL, Chain, StackSlot,
12736 MachinePointerInfo::getFixedStack(SSFI),
12737 false, false, false, 0);
12743 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
12744 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
12745 SelectionDAG &DAG) const {
12746 // This algorithm is not obvious. Here it is what we're trying to output:
12749 punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U }
12750 subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 }
12752 haddpd %xmm0, %xmm0
12754 pshufd $0x4e, %xmm0, %xmm1
12760 LLVMContext *Context = DAG.getContext();
12762 // Build some magic constants.
12763 static const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 };
12764 Constant *C0 = ConstantDataVector::get(*Context, CV0);
12765 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
12767 SmallVector<Constant*,2> CV1;
12769 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
12770 APInt(64, 0x4330000000000000ULL))));
12772 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
12773 APInt(64, 0x4530000000000000ULL))));
12774 Constant *C1 = ConstantVector::get(CV1);
12775 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
12777 // Load the 64-bit value into an XMM register.
12778 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
12780 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
12781 MachinePointerInfo::getConstantPool(),
12782 false, false, false, 16);
12783 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32,
12784 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, XR1),
12787 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
12788 MachinePointerInfo::getConstantPool(),
12789 false, false, false, 16);
12790 SDValue XR2F = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Unpck1);
12791 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
12794 if (Subtarget->hasSSE3()) {
12795 // FIXME: The 'haddpd' instruction may be slower than 'movhlps + addsd'.
12796 Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub);
12798 SDValue S2F = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Sub);
12799 SDValue Shuffle = getTargetShuffleNode(X86ISD::PSHUFD, dl, MVT::v4i32,
12801 Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64,
12802 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Shuffle),
12806 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result,
12807 DAG.getIntPtrConstant(0));
12810 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
12811 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
12812 SelectionDAG &DAG) const {
12814 // FP constant to bias correct the final result.
12815 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
12818 // Load the 32-bit value into an XMM register.
12819 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
12822 // Zero out the upper parts of the register.
12823 Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG);
12825 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12826 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Load),
12827 DAG.getIntPtrConstant(0));
12829 // Or the load with the bias.
12830 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
12831 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
12832 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
12833 MVT::v2f64, Load)),
12834 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
12835 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
12836 MVT::v2f64, Bias)));
12837 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12838 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or),
12839 DAG.getIntPtrConstant(0));
12841 // Subtract the bias.
12842 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
12844 // Handle final rounding.
12845 EVT DestVT = Op.getValueType();
12847 if (DestVT.bitsLT(MVT::f64))
12848 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
12849 DAG.getIntPtrConstant(0));
12850 if (DestVT.bitsGT(MVT::f64))
12851 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
12853 // Handle final rounding.
12857 SDValue X86TargetLowering::lowerUINT_TO_FP_vec(SDValue Op,
12858 SelectionDAG &DAG) const {
12859 SDValue N0 = Op.getOperand(0);
12860 MVT SVT = N0.getSimpleValueType();
12863 assert((SVT == MVT::v4i8 || SVT == MVT::v4i16 ||
12864 SVT == MVT::v8i8 || SVT == MVT::v8i16) &&
12865 "Custom UINT_TO_FP is not supported!");
12867 MVT NVT = MVT::getVectorVT(MVT::i32, SVT.getVectorNumElements());
12868 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
12869 DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0));
12872 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
12873 SelectionDAG &DAG) const {
12874 SDValue N0 = Op.getOperand(0);
12877 if (Op.getValueType().isVector())
12878 return lowerUINT_TO_FP_vec(Op, DAG);
12880 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
12881 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
12882 // the optimization here.
12883 if (DAG.SignBitIsZero(N0))
12884 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
12886 MVT SrcVT = N0.getSimpleValueType();
12887 MVT DstVT = Op.getSimpleValueType();
12888 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
12889 return LowerUINT_TO_FP_i64(Op, DAG);
12890 if (SrcVT == MVT::i32 && X86ScalarSSEf64)
12891 return LowerUINT_TO_FP_i32(Op, DAG);
12892 if (Subtarget->is64Bit() && SrcVT == MVT::i64 && DstVT == MVT::f32)
12895 // Make a 64-bit buffer, and use it to build an FILD.
12896 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
12897 if (SrcVT == MVT::i32) {
12898 SDValue WordOff = DAG.getConstant(4, getPointerTy());
12899 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
12900 getPointerTy(), StackSlot, WordOff);
12901 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
12902 StackSlot, MachinePointerInfo(),
12904 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
12905 OffsetSlot, MachinePointerInfo(),
12907 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
12911 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
12912 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
12913 StackSlot, MachinePointerInfo(),
12915 // For i64 source, we need to add the appropriate power of 2 if the input
12916 // was negative. This is the same as the optimization in
12917 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
12918 // we must be careful to do the computation in x87 extended precision, not
12919 // in SSE. (The generic code can't know it's OK to do this, or how to.)
12920 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
12921 MachineMemOperand *MMO =
12922 DAG.getMachineFunction()
12923 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
12924 MachineMemOperand::MOLoad, 8, 8);
12926 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
12927 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
12928 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops,
12931 APInt FF(32, 0x5F800000ULL);
12933 // Check whether the sign bit is set.
12934 SDValue SignSet = DAG.getSetCC(dl,
12935 getSetCCResultType(*DAG.getContext(), MVT::i64),
12936 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
12939 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
12940 SDValue FudgePtr = DAG.getConstantPool(
12941 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
12944 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
12945 SDValue Zero = DAG.getIntPtrConstant(0);
12946 SDValue Four = DAG.getIntPtrConstant(4);
12947 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
12949 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
12951 // Load the value out, extending it from f32 to f80.
12952 // FIXME: Avoid the extend by constructing the right constant pool?
12953 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
12954 FudgePtr, MachinePointerInfo::getConstantPool(),
12955 MVT::f32, false, false, false, 4);
12956 // Extend everything to 80 bits to force it to be done on x87.
12957 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
12958 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
12961 std::pair<SDValue,SDValue>
12962 X86TargetLowering:: FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
12963 bool IsSigned, bool IsReplace) const {
12966 EVT DstTy = Op.getValueType();
12968 if (!IsSigned && !isIntegerTypeFTOL(DstTy)) {
12969 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
12973 assert(DstTy.getSimpleVT() <= MVT::i64 &&
12974 DstTy.getSimpleVT() >= MVT::i16 &&
12975 "Unknown FP_TO_INT to lower!");
12977 // These are really Legal.
12978 if (DstTy == MVT::i32 &&
12979 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
12980 return std::make_pair(SDValue(), SDValue());
12981 if (Subtarget->is64Bit() &&
12982 DstTy == MVT::i64 &&
12983 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
12984 return std::make_pair(SDValue(), SDValue());
12986 // We lower FP->int64 either into FISTP64 followed by a load from a temporary
12987 // stack slot, or into the FTOL runtime function.
12988 MachineFunction &MF = DAG.getMachineFunction();
12989 unsigned MemSize = DstTy.getSizeInBits()/8;
12990 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
12991 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
12994 if (!IsSigned && isIntegerTypeFTOL(DstTy))
12995 Opc = X86ISD::WIN_FTOL;
12997 switch (DstTy.getSimpleVT().SimpleTy) {
12998 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
12999 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
13000 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
13001 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
13004 SDValue Chain = DAG.getEntryNode();
13005 SDValue Value = Op.getOperand(0);
13006 EVT TheVT = Op.getOperand(0).getValueType();
13007 // FIXME This causes a redundant load/store if the SSE-class value is already
13008 // in memory, such as if it is on the callstack.
13009 if (isScalarFPTypeInSSEReg(TheVT)) {
13010 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
13011 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
13012 MachinePointerInfo::getFixedStack(SSFI),
13014 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
13016 Chain, StackSlot, DAG.getValueType(TheVT)
13019 MachineMemOperand *MMO =
13020 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
13021 MachineMemOperand::MOLoad, MemSize, MemSize);
13022 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, DstTy, MMO);
13023 Chain = Value.getValue(1);
13024 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
13025 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
13028 MachineMemOperand *MMO =
13029 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
13030 MachineMemOperand::MOStore, MemSize, MemSize);
13032 if (Opc != X86ISD::WIN_FTOL) {
13033 // Build the FP_TO_INT*_IN_MEM
13034 SDValue Ops[] = { Chain, Value, StackSlot };
13035 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
13037 return std::make_pair(FIST, StackSlot);
13039 SDValue ftol = DAG.getNode(X86ISD::WIN_FTOL, DL,
13040 DAG.getVTList(MVT::Other, MVT::Glue),
13042 SDValue eax = DAG.getCopyFromReg(ftol, DL, X86::EAX,
13043 MVT::i32, ftol.getValue(1));
13044 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), DL, X86::EDX,
13045 MVT::i32, eax.getValue(2));
13046 SDValue Ops[] = { eax, edx };
13047 SDValue pair = IsReplace
13048 ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops)
13049 : DAG.getMergeValues(Ops, DL);
13050 return std::make_pair(pair, SDValue());
13054 static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
13055 const X86Subtarget *Subtarget) {
13056 MVT VT = Op->getSimpleValueType(0);
13057 SDValue In = Op->getOperand(0);
13058 MVT InVT = In.getSimpleValueType();
13061 // Optimize vectors in AVX mode:
13064 // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32.
13065 // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32.
13066 // Concat upper and lower parts.
13069 // Use vpunpckldq for 4 lower elements v4i32 -> v2i64.
13070 // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64.
13071 // Concat upper and lower parts.
13074 if (((VT != MVT::v16i16) || (InVT != MVT::v16i8)) &&
13075 ((VT != MVT::v8i32) || (InVT != MVT::v8i16)) &&
13076 ((VT != MVT::v4i64) || (InVT != MVT::v4i32)))
13079 if (Subtarget->hasInt256())
13080 return DAG.getNode(X86ISD::VZEXT, dl, VT, In);
13082 SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl);
13083 SDValue Undef = DAG.getUNDEF(InVT);
13084 bool NeedZero = Op.getOpcode() == ISD::ZERO_EXTEND;
13085 SDValue OpLo = getUnpackl(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
13086 SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
13088 MVT HVT = MVT::getVectorVT(VT.getVectorElementType(),
13089 VT.getVectorNumElements()/2);
13091 OpLo = DAG.getNode(ISD::BITCAST, dl, HVT, OpLo);
13092 OpHi = DAG.getNode(ISD::BITCAST, dl, HVT, OpHi);
13094 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
13097 static SDValue LowerZERO_EXTEND_AVX512(SDValue Op,
13098 SelectionDAG &DAG) {
13099 MVT VT = Op->getSimpleValueType(0);
13100 SDValue In = Op->getOperand(0);
13101 MVT InVT = In.getSimpleValueType();
13103 unsigned int NumElts = VT.getVectorNumElements();
13104 if (NumElts != 8 && NumElts != 16)
13107 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1)
13108 return DAG.getNode(X86ISD::VZEXT, DL, VT, In);
13110 EVT ExtVT = (NumElts == 8)? MVT::v8i64 : MVT::v16i32;
13111 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13112 // Now we have only mask extension
13113 assert(InVT.getVectorElementType() == MVT::i1);
13114 SDValue Cst = DAG.getTargetConstant(1, ExtVT.getScalarType());
13115 const Constant *C = (dyn_cast<ConstantSDNode>(Cst))->getConstantIntValue();
13116 SDValue CP = DAG.getConstantPool(C, TLI.getPointerTy());
13117 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
13118 SDValue Ld = DAG.getLoad(Cst.getValueType(), DL, DAG.getEntryNode(), CP,
13119 MachinePointerInfo::getConstantPool(),
13120 false, false, false, Alignment);
13122 SDValue Brcst = DAG.getNode(X86ISD::VBROADCASTM, DL, ExtVT, In, Ld);
13123 if (VT.is512BitVector())
13125 return DAG.getNode(X86ISD::VTRUNC, DL, VT, Brcst);
13128 static SDValue LowerANY_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
13129 SelectionDAG &DAG) {
13130 if (Subtarget->hasFp256()) {
13131 SDValue Res = LowerAVXExtend(Op, DAG, Subtarget);
13139 static SDValue LowerZERO_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
13140 SelectionDAG &DAG) {
13142 MVT VT = Op.getSimpleValueType();
13143 SDValue In = Op.getOperand(0);
13144 MVT SVT = In.getSimpleValueType();
13146 if (VT.is512BitVector() || SVT.getVectorElementType() == MVT::i1)
13147 return LowerZERO_EXTEND_AVX512(Op, DAG);
13149 if (Subtarget->hasFp256()) {
13150 SDValue Res = LowerAVXExtend(Op, DAG, Subtarget);
13155 assert(!VT.is256BitVector() || !SVT.is128BitVector() ||
13156 VT.getVectorNumElements() != SVT.getVectorNumElements());
13160 SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
13162 MVT VT = Op.getSimpleValueType();
13163 SDValue In = Op.getOperand(0);
13164 MVT InVT = In.getSimpleValueType();
13166 if (VT == MVT::i1) {
13167 assert((InVT.isInteger() && (InVT.getSizeInBits() <= 64)) &&
13168 "Invalid scalar TRUNCATE operation");
13169 if (InVT.getSizeInBits() >= 32)
13171 In = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, In);
13172 return DAG.getNode(ISD::TRUNCATE, DL, VT, In);
13174 assert(VT.getVectorNumElements() == InVT.getVectorNumElements() &&
13175 "Invalid TRUNCATE operation");
13177 if (InVT.is512BitVector() || VT.getVectorElementType() == MVT::i1) {
13178 if (VT.getVectorElementType().getSizeInBits() >=8)
13179 return DAG.getNode(X86ISD::VTRUNC, DL, VT, In);
13181 assert(VT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
13182 unsigned NumElts = InVT.getVectorNumElements();
13183 assert ((NumElts == 8 || NumElts == 16) && "Unexpected vector type");
13184 if (InVT.getSizeInBits() < 512) {
13185 MVT ExtVT = (NumElts == 16)? MVT::v16i32 : MVT::v8i64;
13186 In = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, In);
13190 SDValue Cst = DAG.getTargetConstant(1, InVT.getVectorElementType());
13191 const Constant *C = (dyn_cast<ConstantSDNode>(Cst))->getConstantIntValue();
13192 SDValue CP = DAG.getConstantPool(C, getPointerTy());
13193 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
13194 SDValue Ld = DAG.getLoad(Cst.getValueType(), DL, DAG.getEntryNode(), CP,
13195 MachinePointerInfo::getConstantPool(),
13196 false, false, false, Alignment);
13197 SDValue OneV = DAG.getNode(X86ISD::VBROADCAST, DL, InVT, Ld);
13198 SDValue And = DAG.getNode(ISD::AND, DL, InVT, OneV, In);
13199 return DAG.getNode(X86ISD::TESTM, DL, VT, And, And);
13202 if ((VT == MVT::v4i32) && (InVT == MVT::v4i64)) {
13203 // On AVX2, v4i64 -> v4i32 becomes VPERMD.
13204 if (Subtarget->hasInt256()) {
13205 static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1};
13206 In = DAG.getNode(ISD::BITCAST, DL, MVT::v8i32, In);
13207 In = DAG.getVectorShuffle(MVT::v8i32, DL, In, DAG.getUNDEF(MVT::v8i32),
13209 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, In,
13210 DAG.getIntPtrConstant(0));
13213 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13214 DAG.getIntPtrConstant(0));
13215 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13216 DAG.getIntPtrConstant(2));
13217 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpLo);
13218 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpHi);
13219 static const int ShufMask[] = {0, 2, 4, 6};
13220 return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask);
13223 if ((VT == MVT::v8i16) && (InVT == MVT::v8i32)) {
13224 // On AVX2, v8i32 -> v8i16 becomed PSHUFB.
13225 if (Subtarget->hasInt256()) {
13226 In = DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, In);
13228 SmallVector<SDValue,32> pshufbMask;
13229 for (unsigned i = 0; i < 2; ++i) {
13230 pshufbMask.push_back(DAG.getConstant(0x0, MVT::i8));
13231 pshufbMask.push_back(DAG.getConstant(0x1, MVT::i8));
13232 pshufbMask.push_back(DAG.getConstant(0x4, MVT::i8));
13233 pshufbMask.push_back(DAG.getConstant(0x5, MVT::i8));
13234 pshufbMask.push_back(DAG.getConstant(0x8, MVT::i8));
13235 pshufbMask.push_back(DAG.getConstant(0x9, MVT::i8));
13236 pshufbMask.push_back(DAG.getConstant(0xc, MVT::i8));
13237 pshufbMask.push_back(DAG.getConstant(0xd, MVT::i8));
13238 for (unsigned j = 0; j < 8; ++j)
13239 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
13241 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, pshufbMask);
13242 In = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8, In, BV);
13243 In = DAG.getNode(ISD::BITCAST, DL, MVT::v4i64, In);
13245 static const int ShufMask[] = {0, 2, -1, -1};
13246 In = DAG.getVectorShuffle(MVT::v4i64, DL, In, DAG.getUNDEF(MVT::v4i64),
13248 In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13249 DAG.getIntPtrConstant(0));
13250 return DAG.getNode(ISD::BITCAST, DL, VT, In);
13253 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
13254 DAG.getIntPtrConstant(0));
13256 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
13257 DAG.getIntPtrConstant(4));
13259 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, OpLo);
13260 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, OpHi);
13262 // The PSHUFB mask:
13263 static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13,
13264 -1, -1, -1, -1, -1, -1, -1, -1};
13266 SDValue Undef = DAG.getUNDEF(MVT::v16i8);
13267 OpLo = DAG.getVectorShuffle(MVT::v16i8, DL, OpLo, Undef, ShufMask1);
13268 OpHi = DAG.getVectorShuffle(MVT::v16i8, DL, OpHi, Undef, ShufMask1);
13270 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpLo);
13271 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpHi);
13273 // The MOVLHPS Mask:
13274 static const int ShufMask2[] = {0, 1, 4, 5};
13275 SDValue res = DAG.getVectorShuffle(MVT::v4i32, DL, OpLo, OpHi, ShufMask2);
13276 return DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, res);
13279 // Handle truncation of V256 to V128 using shuffles.
13280 if (!VT.is128BitVector() || !InVT.is256BitVector())
13283 assert(Subtarget->hasFp256() && "256-bit vector without AVX!");
13285 unsigned NumElems = VT.getVectorNumElements();
13286 MVT NVT = MVT::getVectorVT(VT.getVectorElementType(), NumElems * 2);
13288 SmallVector<int, 16> MaskVec(NumElems * 2, -1);
13289 // Prepare truncation shuffle mask
13290 for (unsigned i = 0; i != NumElems; ++i)
13291 MaskVec[i] = i * 2;
13292 SDValue V = DAG.getVectorShuffle(NVT, DL,
13293 DAG.getNode(ISD::BITCAST, DL, NVT, In),
13294 DAG.getUNDEF(NVT), &MaskVec[0]);
13295 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V,
13296 DAG.getIntPtrConstant(0));
13299 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
13300 SelectionDAG &DAG) const {
13301 assert(!Op.getSimpleValueType().isVector());
13303 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
13304 /*IsSigned=*/ true, /*IsReplace=*/ false);
13305 SDValue FIST = Vals.first, StackSlot = Vals.second;
13306 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
13307 if (!FIST.getNode()) return Op;
13309 if (StackSlot.getNode())
13310 // Load the result.
13311 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
13312 FIST, StackSlot, MachinePointerInfo(),
13313 false, false, false, 0);
13315 // The node is the result.
13319 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
13320 SelectionDAG &DAG) const {
13321 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
13322 /*IsSigned=*/ false, /*IsReplace=*/ false);
13323 SDValue FIST = Vals.first, StackSlot = Vals.second;
13324 assert(FIST.getNode() && "Unexpected failure");
13326 if (StackSlot.getNode())
13327 // Load the result.
13328 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
13329 FIST, StackSlot, MachinePointerInfo(),
13330 false, false, false, 0);
13332 // The node is the result.
13336 static SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) {
13338 MVT VT = Op.getSimpleValueType();
13339 SDValue In = Op.getOperand(0);
13340 MVT SVT = In.getSimpleValueType();
13342 assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!");
13344 return DAG.getNode(X86ISD::VFPEXT, DL, VT,
13345 DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32,
13346 In, DAG.getUNDEF(SVT)));
13349 // The only differences between FABS and FNEG are the mask and the logic op.
13350 static SDValue LowerFABSorFNEG(SDValue Op, SelectionDAG &DAG) {
13351 assert((Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG) &&
13352 "Wrong opcode for lowering FABS or FNEG.");
13354 bool IsFABS = (Op.getOpcode() == ISD::FABS);
13356 MVT VT = Op.getSimpleValueType();
13357 // Assume scalar op for initialization; update for vector if needed.
13358 // Note that there are no scalar bitwise logical SSE/AVX instructions, so we
13359 // generate a 16-byte vector constant and logic op even for the scalar case.
13360 // Using a 16-byte mask allows folding the load of the mask with
13361 // the logic op, so it can save (~4 bytes) on code size.
13363 unsigned NumElts = VT == MVT::f64 ? 2 : 4;
13364 // FIXME: Use function attribute "OptimizeForSize" and/or CodeGenOpt::Level to
13365 // decide if we should generate a 16-byte constant mask when we only need 4 or
13366 // 8 bytes for the scalar case.
13367 if (VT.isVector()) {
13368 EltVT = VT.getVectorElementType();
13369 NumElts = VT.getVectorNumElements();
13372 unsigned EltBits = EltVT.getSizeInBits();
13373 LLVMContext *Context = DAG.getContext();
13374 // For FABS, mask is 0x7f...; for FNEG, mask is 0x80...
13376 IsFABS ? APInt::getSignedMaxValue(EltBits) : APInt::getSignBit(EltBits);
13377 Constant *C = ConstantInt::get(*Context, MaskElt);
13378 C = ConstantVector::getSplat(NumElts, C);
13379 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13380 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy());
13381 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
13382 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
13383 MachinePointerInfo::getConstantPool(),
13384 false, false, false, Alignment);
13386 if (VT.isVector()) {
13387 // For a vector, cast operands to a vector type, perform the logic op,
13388 // and cast the result back to the original value type.
13389 MVT VecVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
13390 SDValue Op0Casted = DAG.getNode(ISD::BITCAST, dl, VecVT, Op.getOperand(0));
13391 SDValue MaskCasted = DAG.getNode(ISD::BITCAST, dl, VecVT, Mask);
13392 unsigned LogicOp = IsFABS ? ISD::AND : ISD::XOR;
13393 return DAG.getNode(ISD::BITCAST, dl, VT,
13394 DAG.getNode(LogicOp, dl, VecVT, Op0Casted, MaskCasted));
13396 // If not vector, then scalar.
13397 unsigned LogicOp = IsFABS ? X86ISD::FAND : X86ISD::FXOR;
13398 return DAG.getNode(LogicOp, dl, VT, Op.getOperand(0), Mask);
13401 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
13402 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13403 LLVMContext *Context = DAG.getContext();
13404 SDValue Op0 = Op.getOperand(0);
13405 SDValue Op1 = Op.getOperand(1);
13407 MVT VT = Op.getSimpleValueType();
13408 MVT SrcVT = Op1.getSimpleValueType();
13410 // If second operand is smaller, extend it first.
13411 if (SrcVT.bitsLT(VT)) {
13412 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
13415 // And if it is bigger, shrink it first.
13416 if (SrcVT.bitsGT(VT)) {
13417 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
13421 // At this point the operands and the result should have the same
13422 // type, and that won't be f80 since that is not custom lowered.
13424 // First get the sign bit of second operand.
13425 SmallVector<Constant*,4> CV;
13426 if (SrcVT == MVT::f64) {
13427 const fltSemantics &Sem = APFloat::IEEEdouble;
13428 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(64, 1ULL << 63))));
13429 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(64, 0))));
13431 const fltSemantics &Sem = APFloat::IEEEsingle;
13432 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 1U << 31))));
13433 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 0))));
13434 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 0))));
13435 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 0))));
13437 Constant *C = ConstantVector::get(CV);
13438 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(), 16);
13439 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
13440 MachinePointerInfo::getConstantPool(),
13441 false, false, false, 16);
13442 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
13444 // Shift sign bit right or left if the two operands have different types.
13445 if (SrcVT.bitsGT(VT)) {
13446 // Op0 is MVT::f32, Op1 is MVT::f64.
13447 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
13448 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
13449 DAG.getConstant(32, MVT::i32));
13450 SignBit = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, SignBit);
13451 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
13452 DAG.getIntPtrConstant(0));
13455 // Clear first operand sign bit.
13457 if (VT == MVT::f64) {
13458 const fltSemantics &Sem = APFloat::IEEEdouble;
13459 CV.push_back(ConstantFP::get(*Context, APFloat(Sem,
13460 APInt(64, ~(1ULL << 63)))));
13461 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(64, 0))));
13463 const fltSemantics &Sem = APFloat::IEEEsingle;
13464 CV.push_back(ConstantFP::get(*Context, APFloat(Sem,
13465 APInt(32, ~(1U << 31)))));
13466 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 0))));
13467 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 0))));
13468 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 0))));
13470 C = ConstantVector::get(CV);
13471 CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(), 16);
13472 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
13473 MachinePointerInfo::getConstantPool(),
13474 false, false, false, 16);
13475 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
13477 // Or the value with the sign bit.
13478 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
13481 static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) {
13482 SDValue N0 = Op.getOperand(0);
13484 MVT VT = Op.getSimpleValueType();
13486 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
13487 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
13488 DAG.getConstant(1, VT));
13489 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, VT));
13492 // LowerVectorAllZeroTest - Check whether an OR'd tree is PTEST-able.
13494 static SDValue LowerVectorAllZeroTest(SDValue Op, const X86Subtarget *Subtarget,
13495 SelectionDAG &DAG) {
13496 assert(Op.getOpcode() == ISD::OR && "Only check OR'd tree.");
13498 if (!Subtarget->hasSSE41())
13501 if (!Op->hasOneUse())
13504 SDNode *N = Op.getNode();
13507 SmallVector<SDValue, 8> Opnds;
13508 DenseMap<SDValue, unsigned> VecInMap;
13509 SmallVector<SDValue, 8> VecIns;
13510 EVT VT = MVT::Other;
13512 // Recognize a special case where a vector is casted into wide integer to
13514 Opnds.push_back(N->getOperand(0));
13515 Opnds.push_back(N->getOperand(1));
13517 for (unsigned Slot = 0, e = Opnds.size(); Slot < e; ++Slot) {
13518 SmallVectorImpl<SDValue>::const_iterator I = Opnds.begin() + Slot;
13519 // BFS traverse all OR'd operands.
13520 if (I->getOpcode() == ISD::OR) {
13521 Opnds.push_back(I->getOperand(0));
13522 Opnds.push_back(I->getOperand(1));
13523 // Re-evaluate the number of nodes to be traversed.
13524 e += 2; // 2 more nodes (LHS and RHS) are pushed.
13528 // Quit if a non-EXTRACT_VECTOR_ELT
13529 if (I->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
13532 // Quit if without a constant index.
13533 SDValue Idx = I->getOperand(1);
13534 if (!isa<ConstantSDNode>(Idx))
13537 SDValue ExtractedFromVec = I->getOperand(0);
13538 DenseMap<SDValue, unsigned>::iterator M = VecInMap.find(ExtractedFromVec);
13539 if (M == VecInMap.end()) {
13540 VT = ExtractedFromVec.getValueType();
13541 // Quit if not 128/256-bit vector.
13542 if (!VT.is128BitVector() && !VT.is256BitVector())
13544 // Quit if not the same type.
13545 if (VecInMap.begin() != VecInMap.end() &&
13546 VT != VecInMap.begin()->first.getValueType())
13548 M = VecInMap.insert(std::make_pair(ExtractedFromVec, 0)).first;
13549 VecIns.push_back(ExtractedFromVec);
13551 M->second |= 1U << cast<ConstantSDNode>(Idx)->getZExtValue();
13554 assert((VT.is128BitVector() || VT.is256BitVector()) &&
13555 "Not extracted from 128-/256-bit vector.");
13557 unsigned FullMask = (1U << VT.getVectorNumElements()) - 1U;
13559 for (DenseMap<SDValue, unsigned>::const_iterator
13560 I = VecInMap.begin(), E = VecInMap.end(); I != E; ++I) {
13561 // Quit if not all elements are used.
13562 if (I->second != FullMask)
13566 EVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
13568 // Cast all vectors into TestVT for PTEST.
13569 for (unsigned i = 0, e = VecIns.size(); i < e; ++i)
13570 VecIns[i] = DAG.getNode(ISD::BITCAST, DL, TestVT, VecIns[i]);
13572 // If more than one full vectors are evaluated, OR them first before PTEST.
13573 for (unsigned Slot = 0, e = VecIns.size(); e - Slot > 1; Slot += 2, e += 1) {
13574 // Each iteration will OR 2 nodes and append the result until there is only
13575 // 1 node left, i.e. the final OR'd value of all vectors.
13576 SDValue LHS = VecIns[Slot];
13577 SDValue RHS = VecIns[Slot + 1];
13578 VecIns.push_back(DAG.getNode(ISD::OR, DL, TestVT, LHS, RHS));
13581 return DAG.getNode(X86ISD::PTEST, DL, MVT::i32,
13582 VecIns.back(), VecIns.back());
13585 /// \brief return true if \c Op has a use that doesn't just read flags.
13586 static bool hasNonFlagsUse(SDValue Op) {
13587 for (SDNode::use_iterator UI = Op->use_begin(), UE = Op->use_end(); UI != UE;
13589 SDNode *User = *UI;
13590 unsigned UOpNo = UI.getOperandNo();
13591 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
13592 // Look pass truncate.
13593 UOpNo = User->use_begin().getOperandNo();
13594 User = *User->use_begin();
13597 if (User->getOpcode() != ISD::BRCOND && User->getOpcode() != ISD::SETCC &&
13598 !(User->getOpcode() == ISD::SELECT && UOpNo == 0))
13604 /// Emit nodes that will be selected as "test Op0,Op0", or something
13606 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC, SDLoc dl,
13607 SelectionDAG &DAG) const {
13608 if (Op.getValueType() == MVT::i1)
13609 // KORTEST instruction should be selected
13610 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
13611 DAG.getConstant(0, Op.getValueType()));
13613 // CF and OF aren't always set the way we want. Determine which
13614 // of these we need.
13615 bool NeedCF = false;
13616 bool NeedOF = false;
13619 case X86::COND_A: case X86::COND_AE:
13620 case X86::COND_B: case X86::COND_BE:
13623 case X86::COND_G: case X86::COND_GE:
13624 case X86::COND_L: case X86::COND_LE:
13625 case X86::COND_O: case X86::COND_NO: {
13626 // Check if we really need to set the
13627 // Overflow flag. If NoSignedWrap is present
13628 // that is not actually needed.
13629 switch (Op->getOpcode()) {
13634 const BinaryWithFlagsSDNode *BinNode =
13635 cast<BinaryWithFlagsSDNode>(Op.getNode());
13636 if (BinNode->hasNoSignedWrap())
13646 // See if we can use the EFLAGS value from the operand instead of
13647 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
13648 // we prove that the arithmetic won't overflow, we can't use OF or CF.
13649 if (Op.getResNo() != 0 || NeedOF || NeedCF) {
13650 // Emit a CMP with 0, which is the TEST pattern.
13651 //if (Op.getValueType() == MVT::i1)
13652 // return DAG.getNode(X86ISD::CMP, dl, MVT::i1, Op,
13653 // DAG.getConstant(0, MVT::i1));
13654 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
13655 DAG.getConstant(0, Op.getValueType()));
13657 unsigned Opcode = 0;
13658 unsigned NumOperands = 0;
13660 // Truncate operations may prevent the merge of the SETCC instruction
13661 // and the arithmetic instruction before it. Attempt to truncate the operands
13662 // of the arithmetic instruction and use a reduced bit-width instruction.
13663 bool NeedTruncation = false;
13664 SDValue ArithOp = Op;
13665 if (Op->getOpcode() == ISD::TRUNCATE && Op->hasOneUse()) {
13666 SDValue Arith = Op->getOperand(0);
13667 // Both the trunc and the arithmetic op need to have one user each.
13668 if (Arith->hasOneUse())
13669 switch (Arith.getOpcode()) {
13676 NeedTruncation = true;
13682 // NOTICE: In the code below we use ArithOp to hold the arithmetic operation
13683 // which may be the result of a CAST. We use the variable 'Op', which is the
13684 // non-casted variable when we check for possible users.
13685 switch (ArithOp.getOpcode()) {
13687 // Due to an isel shortcoming, be conservative if this add is likely to be
13688 // selected as part of a load-modify-store instruction. When the root node
13689 // in a match is a store, isel doesn't know how to remap non-chain non-flag
13690 // uses of other nodes in the match, such as the ADD in this case. This
13691 // leads to the ADD being left around and reselected, with the result being
13692 // two adds in the output. Alas, even if none our users are stores, that
13693 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
13694 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
13695 // climbing the DAG back to the root, and it doesn't seem to be worth the
13697 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13698 UE = Op.getNode()->use_end(); UI != UE; ++UI)
13699 if (UI->getOpcode() != ISD::CopyToReg &&
13700 UI->getOpcode() != ISD::SETCC &&
13701 UI->getOpcode() != ISD::STORE)
13704 if (ConstantSDNode *C =
13705 dyn_cast<ConstantSDNode>(ArithOp.getNode()->getOperand(1))) {
13706 // An add of one will be selected as an INC.
13707 if (C->getAPIntValue() == 1 && !Subtarget->slowIncDec()) {
13708 Opcode = X86ISD::INC;
13713 // An add of negative one (subtract of one) will be selected as a DEC.
13714 if (C->getAPIntValue().isAllOnesValue() && !Subtarget->slowIncDec()) {
13715 Opcode = X86ISD::DEC;
13721 // Otherwise use a regular EFLAGS-setting add.
13722 Opcode = X86ISD::ADD;
13727 // If we have a constant logical shift that's only used in a comparison
13728 // against zero turn it into an equivalent AND. This allows turning it into
13729 // a TEST instruction later.
13730 if ((X86CC == X86::COND_E || X86CC == X86::COND_NE) && Op->hasOneUse() &&
13731 isa<ConstantSDNode>(Op->getOperand(1)) && !hasNonFlagsUse(Op)) {
13732 EVT VT = Op.getValueType();
13733 unsigned BitWidth = VT.getSizeInBits();
13734 unsigned ShAmt = Op->getConstantOperandVal(1);
13735 if (ShAmt >= BitWidth) // Avoid undefined shifts.
13737 APInt Mask = ArithOp.getOpcode() == ISD::SRL
13738 ? APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt)
13739 : APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt);
13740 if (!Mask.isSignedIntN(32)) // Avoid large immediates.
13742 SDValue New = DAG.getNode(ISD::AND, dl, VT, Op->getOperand(0),
13743 DAG.getConstant(Mask, VT));
13744 DAG.ReplaceAllUsesWith(Op, New);
13750 // If the primary and result isn't used, don't bother using X86ISD::AND,
13751 // because a TEST instruction will be better.
13752 if (!hasNonFlagsUse(Op))
13758 // Due to the ISEL shortcoming noted above, be conservative if this op is
13759 // likely to be selected as part of a load-modify-store instruction.
13760 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13761 UE = Op.getNode()->use_end(); UI != UE; ++UI)
13762 if (UI->getOpcode() == ISD::STORE)
13765 // Otherwise use a regular EFLAGS-setting instruction.
13766 switch (ArithOp.getOpcode()) {
13767 default: llvm_unreachable("unexpected operator!");
13768 case ISD::SUB: Opcode = X86ISD::SUB; break;
13769 case ISD::XOR: Opcode = X86ISD::XOR; break;
13770 case ISD::AND: Opcode = X86ISD::AND; break;
13772 if (!NeedTruncation && (X86CC == X86::COND_E || X86CC == X86::COND_NE)) {
13773 SDValue EFLAGS = LowerVectorAllZeroTest(Op, Subtarget, DAG);
13774 if (EFLAGS.getNode())
13777 Opcode = X86ISD::OR;
13791 return SDValue(Op.getNode(), 1);
13797 // If we found that truncation is beneficial, perform the truncation and
13799 if (NeedTruncation) {
13800 EVT VT = Op.getValueType();
13801 SDValue WideVal = Op->getOperand(0);
13802 EVT WideVT = WideVal.getValueType();
13803 unsigned ConvertedOp = 0;
13804 // Use a target machine opcode to prevent further DAGCombine
13805 // optimizations that may separate the arithmetic operations
13806 // from the setcc node.
13807 switch (WideVal.getOpcode()) {
13809 case ISD::ADD: ConvertedOp = X86ISD::ADD; break;
13810 case ISD::SUB: ConvertedOp = X86ISD::SUB; break;
13811 case ISD::AND: ConvertedOp = X86ISD::AND; break;
13812 case ISD::OR: ConvertedOp = X86ISD::OR; break;
13813 case ISD::XOR: ConvertedOp = X86ISD::XOR; break;
13817 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13818 if (TLI.isOperationLegal(WideVal.getOpcode(), WideVT)) {
13819 SDValue V0 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(0));
13820 SDValue V1 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(1));
13821 Op = DAG.getNode(ConvertedOp, dl, VT, V0, V1);
13827 // Emit a CMP with 0, which is the TEST pattern.
13828 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
13829 DAG.getConstant(0, Op.getValueType()));
13831 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
13832 SmallVector<SDValue, 4> Ops;
13833 for (unsigned i = 0; i != NumOperands; ++i)
13834 Ops.push_back(Op.getOperand(i));
13836 SDValue New = DAG.getNode(Opcode, dl, VTs, Ops);
13837 DAG.ReplaceAllUsesWith(Op, New);
13838 return SDValue(New.getNode(), 1);
13841 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
13843 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
13844 SDLoc dl, SelectionDAG &DAG) const {
13845 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1)) {
13846 if (C->getAPIntValue() == 0)
13847 return EmitTest(Op0, X86CC, dl, DAG);
13849 if (Op0.getValueType() == MVT::i1)
13850 llvm_unreachable("Unexpected comparison operation for MVT::i1 operands");
13853 if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 ||
13854 Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) {
13855 // Do the comparison at i32 if it's smaller, besides the Atom case.
13856 // This avoids subregister aliasing issues. Keep the smaller reference
13857 // if we're optimizing for size, however, as that'll allow better folding
13858 // of memory operations.
13859 if (Op0.getValueType() != MVT::i32 && Op0.getValueType() != MVT::i64 &&
13860 !DAG.getMachineFunction().getFunction()->getAttributes().hasAttribute(
13861 AttributeSet::FunctionIndex, Attribute::MinSize) &&
13862 !Subtarget->isAtom()) {
13863 unsigned ExtendOp =
13864 isX86CCUnsigned(X86CC) ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
13865 Op0 = DAG.getNode(ExtendOp, dl, MVT::i32, Op0);
13866 Op1 = DAG.getNode(ExtendOp, dl, MVT::i32, Op1);
13868 // Use SUB instead of CMP to enable CSE between SUB and CMP.
13869 SDVTList VTs = DAG.getVTList(Op0.getValueType(), MVT::i32);
13870 SDValue Sub = DAG.getNode(X86ISD::SUB, dl, VTs,
13872 return SDValue(Sub.getNode(), 1);
13874 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
13877 /// Convert a comparison if required by the subtarget.
13878 SDValue X86TargetLowering::ConvertCmpIfNecessary(SDValue Cmp,
13879 SelectionDAG &DAG) const {
13880 // If the subtarget does not support the FUCOMI instruction, floating-point
13881 // comparisons have to be converted.
13882 if (Subtarget->hasCMov() ||
13883 Cmp.getOpcode() != X86ISD::CMP ||
13884 !Cmp.getOperand(0).getValueType().isFloatingPoint() ||
13885 !Cmp.getOperand(1).getValueType().isFloatingPoint())
13888 // The instruction selector will select an FUCOM instruction instead of
13889 // FUCOMI, which writes the comparison result to FPSW instead of EFLAGS. Hence
13890 // build an SDNode sequence that transfers the result from FPSW into EFLAGS:
13891 // (X86sahf (trunc (srl (X86fp_stsw (trunc (X86cmp ...)), 8))))
13893 SDValue TruncFPSW = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, Cmp);
13894 SDValue FNStSW = DAG.getNode(X86ISD::FNSTSW16r, dl, MVT::i16, TruncFPSW);
13895 SDValue Srl = DAG.getNode(ISD::SRL, dl, MVT::i16, FNStSW,
13896 DAG.getConstant(8, MVT::i8));
13897 SDValue TruncSrl = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Srl);
13898 return DAG.getNode(X86ISD::SAHF, dl, MVT::i32, TruncSrl);
13901 static bool isAllOnes(SDValue V) {
13902 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
13903 return C && C->isAllOnesValue();
13906 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
13907 /// if it's possible.
13908 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
13909 SDLoc dl, SelectionDAG &DAG) const {
13910 SDValue Op0 = And.getOperand(0);
13911 SDValue Op1 = And.getOperand(1);
13912 if (Op0.getOpcode() == ISD::TRUNCATE)
13913 Op0 = Op0.getOperand(0);
13914 if (Op1.getOpcode() == ISD::TRUNCATE)
13915 Op1 = Op1.getOperand(0);
13918 if (Op1.getOpcode() == ISD::SHL)
13919 std::swap(Op0, Op1);
13920 if (Op0.getOpcode() == ISD::SHL) {
13921 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
13922 if (And00C->getZExtValue() == 1) {
13923 // If we looked past a truncate, check that it's only truncating away
13925 unsigned BitWidth = Op0.getValueSizeInBits();
13926 unsigned AndBitWidth = And.getValueSizeInBits();
13927 if (BitWidth > AndBitWidth) {
13929 DAG.computeKnownBits(Op0, Zeros, Ones);
13930 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
13934 RHS = Op0.getOperand(1);
13936 } else if (Op1.getOpcode() == ISD::Constant) {
13937 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
13938 uint64_t AndRHSVal = AndRHS->getZExtValue();
13939 SDValue AndLHS = Op0;
13941 if (AndRHSVal == 1 && AndLHS.getOpcode() == ISD::SRL) {
13942 LHS = AndLHS.getOperand(0);
13943 RHS = AndLHS.getOperand(1);
13946 // Use BT if the immediate can't be encoded in a TEST instruction.
13947 if (!isUInt<32>(AndRHSVal) && isPowerOf2_64(AndRHSVal)) {
13949 RHS = DAG.getConstant(Log2_64_Ceil(AndRHSVal), LHS.getValueType());
13953 if (LHS.getNode()) {
13954 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
13955 // instruction. Since the shift amount is in-range-or-undefined, we know
13956 // that doing a bittest on the i32 value is ok. We extend to i32 because
13957 // the encoding for the i16 version is larger than the i32 version.
13958 // Also promote i16 to i32 for performance / code size reason.
13959 if (LHS.getValueType() == MVT::i8 ||
13960 LHS.getValueType() == MVT::i16)
13961 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
13963 // If the operand types disagree, extend the shift amount to match. Since
13964 // BT ignores high bits (like shifts) we can use anyextend.
13965 if (LHS.getValueType() != RHS.getValueType())
13966 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
13968 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
13969 X86::CondCode Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
13970 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
13971 DAG.getConstant(Cond, MVT::i8), BT);
13977 /// \brief - Turns an ISD::CondCode into a value suitable for SSE floating point
13979 static int translateX86FSETCC(ISD::CondCode SetCCOpcode, SDValue &Op0,
13984 // SSE Condition code mapping:
13993 switch (SetCCOpcode) {
13994 default: llvm_unreachable("Unexpected SETCC condition");
13996 case ISD::SETEQ: SSECC = 0; break;
13998 case ISD::SETGT: Swap = true; // Fallthrough
14000 case ISD::SETOLT: SSECC = 1; break;
14002 case ISD::SETGE: Swap = true; // Fallthrough
14004 case ISD::SETOLE: SSECC = 2; break;
14005 case ISD::SETUO: SSECC = 3; break;
14007 case ISD::SETNE: SSECC = 4; break;
14008 case ISD::SETULE: Swap = true; // Fallthrough
14009 case ISD::SETUGE: SSECC = 5; break;
14010 case ISD::SETULT: Swap = true; // Fallthrough
14011 case ISD::SETUGT: SSECC = 6; break;
14012 case ISD::SETO: SSECC = 7; break;
14014 case ISD::SETONE: SSECC = 8; break;
14017 std::swap(Op0, Op1);
14022 // Lower256IntVSETCC - Break a VSETCC 256-bit integer VSETCC into two new 128
14023 // ones, and then concatenate the result back.
14024 static SDValue Lower256IntVSETCC(SDValue Op, SelectionDAG &DAG) {
14025 MVT VT = Op.getSimpleValueType();
14027 assert(VT.is256BitVector() && Op.getOpcode() == ISD::SETCC &&
14028 "Unsupported value type for operation");
14030 unsigned NumElems = VT.getVectorNumElements();
14032 SDValue CC = Op.getOperand(2);
14034 // Extract the LHS vectors
14035 SDValue LHS = Op.getOperand(0);
14036 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
14037 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
14039 // Extract the RHS vectors
14040 SDValue RHS = Op.getOperand(1);
14041 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
14042 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
14044 // Issue the operation on the smaller types and concatenate the result back
14045 MVT EltVT = VT.getVectorElementType();
14046 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
14047 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
14048 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC),
14049 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
14052 static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG,
14053 const X86Subtarget *Subtarget) {
14054 SDValue Op0 = Op.getOperand(0);
14055 SDValue Op1 = Op.getOperand(1);
14056 SDValue CC = Op.getOperand(2);
14057 MVT VT = Op.getSimpleValueType();
14060 assert(Op0.getValueType().getVectorElementType().getSizeInBits() >= 8 &&
14061 Op.getValueType().getScalarType() == MVT::i1 &&
14062 "Cannot set masked compare for this operation");
14064 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14066 bool Unsigned = false;
14069 switch (SetCCOpcode) {
14070 default: llvm_unreachable("Unexpected SETCC condition");
14071 case ISD::SETNE: SSECC = 4; break;
14072 case ISD::SETEQ: Opc = X86ISD::PCMPEQM; break;
14073 case ISD::SETUGT: SSECC = 6; Unsigned = true; break;
14074 case ISD::SETLT: Swap = true; //fall-through
14075 case ISD::SETGT: Opc = X86ISD::PCMPGTM; break;
14076 case ISD::SETULT: SSECC = 1; Unsigned = true; break;
14077 case ISD::SETUGE: SSECC = 5; Unsigned = true; break; //NLT
14078 case ISD::SETGE: Swap = true; SSECC = 2; break; // LE + swap
14079 case ISD::SETULE: Unsigned = true; //fall-through
14080 case ISD::SETLE: SSECC = 2; break;
14084 std::swap(Op0, Op1);
14086 return DAG.getNode(Opc, dl, VT, Op0, Op1);
14087 Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM;
14088 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14089 DAG.getConstant(SSECC, MVT::i8));
14092 /// \brief Try to turn a VSETULT into a VSETULE by modifying its second
14093 /// operand \p Op1. If non-trivial (for example because it's not constant)
14094 /// return an empty value.
14095 static SDValue ChangeVSETULTtoVSETULE(SDLoc dl, SDValue Op1, SelectionDAG &DAG)
14097 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Op1.getNode());
14101 MVT VT = Op1.getSimpleValueType();
14102 MVT EVT = VT.getVectorElementType();
14103 unsigned n = VT.getVectorNumElements();
14104 SmallVector<SDValue, 8> ULTOp1;
14106 for (unsigned i = 0; i < n; ++i) {
14107 ConstantSDNode *Elt = dyn_cast<ConstantSDNode>(BV->getOperand(i));
14108 if (!Elt || Elt->isOpaque() || Elt->getValueType(0) != EVT)
14111 // Avoid underflow.
14112 APInt Val = Elt->getAPIntValue();
14116 ULTOp1.push_back(DAG.getConstant(Val - 1, EVT));
14119 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, ULTOp1);
14122 static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
14123 SelectionDAG &DAG) {
14124 SDValue Op0 = Op.getOperand(0);
14125 SDValue Op1 = Op.getOperand(1);
14126 SDValue CC = Op.getOperand(2);
14127 MVT VT = Op.getSimpleValueType();
14128 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14129 bool isFP = Op.getOperand(1).getSimpleValueType().isFloatingPoint();
14134 MVT EltVT = Op0.getSimpleValueType().getVectorElementType();
14135 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
14138 unsigned SSECC = translateX86FSETCC(SetCCOpcode, Op0, Op1);
14139 unsigned Opc = X86ISD::CMPP;
14140 if (Subtarget->hasAVX512() && VT.getVectorElementType() == MVT::i1) {
14141 assert(VT.getVectorNumElements() <= 16);
14142 Opc = X86ISD::CMPM;
14144 // In the two special cases we can't handle, emit two comparisons.
14147 unsigned CombineOpc;
14148 if (SetCCOpcode == ISD::SETUEQ) {
14149 CC0 = 3; CC1 = 0; CombineOpc = ISD::OR;
14151 assert(SetCCOpcode == ISD::SETONE);
14152 CC0 = 7; CC1 = 4; CombineOpc = ISD::AND;
14155 SDValue Cmp0 = DAG.getNode(Opc, dl, VT, Op0, Op1,
14156 DAG.getConstant(CC0, MVT::i8));
14157 SDValue Cmp1 = DAG.getNode(Opc, dl, VT, Op0, Op1,
14158 DAG.getConstant(CC1, MVT::i8));
14159 return DAG.getNode(CombineOpc, dl, VT, Cmp0, Cmp1);
14161 // Handle all other FP comparisons here.
14162 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14163 DAG.getConstant(SSECC, MVT::i8));
14166 // Break 256-bit integer vector compare into smaller ones.
14167 if (VT.is256BitVector() && !Subtarget->hasInt256())
14168 return Lower256IntVSETCC(Op, DAG);
14170 bool MaskResult = (VT.getVectorElementType() == MVT::i1);
14171 EVT OpVT = Op1.getValueType();
14172 if (Subtarget->hasAVX512()) {
14173 if (Op1.getValueType().is512BitVector() ||
14174 (Subtarget->hasBWI() && Subtarget->hasVLX()) ||
14175 (MaskResult && OpVT.getVectorElementType().getSizeInBits() >= 32))
14176 return LowerIntVSETCC_AVX512(Op, DAG, Subtarget);
14178 // In AVX-512 architecture setcc returns mask with i1 elements,
14179 // But there is no compare instruction for i8 and i16 elements in KNL.
14180 // We are not talking about 512-bit operands in this case, these
14181 // types are illegal.
14183 (OpVT.getVectorElementType().getSizeInBits() < 32 &&
14184 OpVT.getVectorElementType().getSizeInBits() >= 8))
14185 return DAG.getNode(ISD::TRUNCATE, dl, VT,
14186 DAG.getNode(ISD::SETCC, dl, OpVT, Op0, Op1, CC));
14189 // We are handling one of the integer comparisons here. Since SSE only has
14190 // GT and EQ comparisons for integer, swapping operands and multiple
14191 // operations may be required for some comparisons.
14193 bool Swap = false, Invert = false, FlipSigns = false, MinMax = false;
14194 bool Subus = false;
14196 switch (SetCCOpcode) {
14197 default: llvm_unreachable("Unexpected SETCC condition");
14198 case ISD::SETNE: Invert = true;
14199 case ISD::SETEQ: Opc = X86ISD::PCMPEQ; break;
14200 case ISD::SETLT: Swap = true;
14201 case ISD::SETGT: Opc = X86ISD::PCMPGT; break;
14202 case ISD::SETGE: Swap = true;
14203 case ISD::SETLE: Opc = X86ISD::PCMPGT;
14204 Invert = true; break;
14205 case ISD::SETULT: Swap = true;
14206 case ISD::SETUGT: Opc = X86ISD::PCMPGT;
14207 FlipSigns = true; break;
14208 case ISD::SETUGE: Swap = true;
14209 case ISD::SETULE: Opc = X86ISD::PCMPGT;
14210 FlipSigns = true; Invert = true; break;
14213 // Special case: Use min/max operations for SETULE/SETUGE
14214 MVT VET = VT.getVectorElementType();
14216 (Subtarget->hasSSE41() && (VET >= MVT::i8 && VET <= MVT::i32))
14217 || (Subtarget->hasSSE2() && (VET == MVT::i8));
14220 switch (SetCCOpcode) {
14222 case ISD::SETULE: Opc = X86ISD::UMIN; MinMax = true; break;
14223 case ISD::SETUGE: Opc = X86ISD::UMAX; MinMax = true; break;
14226 if (MinMax) { Swap = false; Invert = false; FlipSigns = false; }
14229 bool hasSubus = Subtarget->hasSSE2() && (VET == MVT::i8 || VET == MVT::i16);
14230 if (!MinMax && hasSubus) {
14231 // As another special case, use PSUBUS[BW] when it's profitable. E.g. for
14233 // t = psubus Op0, Op1
14234 // pcmpeq t, <0..0>
14235 switch (SetCCOpcode) {
14237 case ISD::SETULT: {
14238 // If the comparison is against a constant we can turn this into a
14239 // setule. With psubus, setule does not require a swap. This is
14240 // beneficial because the constant in the register is no longer
14241 // destructed as the destination so it can be hoisted out of a loop.
14242 // Only do this pre-AVX since vpcmp* is no longer destructive.
14243 if (Subtarget->hasAVX())
14245 SDValue ULEOp1 = ChangeVSETULTtoVSETULE(dl, Op1, DAG);
14246 if (ULEOp1.getNode()) {
14248 Subus = true; Invert = false; Swap = false;
14252 // Psubus is better than flip-sign because it requires no inversion.
14253 case ISD::SETUGE: Subus = true; Invert = false; Swap = true; break;
14254 case ISD::SETULE: Subus = true; Invert = false; Swap = false; break;
14258 Opc = X86ISD::SUBUS;
14264 std::swap(Op0, Op1);
14266 // Check that the operation in question is available (most are plain SSE2,
14267 // but PCMPGTQ and PCMPEQQ have different requirements).
14268 if (VT == MVT::v2i64) {
14269 if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42()) {
14270 assert(Subtarget->hasSSE2() && "Don't know how to lower!");
14272 // First cast everything to the right type.
14273 Op0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op0);
14274 Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op1);
14276 // Since SSE has no unsigned integer comparisons, we need to flip the sign
14277 // bits of the inputs before performing those operations. The lower
14278 // compare is always unsigned.
14281 SB = DAG.getConstant(0x80000000U, MVT::v4i32);
14283 SDValue Sign = DAG.getConstant(0x80000000U, MVT::i32);
14284 SDValue Zero = DAG.getConstant(0x00000000U, MVT::i32);
14285 SB = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
14286 Sign, Zero, Sign, Zero);
14288 Op0 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op0, SB);
14289 Op1 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op1, SB);
14291 // Emulate PCMPGTQ with (hi1 > hi2) | ((hi1 == hi2) & (lo1 > lo2))
14292 SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1);
14293 SDValue EQ = DAG.getNode(X86ISD::PCMPEQ, dl, MVT::v4i32, Op0, Op1);
14295 // Create masks for only the low parts/high parts of the 64 bit integers.
14296 static const int MaskHi[] = { 1, 1, 3, 3 };
14297 static const int MaskLo[] = { 0, 0, 2, 2 };
14298 SDValue EQHi = DAG.getVectorShuffle(MVT::v4i32, dl, EQ, EQ, MaskHi);
14299 SDValue GTLo = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskLo);
14300 SDValue GTHi = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi);
14302 SDValue Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, EQHi, GTLo);
14303 Result = DAG.getNode(ISD::OR, dl, MVT::v4i32, Result, GTHi);
14306 Result = DAG.getNOT(dl, Result, MVT::v4i32);
14308 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
14311 if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) {
14312 // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with
14313 // pcmpeqd + pshufd + pand.
14314 assert(Subtarget->hasSSE2() && !FlipSigns && "Don't know how to lower!");
14316 // First cast everything to the right type.
14317 Op0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op0);
14318 Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op1);
14321 SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1);
14323 // Make sure the lower and upper halves are both all-ones.
14324 static const int Mask[] = { 1, 0, 3, 2 };
14325 SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask);
14326 Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf);
14329 Result = DAG.getNOT(dl, Result, MVT::v4i32);
14331 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
14335 // Since SSE has no unsigned integer comparisons, we need to flip the sign
14336 // bits of the inputs before performing those operations.
14338 EVT EltVT = VT.getVectorElementType();
14339 SDValue SB = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()), VT);
14340 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SB);
14341 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SB);
14344 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
14346 // If the logical-not of the result is required, perform that now.
14348 Result = DAG.getNOT(dl, Result, VT);
14351 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Op0, Result);
14354 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Result,
14355 getZeroVector(VT, Subtarget, DAG, dl));
14360 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
14362 MVT VT = Op.getSimpleValueType();
14364 if (VT.isVector()) return LowerVSETCC(Op, Subtarget, DAG);
14366 assert(((!Subtarget->hasAVX512() && VT == MVT::i8) || (VT == MVT::i1))
14367 && "SetCC type must be 8-bit or 1-bit integer");
14368 SDValue Op0 = Op.getOperand(0);
14369 SDValue Op1 = Op.getOperand(1);
14371 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
14373 // Optimize to BT if possible.
14374 // Lower (X & (1 << N)) == 0 to BT(X, N).
14375 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
14376 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
14377 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
14378 Op1.getOpcode() == ISD::Constant &&
14379 cast<ConstantSDNode>(Op1)->isNullValue() &&
14380 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14381 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
14382 if (NewSetCC.getNode())
14386 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
14388 if (Op1.getOpcode() == ISD::Constant &&
14389 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
14390 cast<ConstantSDNode>(Op1)->isNullValue()) &&
14391 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14393 // If the input is a setcc, then reuse the input setcc or use a new one with
14394 // the inverted condition.
14395 if (Op0.getOpcode() == X86ISD::SETCC) {
14396 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
14397 bool Invert = (CC == ISD::SETNE) ^
14398 cast<ConstantSDNode>(Op1)->isNullValue();
14402 CCode = X86::GetOppositeBranchCondition(CCode);
14403 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14404 DAG.getConstant(CCode, MVT::i8),
14405 Op0.getOperand(1));
14407 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
14411 if ((Op0.getValueType() == MVT::i1) && (Op1.getOpcode() == ISD::Constant) &&
14412 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1) &&
14413 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14415 ISD::CondCode NewCC = ISD::getSetCCInverse(CC, true);
14416 return DAG.getSetCC(dl, VT, Op0, DAG.getConstant(0, MVT::i1), NewCC);
14419 bool isFP = Op1.getSimpleValueType().isFloatingPoint();
14420 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
14421 if (X86CC == X86::COND_INVALID)
14424 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, dl, DAG);
14425 EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
14426 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14427 DAG.getConstant(X86CC, MVT::i8), EFLAGS);
14429 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
14433 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
14434 static bool isX86LogicalCmp(SDValue Op) {
14435 unsigned Opc = Op.getNode()->getOpcode();
14436 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI ||
14437 Opc == X86ISD::SAHF)
14439 if (Op.getResNo() == 1 &&
14440 (Opc == X86ISD::ADD ||
14441 Opc == X86ISD::SUB ||
14442 Opc == X86ISD::ADC ||
14443 Opc == X86ISD::SBB ||
14444 Opc == X86ISD::SMUL ||
14445 Opc == X86ISD::UMUL ||
14446 Opc == X86ISD::INC ||
14447 Opc == X86ISD::DEC ||
14448 Opc == X86ISD::OR ||
14449 Opc == X86ISD::XOR ||
14450 Opc == X86ISD::AND))
14453 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
14459 static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) {
14460 if (V.getOpcode() != ISD::TRUNCATE)
14463 SDValue VOp0 = V.getOperand(0);
14464 unsigned InBits = VOp0.getValueSizeInBits();
14465 unsigned Bits = V.getValueSizeInBits();
14466 return DAG.MaskedValueIsZero(VOp0, APInt::getHighBitsSet(InBits,InBits-Bits));
14469 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
14470 bool addTest = true;
14471 SDValue Cond = Op.getOperand(0);
14472 SDValue Op1 = Op.getOperand(1);
14473 SDValue Op2 = Op.getOperand(2);
14475 EVT VT = Op1.getValueType();
14478 // Lower fp selects into a CMP/AND/ANDN/OR sequence when the necessary SSE ops
14479 // are available. Otherwise fp cmovs get lowered into a less efficient branch
14480 // sequence later on.
14481 if (Cond.getOpcode() == ISD::SETCC &&
14482 ((Subtarget->hasSSE2() && (VT == MVT::f32 || VT == MVT::f64)) ||
14483 (Subtarget->hasSSE1() && VT == MVT::f32)) &&
14484 VT == Cond.getOperand(0).getValueType() && Cond->hasOneUse()) {
14485 SDValue CondOp0 = Cond.getOperand(0), CondOp1 = Cond.getOperand(1);
14486 int SSECC = translateX86FSETCC(
14487 cast<CondCodeSDNode>(Cond.getOperand(2))->get(), CondOp0, CondOp1);
14490 if (Subtarget->hasAVX512()) {
14491 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CondOp0, CondOp1,
14492 DAG.getConstant(SSECC, MVT::i8));
14493 return DAG.getNode(X86ISD::SELECT, DL, VT, Cmp, Op1, Op2);
14495 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, VT, CondOp0, CondOp1,
14496 DAG.getConstant(SSECC, MVT::i8));
14497 SDValue AndN = DAG.getNode(X86ISD::FANDN, DL, VT, Cmp, Op2);
14498 SDValue And = DAG.getNode(X86ISD::FAND, DL, VT, Cmp, Op1);
14499 return DAG.getNode(X86ISD::FOR, DL, VT, AndN, And);
14503 if (Cond.getOpcode() == ISD::SETCC) {
14504 SDValue NewCond = LowerSETCC(Cond, DAG);
14505 if (NewCond.getNode())
14509 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
14510 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
14511 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
14512 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
14513 if (Cond.getOpcode() == X86ISD::SETCC &&
14514 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
14515 isZero(Cond.getOperand(1).getOperand(1))) {
14516 SDValue Cmp = Cond.getOperand(1);
14518 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
14520 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
14521 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
14522 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
14524 SDValue CmpOp0 = Cmp.getOperand(0);
14525 // Apply further optimizations for special cases
14526 // (select (x != 0), -1, 0) -> neg & sbb
14527 // (select (x == 0), 0, -1) -> neg & sbb
14528 if (ConstantSDNode *YC = dyn_cast<ConstantSDNode>(Y))
14529 if (YC->isNullValue() &&
14530 (isAllOnes(Op1) == (CondCode == X86::COND_NE))) {
14531 SDVTList VTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32);
14532 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, VTs,
14533 DAG.getConstant(0, CmpOp0.getValueType()),
14535 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14536 DAG.getConstant(X86::COND_B, MVT::i8),
14537 SDValue(Neg.getNode(), 1));
14541 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
14542 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
14543 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
14545 SDValue Res = // Res = 0 or -1.
14546 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14547 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
14549 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
14550 Res = DAG.getNOT(DL, Res, Res.getValueType());
14552 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
14553 if (!N2C || !N2C->isNullValue())
14554 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
14559 // Look past (and (setcc_carry (cmp ...)), 1).
14560 if (Cond.getOpcode() == ISD::AND &&
14561 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
14562 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
14563 if (C && C->getAPIntValue() == 1)
14564 Cond = Cond.getOperand(0);
14567 // If condition flag is set by a X86ISD::CMP, then use it as the condition
14568 // setting operand in place of the X86ISD::SETCC.
14569 unsigned CondOpcode = Cond.getOpcode();
14570 if (CondOpcode == X86ISD::SETCC ||
14571 CondOpcode == X86ISD::SETCC_CARRY) {
14572 CC = Cond.getOperand(0);
14574 SDValue Cmp = Cond.getOperand(1);
14575 unsigned Opc = Cmp.getOpcode();
14576 MVT VT = Op.getSimpleValueType();
14578 bool IllegalFPCMov = false;
14579 if (VT.isFloatingPoint() && !VT.isVector() &&
14580 !isScalarFPTypeInSSEReg(VT)) // FPStack?
14581 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
14583 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
14584 Opc == X86ISD::BT) { // FIXME
14588 } else if (CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
14589 CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
14590 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
14591 Cond.getOperand(0).getValueType() != MVT::i8)) {
14592 SDValue LHS = Cond.getOperand(0);
14593 SDValue RHS = Cond.getOperand(1);
14594 unsigned X86Opcode;
14597 switch (CondOpcode) {
14598 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
14599 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
14600 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
14601 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
14602 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
14603 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
14604 default: llvm_unreachable("unexpected overflowing operator");
14606 if (CondOpcode == ISD::UMULO)
14607 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
14610 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
14612 SDValue X86Op = DAG.getNode(X86Opcode, DL, VTs, LHS, RHS);
14614 if (CondOpcode == ISD::UMULO)
14615 Cond = X86Op.getValue(2);
14617 Cond = X86Op.getValue(1);
14619 CC = DAG.getConstant(X86Cond, MVT::i8);
14624 // Look pass the truncate if the high bits are known zero.
14625 if (isTruncWithZeroHighBitsInput(Cond, DAG))
14626 Cond = Cond.getOperand(0);
14628 // We know the result of AND is compared against zero. Try to match
14630 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
14631 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
14632 if (NewSetCC.getNode()) {
14633 CC = NewSetCC.getOperand(0);
14634 Cond = NewSetCC.getOperand(1);
14641 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
14642 Cond = EmitTest(Cond, X86::COND_NE, DL, DAG);
14645 // a < b ? -1 : 0 -> RES = ~setcc_carry
14646 // a < b ? 0 : -1 -> RES = setcc_carry
14647 // a >= b ? -1 : 0 -> RES = setcc_carry
14648 // a >= b ? 0 : -1 -> RES = ~setcc_carry
14649 if (Cond.getOpcode() == X86ISD::SUB) {
14650 Cond = ConvertCmpIfNecessary(Cond, DAG);
14651 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
14653 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
14654 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
14655 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14656 DAG.getConstant(X86::COND_B, MVT::i8), Cond);
14657 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
14658 return DAG.getNOT(DL, Res, Res.getValueType());
14663 // X86 doesn't have an i8 cmov. If both operands are the result of a truncate
14664 // widen the cmov and push the truncate through. This avoids introducing a new
14665 // branch during isel and doesn't add any extensions.
14666 if (Op.getValueType() == MVT::i8 &&
14667 Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) {
14668 SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0);
14669 if (T1.getValueType() == T2.getValueType() &&
14670 // Blacklist CopyFromReg to avoid partial register stalls.
14671 T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){
14672 SDVTList VTs = DAG.getVTList(T1.getValueType(), MVT::Glue);
14673 SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, VTs, T2, T1, CC, Cond);
14674 return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov);
14678 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
14679 // condition is true.
14680 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
14681 SDValue Ops[] = { Op2, Op1, CC, Cond };
14682 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops);
14685 static SDValue LowerSIGN_EXTEND_AVX512(SDValue Op, SelectionDAG &DAG) {
14686 MVT VT = Op->getSimpleValueType(0);
14687 SDValue In = Op->getOperand(0);
14688 MVT InVT = In.getSimpleValueType();
14691 unsigned int NumElts = VT.getVectorNumElements();
14692 if (NumElts != 8 && NumElts != 16)
14695 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1)
14696 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14698 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
14699 assert (InVT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
14701 MVT ExtVT = (NumElts == 8) ? MVT::v8i64 : MVT::v16i32;
14702 Constant *C = ConstantInt::get(*DAG.getContext(),
14703 APInt::getAllOnesValue(ExtVT.getScalarType().getSizeInBits()));
14705 SDValue CP = DAG.getConstantPool(C, TLI.getPointerTy());
14706 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
14707 SDValue Ld = DAG.getLoad(ExtVT.getScalarType(), dl, DAG.getEntryNode(), CP,
14708 MachinePointerInfo::getConstantPool(),
14709 false, false, false, Alignment);
14710 SDValue Brcst = DAG.getNode(X86ISD::VBROADCASTM, dl, ExtVT, In, Ld);
14711 if (VT.is512BitVector())
14713 return DAG.getNode(X86ISD::VTRUNC, dl, VT, Brcst);
14716 static SDValue LowerSIGN_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
14717 SelectionDAG &DAG) {
14718 MVT VT = Op->getSimpleValueType(0);
14719 SDValue In = Op->getOperand(0);
14720 MVT InVT = In.getSimpleValueType();
14723 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
14724 return LowerSIGN_EXTEND_AVX512(Op, DAG);
14726 if ((VT != MVT::v4i64 || InVT != MVT::v4i32) &&
14727 (VT != MVT::v8i32 || InVT != MVT::v8i16) &&
14728 (VT != MVT::v16i16 || InVT != MVT::v16i8))
14731 if (Subtarget->hasInt256())
14732 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14734 // Optimize vectors in AVX mode
14735 // Sign extend v8i16 to v8i32 and
14738 // Divide input vector into two parts
14739 // for v4i32 the shuffle mask will be { 0, 1, -1, -1} {2, 3, -1, -1}
14740 // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32
14741 // concat the vectors to original VT
14743 unsigned NumElems = InVT.getVectorNumElements();
14744 SDValue Undef = DAG.getUNDEF(InVT);
14746 SmallVector<int,8> ShufMask1(NumElems, -1);
14747 for (unsigned i = 0; i != NumElems/2; ++i)
14750 SDValue OpLo = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask1[0]);
14752 SmallVector<int,8> ShufMask2(NumElems, -1);
14753 for (unsigned i = 0; i != NumElems/2; ++i)
14754 ShufMask2[i] = i + NumElems/2;
14756 SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask2[0]);
14758 MVT HalfVT = MVT::getVectorVT(VT.getScalarType(),
14759 VT.getVectorNumElements()/2);
14761 OpLo = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpLo);
14762 OpHi = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpHi);
14764 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
14767 // Lower vector extended loads using a shuffle. If SSSE3 is not available we
14768 // may emit an illegal shuffle but the expansion is still better than scalar
14769 // code. We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise
14770 // we'll emit a shuffle and a arithmetic shift.
14771 // TODO: It is possible to support ZExt by zeroing the undef values during
14772 // the shuffle phase or after the shuffle.
14773 static SDValue LowerExtendedLoad(SDValue Op, const X86Subtarget *Subtarget,
14774 SelectionDAG &DAG) {
14775 MVT RegVT = Op.getSimpleValueType();
14776 assert(RegVT.isVector() && "We only custom lower vector sext loads.");
14777 assert(RegVT.isInteger() &&
14778 "We only custom lower integer vector sext loads.");
14780 // Nothing useful we can do without SSE2 shuffles.
14781 assert(Subtarget->hasSSE2() && "We only custom lower sext loads with SSE2.");
14783 LoadSDNode *Ld = cast<LoadSDNode>(Op.getNode());
14785 EVT MemVT = Ld->getMemoryVT();
14786 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
14787 unsigned RegSz = RegVT.getSizeInBits();
14789 ISD::LoadExtType Ext = Ld->getExtensionType();
14791 assert((Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)
14792 && "Only anyext and sext are currently implemented.");
14793 assert(MemVT != RegVT && "Cannot extend to the same type");
14794 assert(MemVT.isVector() && "Must load a vector from memory");
14796 unsigned NumElems = RegVT.getVectorNumElements();
14797 unsigned MemSz = MemVT.getSizeInBits();
14798 assert(RegSz > MemSz && "Register size must be greater than the mem size");
14800 if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256()) {
14801 // The only way in which we have a legal 256-bit vector result but not the
14802 // integer 256-bit operations needed to directly lower a sextload is if we
14803 // have AVX1 but not AVX2. In that case, we can always emit a sextload to
14804 // a 128-bit vector and a normal sign_extend to 256-bits that should get
14805 // correctly legalized. We do this late to allow the canonical form of
14806 // sextload to persist throughout the rest of the DAG combiner -- it wants
14807 // to fold together any extensions it can, and so will fuse a sign_extend
14808 // of an sextload into a sextload targeting a wider value.
14810 if (MemSz == 128) {
14811 // Just switch this to a normal load.
14812 assert(TLI.isTypeLegal(MemVT) && "If the memory type is a 128-bit type, "
14813 "it must be a legal 128-bit vector "
14815 Load = DAG.getLoad(MemVT, dl, Ld->getChain(), Ld->getBasePtr(),
14816 Ld->getPointerInfo(), Ld->isVolatile(), Ld->isNonTemporal(),
14817 Ld->isInvariant(), Ld->getAlignment());
14819 assert(MemSz < 128 &&
14820 "Can't extend a type wider than 128 bits to a 256 bit vector!");
14821 // Do an sext load to a 128-bit vector type. We want to use the same
14822 // number of elements, but elements half as wide. This will end up being
14823 // recursively lowered by this routine, but will succeed as we definitely
14824 // have all the necessary features if we're using AVX1.
14826 EVT::getIntegerVT(*DAG.getContext(), RegVT.getScalarSizeInBits() / 2);
14827 EVT HalfVecVT = EVT::getVectorVT(*DAG.getContext(), HalfEltVT, NumElems);
14829 DAG.getExtLoad(Ext, dl, HalfVecVT, Ld->getChain(), Ld->getBasePtr(),
14830 Ld->getPointerInfo(), MemVT, Ld->isVolatile(),
14831 Ld->isNonTemporal(), Ld->isInvariant(),
14832 Ld->getAlignment());
14835 // Replace chain users with the new chain.
14836 assert(Load->getNumValues() == 2 && "Loads must carry a chain!");
14837 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Load.getValue(1));
14839 // Finally, do a normal sign-extend to the desired register.
14840 return DAG.getSExtOrTrunc(Load, dl, RegVT);
14843 // All sizes must be a power of two.
14844 assert(isPowerOf2_32(RegSz * MemSz * NumElems) &&
14845 "Non-power-of-two elements are not custom lowered!");
14847 // Attempt to load the original value using scalar loads.
14848 // Find the largest scalar type that divides the total loaded size.
14849 MVT SclrLoadTy = MVT::i8;
14850 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
14851 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
14852 MVT Tp = (MVT::SimpleValueType)tp;
14853 if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
14858 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
14859 if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
14861 SclrLoadTy = MVT::f64;
14863 // Calculate the number of scalar loads that we need to perform
14864 // in order to load our vector from memory.
14865 unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
14867 assert((Ext != ISD::SEXTLOAD || NumLoads == 1) &&
14868 "Can only lower sext loads with a single scalar load!");
14870 unsigned loadRegZize = RegSz;
14871 if (Ext == ISD::SEXTLOAD && RegSz == 256)
14874 // Represent our vector as a sequence of elements which are the
14875 // largest scalar that we can load.
14876 EVT LoadUnitVecVT = EVT::getVectorVT(
14877 *DAG.getContext(), SclrLoadTy, loadRegZize / SclrLoadTy.getSizeInBits());
14879 // Represent the data using the same element type that is stored in
14880 // memory. In practice, we ''widen'' MemVT.
14882 EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
14883 loadRegZize / MemVT.getScalarType().getSizeInBits());
14885 assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
14886 "Invalid vector type");
14888 // We can't shuffle using an illegal type.
14889 assert(TLI.isTypeLegal(WideVecVT) &&
14890 "We only lower types that form legal widened vector types");
14892 SmallVector<SDValue, 8> Chains;
14893 SDValue Ptr = Ld->getBasePtr();
14894 SDValue Increment =
14895 DAG.getConstant(SclrLoadTy.getSizeInBits() / 8, TLI.getPointerTy());
14896 SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
14898 for (unsigned i = 0; i < NumLoads; ++i) {
14899 // Perform a single load.
14900 SDValue ScalarLoad =
14901 DAG.getLoad(SclrLoadTy, dl, Ld->getChain(), Ptr, Ld->getPointerInfo(),
14902 Ld->isVolatile(), Ld->isNonTemporal(), Ld->isInvariant(),
14903 Ld->getAlignment());
14904 Chains.push_back(ScalarLoad.getValue(1));
14905 // Create the first element type using SCALAR_TO_VECTOR in order to avoid
14906 // another round of DAGCombining.
14908 Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
14910 Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
14911 ScalarLoad, DAG.getIntPtrConstant(i));
14913 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
14916 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
14918 // Bitcast the loaded value to a vector of the original element type, in
14919 // the size of the target vector type.
14920 SDValue SlicedVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Res);
14921 unsigned SizeRatio = RegSz / MemSz;
14923 if (Ext == ISD::SEXTLOAD) {
14924 // If we have SSE4.1, we can directly emit a VSEXT node.
14925 if (Subtarget->hasSSE41()) {
14926 SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
14927 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
14931 // Otherwise we'll shuffle the small elements in the high bits of the
14932 // larger type and perform an arithmetic shift. If the shift is not legal
14933 // it's better to scalarize.
14934 assert(TLI.isOperationLegalOrCustom(ISD::SRA, RegVT) &&
14935 "We can't implement a sext load without an arithmetic right shift!");
14937 // Redistribute the loaded elements into the different locations.
14938 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
14939 for (unsigned i = 0; i != NumElems; ++i)
14940 ShuffleVec[i * SizeRatio + SizeRatio - 1] = i;
14942 SDValue Shuff = DAG.getVectorShuffle(
14943 WideVecVT, dl, SlicedVec, DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
14945 Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
14947 // Build the arithmetic shift.
14948 unsigned Amt = RegVT.getVectorElementType().getSizeInBits() -
14949 MemVT.getVectorElementType().getSizeInBits();
14951 DAG.getNode(ISD::SRA, dl, RegVT, Shuff, DAG.getConstant(Amt, RegVT));
14953 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
14957 // Redistribute the loaded elements into the different locations.
14958 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
14959 for (unsigned i = 0; i != NumElems; ++i)
14960 ShuffleVec[i * SizeRatio] = i;
14962 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
14963 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
14965 // Bitcast to the requested type.
14966 Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
14967 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
14971 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
14972 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
14973 // from the AND / OR.
14974 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
14975 Opc = Op.getOpcode();
14976 if (Opc != ISD::OR && Opc != ISD::AND)
14978 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
14979 Op.getOperand(0).hasOneUse() &&
14980 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
14981 Op.getOperand(1).hasOneUse());
14984 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
14985 // 1 and that the SETCC node has a single use.
14986 static bool isXor1OfSetCC(SDValue Op) {
14987 if (Op.getOpcode() != ISD::XOR)
14989 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
14990 if (N1C && N1C->getAPIntValue() == 1) {
14991 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
14992 Op.getOperand(0).hasOneUse();
14997 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
14998 bool addTest = true;
14999 SDValue Chain = Op.getOperand(0);
15000 SDValue Cond = Op.getOperand(1);
15001 SDValue Dest = Op.getOperand(2);
15004 bool Inverted = false;
15006 if (Cond.getOpcode() == ISD::SETCC) {
15007 // Check for setcc([su]{add,sub,mul}o == 0).
15008 if (cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETEQ &&
15009 isa<ConstantSDNode>(Cond.getOperand(1)) &&
15010 cast<ConstantSDNode>(Cond.getOperand(1))->isNullValue() &&
15011 Cond.getOperand(0).getResNo() == 1 &&
15012 (Cond.getOperand(0).getOpcode() == ISD::SADDO ||
15013 Cond.getOperand(0).getOpcode() == ISD::UADDO ||
15014 Cond.getOperand(0).getOpcode() == ISD::SSUBO ||
15015 Cond.getOperand(0).getOpcode() == ISD::USUBO ||
15016 Cond.getOperand(0).getOpcode() == ISD::SMULO ||
15017 Cond.getOperand(0).getOpcode() == ISD::UMULO)) {
15019 Cond = Cond.getOperand(0);
15021 SDValue NewCond = LowerSETCC(Cond, DAG);
15022 if (NewCond.getNode())
15027 // FIXME: LowerXALUO doesn't handle these!!
15028 else if (Cond.getOpcode() == X86ISD::ADD ||
15029 Cond.getOpcode() == X86ISD::SUB ||
15030 Cond.getOpcode() == X86ISD::SMUL ||
15031 Cond.getOpcode() == X86ISD::UMUL)
15032 Cond = LowerXALUO(Cond, DAG);
15035 // Look pass (and (setcc_carry (cmp ...)), 1).
15036 if (Cond.getOpcode() == ISD::AND &&
15037 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
15038 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
15039 if (C && C->getAPIntValue() == 1)
15040 Cond = Cond.getOperand(0);
15043 // If condition flag is set by a X86ISD::CMP, then use it as the condition
15044 // setting operand in place of the X86ISD::SETCC.
15045 unsigned CondOpcode = Cond.getOpcode();
15046 if (CondOpcode == X86ISD::SETCC ||
15047 CondOpcode == X86ISD::SETCC_CARRY) {
15048 CC = Cond.getOperand(0);
15050 SDValue Cmp = Cond.getOperand(1);
15051 unsigned Opc = Cmp.getOpcode();
15052 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
15053 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
15057 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
15061 // These can only come from an arithmetic instruction with overflow,
15062 // e.g. SADDO, UADDO.
15063 Cond = Cond.getNode()->getOperand(1);
15069 CondOpcode = Cond.getOpcode();
15070 if (CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
15071 CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
15072 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
15073 Cond.getOperand(0).getValueType() != MVT::i8)) {
15074 SDValue LHS = Cond.getOperand(0);
15075 SDValue RHS = Cond.getOperand(1);
15076 unsigned X86Opcode;
15079 // Keep this in sync with LowerXALUO, otherwise we might create redundant
15080 // instructions that can't be removed afterwards (i.e. X86ISD::ADD and
15082 switch (CondOpcode) {
15083 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
15085 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
15087 X86Opcode = X86ISD::INC; X86Cond = X86::COND_O;
15090 X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
15091 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
15093 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
15095 X86Opcode = X86ISD::DEC; X86Cond = X86::COND_O;
15098 X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
15099 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
15100 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
15101 default: llvm_unreachable("unexpected overflowing operator");
15104 X86Cond = X86::GetOppositeBranchCondition((X86::CondCode)X86Cond);
15105 if (CondOpcode == ISD::UMULO)
15106 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
15109 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
15111 SDValue X86Op = DAG.getNode(X86Opcode, dl, VTs, LHS, RHS);
15113 if (CondOpcode == ISD::UMULO)
15114 Cond = X86Op.getValue(2);
15116 Cond = X86Op.getValue(1);
15118 CC = DAG.getConstant(X86Cond, MVT::i8);
15122 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
15123 SDValue Cmp = Cond.getOperand(0).getOperand(1);
15124 if (CondOpc == ISD::OR) {
15125 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
15126 // two branches instead of an explicit OR instruction with a
15128 if (Cmp == Cond.getOperand(1).getOperand(1) &&
15129 isX86LogicalCmp(Cmp)) {
15130 CC = Cond.getOperand(0).getOperand(0);
15131 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15132 Chain, Dest, CC, Cmp);
15133 CC = Cond.getOperand(1).getOperand(0);
15137 } else { // ISD::AND
15138 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
15139 // two branches instead of an explicit AND instruction with a
15140 // separate test. However, we only do this if this block doesn't
15141 // have a fall-through edge, because this requires an explicit
15142 // jmp when the condition is false.
15143 if (Cmp == Cond.getOperand(1).getOperand(1) &&
15144 isX86LogicalCmp(Cmp) &&
15145 Op.getNode()->hasOneUse()) {
15146 X86::CondCode CCode =
15147 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
15148 CCode = X86::GetOppositeBranchCondition(CCode);
15149 CC = DAG.getConstant(CCode, MVT::i8);
15150 SDNode *User = *Op.getNode()->use_begin();
15151 // Look for an unconditional branch following this conditional branch.
15152 // We need this because we need to reverse the successors in order
15153 // to implement FCMP_OEQ.
15154 if (User->getOpcode() == ISD::BR) {
15155 SDValue FalseBB = User->getOperand(1);
15157 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15158 assert(NewBR == User);
15162 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15163 Chain, Dest, CC, Cmp);
15164 X86::CondCode CCode =
15165 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
15166 CCode = X86::GetOppositeBranchCondition(CCode);
15167 CC = DAG.getConstant(CCode, MVT::i8);
15173 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
15174 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
15175 // It should be transformed during dag combiner except when the condition
15176 // is set by a arithmetics with overflow node.
15177 X86::CondCode CCode =
15178 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
15179 CCode = X86::GetOppositeBranchCondition(CCode);
15180 CC = DAG.getConstant(CCode, MVT::i8);
15181 Cond = Cond.getOperand(0).getOperand(1);
15183 } else if (Cond.getOpcode() == ISD::SETCC &&
15184 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETOEQ) {
15185 // For FCMP_OEQ, we can emit
15186 // two branches instead of an explicit AND instruction with a
15187 // separate test. However, we only do this if this block doesn't
15188 // have a fall-through edge, because this requires an explicit
15189 // jmp when the condition is false.
15190 if (Op.getNode()->hasOneUse()) {
15191 SDNode *User = *Op.getNode()->use_begin();
15192 // Look for an unconditional branch following this conditional branch.
15193 // We need this because we need to reverse the successors in order
15194 // to implement FCMP_OEQ.
15195 if (User->getOpcode() == ISD::BR) {
15196 SDValue FalseBB = User->getOperand(1);
15198 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15199 assert(NewBR == User);
15203 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
15204 Cond.getOperand(0), Cond.getOperand(1));
15205 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
15206 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
15207 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15208 Chain, Dest, CC, Cmp);
15209 CC = DAG.getConstant(X86::COND_P, MVT::i8);
15214 } else if (Cond.getOpcode() == ISD::SETCC &&
15215 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETUNE) {
15216 // For FCMP_UNE, we can emit
15217 // two branches instead of an explicit AND instruction with a
15218 // separate test. However, we only do this if this block doesn't
15219 // have a fall-through edge, because this requires an explicit
15220 // jmp when the condition is false.
15221 if (Op.getNode()->hasOneUse()) {
15222 SDNode *User = *Op.getNode()->use_begin();
15223 // Look for an unconditional branch following this conditional branch.
15224 // We need this because we need to reverse the successors in order
15225 // to implement FCMP_UNE.
15226 if (User->getOpcode() == ISD::BR) {
15227 SDValue FalseBB = User->getOperand(1);
15229 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15230 assert(NewBR == User);
15233 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
15234 Cond.getOperand(0), Cond.getOperand(1));
15235 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
15236 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
15237 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15238 Chain, Dest, CC, Cmp);
15239 CC = DAG.getConstant(X86::COND_NP, MVT::i8);
15249 // Look pass the truncate if the high bits are known zero.
15250 if (isTruncWithZeroHighBitsInput(Cond, DAG))
15251 Cond = Cond.getOperand(0);
15253 // We know the result of AND is compared against zero. Try to match
15255 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
15256 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
15257 if (NewSetCC.getNode()) {
15258 CC = NewSetCC.getOperand(0);
15259 Cond = NewSetCC.getOperand(1);
15266 X86::CondCode X86Cond = Inverted ? X86::COND_E : X86::COND_NE;
15267 CC = DAG.getConstant(X86Cond, MVT::i8);
15268 Cond = EmitTest(Cond, X86Cond, dl, DAG);
15270 Cond = ConvertCmpIfNecessary(Cond, DAG);
15271 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15272 Chain, Dest, CC, Cond);
15275 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
15276 // Calls to _alloca are needed to probe the stack when allocating more than 4k
15277 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
15278 // that the guard pages used by the OS virtual memory manager are allocated in
15279 // correct sequence.
15281 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
15282 SelectionDAG &DAG) const {
15283 MachineFunction &MF = DAG.getMachineFunction();
15284 bool SplitStack = MF.shouldSplitStack();
15285 bool Lower = (Subtarget->isOSWindows() && !Subtarget->isTargetMacho()) ||
15290 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15291 SDNode* Node = Op.getNode();
15293 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
15294 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
15295 " not tell us which reg is the stack pointer!");
15296 EVT VT = Node->getValueType(0);
15297 SDValue Tmp1 = SDValue(Node, 0);
15298 SDValue Tmp2 = SDValue(Node, 1);
15299 SDValue Tmp3 = Node->getOperand(2);
15300 SDValue Chain = Tmp1.getOperand(0);
15302 // Chain the dynamic stack allocation so that it doesn't modify the stack
15303 // pointer when other instructions are using the stack.
15304 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true),
15307 SDValue Size = Tmp2.getOperand(1);
15308 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
15309 Chain = SP.getValue(1);
15310 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
15311 const TargetFrameLowering &TFI = *DAG.getSubtarget().getFrameLowering();
15312 unsigned StackAlign = TFI.getStackAlignment();
15313 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
15314 if (Align > StackAlign)
15315 Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1,
15316 DAG.getConstant(-(uint64_t)Align, VT));
15317 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
15319 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
15320 DAG.getIntPtrConstant(0, true), SDValue(),
15323 SDValue Ops[2] = { Tmp1, Tmp2 };
15324 return DAG.getMergeValues(Ops, dl);
15328 SDValue Chain = Op.getOperand(0);
15329 SDValue Size = Op.getOperand(1);
15330 unsigned Align = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
15331 EVT VT = Op.getNode()->getValueType(0);
15333 bool Is64Bit = Subtarget->is64Bit();
15334 EVT SPTy = getPointerTy();
15337 MachineRegisterInfo &MRI = MF.getRegInfo();
15340 // The 64 bit implementation of segmented stacks needs to clobber both r10
15341 // r11. This makes it impossible to use it along with nested parameters.
15342 const Function *F = MF.getFunction();
15344 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
15346 if (I->hasNestAttr())
15347 report_fatal_error("Cannot use segmented stacks with functions that "
15348 "have nested arguments.");
15351 const TargetRegisterClass *AddrRegClass =
15352 getRegClassFor(getPointerTy());
15353 unsigned Vreg = MRI.createVirtualRegister(AddrRegClass);
15354 Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size);
15355 SDValue Value = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain,
15356 DAG.getRegister(Vreg, SPTy));
15357 SDValue Ops1[2] = { Value, Chain };
15358 return DAG.getMergeValues(Ops1, dl);
15361 const unsigned Reg = (Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX);
15363 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
15364 Flag = Chain.getValue(1);
15365 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
15367 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
15369 const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
15370 DAG.getSubtarget().getRegisterInfo());
15371 unsigned SPReg = RegInfo->getStackRegister();
15372 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, SPTy);
15373 Chain = SP.getValue(1);
15376 SP = DAG.getNode(ISD::AND, dl, VT, SP.getValue(0),
15377 DAG.getConstant(-(uint64_t)Align, VT));
15378 Chain = DAG.getCopyToReg(Chain, dl, SPReg, SP);
15381 SDValue Ops1[2] = { SP, Chain };
15382 return DAG.getMergeValues(Ops1, dl);
15386 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
15387 MachineFunction &MF = DAG.getMachineFunction();
15388 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
15390 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
15393 if (!Subtarget->is64Bit() || Subtarget->isTargetWin64()) {
15394 // vastart just stores the address of the VarArgsFrameIndex slot into the
15395 // memory location argument.
15396 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
15398 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
15399 MachinePointerInfo(SV), false, false, 0);
15403 // gp_offset (0 - 6 * 8)
15404 // fp_offset (48 - 48 + 8 * 16)
15405 // overflow_arg_area (point to parameters coming in memory).
15407 SmallVector<SDValue, 8> MemOps;
15408 SDValue FIN = Op.getOperand(1);
15410 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
15411 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
15413 FIN, MachinePointerInfo(SV), false, false, 0);
15414 MemOps.push_back(Store);
15417 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
15418 FIN, DAG.getIntPtrConstant(4));
15419 Store = DAG.getStore(Op.getOperand(0), DL,
15420 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
15422 FIN, MachinePointerInfo(SV, 4), false, false, 0);
15423 MemOps.push_back(Store);
15425 // Store ptr to overflow_arg_area
15426 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
15427 FIN, DAG.getIntPtrConstant(4));
15428 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
15430 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
15431 MachinePointerInfo(SV, 8),
15433 MemOps.push_back(Store);
15435 // Store ptr to reg_save_area.
15436 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
15437 FIN, DAG.getIntPtrConstant(8));
15438 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
15440 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN,
15441 MachinePointerInfo(SV, 16), false, false, 0);
15442 MemOps.push_back(Store);
15443 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
15446 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
15447 assert(Subtarget->is64Bit() &&
15448 "LowerVAARG only handles 64-bit va_arg!");
15449 assert((Subtarget->isTargetLinux() ||
15450 Subtarget->isTargetDarwin()) &&
15451 "Unhandled target in LowerVAARG");
15452 assert(Op.getNode()->getNumOperands() == 4);
15453 SDValue Chain = Op.getOperand(0);
15454 SDValue SrcPtr = Op.getOperand(1);
15455 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
15456 unsigned Align = Op.getConstantOperandVal(3);
15459 EVT ArgVT = Op.getNode()->getValueType(0);
15460 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
15461 uint32_t ArgSize = getDataLayout()->getTypeAllocSize(ArgTy);
15464 // Decide which area this value should be read from.
15465 // TODO: Implement the AMD64 ABI in its entirety. This simple
15466 // selection mechanism works only for the basic types.
15467 if (ArgVT == MVT::f80) {
15468 llvm_unreachable("va_arg for f80 not yet implemented");
15469 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
15470 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
15471 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
15472 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
15474 llvm_unreachable("Unhandled argument type in LowerVAARG");
15477 if (ArgMode == 2) {
15478 // Sanity Check: Make sure using fp_offset makes sense.
15479 assert(!DAG.getTarget().Options.UseSoftFloat &&
15480 !(DAG.getMachineFunction()
15481 .getFunction()->getAttributes()
15482 .hasAttribute(AttributeSet::FunctionIndex,
15483 Attribute::NoImplicitFloat)) &&
15484 Subtarget->hasSSE1());
15487 // Insert VAARG_64 node into the DAG
15488 // VAARG_64 returns two values: Variable Argument Address, Chain
15489 SmallVector<SDValue, 11> InstOps;
15490 InstOps.push_back(Chain);
15491 InstOps.push_back(SrcPtr);
15492 InstOps.push_back(DAG.getConstant(ArgSize, MVT::i32));
15493 InstOps.push_back(DAG.getConstant(ArgMode, MVT::i8));
15494 InstOps.push_back(DAG.getConstant(Align, MVT::i32));
15495 SDVTList VTs = DAG.getVTList(getPointerTy(), MVT::Other);
15496 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
15497 VTs, InstOps, MVT::i64,
15498 MachinePointerInfo(SV),
15500 /*Volatile=*/false,
15502 /*WriteMem=*/true);
15503 Chain = VAARG.getValue(1);
15505 // Load the next argument and return it
15506 return DAG.getLoad(ArgVT, dl,
15509 MachinePointerInfo(),
15510 false, false, false, 0);
15513 static SDValue LowerVACOPY(SDValue Op, const X86Subtarget *Subtarget,
15514 SelectionDAG &DAG) {
15515 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
15516 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
15517 SDValue Chain = Op.getOperand(0);
15518 SDValue DstPtr = Op.getOperand(1);
15519 SDValue SrcPtr = Op.getOperand(2);
15520 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
15521 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
15524 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
15525 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
15527 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
15530 // getTargetVShiftByConstNode - Handle vector element shifts where the shift
15531 // amount is a constant. Takes immediate version of shift as input.
15532 static SDValue getTargetVShiftByConstNode(unsigned Opc, SDLoc dl, MVT VT,
15533 SDValue SrcOp, uint64_t ShiftAmt,
15534 SelectionDAG &DAG) {
15535 MVT ElementType = VT.getVectorElementType();
15537 // Fold this packed shift into its first operand if ShiftAmt is 0.
15541 // Check for ShiftAmt >= element width
15542 if (ShiftAmt >= ElementType.getSizeInBits()) {
15543 if (Opc == X86ISD::VSRAI)
15544 ShiftAmt = ElementType.getSizeInBits() - 1;
15546 return DAG.getConstant(0, VT);
15549 assert((Opc == X86ISD::VSHLI || Opc == X86ISD::VSRLI || Opc == X86ISD::VSRAI)
15550 && "Unknown target vector shift-by-constant node");
15552 // Fold this packed vector shift into a build vector if SrcOp is a
15553 // vector of Constants or UNDEFs, and SrcOp valuetype is the same as VT.
15554 if (VT == SrcOp.getSimpleValueType() &&
15555 ISD::isBuildVectorOfConstantSDNodes(SrcOp.getNode())) {
15556 SmallVector<SDValue, 8> Elts;
15557 unsigned NumElts = SrcOp->getNumOperands();
15558 ConstantSDNode *ND;
15561 default: llvm_unreachable(nullptr);
15562 case X86ISD::VSHLI:
15563 for (unsigned i=0; i!=NumElts; ++i) {
15564 SDValue CurrentOp = SrcOp->getOperand(i);
15565 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15566 Elts.push_back(CurrentOp);
15569 ND = cast<ConstantSDNode>(CurrentOp);
15570 const APInt &C = ND->getAPIntValue();
15571 Elts.push_back(DAG.getConstant(C.shl(ShiftAmt), ElementType));
15574 case X86ISD::VSRLI:
15575 for (unsigned i=0; i!=NumElts; ++i) {
15576 SDValue CurrentOp = SrcOp->getOperand(i);
15577 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15578 Elts.push_back(CurrentOp);
15581 ND = cast<ConstantSDNode>(CurrentOp);
15582 const APInt &C = ND->getAPIntValue();
15583 Elts.push_back(DAG.getConstant(C.lshr(ShiftAmt), ElementType));
15586 case X86ISD::VSRAI:
15587 for (unsigned i=0; i!=NumElts; ++i) {
15588 SDValue CurrentOp = SrcOp->getOperand(i);
15589 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15590 Elts.push_back(CurrentOp);
15593 ND = cast<ConstantSDNode>(CurrentOp);
15594 const APInt &C = ND->getAPIntValue();
15595 Elts.push_back(DAG.getConstant(C.ashr(ShiftAmt), ElementType));
15600 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
15603 return DAG.getNode(Opc, dl, VT, SrcOp, DAG.getConstant(ShiftAmt, MVT::i8));
15606 // getTargetVShiftNode - Handle vector element shifts where the shift amount
15607 // may or may not be a constant. Takes immediate version of shift as input.
15608 static SDValue getTargetVShiftNode(unsigned Opc, SDLoc dl, MVT VT,
15609 SDValue SrcOp, SDValue ShAmt,
15610 SelectionDAG &DAG) {
15611 assert(ShAmt.getValueType() == MVT::i32 && "ShAmt is not i32");
15613 // Catch shift-by-constant.
15614 if (ConstantSDNode *CShAmt = dyn_cast<ConstantSDNode>(ShAmt))
15615 return getTargetVShiftByConstNode(Opc, dl, VT, SrcOp,
15616 CShAmt->getZExtValue(), DAG);
15618 // Change opcode to non-immediate version
15620 default: llvm_unreachable("Unknown target vector shift node");
15621 case X86ISD::VSHLI: Opc = X86ISD::VSHL; break;
15622 case X86ISD::VSRLI: Opc = X86ISD::VSRL; break;
15623 case X86ISD::VSRAI: Opc = X86ISD::VSRA; break;
15626 // Need to build a vector containing shift amount
15627 // Shift amount is 32-bits, but SSE instructions read 64-bit, so fill with 0
15630 ShOps[1] = DAG.getConstant(0, MVT::i32);
15631 ShOps[2] = ShOps[3] = DAG.getUNDEF(MVT::i32);
15632 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, ShOps);
15634 // The return type has to be a 128-bit type with the same element
15635 // type as the input type.
15636 MVT EltVT = VT.getVectorElementType();
15637 EVT ShVT = MVT::getVectorVT(EltVT, 128/EltVT.getSizeInBits());
15639 ShAmt = DAG.getNode(ISD::BITCAST, dl, ShVT, ShAmt);
15640 return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt);
15643 /// \brief Return (vselect \p Mask, \p Op, \p PreservedSrc) along with the
15644 /// necessary casting for \p Mask when lowering masking intrinsics.
15645 static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask,
15646 SDValue PreservedSrc, SelectionDAG &DAG) {
15647 EVT VT = Op.getValueType();
15648 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(),
15649 MVT::i1, VT.getVectorNumElements());
15652 assert(MaskVT.isSimple() && "invalid mask type");
15653 return DAG.getNode(ISD::VSELECT, dl, VT,
15654 DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask),
15658 static unsigned getOpcodeForFMAIntrinsic(unsigned IntNo) {
15660 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
15661 case Intrinsic::x86_fma_vfmadd_ps:
15662 case Intrinsic::x86_fma_vfmadd_pd:
15663 case Intrinsic::x86_fma_vfmadd_ps_256:
15664 case Intrinsic::x86_fma_vfmadd_pd_256:
15665 case Intrinsic::x86_fma_mask_vfmadd_ps_512:
15666 case Intrinsic::x86_fma_mask_vfmadd_pd_512:
15667 return X86ISD::FMADD;
15668 case Intrinsic::x86_fma_vfmsub_ps:
15669 case Intrinsic::x86_fma_vfmsub_pd:
15670 case Intrinsic::x86_fma_vfmsub_ps_256:
15671 case Intrinsic::x86_fma_vfmsub_pd_256:
15672 case Intrinsic::x86_fma_mask_vfmsub_ps_512:
15673 case Intrinsic::x86_fma_mask_vfmsub_pd_512:
15674 return X86ISD::FMSUB;
15675 case Intrinsic::x86_fma_vfnmadd_ps:
15676 case Intrinsic::x86_fma_vfnmadd_pd:
15677 case Intrinsic::x86_fma_vfnmadd_ps_256:
15678 case Intrinsic::x86_fma_vfnmadd_pd_256:
15679 case Intrinsic::x86_fma_mask_vfnmadd_ps_512:
15680 case Intrinsic::x86_fma_mask_vfnmadd_pd_512:
15681 return X86ISD::FNMADD;
15682 case Intrinsic::x86_fma_vfnmsub_ps:
15683 case Intrinsic::x86_fma_vfnmsub_pd:
15684 case Intrinsic::x86_fma_vfnmsub_ps_256:
15685 case Intrinsic::x86_fma_vfnmsub_pd_256:
15686 case Intrinsic::x86_fma_mask_vfnmsub_ps_512:
15687 case Intrinsic::x86_fma_mask_vfnmsub_pd_512:
15688 return X86ISD::FNMSUB;
15689 case Intrinsic::x86_fma_vfmaddsub_ps:
15690 case Intrinsic::x86_fma_vfmaddsub_pd:
15691 case Intrinsic::x86_fma_vfmaddsub_ps_256:
15692 case Intrinsic::x86_fma_vfmaddsub_pd_256:
15693 case Intrinsic::x86_fma_mask_vfmaddsub_ps_512:
15694 case Intrinsic::x86_fma_mask_vfmaddsub_pd_512:
15695 return X86ISD::FMADDSUB;
15696 case Intrinsic::x86_fma_vfmsubadd_ps:
15697 case Intrinsic::x86_fma_vfmsubadd_pd:
15698 case Intrinsic::x86_fma_vfmsubadd_ps_256:
15699 case Intrinsic::x86_fma_vfmsubadd_pd_256:
15700 case Intrinsic::x86_fma_mask_vfmsubadd_ps_512:
15701 case Intrinsic::x86_fma_mask_vfmsubadd_pd_512:
15702 return X86ISD::FMSUBADD;
15706 static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
15708 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
15710 const IntrinsicData* IntrData = getIntrinsicWithoutChain(IntNo);
15712 switch(IntrData->Type) {
15713 case INTR_TYPE_1OP:
15714 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1));
15715 case INTR_TYPE_2OP:
15716 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
15718 case INTR_TYPE_3OP:
15719 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
15720 Op.getOperand(2), Op.getOperand(3));
15721 case COMI: { // Comparison intrinsics
15722 ISD::CondCode CC = (ISD::CondCode)IntrData->Opc1;
15723 SDValue LHS = Op.getOperand(1);
15724 SDValue RHS = Op.getOperand(2);
15725 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
15726 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
15727 SDValue Cond = DAG.getNode(IntrData->Opc0, dl, MVT::i32, LHS, RHS);
15728 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15729 DAG.getConstant(X86CC, MVT::i8), Cond);
15730 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
15733 return getTargetVShiftNode(IntrData->Opc0, dl, Op.getSimpleValueType(),
15734 Op.getOperand(1), Op.getOperand(2), DAG);
15741 default: return SDValue(); // Don't custom lower most intrinsics.
15743 // Arithmetic intrinsics.
15744 case Intrinsic::x86_sse2_pmulu_dq:
15745 case Intrinsic::x86_avx2_pmulu_dq:
15746 return DAG.getNode(X86ISD::PMULUDQ, dl, Op.getValueType(),
15747 Op.getOperand(1), Op.getOperand(2));
15749 case Intrinsic::x86_sse41_pmuldq:
15750 case Intrinsic::x86_avx2_pmul_dq:
15751 return DAG.getNode(X86ISD::PMULDQ, dl, Op.getValueType(),
15752 Op.getOperand(1), Op.getOperand(2));
15754 case Intrinsic::x86_sse2_pmulhu_w:
15755 case Intrinsic::x86_avx2_pmulhu_w:
15756 return DAG.getNode(ISD::MULHU, dl, Op.getValueType(),
15757 Op.getOperand(1), Op.getOperand(2));
15759 case Intrinsic::x86_sse2_pmulh_w:
15760 case Intrinsic::x86_avx2_pmulh_w:
15761 return DAG.getNode(ISD::MULHS, dl, Op.getValueType(),
15762 Op.getOperand(1), Op.getOperand(2));
15764 // SSE/SSE2/AVX floating point max/min intrinsics.
15765 case Intrinsic::x86_sse_max_ps:
15766 case Intrinsic::x86_sse2_max_pd:
15767 case Intrinsic::x86_avx_max_ps_256:
15768 case Intrinsic::x86_avx_max_pd_256:
15769 case Intrinsic::x86_sse_min_ps:
15770 case Intrinsic::x86_sse2_min_pd:
15771 case Intrinsic::x86_avx_min_ps_256:
15772 case Intrinsic::x86_avx_min_pd_256: {
15775 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
15776 case Intrinsic::x86_sse_max_ps:
15777 case Intrinsic::x86_sse2_max_pd:
15778 case Intrinsic::x86_avx_max_ps_256:
15779 case Intrinsic::x86_avx_max_pd_256:
15780 Opcode = X86ISD::FMAX;
15782 case Intrinsic::x86_sse_min_ps:
15783 case Intrinsic::x86_sse2_min_pd:
15784 case Intrinsic::x86_avx_min_ps_256:
15785 case Intrinsic::x86_avx_min_pd_256:
15786 Opcode = X86ISD::FMIN;
15789 return DAG.getNode(Opcode, dl, Op.getValueType(),
15790 Op.getOperand(1), Op.getOperand(2));
15793 // AVX2 variable shift intrinsics
15794 case Intrinsic::x86_avx2_psllv_d:
15795 case Intrinsic::x86_avx2_psllv_q:
15796 case Intrinsic::x86_avx2_psllv_d_256:
15797 case Intrinsic::x86_avx2_psllv_q_256:
15798 case Intrinsic::x86_avx2_psrlv_d:
15799 case Intrinsic::x86_avx2_psrlv_q:
15800 case Intrinsic::x86_avx2_psrlv_d_256:
15801 case Intrinsic::x86_avx2_psrlv_q_256:
15802 case Intrinsic::x86_avx2_psrav_d:
15803 case Intrinsic::x86_avx2_psrav_d_256: {
15806 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
15807 case Intrinsic::x86_avx2_psllv_d:
15808 case Intrinsic::x86_avx2_psllv_q:
15809 case Intrinsic::x86_avx2_psllv_d_256:
15810 case Intrinsic::x86_avx2_psllv_q_256:
15813 case Intrinsic::x86_avx2_psrlv_d:
15814 case Intrinsic::x86_avx2_psrlv_q:
15815 case Intrinsic::x86_avx2_psrlv_d_256:
15816 case Intrinsic::x86_avx2_psrlv_q_256:
15819 case Intrinsic::x86_avx2_psrav_d:
15820 case Intrinsic::x86_avx2_psrav_d_256:
15824 return DAG.getNode(Opcode, dl, Op.getValueType(),
15825 Op.getOperand(1), Op.getOperand(2));
15828 case Intrinsic::x86_sse2_packssdw_128:
15829 case Intrinsic::x86_sse2_packsswb_128:
15830 case Intrinsic::x86_avx2_packssdw:
15831 case Intrinsic::x86_avx2_packsswb:
15832 return DAG.getNode(X86ISD::PACKSS, dl, Op.getValueType(),
15833 Op.getOperand(1), Op.getOperand(2));
15835 case Intrinsic::x86_sse2_packuswb_128:
15836 case Intrinsic::x86_sse41_packusdw:
15837 case Intrinsic::x86_avx2_packuswb:
15838 case Intrinsic::x86_avx2_packusdw:
15839 return DAG.getNode(X86ISD::PACKUS, dl, Op.getValueType(),
15840 Op.getOperand(1), Op.getOperand(2));
15842 case Intrinsic::x86_ssse3_pshuf_b_128:
15843 case Intrinsic::x86_avx2_pshuf_b:
15844 return DAG.getNode(X86ISD::PSHUFB, dl, Op.getValueType(),
15845 Op.getOperand(1), Op.getOperand(2));
15847 case Intrinsic::x86_sse2_pshuf_d:
15848 return DAG.getNode(X86ISD::PSHUFD, dl, Op.getValueType(),
15849 Op.getOperand(1), Op.getOperand(2));
15851 case Intrinsic::x86_sse2_pshufl_w:
15852 return DAG.getNode(X86ISD::PSHUFLW, dl, Op.getValueType(),
15853 Op.getOperand(1), Op.getOperand(2));
15855 case Intrinsic::x86_sse2_pshufh_w:
15856 return DAG.getNode(X86ISD::PSHUFHW, dl, Op.getValueType(),
15857 Op.getOperand(1), Op.getOperand(2));
15859 case Intrinsic::x86_ssse3_psign_b_128:
15860 case Intrinsic::x86_ssse3_psign_w_128:
15861 case Intrinsic::x86_ssse3_psign_d_128:
15862 case Intrinsic::x86_avx2_psign_b:
15863 case Intrinsic::x86_avx2_psign_w:
15864 case Intrinsic::x86_avx2_psign_d:
15865 return DAG.getNode(X86ISD::PSIGN, dl, Op.getValueType(),
15866 Op.getOperand(1), Op.getOperand(2));
15868 case Intrinsic::x86_avx2_permd:
15869 case Intrinsic::x86_avx2_permps:
15870 // Operands intentionally swapped. Mask is last operand to intrinsic,
15871 // but second operand for node/instruction.
15872 return DAG.getNode(X86ISD::VPERMV, dl, Op.getValueType(),
15873 Op.getOperand(2), Op.getOperand(1));
15875 case Intrinsic::x86_avx512_mask_valign_q_512:
15876 case Intrinsic::x86_avx512_mask_valign_d_512:
15877 // Vector source operands are swapped.
15878 return getVectorMaskingNode(DAG.getNode(X86ISD::VALIGN, dl,
15879 Op.getValueType(), Op.getOperand(2),
15882 Op.getOperand(5), Op.getOperand(4), DAG);
15884 // ptest and testp intrinsics. The intrinsic these come from are designed to
15885 // return an integer value, not just an instruction so lower it to the ptest
15886 // or testp pattern and a setcc for the result.
15887 case Intrinsic::x86_sse41_ptestz:
15888 case Intrinsic::x86_sse41_ptestc:
15889 case Intrinsic::x86_sse41_ptestnzc:
15890 case Intrinsic::x86_avx_ptestz_256:
15891 case Intrinsic::x86_avx_ptestc_256:
15892 case Intrinsic::x86_avx_ptestnzc_256:
15893 case Intrinsic::x86_avx_vtestz_ps:
15894 case Intrinsic::x86_avx_vtestc_ps:
15895 case Intrinsic::x86_avx_vtestnzc_ps:
15896 case Intrinsic::x86_avx_vtestz_pd:
15897 case Intrinsic::x86_avx_vtestc_pd:
15898 case Intrinsic::x86_avx_vtestnzc_pd:
15899 case Intrinsic::x86_avx_vtestz_ps_256:
15900 case Intrinsic::x86_avx_vtestc_ps_256:
15901 case Intrinsic::x86_avx_vtestnzc_ps_256:
15902 case Intrinsic::x86_avx_vtestz_pd_256:
15903 case Intrinsic::x86_avx_vtestc_pd_256:
15904 case Intrinsic::x86_avx_vtestnzc_pd_256: {
15905 bool IsTestPacked = false;
15908 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
15909 case Intrinsic::x86_avx_vtestz_ps:
15910 case Intrinsic::x86_avx_vtestz_pd:
15911 case Intrinsic::x86_avx_vtestz_ps_256:
15912 case Intrinsic::x86_avx_vtestz_pd_256:
15913 IsTestPacked = true; // Fallthrough
15914 case Intrinsic::x86_sse41_ptestz:
15915 case Intrinsic::x86_avx_ptestz_256:
15917 X86CC = X86::COND_E;
15919 case Intrinsic::x86_avx_vtestc_ps:
15920 case Intrinsic::x86_avx_vtestc_pd:
15921 case Intrinsic::x86_avx_vtestc_ps_256:
15922 case Intrinsic::x86_avx_vtestc_pd_256:
15923 IsTestPacked = true; // Fallthrough
15924 case Intrinsic::x86_sse41_ptestc:
15925 case Intrinsic::x86_avx_ptestc_256:
15927 X86CC = X86::COND_B;
15929 case Intrinsic::x86_avx_vtestnzc_ps:
15930 case Intrinsic::x86_avx_vtestnzc_pd:
15931 case Intrinsic::x86_avx_vtestnzc_ps_256:
15932 case Intrinsic::x86_avx_vtestnzc_pd_256:
15933 IsTestPacked = true; // Fallthrough
15934 case Intrinsic::x86_sse41_ptestnzc:
15935 case Intrinsic::x86_avx_ptestnzc_256:
15937 X86CC = X86::COND_A;
15941 SDValue LHS = Op.getOperand(1);
15942 SDValue RHS = Op.getOperand(2);
15943 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
15944 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
15945 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
15946 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
15947 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
15949 case Intrinsic::x86_avx512_kortestz_w:
15950 case Intrinsic::x86_avx512_kortestc_w: {
15951 unsigned X86CC = (IntNo == Intrinsic::x86_avx512_kortestz_w)? X86::COND_E: X86::COND_B;
15952 SDValue LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1, Op.getOperand(1));
15953 SDValue RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1, Op.getOperand(2));
15954 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
15955 SDValue Test = DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, LHS, RHS);
15956 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i1, CC, Test);
15957 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
15960 case Intrinsic::x86_sse42_pcmpistria128:
15961 case Intrinsic::x86_sse42_pcmpestria128:
15962 case Intrinsic::x86_sse42_pcmpistric128:
15963 case Intrinsic::x86_sse42_pcmpestric128:
15964 case Intrinsic::x86_sse42_pcmpistrio128:
15965 case Intrinsic::x86_sse42_pcmpestrio128:
15966 case Intrinsic::x86_sse42_pcmpistris128:
15967 case Intrinsic::x86_sse42_pcmpestris128:
15968 case Intrinsic::x86_sse42_pcmpistriz128:
15969 case Intrinsic::x86_sse42_pcmpestriz128: {
15973 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
15974 case Intrinsic::x86_sse42_pcmpistria128:
15975 Opcode = X86ISD::PCMPISTRI;
15976 X86CC = X86::COND_A;
15978 case Intrinsic::x86_sse42_pcmpestria128:
15979 Opcode = X86ISD::PCMPESTRI;
15980 X86CC = X86::COND_A;
15982 case Intrinsic::x86_sse42_pcmpistric128:
15983 Opcode = X86ISD::PCMPISTRI;
15984 X86CC = X86::COND_B;
15986 case Intrinsic::x86_sse42_pcmpestric128:
15987 Opcode = X86ISD::PCMPESTRI;
15988 X86CC = X86::COND_B;
15990 case Intrinsic::x86_sse42_pcmpistrio128:
15991 Opcode = X86ISD::PCMPISTRI;
15992 X86CC = X86::COND_O;
15994 case Intrinsic::x86_sse42_pcmpestrio128:
15995 Opcode = X86ISD::PCMPESTRI;
15996 X86CC = X86::COND_O;
15998 case Intrinsic::x86_sse42_pcmpistris128:
15999 Opcode = X86ISD::PCMPISTRI;
16000 X86CC = X86::COND_S;
16002 case Intrinsic::x86_sse42_pcmpestris128:
16003 Opcode = X86ISD::PCMPESTRI;
16004 X86CC = X86::COND_S;
16006 case Intrinsic::x86_sse42_pcmpistriz128:
16007 Opcode = X86ISD::PCMPISTRI;
16008 X86CC = X86::COND_E;
16010 case Intrinsic::x86_sse42_pcmpestriz128:
16011 Opcode = X86ISD::PCMPESTRI;
16012 X86CC = X86::COND_E;
16015 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
16016 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
16017 SDValue PCMP = DAG.getNode(Opcode, dl, VTs, NewOps);
16018 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16019 DAG.getConstant(X86CC, MVT::i8),
16020 SDValue(PCMP.getNode(), 1));
16021 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16024 case Intrinsic::x86_sse42_pcmpistri128:
16025 case Intrinsic::x86_sse42_pcmpestri128: {
16027 if (IntNo == Intrinsic::x86_sse42_pcmpistri128)
16028 Opcode = X86ISD::PCMPISTRI;
16030 Opcode = X86ISD::PCMPESTRI;
16032 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
16033 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
16034 return DAG.getNode(Opcode, dl, VTs, NewOps);
16037 case Intrinsic::x86_fma_mask_vfmadd_ps_512:
16038 case Intrinsic::x86_fma_mask_vfmadd_pd_512:
16039 case Intrinsic::x86_fma_mask_vfmsub_ps_512:
16040 case Intrinsic::x86_fma_mask_vfmsub_pd_512:
16041 case Intrinsic::x86_fma_mask_vfnmadd_ps_512:
16042 case Intrinsic::x86_fma_mask_vfnmadd_pd_512:
16043 case Intrinsic::x86_fma_mask_vfnmsub_ps_512:
16044 case Intrinsic::x86_fma_mask_vfnmsub_pd_512:
16045 case Intrinsic::x86_fma_mask_vfmaddsub_ps_512:
16046 case Intrinsic::x86_fma_mask_vfmaddsub_pd_512:
16047 case Intrinsic::x86_fma_mask_vfmsubadd_ps_512:
16048 case Intrinsic::x86_fma_mask_vfmsubadd_pd_512: {
16049 auto *SAE = cast<ConstantSDNode>(Op.getOperand(5));
16050 if (SAE->getZExtValue() == X86::STATIC_ROUNDING::CUR_DIRECTION)
16051 return getVectorMaskingNode(DAG.getNode(getOpcodeForFMAIntrinsic(IntNo),
16052 dl, Op.getValueType(),
16056 Op.getOperand(4), Op.getOperand(1), DAG);
16061 case Intrinsic::x86_fma_vfmadd_ps:
16062 case Intrinsic::x86_fma_vfmadd_pd:
16063 case Intrinsic::x86_fma_vfmsub_ps:
16064 case Intrinsic::x86_fma_vfmsub_pd:
16065 case Intrinsic::x86_fma_vfnmadd_ps:
16066 case Intrinsic::x86_fma_vfnmadd_pd:
16067 case Intrinsic::x86_fma_vfnmsub_ps:
16068 case Intrinsic::x86_fma_vfnmsub_pd:
16069 case Intrinsic::x86_fma_vfmaddsub_ps:
16070 case Intrinsic::x86_fma_vfmaddsub_pd:
16071 case Intrinsic::x86_fma_vfmsubadd_ps:
16072 case Intrinsic::x86_fma_vfmsubadd_pd:
16073 case Intrinsic::x86_fma_vfmadd_ps_256:
16074 case Intrinsic::x86_fma_vfmadd_pd_256:
16075 case Intrinsic::x86_fma_vfmsub_ps_256:
16076 case Intrinsic::x86_fma_vfmsub_pd_256:
16077 case Intrinsic::x86_fma_vfnmadd_ps_256:
16078 case Intrinsic::x86_fma_vfnmadd_pd_256:
16079 case Intrinsic::x86_fma_vfnmsub_ps_256:
16080 case Intrinsic::x86_fma_vfnmsub_pd_256:
16081 case Intrinsic::x86_fma_vfmaddsub_ps_256:
16082 case Intrinsic::x86_fma_vfmaddsub_pd_256:
16083 case Intrinsic::x86_fma_vfmsubadd_ps_256:
16084 case Intrinsic::x86_fma_vfmsubadd_pd_256:
16085 return DAG.getNode(getOpcodeForFMAIntrinsic(IntNo), dl, Op.getValueType(),
16086 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
16090 static SDValue getGatherNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16091 SDValue Src, SDValue Mask, SDValue Base,
16092 SDValue Index, SDValue ScaleOp, SDValue Chain,
16093 const X86Subtarget * Subtarget) {
16095 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
16096 assert(C && "Invalid scale type");
16097 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), MVT::i8);
16098 EVT MaskVT = MVT::getVectorVT(MVT::i1,
16099 Index.getSimpleValueType().getVectorNumElements());
16101 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16103 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), MaskVT);
16105 MaskInReg = DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask);
16106 SDVTList VTs = DAG.getVTList(Op.getValueType(), MaskVT, MVT::Other);
16107 SDValue Disp = DAG.getTargetConstant(0, MVT::i32);
16108 SDValue Segment = DAG.getRegister(0, MVT::i32);
16109 if (Src.getOpcode() == ISD::UNDEF)
16110 Src = getZeroVector(Op.getValueType(), Subtarget, DAG, dl);
16111 SDValue Ops[] = {Src, MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
16112 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
16113 SDValue RetOps[] = { SDValue(Res, 0), SDValue(Res, 2) };
16114 return DAG.getMergeValues(RetOps, dl);
16117 static SDValue getScatterNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16118 SDValue Src, SDValue Mask, SDValue Base,
16119 SDValue Index, SDValue ScaleOp, SDValue Chain) {
16121 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
16122 assert(C && "Invalid scale type");
16123 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), MVT::i8);
16124 SDValue Disp = DAG.getTargetConstant(0, MVT::i32);
16125 SDValue Segment = DAG.getRegister(0, MVT::i32);
16126 EVT MaskVT = MVT::getVectorVT(MVT::i1,
16127 Index.getSimpleValueType().getVectorNumElements());
16129 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16131 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), MaskVT);
16133 MaskInReg = DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask);
16134 SDVTList VTs = DAG.getVTList(MaskVT, MVT::Other);
16135 SDValue Ops[] = {Base, Scale, Index, Disp, Segment, MaskInReg, Src, Chain};
16136 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
16137 return SDValue(Res, 1);
16140 static SDValue getPrefetchNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16141 SDValue Mask, SDValue Base, SDValue Index,
16142 SDValue ScaleOp, SDValue Chain) {
16144 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
16145 assert(C && "Invalid scale type");
16146 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), MVT::i8);
16147 SDValue Disp = DAG.getTargetConstant(0, MVT::i32);
16148 SDValue Segment = DAG.getRegister(0, MVT::i32);
16150 MVT::getVectorVT(MVT::i1, Index.getSimpleValueType().getVectorNumElements());
16152 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16154 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), MaskVT);
16156 MaskInReg = DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask);
16157 //SDVTList VTs = DAG.getVTList(MVT::Other);
16158 SDValue Ops[] = {MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
16159 SDNode *Res = DAG.getMachineNode(Opc, dl, MVT::Other, Ops);
16160 return SDValue(Res, 0);
16163 // getReadPerformanceCounter - Handles the lowering of builtin intrinsics that
16164 // read performance monitor counters (x86_rdpmc).
16165 static void getReadPerformanceCounter(SDNode *N, SDLoc DL,
16166 SelectionDAG &DAG, const X86Subtarget *Subtarget,
16167 SmallVectorImpl<SDValue> &Results) {
16168 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
16169 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
16172 // The ECX register is used to select the index of the performance counter
16174 SDValue Chain = DAG.getCopyToReg(N->getOperand(0), DL, X86::ECX,
16176 SDValue rd = DAG.getNode(X86ISD::RDPMC_DAG, DL, Tys, Chain);
16178 // Reads the content of a 64-bit performance counter and returns it in the
16179 // registers EDX:EAX.
16180 if (Subtarget->is64Bit()) {
16181 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
16182 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
16185 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
16186 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
16189 Chain = HI.getValue(1);
16191 if (Subtarget->is64Bit()) {
16192 // The EAX register is loaded with the low-order 32 bits. The EDX register
16193 // is loaded with the supported high-order bits of the counter.
16194 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
16195 DAG.getConstant(32, MVT::i8));
16196 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
16197 Results.push_back(Chain);
16201 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
16202 SDValue Ops[] = { LO, HI };
16203 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
16204 Results.push_back(Pair);
16205 Results.push_back(Chain);
16208 // getReadTimeStampCounter - Handles the lowering of builtin intrinsics that
16209 // read the time stamp counter (x86_rdtsc and x86_rdtscp). This function is
16210 // also used to custom lower READCYCLECOUNTER nodes.
16211 static void getReadTimeStampCounter(SDNode *N, SDLoc DL, unsigned Opcode,
16212 SelectionDAG &DAG, const X86Subtarget *Subtarget,
16213 SmallVectorImpl<SDValue> &Results) {
16214 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
16215 SDValue rd = DAG.getNode(Opcode, DL, Tys, N->getOperand(0));
16218 // The processor's time-stamp counter (a 64-bit MSR) is stored into the
16219 // EDX:EAX registers. EDX is loaded with the high-order 32 bits of the MSR
16220 // and the EAX register is loaded with the low-order 32 bits.
16221 if (Subtarget->is64Bit()) {
16222 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
16223 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
16226 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
16227 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
16230 SDValue Chain = HI.getValue(1);
16232 if (Opcode == X86ISD::RDTSCP_DAG) {
16233 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
16235 // Instruction RDTSCP loads the IA32:TSC_AUX_MSR (address C000_0103H) into
16236 // the ECX register. Add 'ecx' explicitly to the chain.
16237 SDValue ecx = DAG.getCopyFromReg(Chain, DL, X86::ECX, MVT::i32,
16239 // Explicitly store the content of ECX at the location passed in input
16240 // to the 'rdtscp' intrinsic.
16241 Chain = DAG.getStore(ecx.getValue(1), DL, ecx, N->getOperand(2),
16242 MachinePointerInfo(), false, false, 0);
16245 if (Subtarget->is64Bit()) {
16246 // The EDX register is loaded with the high-order 32 bits of the MSR, and
16247 // the EAX register is loaded with the low-order 32 bits.
16248 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
16249 DAG.getConstant(32, MVT::i8));
16250 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
16251 Results.push_back(Chain);
16255 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
16256 SDValue Ops[] = { LO, HI };
16257 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
16258 Results.push_back(Pair);
16259 Results.push_back(Chain);
16262 static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget *Subtarget,
16263 SelectionDAG &DAG) {
16264 SmallVector<SDValue, 2> Results;
16266 getReadTimeStampCounter(Op.getNode(), DL, X86ISD::RDTSC_DAG, DAG, Subtarget,
16268 return DAG.getMergeValues(Results, DL);
16272 static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
16273 SelectionDAG &DAG) {
16274 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
16276 const IntrinsicData* IntrData = getIntrinsicWithChain(IntNo);
16281 switch(IntrData->Type) {
16283 llvm_unreachable("Unknown Intrinsic Type");
16287 // Emit the node with the right value type.
16288 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other);
16289 SDValue Result = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
16291 // If the value returned by RDRAND/RDSEED was valid (CF=1), return 1.
16292 // Otherwise return the value from Rand, which is always 0, casted to i32.
16293 SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)),
16294 DAG.getConstant(1, Op->getValueType(1)),
16295 DAG.getConstant(X86::COND_B, MVT::i32),
16296 SDValue(Result.getNode(), 1) };
16297 SDValue isValid = DAG.getNode(X86ISD::CMOV, dl,
16298 DAG.getVTList(Op->getValueType(1), MVT::Glue),
16301 // Return { result, isValid, chain }.
16302 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid,
16303 SDValue(Result.getNode(), 2));
16306 //gather(v1, mask, index, base, scale);
16307 SDValue Chain = Op.getOperand(0);
16308 SDValue Src = Op.getOperand(2);
16309 SDValue Base = Op.getOperand(3);
16310 SDValue Index = Op.getOperand(4);
16311 SDValue Mask = Op.getOperand(5);
16312 SDValue Scale = Op.getOperand(6);
16313 return getGatherNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale, Chain,
16317 //scatter(base, mask, index, v1, scale);
16318 SDValue Chain = Op.getOperand(0);
16319 SDValue Base = Op.getOperand(2);
16320 SDValue Mask = Op.getOperand(3);
16321 SDValue Index = Op.getOperand(4);
16322 SDValue Src = Op.getOperand(5);
16323 SDValue Scale = Op.getOperand(6);
16324 return getScatterNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale, Chain);
16327 SDValue Hint = Op.getOperand(6);
16329 if (dyn_cast<ConstantSDNode> (Hint) == nullptr ||
16330 (HintVal = dyn_cast<ConstantSDNode> (Hint)->getZExtValue()) > 1)
16331 llvm_unreachable("Wrong prefetch hint in intrinsic: should be 0 or 1");
16332 unsigned Opcode = (HintVal ? IntrData->Opc1 : IntrData->Opc0);
16333 SDValue Chain = Op.getOperand(0);
16334 SDValue Mask = Op.getOperand(2);
16335 SDValue Index = Op.getOperand(3);
16336 SDValue Base = Op.getOperand(4);
16337 SDValue Scale = Op.getOperand(5);
16338 return getPrefetchNode(Opcode, Op, DAG, Mask, Base, Index, Scale, Chain);
16340 // Read Time Stamp Counter (RDTSC) and Processor ID (RDTSCP).
16342 SmallVector<SDValue, 2> Results;
16343 getReadTimeStampCounter(Op.getNode(), dl, IntrData->Opc0, DAG, Subtarget, Results);
16344 return DAG.getMergeValues(Results, dl);
16346 // Read Performance Monitoring Counters.
16348 SmallVector<SDValue, 2> Results;
16349 getReadPerformanceCounter(Op.getNode(), dl, DAG, Subtarget, Results);
16350 return DAG.getMergeValues(Results, dl);
16352 // XTEST intrinsics.
16354 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
16355 SDValue InTrans = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
16356 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16357 DAG.getConstant(X86::COND_NE, MVT::i8),
16359 SDValue Ret = DAG.getNode(ISD::ZERO_EXTEND, dl, Op->getValueType(0), SetCC);
16360 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(),
16361 Ret, SDValue(InTrans.getNode(), 1));
16365 SmallVector<SDValue, 2> Results;
16366 SDVTList CFVTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
16367 SDVTList VTs = DAG.getVTList(Op.getOperand(3)->getValueType(0), MVT::Other);
16368 SDValue GenCF = DAG.getNode(X86ISD::ADD, dl, CFVTs, Op.getOperand(2),
16369 DAG.getConstant(-1, MVT::i8));
16370 SDValue Res = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(3),
16371 Op.getOperand(4), GenCF.getValue(1));
16372 SDValue Store = DAG.getStore(Op.getOperand(0), dl, Res.getValue(0),
16373 Op.getOperand(5), MachinePointerInfo(),
16375 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16376 DAG.getConstant(X86::COND_B, MVT::i8),
16378 Results.push_back(SetCC);
16379 Results.push_back(Store);
16380 return DAG.getMergeValues(Results, dl);
16385 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
16386 SelectionDAG &DAG) const {
16387 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
16388 MFI->setReturnAddressIsTaken(true);
16390 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
16393 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
16395 EVT PtrVT = getPointerTy();
16398 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
16399 const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
16400 DAG.getSubtarget().getRegisterInfo());
16401 SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), PtrVT);
16402 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
16403 DAG.getNode(ISD::ADD, dl, PtrVT,
16404 FrameAddr, Offset),
16405 MachinePointerInfo(), false, false, false, 0);
16408 // Just load the return address.
16409 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
16410 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
16411 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
16414 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
16415 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
16416 MFI->setFrameAddressIsTaken(true);
16418 EVT VT = Op.getValueType();
16419 SDLoc dl(Op); // FIXME probably not meaningful
16420 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
16421 const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
16422 DAG.getSubtarget().getRegisterInfo());
16423 unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
16424 assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
16425 (FrameReg == X86::EBP && VT == MVT::i32)) &&
16426 "Invalid Frame Register!");
16427 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
16429 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
16430 MachinePointerInfo(),
16431 false, false, false, 0);
16435 // FIXME? Maybe this could be a TableGen attribute on some registers and
16436 // this table could be generated automatically from RegInfo.
16437 unsigned X86TargetLowering::getRegisterByName(const char* RegName,
16439 unsigned Reg = StringSwitch<unsigned>(RegName)
16440 .Case("esp", X86::ESP)
16441 .Case("rsp", X86::RSP)
16445 report_fatal_error("Invalid register name global variable");
16448 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
16449 SelectionDAG &DAG) const {
16450 const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
16451 DAG.getSubtarget().getRegisterInfo());
16452 return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize());
16455 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
16456 SDValue Chain = Op.getOperand(0);
16457 SDValue Offset = Op.getOperand(1);
16458 SDValue Handler = Op.getOperand(2);
16461 EVT PtrVT = getPointerTy();
16462 const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
16463 DAG.getSubtarget().getRegisterInfo());
16464 unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
16465 assert(((FrameReg == X86::RBP && PtrVT == MVT::i64) ||
16466 (FrameReg == X86::EBP && PtrVT == MVT::i32)) &&
16467 "Invalid Frame Register!");
16468 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, PtrVT);
16469 unsigned StoreAddrReg = (PtrVT == MVT::i64) ? X86::RCX : X86::ECX;
16471 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Frame,
16472 DAG.getIntPtrConstant(RegInfo->getSlotSize()));
16473 StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, StoreAddr, Offset);
16474 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
16476 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
16478 return DAG.getNode(X86ISD::EH_RETURN, dl, MVT::Other, Chain,
16479 DAG.getRegister(StoreAddrReg, PtrVT));
16482 SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
16483 SelectionDAG &DAG) const {
16485 return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL,
16486 DAG.getVTList(MVT::i32, MVT::Other),
16487 Op.getOperand(0), Op.getOperand(1));
16490 SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
16491 SelectionDAG &DAG) const {
16493 return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
16494 Op.getOperand(0), Op.getOperand(1));
16497 static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
16498 return Op.getOperand(0);
16501 SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
16502 SelectionDAG &DAG) const {
16503 SDValue Root = Op.getOperand(0);
16504 SDValue Trmp = Op.getOperand(1); // trampoline
16505 SDValue FPtr = Op.getOperand(2); // nested function
16506 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
16509 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
16510 const TargetRegisterInfo *TRI = DAG.getSubtarget().getRegisterInfo();
16512 if (Subtarget->is64Bit()) {
16513 SDValue OutChains[6];
16515 // Large code-model.
16516 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
16517 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
16519 const unsigned char N86R10 = TRI->getEncodingValue(X86::R10) & 0x7;
16520 const unsigned char N86R11 = TRI->getEncodingValue(X86::R11) & 0x7;
16522 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
16524 // Load the pointer to the nested function into R11.
16525 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
16526 SDValue Addr = Trmp;
16527 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
16528 Addr, MachinePointerInfo(TrmpAddr),
16531 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
16532 DAG.getConstant(2, MVT::i64));
16533 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
16534 MachinePointerInfo(TrmpAddr, 2),
16537 // Load the 'nest' parameter value into R10.
16538 // R10 is specified in X86CallingConv.td
16539 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
16540 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
16541 DAG.getConstant(10, MVT::i64));
16542 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
16543 Addr, MachinePointerInfo(TrmpAddr, 10),
16546 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
16547 DAG.getConstant(12, MVT::i64));
16548 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
16549 MachinePointerInfo(TrmpAddr, 12),
16552 // Jump to the nested function.
16553 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
16554 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
16555 DAG.getConstant(20, MVT::i64));
16556 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
16557 Addr, MachinePointerInfo(TrmpAddr, 20),
16560 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
16561 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
16562 DAG.getConstant(22, MVT::i64));
16563 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
16564 MachinePointerInfo(TrmpAddr, 22),
16567 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
16569 const Function *Func =
16570 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
16571 CallingConv::ID CC = Func->getCallingConv();
16576 llvm_unreachable("Unsupported calling convention");
16577 case CallingConv::C:
16578 case CallingConv::X86_StdCall: {
16579 // Pass 'nest' parameter in ECX.
16580 // Must be kept in sync with X86CallingConv.td
16581 NestReg = X86::ECX;
16583 // Check that ECX wasn't needed by an 'inreg' parameter.
16584 FunctionType *FTy = Func->getFunctionType();
16585 const AttributeSet &Attrs = Func->getAttributes();
16587 if (!Attrs.isEmpty() && !Func->isVarArg()) {
16588 unsigned InRegCount = 0;
16591 for (FunctionType::param_iterator I = FTy->param_begin(),
16592 E = FTy->param_end(); I != E; ++I, ++Idx)
16593 if (Attrs.hasAttribute(Idx, Attribute::InReg))
16594 // FIXME: should only count parameters that are lowered to integers.
16595 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
16597 if (InRegCount > 2) {
16598 report_fatal_error("Nest register in use - reduce number of inreg"
16604 case CallingConv::X86_FastCall:
16605 case CallingConv::X86_ThisCall:
16606 case CallingConv::Fast:
16607 // Pass 'nest' parameter in EAX.
16608 // Must be kept in sync with X86CallingConv.td
16609 NestReg = X86::EAX;
16613 SDValue OutChains[4];
16614 SDValue Addr, Disp;
16616 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
16617 DAG.getConstant(10, MVT::i32));
16618 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
16620 // This is storing the opcode for MOV32ri.
16621 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
16622 const unsigned char N86Reg = TRI->getEncodingValue(NestReg) & 0x7;
16623 OutChains[0] = DAG.getStore(Root, dl,
16624 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
16625 Trmp, MachinePointerInfo(TrmpAddr),
16628 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
16629 DAG.getConstant(1, MVT::i32));
16630 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
16631 MachinePointerInfo(TrmpAddr, 1),
16634 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
16635 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
16636 DAG.getConstant(5, MVT::i32));
16637 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
16638 MachinePointerInfo(TrmpAddr, 5),
16641 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
16642 DAG.getConstant(6, MVT::i32));
16643 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
16644 MachinePointerInfo(TrmpAddr, 6),
16647 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
16651 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
16652 SelectionDAG &DAG) const {
16654 The rounding mode is in bits 11:10 of FPSR, and has the following
16656 00 Round to nearest
16661 FLT_ROUNDS, on the other hand, expects the following:
16668 To perform the conversion, we do:
16669 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
16672 MachineFunction &MF = DAG.getMachineFunction();
16673 const TargetMachine &TM = MF.getTarget();
16674 const TargetFrameLowering &TFI = *TM.getSubtargetImpl()->getFrameLowering();
16675 unsigned StackAlignment = TFI.getStackAlignment();
16676 MVT VT = Op.getSimpleValueType();
16679 // Save FP Control Word to stack slot
16680 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
16681 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
16683 MachineMemOperand *MMO =
16684 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
16685 MachineMemOperand::MOStore, 2, 2);
16687 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
16688 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
16689 DAG.getVTList(MVT::Other),
16690 Ops, MVT::i16, MMO);
16692 // Load FP Control Word from stack slot
16693 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
16694 MachinePointerInfo(), false, false, false, 0);
16696 // Transform as necessary
16698 DAG.getNode(ISD::SRL, DL, MVT::i16,
16699 DAG.getNode(ISD::AND, DL, MVT::i16,
16700 CWD, DAG.getConstant(0x800, MVT::i16)),
16701 DAG.getConstant(11, MVT::i8));
16703 DAG.getNode(ISD::SRL, DL, MVT::i16,
16704 DAG.getNode(ISD::AND, DL, MVT::i16,
16705 CWD, DAG.getConstant(0x400, MVT::i16)),
16706 DAG.getConstant(9, MVT::i8));
16709 DAG.getNode(ISD::AND, DL, MVT::i16,
16710 DAG.getNode(ISD::ADD, DL, MVT::i16,
16711 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
16712 DAG.getConstant(1, MVT::i16)),
16713 DAG.getConstant(3, MVT::i16));
16715 return DAG.getNode((VT.getSizeInBits() < 16 ?
16716 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
16719 static SDValue LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
16720 MVT VT = Op.getSimpleValueType();
16722 unsigned NumBits = VT.getSizeInBits();
16725 Op = Op.getOperand(0);
16726 if (VT == MVT::i8) {
16727 // Zero extend to i32 since there is not an i8 bsr.
16729 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
16732 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
16733 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
16734 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
16736 // If src is zero (i.e. bsr sets ZF), returns NumBits.
16739 DAG.getConstant(NumBits+NumBits-1, OpVT),
16740 DAG.getConstant(X86::COND_E, MVT::i8),
16743 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops);
16745 // Finally xor with NumBits-1.
16746 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
16749 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
16753 static SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, SelectionDAG &DAG) {
16754 MVT VT = Op.getSimpleValueType();
16756 unsigned NumBits = VT.getSizeInBits();
16759 Op = Op.getOperand(0);
16760 if (VT == MVT::i8) {
16761 // Zero extend to i32 since there is not an i8 bsr.
16763 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
16766 // Issue a bsr (scan bits in reverse).
16767 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
16768 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
16770 // And xor with NumBits-1.
16771 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
16774 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
16778 static SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
16779 MVT VT = Op.getSimpleValueType();
16780 unsigned NumBits = VT.getSizeInBits();
16782 Op = Op.getOperand(0);
16784 // Issue a bsf (scan bits forward) which also sets EFLAGS.
16785 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
16786 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
16788 // If src is zero (i.e. bsf sets ZF), returns NumBits.
16791 DAG.getConstant(NumBits, VT),
16792 DAG.getConstant(X86::COND_E, MVT::i8),
16795 return DAG.getNode(X86ISD::CMOV, dl, VT, Ops);
16798 // Lower256IntArith - Break a 256-bit integer operation into two new 128-bit
16799 // ones, and then concatenate the result back.
16800 static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) {
16801 MVT VT = Op.getSimpleValueType();
16803 assert(VT.is256BitVector() && VT.isInteger() &&
16804 "Unsupported value type for operation");
16806 unsigned NumElems = VT.getVectorNumElements();
16809 // Extract the LHS vectors
16810 SDValue LHS = Op.getOperand(0);
16811 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
16812 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
16814 // Extract the RHS vectors
16815 SDValue RHS = Op.getOperand(1);
16816 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
16817 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
16819 MVT EltVT = VT.getVectorElementType();
16820 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
16822 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
16823 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1),
16824 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
16827 static SDValue LowerADD(SDValue Op, SelectionDAG &DAG) {
16828 assert(Op.getSimpleValueType().is256BitVector() &&
16829 Op.getSimpleValueType().isInteger() &&
16830 "Only handle AVX 256-bit vector integer operation");
16831 return Lower256IntArith(Op, DAG);
16834 static SDValue LowerSUB(SDValue Op, SelectionDAG &DAG) {
16835 assert(Op.getSimpleValueType().is256BitVector() &&
16836 Op.getSimpleValueType().isInteger() &&
16837 "Only handle AVX 256-bit vector integer operation");
16838 return Lower256IntArith(Op, DAG);
16841 static SDValue LowerMUL(SDValue Op, const X86Subtarget *Subtarget,
16842 SelectionDAG &DAG) {
16844 MVT VT = Op.getSimpleValueType();
16846 // Decompose 256-bit ops into smaller 128-bit ops.
16847 if (VT.is256BitVector() && !Subtarget->hasInt256())
16848 return Lower256IntArith(Op, DAG);
16850 SDValue A = Op.getOperand(0);
16851 SDValue B = Op.getOperand(1);
16853 // Lower v4i32 mul as 2x shuffle, 2x pmuludq, 2x shuffle.
16854 if (VT == MVT::v4i32) {
16855 assert(Subtarget->hasSSE2() && !Subtarget->hasSSE41() &&
16856 "Should not custom lower when pmuldq is available!");
16858 // Extract the odd parts.
16859 static const int UnpackMask[] = { 1, -1, 3, -1 };
16860 SDValue Aodds = DAG.getVectorShuffle(VT, dl, A, A, UnpackMask);
16861 SDValue Bodds = DAG.getVectorShuffle(VT, dl, B, B, UnpackMask);
16863 // Multiply the even parts.
16864 SDValue Evens = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, A, B);
16865 // Now multiply odd parts.
16866 SDValue Odds = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, Aodds, Bodds);
16868 Evens = DAG.getNode(ISD::BITCAST, dl, VT, Evens);
16869 Odds = DAG.getNode(ISD::BITCAST, dl, VT, Odds);
16871 // Merge the two vectors back together with a shuffle. This expands into 2
16873 static const int ShufMask[] = { 0, 4, 2, 6 };
16874 return DAG.getVectorShuffle(VT, dl, Evens, Odds, ShufMask);
16877 assert((VT == MVT::v2i64 || VT == MVT::v4i64 || VT == MVT::v8i64) &&
16878 "Only know how to lower V2I64/V4I64/V8I64 multiply");
16880 // Ahi = psrlqi(a, 32);
16881 // Bhi = psrlqi(b, 32);
16883 // AloBlo = pmuludq(a, b);
16884 // AloBhi = pmuludq(a, Bhi);
16885 // AhiBlo = pmuludq(Ahi, b);
16887 // AloBhi = psllqi(AloBhi, 32);
16888 // AhiBlo = psllqi(AhiBlo, 32);
16889 // return AloBlo + AloBhi + AhiBlo;
16891 SDValue Ahi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, A, 32, DAG);
16892 SDValue Bhi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, B, 32, DAG);
16894 // Bit cast to 32-bit vectors for MULUDQ
16895 EVT MulVT = (VT == MVT::v2i64) ? MVT::v4i32 :
16896 (VT == MVT::v4i64) ? MVT::v8i32 : MVT::v16i32;
16897 A = DAG.getNode(ISD::BITCAST, dl, MulVT, A);
16898 B = DAG.getNode(ISD::BITCAST, dl, MulVT, B);
16899 Ahi = DAG.getNode(ISD::BITCAST, dl, MulVT, Ahi);
16900 Bhi = DAG.getNode(ISD::BITCAST, dl, MulVT, Bhi);
16902 SDValue AloBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, B);
16903 SDValue AloBhi = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, Bhi);
16904 SDValue AhiBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Ahi, B);
16906 AloBhi = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AloBhi, 32, DAG);
16907 AhiBlo = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AhiBlo, 32, DAG);
16909 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
16910 return DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
16913 SDValue X86TargetLowering::LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const {
16914 assert(Subtarget->isTargetWin64() && "Unexpected target");
16915 EVT VT = Op.getValueType();
16916 assert(VT.isInteger() && VT.getSizeInBits() == 128 &&
16917 "Unexpected return type for lowering");
16921 switch (Op->getOpcode()) {
16922 default: llvm_unreachable("Unexpected request for libcall!");
16923 case ISD::SDIV: isSigned = true; LC = RTLIB::SDIV_I128; break;
16924 case ISD::UDIV: isSigned = false; LC = RTLIB::UDIV_I128; break;
16925 case ISD::SREM: isSigned = true; LC = RTLIB::SREM_I128; break;
16926 case ISD::UREM: isSigned = false; LC = RTLIB::UREM_I128; break;
16927 case ISD::SDIVREM: isSigned = true; LC = RTLIB::SDIVREM_I128; break;
16928 case ISD::UDIVREM: isSigned = false; LC = RTLIB::UDIVREM_I128; break;
16932 SDValue InChain = DAG.getEntryNode();
16934 TargetLowering::ArgListTy Args;
16935 TargetLowering::ArgListEntry Entry;
16936 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
16937 EVT ArgVT = Op->getOperand(i).getValueType();
16938 assert(ArgVT.isInteger() && ArgVT.getSizeInBits() == 128 &&
16939 "Unexpected argument type for lowering");
16940 SDValue StackPtr = DAG.CreateStackTemporary(ArgVT, 16);
16941 Entry.Node = StackPtr;
16942 InChain = DAG.getStore(InChain, dl, Op->getOperand(i), StackPtr, MachinePointerInfo(),
16944 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
16945 Entry.Ty = PointerType::get(ArgTy,0);
16946 Entry.isSExt = false;
16947 Entry.isZExt = false;
16948 Args.push_back(Entry);
16951 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
16954 TargetLowering::CallLoweringInfo CLI(DAG);
16955 CLI.setDebugLoc(dl).setChain(InChain)
16956 .setCallee(getLibcallCallingConv(LC),
16957 static_cast<EVT>(MVT::v2i64).getTypeForEVT(*DAG.getContext()),
16958 Callee, std::move(Args), 0)
16959 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
16961 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
16962 return DAG.getNode(ISD::BITCAST, dl, VT, CallInfo.first);
16965 static SDValue LowerMUL_LOHI(SDValue Op, const X86Subtarget *Subtarget,
16966 SelectionDAG &DAG) {
16967 SDValue Op0 = Op.getOperand(0), Op1 = Op.getOperand(1);
16968 EVT VT = Op0.getValueType();
16971 assert((VT == MVT::v4i32 && Subtarget->hasSSE2()) ||
16972 (VT == MVT::v8i32 && Subtarget->hasInt256()));
16974 // PMULxD operations multiply each even value (starting at 0) of LHS with
16975 // the related value of RHS and produce a widen result.
16976 // E.g., PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
16977 // => <2 x i64> <ae|cg>
16979 // In other word, to have all the results, we need to perform two PMULxD:
16980 // 1. one with the even values.
16981 // 2. one with the odd values.
16982 // To achieve #2, with need to place the odd values at an even position.
16984 // Place the odd value at an even position (basically, shift all values 1
16985 // step to the left):
16986 const int Mask[] = {1, -1, 3, -1, 5, -1, 7, -1};
16987 // <a|b|c|d> => <b|undef|d|undef>
16988 SDValue Odd0 = DAG.getVectorShuffle(VT, dl, Op0, Op0, Mask);
16989 // <e|f|g|h> => <f|undef|h|undef>
16990 SDValue Odd1 = DAG.getVectorShuffle(VT, dl, Op1, Op1, Mask);
16992 // Emit two multiplies, one for the lower 2 ints and one for the higher 2
16994 MVT MulVT = VT == MVT::v4i32 ? MVT::v2i64 : MVT::v4i64;
16995 bool IsSigned = Op->getOpcode() == ISD::SMUL_LOHI;
16997 (!IsSigned || !Subtarget->hasSSE41()) ? X86ISD::PMULUDQ : X86ISD::PMULDQ;
16998 // PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
16999 // => <2 x i64> <ae|cg>
17000 SDValue Mul1 = DAG.getNode(ISD::BITCAST, dl, VT,
17001 DAG.getNode(Opcode, dl, MulVT, Op0, Op1));
17002 // PMULUDQ <4 x i32> <b|undef|d|undef>, <4 x i32> <f|undef|h|undef>
17003 // => <2 x i64> <bf|dh>
17004 SDValue Mul2 = DAG.getNode(ISD::BITCAST, dl, VT,
17005 DAG.getNode(Opcode, dl, MulVT, Odd0, Odd1));
17007 // Shuffle it back into the right order.
17008 SDValue Highs, Lows;
17009 if (VT == MVT::v8i32) {
17010 const int HighMask[] = {1, 9, 3, 11, 5, 13, 7, 15};
17011 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
17012 const int LowMask[] = {0, 8, 2, 10, 4, 12, 6, 14};
17013 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
17015 const int HighMask[] = {1, 5, 3, 7};
17016 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
17017 const int LowMask[] = {0, 4, 2, 6};
17018 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
17021 // If we have a signed multiply but no PMULDQ fix up the high parts of a
17022 // unsigned multiply.
17023 if (IsSigned && !Subtarget->hasSSE41()) {
17025 DAG.getConstant(31, DAG.getTargetLoweringInfo().getShiftAmountTy(VT));
17026 SDValue T1 = DAG.getNode(ISD::AND, dl, VT,
17027 DAG.getNode(ISD::SRA, dl, VT, Op0, ShAmt), Op1);
17028 SDValue T2 = DAG.getNode(ISD::AND, dl, VT,
17029 DAG.getNode(ISD::SRA, dl, VT, Op1, ShAmt), Op0);
17031 SDValue Fixup = DAG.getNode(ISD::ADD, dl, VT, T1, T2);
17032 Highs = DAG.getNode(ISD::SUB, dl, VT, Highs, Fixup);
17035 // The first result of MUL_LOHI is actually the low value, followed by the
17037 SDValue Ops[] = {Lows, Highs};
17038 return DAG.getMergeValues(Ops, dl);
17041 static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
17042 const X86Subtarget *Subtarget) {
17043 MVT VT = Op.getSimpleValueType();
17045 SDValue R = Op.getOperand(0);
17046 SDValue Amt = Op.getOperand(1);
17048 // Optimize shl/srl/sra with constant shift amount.
17049 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
17050 if (auto *ShiftConst = BVAmt->getConstantSplatNode()) {
17051 uint64_t ShiftAmt = ShiftConst->getZExtValue();
17053 if (VT == MVT::v2i64 || VT == MVT::v4i32 || VT == MVT::v8i16 ||
17054 (Subtarget->hasInt256() &&
17055 (VT == MVT::v4i64 || VT == MVT::v8i32 || VT == MVT::v16i16)) ||
17056 (Subtarget->hasAVX512() &&
17057 (VT == MVT::v8i64 || VT == MVT::v16i32))) {
17058 if (Op.getOpcode() == ISD::SHL)
17059 return getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, R, ShiftAmt,
17061 if (Op.getOpcode() == ISD::SRL)
17062 return getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt,
17064 if (Op.getOpcode() == ISD::SRA && VT != MVT::v2i64 && VT != MVT::v4i64)
17065 return getTargetVShiftByConstNode(X86ISD::VSRAI, dl, VT, R, ShiftAmt,
17069 if (VT == MVT::v16i8) {
17070 if (Op.getOpcode() == ISD::SHL) {
17071 // Make a large shift.
17072 SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl,
17073 MVT::v8i16, R, ShiftAmt,
17075 SHL = DAG.getNode(ISD::BITCAST, dl, VT, SHL);
17076 // Zero out the rightmost bits.
17077 SmallVector<SDValue, 16> V(16,
17078 DAG.getConstant(uint8_t(-1U << ShiftAmt),
17080 return DAG.getNode(ISD::AND, dl, VT, SHL,
17081 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
17083 if (Op.getOpcode() == ISD::SRL) {
17084 // Make a large shift.
17085 SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl,
17086 MVT::v8i16, R, ShiftAmt,
17088 SRL = DAG.getNode(ISD::BITCAST, dl, VT, SRL);
17089 // Zero out the leftmost bits.
17090 SmallVector<SDValue, 16> V(16,
17091 DAG.getConstant(uint8_t(-1U) >> ShiftAmt,
17093 return DAG.getNode(ISD::AND, dl, VT, SRL,
17094 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
17096 if (Op.getOpcode() == ISD::SRA) {
17097 if (ShiftAmt == 7) {
17098 // R s>> 7 === R s< 0
17099 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
17100 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
17103 // R s>> a === ((R u>> a) ^ m) - m
17104 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
17105 SmallVector<SDValue, 16> V(16, DAG.getConstant(128 >> ShiftAmt,
17107 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V);
17108 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
17109 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
17112 llvm_unreachable("Unknown shift opcode.");
17115 if (Subtarget->hasInt256() && VT == MVT::v32i8) {
17116 if (Op.getOpcode() == ISD::SHL) {
17117 // Make a large shift.
17118 SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl,
17119 MVT::v16i16, R, ShiftAmt,
17121 SHL = DAG.getNode(ISD::BITCAST, dl, VT, SHL);
17122 // Zero out the rightmost bits.
17123 SmallVector<SDValue, 32> V(32,
17124 DAG.getConstant(uint8_t(-1U << ShiftAmt),
17126 return DAG.getNode(ISD::AND, dl, VT, SHL,
17127 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
17129 if (Op.getOpcode() == ISD::SRL) {
17130 // Make a large shift.
17131 SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl,
17132 MVT::v16i16, R, ShiftAmt,
17134 SRL = DAG.getNode(ISD::BITCAST, dl, VT, SRL);
17135 // Zero out the leftmost bits.
17136 SmallVector<SDValue, 32> V(32,
17137 DAG.getConstant(uint8_t(-1U) >> ShiftAmt,
17139 return DAG.getNode(ISD::AND, dl, VT, SRL,
17140 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
17142 if (Op.getOpcode() == ISD::SRA) {
17143 if (ShiftAmt == 7) {
17144 // R s>> 7 === R s< 0
17145 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
17146 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
17149 // R s>> a === ((R u>> a) ^ m) - m
17150 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
17151 SmallVector<SDValue, 32> V(32, DAG.getConstant(128 >> ShiftAmt,
17153 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V);
17154 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
17155 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
17158 llvm_unreachable("Unknown shift opcode.");
17163 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
17164 if (!Subtarget->is64Bit() &&
17165 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
17166 Amt.getOpcode() == ISD::BITCAST &&
17167 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
17168 Amt = Amt.getOperand(0);
17169 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
17170 VT.getVectorNumElements();
17171 unsigned RatioInLog2 = Log2_32_Ceil(Ratio);
17172 uint64_t ShiftAmt = 0;
17173 for (unsigned i = 0; i != Ratio; ++i) {
17174 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i));
17178 ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2)));
17180 // Check remaining shift amounts.
17181 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
17182 uint64_t ShAmt = 0;
17183 for (unsigned j = 0; j != Ratio; ++j) {
17184 ConstantSDNode *C =
17185 dyn_cast<ConstantSDNode>(Amt.getOperand(i + j));
17189 ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2)));
17191 if (ShAmt != ShiftAmt)
17194 switch (Op.getOpcode()) {
17196 llvm_unreachable("Unknown shift opcode!");
17198 return getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, R, ShiftAmt,
17201 return getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt,
17204 return getTargetVShiftByConstNode(X86ISD::VSRAI, dl, VT, R, ShiftAmt,
17212 static SDValue LowerScalarVariableShift(SDValue Op, SelectionDAG &DAG,
17213 const X86Subtarget* Subtarget) {
17214 MVT VT = Op.getSimpleValueType();
17216 SDValue R = Op.getOperand(0);
17217 SDValue Amt = Op.getOperand(1);
17219 if ((VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) ||
17220 VT == MVT::v4i32 || VT == MVT::v8i16 ||
17221 (Subtarget->hasInt256() &&
17222 ((VT == MVT::v4i64 && Op.getOpcode() != ISD::SRA) ||
17223 VT == MVT::v8i32 || VT == MVT::v16i16)) ||
17224 (Subtarget->hasAVX512() && (VT == MVT::v8i64 || VT == MVT::v16i32))) {
17226 EVT EltVT = VT.getVectorElementType();
17228 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
17229 unsigned NumElts = VT.getVectorNumElements();
17231 for (i = 0; i != NumElts; ++i) {
17232 if (Amt.getOperand(i).getOpcode() == ISD::UNDEF)
17236 for (j = i; j != NumElts; ++j) {
17237 SDValue Arg = Amt.getOperand(j);
17238 if (Arg.getOpcode() == ISD::UNDEF) continue;
17239 if (Arg != Amt.getOperand(i))
17242 if (i != NumElts && j == NumElts)
17243 BaseShAmt = Amt.getOperand(i);
17245 if (Amt.getOpcode() == ISD::EXTRACT_SUBVECTOR)
17246 Amt = Amt.getOperand(0);
17247 if (Amt.getOpcode() == ISD::VECTOR_SHUFFLE &&
17248 cast<ShuffleVectorSDNode>(Amt)->isSplat()) {
17249 SDValue InVec = Amt.getOperand(0);
17250 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
17251 unsigned NumElts = InVec.getValueType().getVectorNumElements();
17253 for (; i != NumElts; ++i) {
17254 SDValue Arg = InVec.getOperand(i);
17255 if (Arg.getOpcode() == ISD::UNDEF) continue;
17259 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
17260 if (ConstantSDNode *C =
17261 dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
17262 unsigned SplatIdx =
17263 cast<ShuffleVectorSDNode>(Amt)->getSplatIndex();
17264 if (C->getZExtValue() == SplatIdx)
17265 BaseShAmt = InVec.getOperand(1);
17268 if (!BaseShAmt.getNode())
17269 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, Amt,
17270 DAG.getIntPtrConstant(0));
17274 if (BaseShAmt.getNode()) {
17275 if (EltVT.bitsGT(MVT::i32))
17276 BaseShAmt = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BaseShAmt);
17277 else if (EltVT.bitsLT(MVT::i32))
17278 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt);
17280 switch (Op.getOpcode()) {
17282 llvm_unreachable("Unknown shift opcode!");
17284 switch (VT.SimpleTy) {
17285 default: return SDValue();
17294 return getTargetVShiftNode(X86ISD::VSHLI, dl, VT, R, BaseShAmt, DAG);
17297 switch (VT.SimpleTy) {
17298 default: return SDValue();
17305 return getTargetVShiftNode(X86ISD::VSRAI, dl, VT, R, BaseShAmt, DAG);
17308 switch (VT.SimpleTy) {
17309 default: return SDValue();
17318 return getTargetVShiftNode(X86ISD::VSRLI, dl, VT, R, BaseShAmt, DAG);
17324 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
17325 if (!Subtarget->is64Bit() &&
17326 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64) ||
17327 (Subtarget->hasAVX512() && VT == MVT::v8i64)) &&
17328 Amt.getOpcode() == ISD::BITCAST &&
17329 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
17330 Amt = Amt.getOperand(0);
17331 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
17332 VT.getVectorNumElements();
17333 std::vector<SDValue> Vals(Ratio);
17334 for (unsigned i = 0; i != Ratio; ++i)
17335 Vals[i] = Amt.getOperand(i);
17336 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
17337 for (unsigned j = 0; j != Ratio; ++j)
17338 if (Vals[j] != Amt.getOperand(i + j))
17341 switch (Op.getOpcode()) {
17343 llvm_unreachable("Unknown shift opcode!");
17345 return DAG.getNode(X86ISD::VSHL, dl, VT, R, Op.getOperand(1));
17347 return DAG.getNode(X86ISD::VSRL, dl, VT, R, Op.getOperand(1));
17349 return DAG.getNode(X86ISD::VSRA, dl, VT, R, Op.getOperand(1));
17356 static SDValue LowerShift(SDValue Op, const X86Subtarget* Subtarget,
17357 SelectionDAG &DAG) {
17358 MVT VT = Op.getSimpleValueType();
17360 SDValue R = Op.getOperand(0);
17361 SDValue Amt = Op.getOperand(1);
17364 assert(VT.isVector() && "Custom lowering only for vector shifts!");
17365 assert(Subtarget->hasSSE2() && "Only custom lower when we have SSE2!");
17367 V = LowerScalarImmediateShift(Op, DAG, Subtarget);
17371 V = LowerScalarVariableShift(Op, DAG, Subtarget);
17375 if (Subtarget->hasAVX512() && (VT == MVT::v16i32 || VT == MVT::v8i64))
17377 // AVX2 has VPSLLV/VPSRAV/VPSRLV.
17378 if (Subtarget->hasInt256()) {
17379 if (Op.getOpcode() == ISD::SRL &&
17380 (VT == MVT::v2i64 || VT == MVT::v4i32 ||
17381 VT == MVT::v4i64 || VT == MVT::v8i32))
17383 if (Op.getOpcode() == ISD::SHL &&
17384 (VT == MVT::v2i64 || VT == MVT::v4i32 ||
17385 VT == MVT::v4i64 || VT == MVT::v8i32))
17387 if (Op.getOpcode() == ISD::SRA && (VT == MVT::v4i32 || VT == MVT::v8i32))
17391 // If possible, lower this packed shift into a vector multiply instead of
17392 // expanding it into a sequence of scalar shifts.
17393 // Do this only if the vector shift count is a constant build_vector.
17394 if (Op.getOpcode() == ISD::SHL &&
17395 (VT == MVT::v8i16 || VT == MVT::v4i32 ||
17396 (Subtarget->hasInt256() && VT == MVT::v16i16)) &&
17397 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
17398 SmallVector<SDValue, 8> Elts;
17399 EVT SVT = VT.getScalarType();
17400 unsigned SVTBits = SVT.getSizeInBits();
17401 const APInt &One = APInt(SVTBits, 1);
17402 unsigned NumElems = VT.getVectorNumElements();
17404 for (unsigned i=0; i !=NumElems; ++i) {
17405 SDValue Op = Amt->getOperand(i);
17406 if (Op->getOpcode() == ISD::UNDEF) {
17407 Elts.push_back(Op);
17411 ConstantSDNode *ND = cast<ConstantSDNode>(Op);
17412 const APInt &C = APInt(SVTBits, ND->getAPIntValue().getZExtValue());
17413 uint64_t ShAmt = C.getZExtValue();
17414 if (ShAmt >= SVTBits) {
17415 Elts.push_back(DAG.getUNDEF(SVT));
17418 Elts.push_back(DAG.getConstant(One.shl(ShAmt), SVT));
17420 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
17421 return DAG.getNode(ISD::MUL, dl, VT, R, BV);
17424 // Lower SHL with variable shift amount.
17425 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
17426 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, VT));
17428 Op = DAG.getNode(ISD::ADD, dl, VT, Op, DAG.getConstant(0x3f800000U, VT));
17429 Op = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, Op);
17430 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
17431 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
17434 // If possible, lower this shift as a sequence of two shifts by
17435 // constant plus a MOVSS/MOVSD instead of scalarizing it.
17437 // (v4i32 (srl A, (build_vector < X, Y, Y, Y>)))
17439 // Could be rewritten as:
17440 // (v4i32 (MOVSS (srl A, <Y,Y,Y,Y>), (srl A, <X,X,X,X>)))
17442 // The advantage is that the two shifts from the example would be
17443 // lowered as X86ISD::VSRLI nodes. This would be cheaper than scalarizing
17444 // the vector shift into four scalar shifts plus four pairs of vector
17446 if ((VT == MVT::v8i16 || VT == MVT::v4i32) &&
17447 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
17448 unsigned TargetOpcode = X86ISD::MOVSS;
17449 bool CanBeSimplified;
17450 // The splat value for the first packed shift (the 'X' from the example).
17451 SDValue Amt1 = Amt->getOperand(0);
17452 // The splat value for the second packed shift (the 'Y' from the example).
17453 SDValue Amt2 = (VT == MVT::v4i32) ? Amt->getOperand(1) :
17454 Amt->getOperand(2);
17456 // See if it is possible to replace this node with a sequence of
17457 // two shifts followed by a MOVSS/MOVSD
17458 if (VT == MVT::v4i32) {
17459 // Check if it is legal to use a MOVSS.
17460 CanBeSimplified = Amt2 == Amt->getOperand(2) &&
17461 Amt2 == Amt->getOperand(3);
17462 if (!CanBeSimplified) {
17463 // Otherwise, check if we can still simplify this node using a MOVSD.
17464 CanBeSimplified = Amt1 == Amt->getOperand(1) &&
17465 Amt->getOperand(2) == Amt->getOperand(3);
17466 TargetOpcode = X86ISD::MOVSD;
17467 Amt2 = Amt->getOperand(2);
17470 // Do similar checks for the case where the machine value type
17472 CanBeSimplified = Amt1 == Amt->getOperand(1);
17473 for (unsigned i=3; i != 8 && CanBeSimplified; ++i)
17474 CanBeSimplified = Amt2 == Amt->getOperand(i);
17476 if (!CanBeSimplified) {
17477 TargetOpcode = X86ISD::MOVSD;
17478 CanBeSimplified = true;
17479 Amt2 = Amt->getOperand(4);
17480 for (unsigned i=0; i != 4 && CanBeSimplified; ++i)
17481 CanBeSimplified = Amt1 == Amt->getOperand(i);
17482 for (unsigned j=4; j != 8 && CanBeSimplified; ++j)
17483 CanBeSimplified = Amt2 == Amt->getOperand(j);
17487 if (CanBeSimplified && isa<ConstantSDNode>(Amt1) &&
17488 isa<ConstantSDNode>(Amt2)) {
17489 // Replace this node with two shifts followed by a MOVSS/MOVSD.
17490 EVT CastVT = MVT::v4i32;
17492 DAG.getConstant(cast<ConstantSDNode>(Amt1)->getAPIntValue(), VT);
17493 SDValue Shift1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat1);
17495 DAG.getConstant(cast<ConstantSDNode>(Amt2)->getAPIntValue(), VT);
17496 SDValue Shift2 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat2);
17497 if (TargetOpcode == X86ISD::MOVSD)
17498 CastVT = MVT::v2i64;
17499 SDValue BitCast1 = DAG.getNode(ISD::BITCAST, dl, CastVT, Shift1);
17500 SDValue BitCast2 = DAG.getNode(ISD::BITCAST, dl, CastVT, Shift2);
17501 SDValue Result = getTargetShuffleNode(TargetOpcode, dl, CastVT, BitCast2,
17503 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
17507 if (VT == MVT::v16i8 && Op->getOpcode() == ISD::SHL) {
17508 assert(Subtarget->hasSSE2() && "Need SSE2 for pslli/pcmpeq.");
17511 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(5, VT));
17512 Op = DAG.getNode(ISD::BITCAST, dl, VT, Op);
17514 // Turn 'a' into a mask suitable for VSELECT
17515 SDValue VSelM = DAG.getConstant(0x80, VT);
17516 SDValue OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
17517 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
17519 SDValue CM1 = DAG.getConstant(0x0f, VT);
17520 SDValue CM2 = DAG.getConstant(0x3f, VT);
17522 // r = VSELECT(r, psllw(r & (char16)15, 4), a);
17523 SDValue M = DAG.getNode(ISD::AND, dl, VT, R, CM1);
17524 M = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, MVT::v8i16, M, 4, DAG);
17525 M = DAG.getNode(ISD::BITCAST, dl, VT, M);
17526 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, M, R);
17529 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
17530 OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
17531 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
17533 // r = VSELECT(r, psllw(r & (char16)63, 2), a);
17534 M = DAG.getNode(ISD::AND, dl, VT, R, CM2);
17535 M = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, MVT::v8i16, M, 2, DAG);
17536 M = DAG.getNode(ISD::BITCAST, dl, VT, M);
17537 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, M, R);
17540 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
17541 OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
17542 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
17544 // return VSELECT(r, r+r, a);
17545 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel,
17546 DAG.getNode(ISD::ADD, dl, VT, R, R), R);
17550 // It's worth extending once and using the v8i32 shifts for 16-bit types, but
17551 // the extra overheads to get from v16i8 to v8i32 make the existing SSE
17552 // solution better.
17553 if (Subtarget->hasInt256() && VT == MVT::v8i16) {
17554 MVT NewVT = VT == MVT::v8i16 ? MVT::v8i32 : MVT::v16i16;
17556 Op.getOpcode() == ISD::SRA ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
17557 R = DAG.getNode(ExtOpc, dl, NewVT, R);
17558 Amt = DAG.getNode(ISD::ANY_EXTEND, dl, NewVT, Amt);
17559 return DAG.getNode(ISD::TRUNCATE, dl, VT,
17560 DAG.getNode(Op.getOpcode(), dl, NewVT, R, Amt));
17563 // Decompose 256-bit shifts into smaller 128-bit shifts.
17564 if (VT.is256BitVector()) {
17565 unsigned NumElems = VT.getVectorNumElements();
17566 MVT EltVT = VT.getVectorElementType();
17567 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
17569 // Extract the two vectors
17570 SDValue V1 = Extract128BitVector(R, 0, DAG, dl);
17571 SDValue V2 = Extract128BitVector(R, NumElems/2, DAG, dl);
17573 // Recreate the shift amount vectors
17574 SDValue Amt1, Amt2;
17575 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
17576 // Constant shift amount
17577 SmallVector<SDValue, 4> Amt1Csts;
17578 SmallVector<SDValue, 4> Amt2Csts;
17579 for (unsigned i = 0; i != NumElems/2; ++i)
17580 Amt1Csts.push_back(Amt->getOperand(i));
17581 for (unsigned i = NumElems/2; i != NumElems; ++i)
17582 Amt2Csts.push_back(Amt->getOperand(i));
17584 Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt1Csts);
17585 Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt2Csts);
17587 // Variable shift amount
17588 Amt1 = Extract128BitVector(Amt, 0, DAG, dl);
17589 Amt2 = Extract128BitVector(Amt, NumElems/2, DAG, dl);
17592 // Issue new vector shifts for the smaller types
17593 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
17594 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
17596 // Concatenate the result back
17597 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
17603 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
17604 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
17605 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
17606 // looks for this combo and may remove the "setcc" instruction if the "setcc"
17607 // has only one use.
17608 SDNode *N = Op.getNode();
17609 SDValue LHS = N->getOperand(0);
17610 SDValue RHS = N->getOperand(1);
17611 unsigned BaseOp = 0;
17614 switch (Op.getOpcode()) {
17615 default: llvm_unreachable("Unknown ovf instruction!");
17617 // A subtract of one will be selected as a INC. Note that INC doesn't
17618 // set CF, so we can't do this for UADDO.
17619 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
17621 BaseOp = X86ISD::INC;
17622 Cond = X86::COND_O;
17625 BaseOp = X86ISD::ADD;
17626 Cond = X86::COND_O;
17629 BaseOp = X86ISD::ADD;
17630 Cond = X86::COND_B;
17633 // A subtract of one will be selected as a DEC. Note that DEC doesn't
17634 // set CF, so we can't do this for USUBO.
17635 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
17637 BaseOp = X86ISD::DEC;
17638 Cond = X86::COND_O;
17641 BaseOp = X86ISD::SUB;
17642 Cond = X86::COND_O;
17645 BaseOp = X86ISD::SUB;
17646 Cond = X86::COND_B;
17649 BaseOp = X86ISD::SMUL;
17650 Cond = X86::COND_O;
17652 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
17653 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
17655 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
17658 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
17659 DAG.getConstant(X86::COND_O, MVT::i32),
17660 SDValue(Sum.getNode(), 2));
17662 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
17666 // Also sets EFLAGS.
17667 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
17668 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
17671 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
17672 DAG.getConstant(Cond, MVT::i32),
17673 SDValue(Sum.getNode(), 1));
17675 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
17678 // Sign extension of the low part of vector elements. This may be used either
17679 // when sign extend instructions are not available or if the vector element
17680 // sizes already match the sign-extended size. If the vector elements are in
17681 // their pre-extended size and sign extend instructions are available, that will
17682 // be handled by LowerSIGN_EXTEND.
17683 SDValue X86TargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
17684 SelectionDAG &DAG) const {
17686 EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
17687 MVT VT = Op.getSimpleValueType();
17689 if (!Subtarget->hasSSE2() || !VT.isVector())
17692 unsigned BitsDiff = VT.getScalarType().getSizeInBits() -
17693 ExtraVT.getScalarType().getSizeInBits();
17695 switch (VT.SimpleTy) {
17696 default: return SDValue();
17699 if (!Subtarget->hasFp256())
17701 if (!Subtarget->hasInt256()) {
17702 // needs to be split
17703 unsigned NumElems = VT.getVectorNumElements();
17705 // Extract the LHS vectors
17706 SDValue LHS = Op.getOperand(0);
17707 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
17708 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
17710 MVT EltVT = VT.getVectorElementType();
17711 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
17713 EVT ExtraEltVT = ExtraVT.getVectorElementType();
17714 unsigned ExtraNumElems = ExtraVT.getVectorNumElements();
17715 ExtraVT = EVT::getVectorVT(*DAG.getContext(), ExtraEltVT,
17717 SDValue Extra = DAG.getValueType(ExtraVT);
17719 LHS1 = DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, Extra);
17720 LHS2 = DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, Extra);
17722 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, LHS1, LHS2);
17727 SDValue Op0 = Op.getOperand(0);
17729 // This is a sign extension of some low part of vector elements without
17730 // changing the size of the vector elements themselves:
17731 // Shift-Left + Shift-Right-Algebraic.
17732 SDValue Shl = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, Op0,
17734 return getTargetVShiftByConstNode(X86ISD::VSRAI, dl, VT, Shl, BitsDiff,
17740 /// Returns true if the operand type is exactly twice the native width, and
17741 /// the corresponding cmpxchg8b or cmpxchg16b instruction is available.
17742 /// Used to know whether to use cmpxchg8/16b when expanding atomic operations
17743 /// (otherwise we leave them alone to become __sync_fetch_and_... calls).
17744 bool X86TargetLowering::needsCmpXchgNb(const Type *MemType) const {
17745 const X86Subtarget &Subtarget =
17746 getTargetMachine().getSubtarget<X86Subtarget>();
17747 unsigned OpWidth = MemType->getPrimitiveSizeInBits();
17750 return !Subtarget.is64Bit(); // FIXME this should be Subtarget.hasCmpxchg8b
17751 else if (OpWidth == 128)
17752 return Subtarget.hasCmpxchg16b();
17757 bool X86TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
17758 return needsCmpXchgNb(SI->getValueOperand()->getType());
17761 // Note: this turns large loads into lock cmpxchg8b/16b.
17762 // FIXME: On 32 bits x86, fild/movq might be faster than lock cmpxchg8b.
17763 bool X86TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
17764 auto PTy = cast<PointerType>(LI->getPointerOperand()->getType());
17765 return needsCmpXchgNb(PTy->getElementType());
17768 bool X86TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
17769 const X86Subtarget &Subtarget =
17770 getTargetMachine().getSubtarget<X86Subtarget>();
17771 unsigned NativeWidth = Subtarget.is64Bit() ? 64 : 32;
17772 const Type *MemType = AI->getType();
17774 // If the operand is too big, we must see if cmpxchg8/16b is available
17775 // and default to library calls otherwise.
17776 if (MemType->getPrimitiveSizeInBits() > NativeWidth)
17777 return needsCmpXchgNb(MemType);
17779 AtomicRMWInst::BinOp Op = AI->getOperation();
17782 llvm_unreachable("Unknown atomic operation");
17783 case AtomicRMWInst::Xchg:
17784 case AtomicRMWInst::Add:
17785 case AtomicRMWInst::Sub:
17786 // It's better to use xadd, xsub or xchg for these in all cases.
17788 case AtomicRMWInst::Or:
17789 case AtomicRMWInst::And:
17790 case AtomicRMWInst::Xor:
17791 // If the atomicrmw's result isn't actually used, we can just add a "lock"
17792 // prefix to a normal instruction for these operations.
17793 return !AI->use_empty();
17794 case AtomicRMWInst::Nand:
17795 case AtomicRMWInst::Max:
17796 case AtomicRMWInst::Min:
17797 case AtomicRMWInst::UMax:
17798 case AtomicRMWInst::UMin:
17799 // These always require a non-trivial set of data operations on x86. We must
17800 // use a cmpxchg loop.
17805 static bool hasMFENCE(const X86Subtarget& Subtarget) {
17806 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
17807 // no-sse2). There isn't any reason to disable it if the target processor
17809 return Subtarget.hasSSE2() || Subtarget.is64Bit();
17813 X86TargetLowering::lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const {
17814 const X86Subtarget &Subtarget =
17815 getTargetMachine().getSubtarget<X86Subtarget>();
17816 unsigned NativeWidth = Subtarget.is64Bit() ? 64 : 32;
17817 const Type *MemType = AI->getType();
17818 // Accesses larger than the native width are turned into cmpxchg/libcalls, so
17819 // there is no benefit in turning such RMWs into loads, and it is actually
17820 // harmful as it introduces a mfence.
17821 if (MemType->getPrimitiveSizeInBits() > NativeWidth)
17824 auto Builder = IRBuilder<>(AI);
17825 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
17826 auto SynchScope = AI->getSynchScope();
17827 // We must restrict the ordering to avoid generating loads with Release or
17828 // ReleaseAcquire orderings.
17829 auto Order = AtomicCmpXchgInst::getStrongestFailureOrdering(AI->getOrdering());
17830 auto Ptr = AI->getPointerOperand();
17832 // Before the load we need a fence. Here is an example lifted from
17833 // http://www.hpl.hp.com/techreports/2012/HPL-2012-68.pdf showing why a fence
17836 // x.store(1, relaxed);
17837 // r1 = y.fetch_add(0, release);
17839 // y.fetch_add(42, acquire);
17840 // r2 = x.load(relaxed);
17841 // r1 = r2 = 0 is impossible, but becomes possible if the idempotent rmw is
17842 // lowered to just a load without a fence. A mfence flushes the store buffer,
17843 // making the optimization clearly correct.
17844 // FIXME: it is required if isAtLeastRelease(Order) but it is not clear
17845 // otherwise, we might be able to be more agressive on relaxed idempotent
17846 // rmw. In practice, they do not look useful, so we don't try to be
17847 // especially clever.
17848 if (SynchScope == SingleThread) {
17849 // FIXME: we could just insert an X86ISD::MEMBARRIER here, except we are at
17850 // the IR level, so we must wrap it in an intrinsic.
17852 } else if (hasMFENCE(Subtarget)) {
17853 Function *MFence = llvm::Intrinsic::getDeclaration(M,
17854 Intrinsic::x86_sse2_mfence);
17855 Builder.CreateCall(MFence);
17857 // FIXME: it might make sense to use a locked operation here but on a
17858 // different cache-line to prevent cache-line bouncing. In practice it
17859 // is probably a small win, and x86 processors without mfence are rare
17860 // enough that we do not bother.
17864 // Finally we can emit the atomic load.
17865 LoadInst *Loaded = Builder.CreateAlignedLoad(Ptr,
17866 AI->getType()->getPrimitiveSizeInBits());
17867 Loaded->setAtomic(Order, SynchScope);
17868 AI->replaceAllUsesWith(Loaded);
17869 AI->eraseFromParent();
17873 static SDValue LowerATOMIC_FENCE(SDValue Op, const X86Subtarget *Subtarget,
17874 SelectionDAG &DAG) {
17876 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
17877 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
17878 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
17879 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
17881 // The only fence that needs an instruction is a sequentially-consistent
17882 // cross-thread fence.
17883 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
17884 if (hasMFENCE(*Subtarget))
17885 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
17887 SDValue Chain = Op.getOperand(0);
17888 SDValue Zero = DAG.getConstant(0, MVT::i32);
17890 DAG.getRegister(X86::ESP, MVT::i32), // Base
17891 DAG.getTargetConstant(1, MVT::i8), // Scale
17892 DAG.getRegister(0, MVT::i32), // Index
17893 DAG.getTargetConstant(0, MVT::i32), // Disp
17894 DAG.getRegister(0, MVT::i32), // Segment.
17898 SDNode *Res = DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops);
17899 return SDValue(Res, 0);
17902 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
17903 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
17906 static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget *Subtarget,
17907 SelectionDAG &DAG) {
17908 MVT T = Op.getSimpleValueType();
17912 switch(T.SimpleTy) {
17913 default: llvm_unreachable("Invalid value type!");
17914 case MVT::i8: Reg = X86::AL; size = 1; break;
17915 case MVT::i16: Reg = X86::AX; size = 2; break;
17916 case MVT::i32: Reg = X86::EAX; size = 4; break;
17918 assert(Subtarget->is64Bit() && "Node not type legal!");
17919 Reg = X86::RAX; size = 8;
17922 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
17923 Op.getOperand(2), SDValue());
17924 SDValue Ops[] = { cpIn.getValue(0),
17927 DAG.getTargetConstant(size, MVT::i8),
17928 cpIn.getValue(1) };
17929 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
17930 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
17931 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
17935 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
17936 SDValue EFLAGS = DAG.getCopyFromReg(cpOut.getValue(1), DL, X86::EFLAGS,
17937 MVT::i32, cpOut.getValue(2));
17938 SDValue Success = DAG.getNode(X86ISD::SETCC, DL, Op->getValueType(1),
17939 DAG.getConstant(X86::COND_E, MVT::i8), EFLAGS);
17941 DAG.ReplaceAllUsesOfValueWith(Op.getValue(0), cpOut);
17942 DAG.ReplaceAllUsesOfValueWith(Op.getValue(1), Success);
17943 DAG.ReplaceAllUsesOfValueWith(Op.getValue(2), EFLAGS.getValue(1));
17947 static SDValue LowerBITCAST(SDValue Op, const X86Subtarget *Subtarget,
17948 SelectionDAG &DAG) {
17949 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
17950 MVT DstVT = Op.getSimpleValueType();
17952 if (SrcVT == MVT::v2i32 || SrcVT == MVT::v4i16 || SrcVT == MVT::v8i8) {
17953 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
17954 if (DstVT != MVT::f64)
17955 // This conversion needs to be expanded.
17958 SDValue InVec = Op->getOperand(0);
17960 unsigned NumElts = SrcVT.getVectorNumElements();
17961 EVT SVT = SrcVT.getVectorElementType();
17963 // Widen the vector in input in the case of MVT::v2i32.
17964 // Example: from MVT::v2i32 to MVT::v4i32.
17965 SmallVector<SDValue, 16> Elts;
17966 for (unsigned i = 0, e = NumElts; i != e; ++i)
17967 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT, InVec,
17968 DAG.getIntPtrConstant(i)));
17970 // Explicitly mark the extra elements as Undef.
17971 SDValue Undef = DAG.getUNDEF(SVT);
17972 for (unsigned i = NumElts, e = NumElts * 2; i != e; ++i)
17973 Elts.push_back(Undef);
17975 EVT NewVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
17976 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Elts);
17977 SDValue ToV2F64 = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, BV);
17978 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, ToV2F64,
17979 DAG.getIntPtrConstant(0));
17982 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
17983 Subtarget->hasMMX() && "Unexpected custom BITCAST");
17984 assert((DstVT == MVT::i64 ||
17985 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
17986 "Unexpected custom BITCAST");
17987 // i64 <=> MMX conversions are Legal.
17988 if (SrcVT==MVT::i64 && DstVT.isVector())
17990 if (DstVT==MVT::i64 && SrcVT.isVector())
17992 // MMX <=> MMX conversions are Legal.
17993 if (SrcVT.isVector() && DstVT.isVector())
17995 // All other conversions need to be expanded.
17999 static SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
18000 SDNode *Node = Op.getNode();
18002 EVT T = Node->getValueType(0);
18003 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
18004 DAG.getConstant(0, T), Node->getOperand(2));
18005 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
18006 cast<AtomicSDNode>(Node)->getMemoryVT(),
18007 Node->getOperand(0),
18008 Node->getOperand(1), negOp,
18009 cast<AtomicSDNode>(Node)->getMemOperand(),
18010 cast<AtomicSDNode>(Node)->getOrdering(),
18011 cast<AtomicSDNode>(Node)->getSynchScope());
18014 static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) {
18015 SDNode *Node = Op.getNode();
18017 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
18019 // Convert seq_cst store -> xchg
18020 // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b)
18021 // FIXME: On 32-bit, store -> fist or movq would be more efficient
18022 // (The only way to get a 16-byte store is cmpxchg16b)
18023 // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment.
18024 if (cast<AtomicSDNode>(Node)->getOrdering() == SequentiallyConsistent ||
18025 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
18026 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
18027 cast<AtomicSDNode>(Node)->getMemoryVT(),
18028 Node->getOperand(0),
18029 Node->getOperand(1), Node->getOperand(2),
18030 cast<AtomicSDNode>(Node)->getMemOperand(),
18031 cast<AtomicSDNode>(Node)->getOrdering(),
18032 cast<AtomicSDNode>(Node)->getSynchScope());
18033 return Swap.getValue(1);
18035 // Other atomic stores have a simple pattern.
18039 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
18040 EVT VT = Op.getNode()->getSimpleValueType(0);
18042 // Let legalize expand this if it isn't a legal type yet.
18043 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
18046 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
18049 bool ExtraOp = false;
18050 switch (Op.getOpcode()) {
18051 default: llvm_unreachable("Invalid code");
18052 case ISD::ADDC: Opc = X86ISD::ADD; break;
18053 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
18054 case ISD::SUBC: Opc = X86ISD::SUB; break;
18055 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
18059 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
18061 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
18062 Op.getOperand(1), Op.getOperand(2));
18065 static SDValue LowerFSINCOS(SDValue Op, const X86Subtarget *Subtarget,
18066 SelectionDAG &DAG) {
18067 assert(Subtarget->isTargetDarwin() && Subtarget->is64Bit());
18069 // For MacOSX, we want to call an alternative entry point: __sincos_stret,
18070 // which returns the values as { float, float } (in XMM0) or
18071 // { double, double } (which is returned in XMM0, XMM1).
18073 SDValue Arg = Op.getOperand(0);
18074 EVT ArgVT = Arg.getValueType();
18075 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
18077 TargetLowering::ArgListTy Args;
18078 TargetLowering::ArgListEntry Entry;
18082 Entry.isSExt = false;
18083 Entry.isZExt = false;
18084 Args.push_back(Entry);
18086 bool isF64 = ArgVT == MVT::f64;
18087 // Only optimize x86_64 for now. i386 is a bit messy. For f32,
18088 // the small struct {f32, f32} is returned in (eax, edx). For f64,
18089 // the results are returned via SRet in memory.
18090 const char *LibcallName = isF64 ? "__sincos_stret" : "__sincosf_stret";
18091 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
18092 SDValue Callee = DAG.getExternalSymbol(LibcallName, TLI.getPointerTy());
18094 Type *RetTy = isF64
18095 ? (Type*)StructType::get(ArgTy, ArgTy, NULL)
18096 : (Type*)VectorType::get(ArgTy, 4);
18098 TargetLowering::CallLoweringInfo CLI(DAG);
18099 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
18100 .setCallee(CallingConv::C, RetTy, Callee, std::move(Args), 0);
18102 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
18105 // Returned in xmm0 and xmm1.
18106 return CallResult.first;
18108 // Returned in bits 0:31 and 32:64 xmm0.
18109 SDValue SinVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
18110 CallResult.first, DAG.getIntPtrConstant(0));
18111 SDValue CosVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
18112 CallResult.first, DAG.getIntPtrConstant(1));
18113 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
18114 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys, SinVal, CosVal);
18117 /// LowerOperation - Provide custom lowering hooks for some operations.
18119 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
18120 switch (Op.getOpcode()) {
18121 default: llvm_unreachable("Should not custom lower this!");
18122 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op,DAG);
18123 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, Subtarget, DAG);
18124 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
18125 return LowerCMP_SWAP(Op, Subtarget, DAG);
18126 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
18127 case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG);
18128 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
18129 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
18130 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
18131 case ISD::VSELECT: return LowerVSELECT(Op, DAG);
18132 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
18133 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
18134 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op,Subtarget,DAG);
18135 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, Subtarget,DAG);
18136 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
18137 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
18138 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
18139 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
18140 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
18141 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
18142 case ISD::SHL_PARTS:
18143 case ISD::SRA_PARTS:
18144 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
18145 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
18146 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
18147 case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG);
18148 case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, Subtarget, DAG);
18149 case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, Subtarget, DAG);
18150 case ISD::ANY_EXTEND: return LowerANY_EXTEND(Op, Subtarget, DAG);
18151 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
18152 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
18153 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
18154 case ISD::LOAD: return LowerExtendedLoad(Op, Subtarget, DAG);
18156 case ISD::FNEG: return LowerFABSorFNEG(Op, DAG);
18157 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
18158 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
18159 case ISD::SETCC: return LowerSETCC(Op, DAG);
18160 case ISD::SELECT: return LowerSELECT(Op, DAG);
18161 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
18162 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
18163 case ISD::VASTART: return LowerVASTART(Op, DAG);
18164 case ISD::VAARG: return LowerVAARG(Op, DAG);
18165 case ISD::VACOPY: return LowerVACOPY(Op, Subtarget, DAG);
18166 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
18167 case ISD::INTRINSIC_VOID:
18168 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, Subtarget, DAG);
18169 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
18170 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
18171 case ISD::FRAME_TO_ARGS_OFFSET:
18172 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
18173 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
18174 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
18175 case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
18176 case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
18177 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
18178 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
18179 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
18180 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
18181 case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, DAG);
18182 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
18183 case ISD::MUL: return LowerMUL(Op, Subtarget, DAG);
18184 case ISD::UMUL_LOHI:
18185 case ISD::SMUL_LOHI: return LowerMUL_LOHI(Op, Subtarget, DAG);
18188 case ISD::SHL: return LowerShift(Op, Subtarget, DAG);
18194 case ISD::UMULO: return LowerXALUO(Op, DAG);
18195 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, Subtarget,DAG);
18196 case ISD::BITCAST: return LowerBITCAST(Op, Subtarget, DAG);
18200 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
18201 case ISD::ADD: return LowerADD(Op, DAG);
18202 case ISD::SUB: return LowerSUB(Op, DAG);
18203 case ISD::FSINCOS: return LowerFSINCOS(Op, Subtarget, DAG);
18207 /// ReplaceNodeResults - Replace a node with an illegal result type
18208 /// with a new node built out of custom code.
18209 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
18210 SmallVectorImpl<SDValue>&Results,
18211 SelectionDAG &DAG) const {
18213 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
18214 switch (N->getOpcode()) {
18216 llvm_unreachable("Do not know how to custom type legalize this operation!");
18217 case ISD::SIGN_EXTEND_INREG:
18222 // We don't want to expand or promote these.
18229 case ISD::UDIVREM: {
18230 SDValue V = LowerWin64_i128OP(SDValue(N,0), DAG);
18231 Results.push_back(V);
18234 case ISD::FP_TO_SINT:
18235 case ISD::FP_TO_UINT: {
18236 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
18238 if (!IsSigned && !isIntegerTypeFTOL(SDValue(N, 0).getValueType()))
18241 std::pair<SDValue,SDValue> Vals =
18242 FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true);
18243 SDValue FIST = Vals.first, StackSlot = Vals.second;
18244 if (FIST.getNode()) {
18245 EVT VT = N->getValueType(0);
18246 // Return a load from the stack slot.
18247 if (StackSlot.getNode())
18248 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
18249 MachinePointerInfo(),
18250 false, false, false, 0));
18252 Results.push_back(FIST);
18256 case ISD::UINT_TO_FP: {
18257 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
18258 if (N->getOperand(0).getValueType() != MVT::v2i32 ||
18259 N->getValueType(0) != MVT::v2f32)
18261 SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64,
18263 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
18265 SDValue VBias = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2f64, Bias, Bias);
18266 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn,
18267 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, VBias));
18268 Or = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or);
18269 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias);
18270 Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub));
18273 case ISD::FP_ROUND: {
18274 if (!TLI.isTypeLegal(N->getOperand(0).getValueType()))
18276 SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0));
18277 Results.push_back(V);
18280 case ISD::INTRINSIC_W_CHAIN: {
18281 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
18283 default : llvm_unreachable("Do not know how to custom type "
18284 "legalize this intrinsic operation!");
18285 case Intrinsic::x86_rdtsc:
18286 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
18288 case Intrinsic::x86_rdtscp:
18289 return getReadTimeStampCounter(N, dl, X86ISD::RDTSCP_DAG, DAG, Subtarget,
18291 case Intrinsic::x86_rdpmc:
18292 return getReadPerformanceCounter(N, dl, DAG, Subtarget, Results);
18295 case ISD::READCYCLECOUNTER: {
18296 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
18299 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
18300 EVT T = N->getValueType(0);
18301 assert((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair");
18302 bool Regs64bit = T == MVT::i128;
18303 EVT HalfT = Regs64bit ? MVT::i64 : MVT::i32;
18304 SDValue cpInL, cpInH;
18305 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
18306 DAG.getConstant(0, HalfT));
18307 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
18308 DAG.getConstant(1, HalfT));
18309 cpInL = DAG.getCopyToReg(N->getOperand(0), dl,
18310 Regs64bit ? X86::RAX : X86::EAX,
18312 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl,
18313 Regs64bit ? X86::RDX : X86::EDX,
18314 cpInH, cpInL.getValue(1));
18315 SDValue swapInL, swapInH;
18316 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
18317 DAG.getConstant(0, HalfT));
18318 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
18319 DAG.getConstant(1, HalfT));
18320 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl,
18321 Regs64bit ? X86::RBX : X86::EBX,
18322 swapInL, cpInH.getValue(1));
18323 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl,
18324 Regs64bit ? X86::RCX : X86::ECX,
18325 swapInH, swapInL.getValue(1));
18326 SDValue Ops[] = { swapInH.getValue(0),
18328 swapInH.getValue(1) };
18329 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
18330 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
18331 unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_DAG :
18332 X86ISD::LCMPXCHG8_DAG;
18333 SDValue Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys, Ops, T, MMO);
18334 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl,
18335 Regs64bit ? X86::RAX : X86::EAX,
18336 HalfT, Result.getValue(1));
18337 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl,
18338 Regs64bit ? X86::RDX : X86::EDX,
18339 HalfT, cpOutL.getValue(2));
18340 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
18342 SDValue EFLAGS = DAG.getCopyFromReg(cpOutH.getValue(1), dl, X86::EFLAGS,
18343 MVT::i32, cpOutH.getValue(2));
18345 DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
18346 DAG.getConstant(X86::COND_E, MVT::i8), EFLAGS);
18347 Success = DAG.getZExtOrTrunc(Success, dl, N->getValueType(1));
18349 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF));
18350 Results.push_back(Success);
18351 Results.push_back(EFLAGS.getValue(1));
18354 case ISD::ATOMIC_SWAP:
18355 case ISD::ATOMIC_LOAD_ADD:
18356 case ISD::ATOMIC_LOAD_SUB:
18357 case ISD::ATOMIC_LOAD_AND:
18358 case ISD::ATOMIC_LOAD_OR:
18359 case ISD::ATOMIC_LOAD_XOR:
18360 case ISD::ATOMIC_LOAD_NAND:
18361 case ISD::ATOMIC_LOAD_MIN:
18362 case ISD::ATOMIC_LOAD_MAX:
18363 case ISD::ATOMIC_LOAD_UMIN:
18364 case ISD::ATOMIC_LOAD_UMAX:
18365 case ISD::ATOMIC_LOAD: {
18366 // Delegate to generic TypeLegalization. Situations we can really handle
18367 // should have already been dealt with by AtomicExpandPass.cpp.
18370 case ISD::BITCAST: {
18371 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
18372 EVT DstVT = N->getValueType(0);
18373 EVT SrcVT = N->getOperand(0)->getValueType(0);
18375 if (SrcVT != MVT::f64 ||
18376 (DstVT != MVT::v2i32 && DstVT != MVT::v4i16 && DstVT != MVT::v8i8))
18379 unsigned NumElts = DstVT.getVectorNumElements();
18380 EVT SVT = DstVT.getVectorElementType();
18381 EVT WiderVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
18382 SDValue Expanded = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
18383 MVT::v2f64, N->getOperand(0));
18384 SDValue ToVecInt = DAG.getNode(ISD::BITCAST, dl, WiderVT, Expanded);
18386 if (ExperimentalVectorWideningLegalization) {
18387 // If we are legalizing vectors by widening, we already have the desired
18388 // legal vector type, just return it.
18389 Results.push_back(ToVecInt);
18393 SmallVector<SDValue, 8> Elts;
18394 for (unsigned i = 0, e = NumElts; i != e; ++i)
18395 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT,
18396 ToVecInt, DAG.getIntPtrConstant(i)));
18398 Results.push_back(DAG.getNode(ISD::BUILD_VECTOR, dl, DstVT, Elts));
18403 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
18405 default: return nullptr;
18406 case X86ISD::BSF: return "X86ISD::BSF";
18407 case X86ISD::BSR: return "X86ISD::BSR";
18408 case X86ISD::SHLD: return "X86ISD::SHLD";
18409 case X86ISD::SHRD: return "X86ISD::SHRD";
18410 case X86ISD::FAND: return "X86ISD::FAND";
18411 case X86ISD::FANDN: return "X86ISD::FANDN";
18412 case X86ISD::FOR: return "X86ISD::FOR";
18413 case X86ISD::FXOR: return "X86ISD::FXOR";
18414 case X86ISD::FSRL: return "X86ISD::FSRL";
18415 case X86ISD::FILD: return "X86ISD::FILD";
18416 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
18417 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
18418 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
18419 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
18420 case X86ISD::FLD: return "X86ISD::FLD";
18421 case X86ISD::FST: return "X86ISD::FST";
18422 case X86ISD::CALL: return "X86ISD::CALL";
18423 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
18424 case X86ISD::RDTSCP_DAG: return "X86ISD::RDTSCP_DAG";
18425 case X86ISD::RDPMC_DAG: return "X86ISD::RDPMC_DAG";
18426 case X86ISD::BT: return "X86ISD::BT";
18427 case X86ISD::CMP: return "X86ISD::CMP";
18428 case X86ISD::COMI: return "X86ISD::COMI";
18429 case X86ISD::UCOMI: return "X86ISD::UCOMI";
18430 case X86ISD::CMPM: return "X86ISD::CMPM";
18431 case X86ISD::CMPMU: return "X86ISD::CMPMU";
18432 case X86ISD::SETCC: return "X86ISD::SETCC";
18433 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
18434 case X86ISD::FSETCC: return "X86ISD::FSETCC";
18435 case X86ISD::CMOV: return "X86ISD::CMOV";
18436 case X86ISD::BRCOND: return "X86ISD::BRCOND";
18437 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
18438 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
18439 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
18440 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
18441 case X86ISD::Wrapper: return "X86ISD::Wrapper";
18442 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
18443 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
18444 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
18445 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
18446 case X86ISD::PINSRB: return "X86ISD::PINSRB";
18447 case X86ISD::PINSRW: return "X86ISD::PINSRW";
18448 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
18449 case X86ISD::ANDNP: return "X86ISD::ANDNP";
18450 case X86ISD::PSIGN: return "X86ISD::PSIGN";
18451 case X86ISD::BLENDI: return "X86ISD::BLENDI";
18452 case X86ISD::SUBUS: return "X86ISD::SUBUS";
18453 case X86ISD::HADD: return "X86ISD::HADD";
18454 case X86ISD::HSUB: return "X86ISD::HSUB";
18455 case X86ISD::FHADD: return "X86ISD::FHADD";
18456 case X86ISD::FHSUB: return "X86ISD::FHSUB";
18457 case X86ISD::UMAX: return "X86ISD::UMAX";
18458 case X86ISD::UMIN: return "X86ISD::UMIN";
18459 case X86ISD::SMAX: return "X86ISD::SMAX";
18460 case X86ISD::SMIN: return "X86ISD::SMIN";
18461 case X86ISD::FMAX: return "X86ISD::FMAX";
18462 case X86ISD::FMIN: return "X86ISD::FMIN";
18463 case X86ISD::FMAXC: return "X86ISD::FMAXC";
18464 case X86ISD::FMINC: return "X86ISD::FMINC";
18465 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
18466 case X86ISD::FRCP: return "X86ISD::FRCP";
18467 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
18468 case X86ISD::TLSBASEADDR: return "X86ISD::TLSBASEADDR";
18469 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
18470 case X86ISD::EH_SJLJ_SETJMP: return "X86ISD::EH_SJLJ_SETJMP";
18471 case X86ISD::EH_SJLJ_LONGJMP: return "X86ISD::EH_SJLJ_LONGJMP";
18472 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
18473 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
18474 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
18475 case X86ISD::FNSTSW16r: return "X86ISD::FNSTSW16r";
18476 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
18477 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
18478 case X86ISD::LCMPXCHG16_DAG: return "X86ISD::LCMPXCHG16_DAG";
18479 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
18480 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
18481 case X86ISD::VZEXT: return "X86ISD::VZEXT";
18482 case X86ISD::VSEXT: return "X86ISD::VSEXT";
18483 case X86ISD::VTRUNC: return "X86ISD::VTRUNC";
18484 case X86ISD::VTRUNCM: return "X86ISD::VTRUNCM";
18485 case X86ISD::VINSERT: return "X86ISD::VINSERT";
18486 case X86ISD::VFPEXT: return "X86ISD::VFPEXT";
18487 case X86ISD::VFPROUND: return "X86ISD::VFPROUND";
18488 case X86ISD::VSHLDQ: return "X86ISD::VSHLDQ";
18489 case X86ISD::VSRLDQ: return "X86ISD::VSRLDQ";
18490 case X86ISD::VSHL: return "X86ISD::VSHL";
18491 case X86ISD::VSRL: return "X86ISD::VSRL";
18492 case X86ISD::VSRA: return "X86ISD::VSRA";
18493 case X86ISD::VSHLI: return "X86ISD::VSHLI";
18494 case X86ISD::VSRLI: return "X86ISD::VSRLI";
18495 case X86ISD::VSRAI: return "X86ISD::VSRAI";
18496 case X86ISD::CMPP: return "X86ISD::CMPP";
18497 case X86ISD::PCMPEQ: return "X86ISD::PCMPEQ";
18498 case X86ISD::PCMPGT: return "X86ISD::PCMPGT";
18499 case X86ISD::PCMPEQM: return "X86ISD::PCMPEQM";
18500 case X86ISD::PCMPGTM: return "X86ISD::PCMPGTM";
18501 case X86ISD::ADD: return "X86ISD::ADD";
18502 case X86ISD::SUB: return "X86ISD::SUB";
18503 case X86ISD::ADC: return "X86ISD::ADC";
18504 case X86ISD::SBB: return "X86ISD::SBB";
18505 case X86ISD::SMUL: return "X86ISD::SMUL";
18506 case X86ISD::UMUL: return "X86ISD::UMUL";
18507 case X86ISD::INC: return "X86ISD::INC";
18508 case X86ISD::DEC: return "X86ISD::DEC";
18509 case X86ISD::OR: return "X86ISD::OR";
18510 case X86ISD::XOR: return "X86ISD::XOR";
18511 case X86ISD::AND: return "X86ISD::AND";
18512 case X86ISD::BEXTR: return "X86ISD::BEXTR";
18513 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
18514 case X86ISD::PTEST: return "X86ISD::PTEST";
18515 case X86ISD::TESTP: return "X86ISD::TESTP";
18516 case X86ISD::TESTM: return "X86ISD::TESTM";
18517 case X86ISD::TESTNM: return "X86ISD::TESTNM";
18518 case X86ISD::KORTEST: return "X86ISD::KORTEST";
18519 case X86ISD::PACKSS: return "X86ISD::PACKSS";
18520 case X86ISD::PACKUS: return "X86ISD::PACKUS";
18521 case X86ISD::PALIGNR: return "X86ISD::PALIGNR";
18522 case X86ISD::VALIGN: return "X86ISD::VALIGN";
18523 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
18524 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
18525 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
18526 case X86ISD::SHUFP: return "X86ISD::SHUFP";
18527 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
18528 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
18529 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
18530 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
18531 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
18532 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
18533 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
18534 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
18535 case X86ISD::MOVSD: return "X86ISD::MOVSD";
18536 case X86ISD::MOVSS: return "X86ISD::MOVSS";
18537 case X86ISD::UNPCKL: return "X86ISD::UNPCKL";
18538 case X86ISD::UNPCKH: return "X86ISD::UNPCKH";
18539 case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
18540 case X86ISD::VBROADCASTM: return "X86ISD::VBROADCASTM";
18541 case X86ISD::VEXTRACT: return "X86ISD::VEXTRACT";
18542 case X86ISD::VPERMILPI: return "X86ISD::VPERMILPI";
18543 case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128";
18544 case X86ISD::VPERMV: return "X86ISD::VPERMV";
18545 case X86ISD::VPERMV3: return "X86ISD::VPERMV3";
18546 case X86ISD::VPERMIV3: return "X86ISD::VPERMIV3";
18547 case X86ISD::VPERMI: return "X86ISD::VPERMI";
18548 case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
18549 case X86ISD::PMULDQ: return "X86ISD::PMULDQ";
18550 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
18551 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
18552 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
18553 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
18554 case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA";
18555 case X86ISD::WIN_FTOL: return "X86ISD::WIN_FTOL";
18556 case X86ISD::SAHF: return "X86ISD::SAHF";
18557 case X86ISD::RDRAND: return "X86ISD::RDRAND";
18558 case X86ISD::RDSEED: return "X86ISD::RDSEED";
18559 case X86ISD::FMADD: return "X86ISD::FMADD";
18560 case X86ISD::FMSUB: return "X86ISD::FMSUB";
18561 case X86ISD::FNMADD: return "X86ISD::FNMADD";
18562 case X86ISD::FNMSUB: return "X86ISD::FNMSUB";
18563 case X86ISD::FMADDSUB: return "X86ISD::FMADDSUB";
18564 case X86ISD::FMSUBADD: return "X86ISD::FMSUBADD";
18565 case X86ISD::PCMPESTRI: return "X86ISD::PCMPESTRI";
18566 case X86ISD::PCMPISTRI: return "X86ISD::PCMPISTRI";
18567 case X86ISD::XTEST: return "X86ISD::XTEST";
18571 // isLegalAddressingMode - Return true if the addressing mode represented
18572 // by AM is legal for this target, for a load/store of the specified type.
18573 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
18575 // X86 supports extremely general addressing modes.
18576 CodeModel::Model M = getTargetMachine().getCodeModel();
18577 Reloc::Model R = getTargetMachine().getRelocationModel();
18579 // X86 allows a sign-extended 32-bit immediate field as a displacement.
18580 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != nullptr))
18585 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
18587 // If a reference to this global requires an extra load, we can't fold it.
18588 if (isGlobalStubReference(GVFlags))
18591 // If BaseGV requires a register for the PIC base, we cannot also have a
18592 // BaseReg specified.
18593 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
18596 // If lower 4G is not available, then we must use rip-relative addressing.
18597 if ((M != CodeModel::Small || R != Reloc::Static) &&
18598 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
18602 switch (AM.Scale) {
18608 // These scales always work.
18613 // These scales are formed with basereg+scalereg. Only accept if there is
18618 default: // Other stuff never works.
18625 bool X86TargetLowering::isVectorShiftByScalarCheap(Type *Ty) const {
18626 unsigned Bits = Ty->getScalarSizeInBits();
18628 // 8-bit shifts are always expensive, but versions with a scalar amount aren't
18629 // particularly cheaper than those without.
18633 // On AVX2 there are new vpsllv[dq] instructions (and other shifts), that make
18634 // variable shifts just as cheap as scalar ones.
18635 if (Subtarget->hasInt256() && (Bits == 32 || Bits == 64))
18638 // Otherwise, it's significantly cheaper to shift by a scalar amount than by a
18639 // fully general vector.
18643 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
18644 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
18646 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
18647 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
18648 return NumBits1 > NumBits2;
18651 bool X86TargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
18652 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
18655 if (!isTypeLegal(EVT::getEVT(Ty1)))
18658 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
18660 // Assuming the caller doesn't have a zeroext or signext return parameter,
18661 // truncation all the way down to i1 is valid.
18665 bool X86TargetLowering::isLegalICmpImmediate(int64_t Imm) const {
18666 return isInt<32>(Imm);
18669 bool X86TargetLowering::isLegalAddImmediate(int64_t Imm) const {
18670 // Can also use sub to handle negated immediates.
18671 return isInt<32>(Imm);
18674 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
18675 if (!VT1.isInteger() || !VT2.isInteger())
18677 unsigned NumBits1 = VT1.getSizeInBits();
18678 unsigned NumBits2 = VT2.getSizeInBits();
18679 return NumBits1 > NumBits2;
18682 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
18683 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
18684 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
18687 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
18688 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
18689 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
18692 bool X86TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
18693 EVT VT1 = Val.getValueType();
18694 if (isZExtFree(VT1, VT2))
18697 if (Val.getOpcode() != ISD::LOAD)
18700 if (!VT1.isSimple() || !VT1.isInteger() ||
18701 !VT2.isSimple() || !VT2.isInteger())
18704 switch (VT1.getSimpleVT().SimpleTy) {
18709 // X86 has 8, 16, and 32-bit zero-extending loads.
18717 X86TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
18718 if (!(Subtarget->hasFMA() || Subtarget->hasFMA4()))
18721 VT = VT.getScalarType();
18723 if (!VT.isSimple())
18726 switch (VT.getSimpleVT().SimpleTy) {
18737 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
18738 // i16 instructions are longer (0x66 prefix) and potentially slower.
18739 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
18742 /// isShuffleMaskLegal - Targets can use this to indicate that they only
18743 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
18744 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
18745 /// are assumed to be legal.
18747 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
18749 if (!VT.isSimple())
18752 MVT SVT = VT.getSimpleVT();
18754 // Very little shuffling can be done for 64-bit vectors right now.
18755 if (VT.getSizeInBits() == 64)
18758 // If this is a single-input shuffle with no 128 bit lane crossings we can
18759 // lower it into pshufb.
18760 if ((SVT.is128BitVector() && Subtarget->hasSSSE3()) ||
18761 (SVT.is256BitVector() && Subtarget->hasInt256())) {
18762 bool isLegal = true;
18763 for (unsigned I = 0, E = M.size(); I != E; ++I) {
18764 if (M[I] >= (int)SVT.getVectorNumElements() ||
18765 ShuffleCrosses128bitLane(SVT, I, M[I])) {
18774 // FIXME: blends, shifts.
18775 return (SVT.getVectorNumElements() == 2 ||
18776 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
18777 isMOVLMask(M, SVT) ||
18778 isMOVHLPSMask(M, SVT) ||
18779 isSHUFPMask(M, SVT) ||
18780 isPSHUFDMask(M, SVT) ||
18781 isPSHUFHWMask(M, SVT, Subtarget->hasInt256()) ||
18782 isPSHUFLWMask(M, SVT, Subtarget->hasInt256()) ||
18783 isPALIGNRMask(M, SVT, Subtarget) ||
18784 isUNPCKLMask(M, SVT, Subtarget->hasInt256()) ||
18785 isUNPCKHMask(M, SVT, Subtarget->hasInt256()) ||
18786 isUNPCKL_v_undef_Mask(M, SVT, Subtarget->hasInt256()) ||
18787 isUNPCKH_v_undef_Mask(M, SVT, Subtarget->hasInt256()) ||
18788 isBlendMask(M, SVT, Subtarget->hasSSE41(), Subtarget->hasInt256()));
18792 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
18794 if (!VT.isSimple())
18797 MVT SVT = VT.getSimpleVT();
18798 unsigned NumElts = SVT.getVectorNumElements();
18799 // FIXME: This collection of masks seems suspect.
18802 if (NumElts == 4 && SVT.is128BitVector()) {
18803 return (isMOVLMask(Mask, SVT) ||
18804 isCommutedMOVLMask(Mask, SVT, true) ||
18805 isSHUFPMask(Mask, SVT) ||
18806 isSHUFPMask(Mask, SVT, /* Commuted */ true));
18811 //===----------------------------------------------------------------------===//
18812 // X86 Scheduler Hooks
18813 //===----------------------------------------------------------------------===//
18815 /// Utility function to emit xbegin specifying the start of an RTM region.
18816 static MachineBasicBlock *EmitXBegin(MachineInstr *MI, MachineBasicBlock *MBB,
18817 const TargetInstrInfo *TII) {
18818 DebugLoc DL = MI->getDebugLoc();
18820 const BasicBlock *BB = MBB->getBasicBlock();
18821 MachineFunction::iterator I = MBB;
18824 // For the v = xbegin(), we generate
18835 MachineBasicBlock *thisMBB = MBB;
18836 MachineFunction *MF = MBB->getParent();
18837 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
18838 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
18839 MF->insert(I, mainMBB);
18840 MF->insert(I, sinkMBB);
18842 // Transfer the remainder of BB and its successor edges to sinkMBB.
18843 sinkMBB->splice(sinkMBB->begin(), MBB,
18844 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
18845 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
18849 // # fallthrough to mainMBB
18850 // # abortion to sinkMBB
18851 BuildMI(thisMBB, DL, TII->get(X86::XBEGIN_4)).addMBB(sinkMBB);
18852 thisMBB->addSuccessor(mainMBB);
18853 thisMBB->addSuccessor(sinkMBB);
18857 BuildMI(mainMBB, DL, TII->get(X86::MOV32ri), X86::EAX).addImm(-1);
18858 mainMBB->addSuccessor(sinkMBB);
18861 // EAX is live into the sinkMBB
18862 sinkMBB->addLiveIn(X86::EAX);
18863 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
18864 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
18867 MI->eraseFromParent();
18871 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
18872 // or XMM0_V32I8 in AVX all of this code can be replaced with that
18873 // in the .td file.
18874 static MachineBasicBlock *EmitPCMPSTRM(MachineInstr *MI, MachineBasicBlock *BB,
18875 const TargetInstrInfo *TII) {
18877 switch (MI->getOpcode()) {
18878 default: llvm_unreachable("illegal opcode!");
18879 case X86::PCMPISTRM128REG: Opc = X86::PCMPISTRM128rr; break;
18880 case X86::VPCMPISTRM128REG: Opc = X86::VPCMPISTRM128rr; break;
18881 case X86::PCMPISTRM128MEM: Opc = X86::PCMPISTRM128rm; break;
18882 case X86::VPCMPISTRM128MEM: Opc = X86::VPCMPISTRM128rm; break;
18883 case X86::PCMPESTRM128REG: Opc = X86::PCMPESTRM128rr; break;
18884 case X86::VPCMPESTRM128REG: Opc = X86::VPCMPESTRM128rr; break;
18885 case X86::PCMPESTRM128MEM: Opc = X86::PCMPESTRM128rm; break;
18886 case X86::VPCMPESTRM128MEM: Opc = X86::VPCMPESTRM128rm; break;
18889 DebugLoc dl = MI->getDebugLoc();
18890 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
18892 unsigned NumArgs = MI->getNumOperands();
18893 for (unsigned i = 1; i < NumArgs; ++i) {
18894 MachineOperand &Op = MI->getOperand(i);
18895 if (!(Op.isReg() && Op.isImplicit()))
18896 MIB.addOperand(Op);
18898 if (MI->hasOneMemOperand())
18899 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
18901 BuildMI(*BB, MI, dl,
18902 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
18903 .addReg(X86::XMM0);
18905 MI->eraseFromParent();
18909 // FIXME: Custom handling because TableGen doesn't support multiple implicit
18910 // defs in an instruction pattern
18911 static MachineBasicBlock *EmitPCMPSTRI(MachineInstr *MI, MachineBasicBlock *BB,
18912 const TargetInstrInfo *TII) {
18914 switch (MI->getOpcode()) {
18915 default: llvm_unreachable("illegal opcode!");
18916 case X86::PCMPISTRIREG: Opc = X86::PCMPISTRIrr; break;
18917 case X86::VPCMPISTRIREG: Opc = X86::VPCMPISTRIrr; break;
18918 case X86::PCMPISTRIMEM: Opc = X86::PCMPISTRIrm; break;
18919 case X86::VPCMPISTRIMEM: Opc = X86::VPCMPISTRIrm; break;
18920 case X86::PCMPESTRIREG: Opc = X86::PCMPESTRIrr; break;
18921 case X86::VPCMPESTRIREG: Opc = X86::VPCMPESTRIrr; break;
18922 case X86::PCMPESTRIMEM: Opc = X86::PCMPESTRIrm; break;
18923 case X86::VPCMPESTRIMEM: Opc = X86::VPCMPESTRIrm; break;
18926 DebugLoc dl = MI->getDebugLoc();
18927 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
18929 unsigned NumArgs = MI->getNumOperands(); // remove the results
18930 for (unsigned i = 1; i < NumArgs; ++i) {
18931 MachineOperand &Op = MI->getOperand(i);
18932 if (!(Op.isReg() && Op.isImplicit()))
18933 MIB.addOperand(Op);
18935 if (MI->hasOneMemOperand())
18936 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
18938 BuildMI(*BB, MI, dl,
18939 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
18942 MI->eraseFromParent();
18946 static MachineBasicBlock * EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB,
18947 const TargetInstrInfo *TII,
18948 const X86Subtarget* Subtarget) {
18949 DebugLoc dl = MI->getDebugLoc();
18951 // Address into RAX/EAX, other two args into ECX, EDX.
18952 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
18953 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
18954 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
18955 for (int i = 0; i < X86::AddrNumOperands; ++i)
18956 MIB.addOperand(MI->getOperand(i));
18958 unsigned ValOps = X86::AddrNumOperands;
18959 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
18960 .addReg(MI->getOperand(ValOps).getReg());
18961 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
18962 .addReg(MI->getOperand(ValOps+1).getReg());
18964 // The instruction doesn't actually take any operands though.
18965 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
18967 MI->eraseFromParent(); // The pseudo is gone now.
18971 MachineBasicBlock *
18972 X86TargetLowering::EmitVAARG64WithCustomInserter(
18974 MachineBasicBlock *MBB) const {
18975 // Emit va_arg instruction on X86-64.
18977 // Operands to this pseudo-instruction:
18978 // 0 ) Output : destination address (reg)
18979 // 1-5) Input : va_list address (addr, i64mem)
18980 // 6 ) ArgSize : Size (in bytes) of vararg type
18981 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
18982 // 8 ) Align : Alignment of type
18983 // 9 ) EFLAGS (implicit-def)
18985 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
18986 assert(X86::AddrNumOperands == 5 && "VAARG_64 assumes 5 address operands");
18988 unsigned DestReg = MI->getOperand(0).getReg();
18989 MachineOperand &Base = MI->getOperand(1);
18990 MachineOperand &Scale = MI->getOperand(2);
18991 MachineOperand &Index = MI->getOperand(3);
18992 MachineOperand &Disp = MI->getOperand(4);
18993 MachineOperand &Segment = MI->getOperand(5);
18994 unsigned ArgSize = MI->getOperand(6).getImm();
18995 unsigned ArgMode = MI->getOperand(7).getImm();
18996 unsigned Align = MI->getOperand(8).getImm();
18998 // Memory Reference
18999 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
19000 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
19001 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
19003 // Machine Information
19004 const TargetInstrInfo *TII = MBB->getParent()->getSubtarget().getInstrInfo();
19005 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
19006 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
19007 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
19008 DebugLoc DL = MI->getDebugLoc();
19010 // struct va_list {
19013 // i64 overflow_area (address)
19014 // i64 reg_save_area (address)
19016 // sizeof(va_list) = 24
19017 // alignment(va_list) = 8
19019 unsigned TotalNumIntRegs = 6;
19020 unsigned TotalNumXMMRegs = 8;
19021 bool UseGPOffset = (ArgMode == 1);
19022 bool UseFPOffset = (ArgMode == 2);
19023 unsigned MaxOffset = TotalNumIntRegs * 8 +
19024 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
19026 /* Align ArgSize to a multiple of 8 */
19027 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
19028 bool NeedsAlign = (Align > 8);
19030 MachineBasicBlock *thisMBB = MBB;
19031 MachineBasicBlock *overflowMBB;
19032 MachineBasicBlock *offsetMBB;
19033 MachineBasicBlock *endMBB;
19035 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
19036 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
19037 unsigned OffsetReg = 0;
19039 if (!UseGPOffset && !UseFPOffset) {
19040 // If we only pull from the overflow region, we don't create a branch.
19041 // We don't need to alter control flow.
19042 OffsetDestReg = 0; // unused
19043 OverflowDestReg = DestReg;
19045 offsetMBB = nullptr;
19046 overflowMBB = thisMBB;
19049 // First emit code to check if gp_offset (or fp_offset) is below the bound.
19050 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
19051 // If not, pull from overflow_area. (branch to overflowMBB)
19056 // offsetMBB overflowMBB
19061 // Registers for the PHI in endMBB
19062 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
19063 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
19065 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
19066 MachineFunction *MF = MBB->getParent();
19067 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
19068 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
19069 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
19071 MachineFunction::iterator MBBIter = MBB;
19074 // Insert the new basic blocks
19075 MF->insert(MBBIter, offsetMBB);
19076 MF->insert(MBBIter, overflowMBB);
19077 MF->insert(MBBIter, endMBB);
19079 // Transfer the remainder of MBB and its successor edges to endMBB.
19080 endMBB->splice(endMBB->begin(), thisMBB,
19081 std::next(MachineBasicBlock::iterator(MI)), thisMBB->end());
19082 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
19084 // Make offsetMBB and overflowMBB successors of thisMBB
19085 thisMBB->addSuccessor(offsetMBB);
19086 thisMBB->addSuccessor(overflowMBB);
19088 // endMBB is a successor of both offsetMBB and overflowMBB
19089 offsetMBB->addSuccessor(endMBB);
19090 overflowMBB->addSuccessor(endMBB);
19092 // Load the offset value into a register
19093 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
19094 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
19098 .addDisp(Disp, UseFPOffset ? 4 : 0)
19099 .addOperand(Segment)
19100 .setMemRefs(MMOBegin, MMOEnd);
19102 // Check if there is enough room left to pull this argument.
19103 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
19105 .addImm(MaxOffset + 8 - ArgSizeA8);
19107 // Branch to "overflowMBB" if offset >= max
19108 // Fall through to "offsetMBB" otherwise
19109 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
19110 .addMBB(overflowMBB);
19113 // In offsetMBB, emit code to use the reg_save_area.
19115 assert(OffsetReg != 0);
19117 // Read the reg_save_area address.
19118 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
19119 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
19124 .addOperand(Segment)
19125 .setMemRefs(MMOBegin, MMOEnd);
19127 // Zero-extend the offset
19128 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
19129 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
19132 .addImm(X86::sub_32bit);
19134 // Add the offset to the reg_save_area to get the final address.
19135 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
19136 .addReg(OffsetReg64)
19137 .addReg(RegSaveReg);
19139 // Compute the offset for the next argument
19140 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
19141 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
19143 .addImm(UseFPOffset ? 16 : 8);
19145 // Store it back into the va_list.
19146 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
19150 .addDisp(Disp, UseFPOffset ? 4 : 0)
19151 .addOperand(Segment)
19152 .addReg(NextOffsetReg)
19153 .setMemRefs(MMOBegin, MMOEnd);
19156 BuildMI(offsetMBB, DL, TII->get(X86::JMP_4))
19161 // Emit code to use overflow area
19164 // Load the overflow_area address into a register.
19165 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
19166 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
19171 .addOperand(Segment)
19172 .setMemRefs(MMOBegin, MMOEnd);
19174 // If we need to align it, do so. Otherwise, just copy the address
19175 // to OverflowDestReg.
19177 // Align the overflow address
19178 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
19179 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
19181 // aligned_addr = (addr + (align-1)) & ~(align-1)
19182 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
19183 .addReg(OverflowAddrReg)
19186 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
19188 .addImm(~(uint64_t)(Align-1));
19190 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
19191 .addReg(OverflowAddrReg);
19194 // Compute the next overflow address after this argument.
19195 // (the overflow address should be kept 8-byte aligned)
19196 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
19197 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
19198 .addReg(OverflowDestReg)
19199 .addImm(ArgSizeA8);
19201 // Store the new overflow address.
19202 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
19207 .addOperand(Segment)
19208 .addReg(NextAddrReg)
19209 .setMemRefs(MMOBegin, MMOEnd);
19211 // If we branched, emit the PHI to the front of endMBB.
19213 BuildMI(*endMBB, endMBB->begin(), DL,
19214 TII->get(X86::PHI), DestReg)
19215 .addReg(OffsetDestReg).addMBB(offsetMBB)
19216 .addReg(OverflowDestReg).addMBB(overflowMBB);
19219 // Erase the pseudo instruction
19220 MI->eraseFromParent();
19225 MachineBasicBlock *
19226 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
19228 MachineBasicBlock *MBB) const {
19229 // Emit code to save XMM registers to the stack. The ABI says that the
19230 // number of registers to save is given in %al, so it's theoretically
19231 // possible to do an indirect jump trick to avoid saving all of them,
19232 // however this code takes a simpler approach and just executes all
19233 // of the stores if %al is non-zero. It's less code, and it's probably
19234 // easier on the hardware branch predictor, and stores aren't all that
19235 // expensive anyway.
19237 // Create the new basic blocks. One block contains all the XMM stores,
19238 // and one block is the final destination regardless of whether any
19239 // stores were performed.
19240 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
19241 MachineFunction *F = MBB->getParent();
19242 MachineFunction::iterator MBBIter = MBB;
19244 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
19245 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
19246 F->insert(MBBIter, XMMSaveMBB);
19247 F->insert(MBBIter, EndMBB);
19249 // Transfer the remainder of MBB and its successor edges to EndMBB.
19250 EndMBB->splice(EndMBB->begin(), MBB,
19251 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
19252 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
19254 // The original block will now fall through to the XMM save block.
19255 MBB->addSuccessor(XMMSaveMBB);
19256 // The XMMSaveMBB will fall through to the end block.
19257 XMMSaveMBB->addSuccessor(EndMBB);
19259 // Now add the instructions.
19260 const TargetInstrInfo *TII = MBB->getParent()->getSubtarget().getInstrInfo();
19261 DebugLoc DL = MI->getDebugLoc();
19263 unsigned CountReg = MI->getOperand(0).getReg();
19264 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
19265 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
19267 if (!Subtarget->isTargetWin64()) {
19268 // If %al is 0, branch around the XMM save block.
19269 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
19270 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
19271 MBB->addSuccessor(EndMBB);
19274 // Make sure the last operand is EFLAGS, which gets clobbered by the branch
19275 // that was just emitted, but clearly shouldn't be "saved".
19276 assert((MI->getNumOperands() <= 3 ||
19277 !MI->getOperand(MI->getNumOperands() - 1).isReg() ||
19278 MI->getOperand(MI->getNumOperands() - 1).getReg() == X86::EFLAGS)
19279 && "Expected last argument to be EFLAGS");
19280 unsigned MOVOpc = Subtarget->hasFp256() ? X86::VMOVAPSmr : X86::MOVAPSmr;
19281 // In the XMM save block, save all the XMM argument registers.
19282 for (int i = 3, e = MI->getNumOperands() - 1; i != e; ++i) {
19283 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
19284 MachineMemOperand *MMO =
19285 F->getMachineMemOperand(
19286 MachinePointerInfo::getFixedStack(RegSaveFrameIndex, Offset),
19287 MachineMemOperand::MOStore,
19288 /*Size=*/16, /*Align=*/16);
19289 BuildMI(XMMSaveMBB, DL, TII->get(MOVOpc))
19290 .addFrameIndex(RegSaveFrameIndex)
19291 .addImm(/*Scale=*/1)
19292 .addReg(/*IndexReg=*/0)
19293 .addImm(/*Disp=*/Offset)
19294 .addReg(/*Segment=*/0)
19295 .addReg(MI->getOperand(i).getReg())
19296 .addMemOperand(MMO);
19299 MI->eraseFromParent(); // The pseudo instruction is gone now.
19304 // The EFLAGS operand of SelectItr might be missing a kill marker
19305 // because there were multiple uses of EFLAGS, and ISel didn't know
19306 // which to mark. Figure out whether SelectItr should have had a
19307 // kill marker, and set it if it should. Returns the correct kill
19309 static bool checkAndUpdateEFLAGSKill(MachineBasicBlock::iterator SelectItr,
19310 MachineBasicBlock* BB,
19311 const TargetRegisterInfo* TRI) {
19312 // Scan forward through BB for a use/def of EFLAGS.
19313 MachineBasicBlock::iterator miI(std::next(SelectItr));
19314 for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) {
19315 const MachineInstr& mi = *miI;
19316 if (mi.readsRegister(X86::EFLAGS))
19318 if (mi.definesRegister(X86::EFLAGS))
19319 break; // Should have kill-flag - update below.
19322 // If we hit the end of the block, check whether EFLAGS is live into a
19324 if (miI == BB->end()) {
19325 for (MachineBasicBlock::succ_iterator sItr = BB->succ_begin(),
19326 sEnd = BB->succ_end();
19327 sItr != sEnd; ++sItr) {
19328 MachineBasicBlock* succ = *sItr;
19329 if (succ->isLiveIn(X86::EFLAGS))
19334 // We found a def, or hit the end of the basic block and EFLAGS wasn't live
19335 // out. SelectMI should have a kill flag on EFLAGS.
19336 SelectItr->addRegisterKilled(X86::EFLAGS, TRI);
19340 MachineBasicBlock *
19341 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
19342 MachineBasicBlock *BB) const {
19343 const TargetInstrInfo *TII = BB->getParent()->getSubtarget().getInstrInfo();
19344 DebugLoc DL = MI->getDebugLoc();
19346 // To "insert" a SELECT_CC instruction, we actually have to insert the
19347 // diamond control-flow pattern. The incoming instruction knows the
19348 // destination vreg to set, the condition code register to branch on, the
19349 // true/false values to select between, and a branch opcode to use.
19350 const BasicBlock *LLVM_BB = BB->getBasicBlock();
19351 MachineFunction::iterator It = BB;
19357 // cmpTY ccX, r1, r2
19359 // fallthrough --> copy0MBB
19360 MachineBasicBlock *thisMBB = BB;
19361 MachineFunction *F = BB->getParent();
19362 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
19363 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
19364 F->insert(It, copy0MBB);
19365 F->insert(It, sinkMBB);
19367 // If the EFLAGS register isn't dead in the terminator, then claim that it's
19368 // live into the sink and copy blocks.
19369 const TargetRegisterInfo *TRI =
19370 BB->getParent()->getSubtarget().getRegisterInfo();
19371 if (!MI->killsRegister(X86::EFLAGS) &&
19372 !checkAndUpdateEFLAGSKill(MI, BB, TRI)) {
19373 copy0MBB->addLiveIn(X86::EFLAGS);
19374 sinkMBB->addLiveIn(X86::EFLAGS);
19377 // Transfer the remainder of BB and its successor edges to sinkMBB.
19378 sinkMBB->splice(sinkMBB->begin(), BB,
19379 std::next(MachineBasicBlock::iterator(MI)), BB->end());
19380 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
19382 // Add the true and fallthrough blocks as its successors.
19383 BB->addSuccessor(copy0MBB);
19384 BB->addSuccessor(sinkMBB);
19386 // Create the conditional branch instruction.
19388 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
19389 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
19392 // %FalseValue = ...
19393 // # fallthrough to sinkMBB
19394 copy0MBB->addSuccessor(sinkMBB);
19397 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
19399 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
19400 TII->get(X86::PHI), MI->getOperand(0).getReg())
19401 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
19402 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
19404 MI->eraseFromParent(); // The pseudo instruction is gone now.
19408 MachineBasicBlock *
19409 X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI,
19410 MachineBasicBlock *BB) const {
19411 MachineFunction *MF = BB->getParent();
19412 const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo();
19413 DebugLoc DL = MI->getDebugLoc();
19414 const BasicBlock *LLVM_BB = BB->getBasicBlock();
19416 assert(MF->shouldSplitStack());
19418 const bool Is64Bit = Subtarget->is64Bit();
19419 const bool IsLP64 = Subtarget->isTarget64BitLP64();
19421 const unsigned TlsReg = Is64Bit ? X86::FS : X86::GS;
19422 const unsigned TlsOffset = IsLP64 ? 0x70 : Is64Bit ? 0x40 : 0x30;
19425 // ... [Till the alloca]
19426 // If stacklet is not large enough, jump to mallocMBB
19429 // Allocate by subtracting from RSP
19430 // Jump to continueMBB
19433 // Allocate by call to runtime
19437 // [rest of original BB]
19440 MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB);
19441 MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB);
19442 MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB);
19444 MachineRegisterInfo &MRI = MF->getRegInfo();
19445 const TargetRegisterClass *AddrRegClass =
19446 getRegClassFor(getPointerTy());
19448 unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass),
19449 bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass),
19450 tmpSPVReg = MRI.createVirtualRegister(AddrRegClass),
19451 SPLimitVReg = MRI.createVirtualRegister(AddrRegClass),
19452 sizeVReg = MI->getOperand(1).getReg(),
19453 physSPReg = IsLP64 || Subtarget->isTargetNaCl64() ? X86::RSP : X86::ESP;
19455 MachineFunction::iterator MBBIter = BB;
19458 MF->insert(MBBIter, bumpMBB);
19459 MF->insert(MBBIter, mallocMBB);
19460 MF->insert(MBBIter, continueMBB);
19462 continueMBB->splice(continueMBB->begin(), BB,
19463 std::next(MachineBasicBlock::iterator(MI)), BB->end());
19464 continueMBB->transferSuccessorsAndUpdatePHIs(BB);
19466 // Add code to the main basic block to check if the stack limit has been hit,
19467 // and if so, jump to mallocMBB otherwise to bumpMBB.
19468 BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg);
19469 BuildMI(BB, DL, TII->get(IsLP64 ? X86::SUB64rr:X86::SUB32rr), SPLimitVReg)
19470 .addReg(tmpSPVReg).addReg(sizeVReg);
19471 BuildMI(BB, DL, TII->get(IsLP64 ? X86::CMP64mr:X86::CMP32mr))
19472 .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg)
19473 .addReg(SPLimitVReg);
19474 BuildMI(BB, DL, TII->get(X86::JG_4)).addMBB(mallocMBB);
19476 // bumpMBB simply decreases the stack pointer, since we know the current
19477 // stacklet has enough space.
19478 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg)
19479 .addReg(SPLimitVReg);
19480 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg)
19481 .addReg(SPLimitVReg);
19482 BuildMI(bumpMBB, DL, TII->get(X86::JMP_4)).addMBB(continueMBB);
19484 // Calls into a routine in libgcc to allocate more space from the heap.
19485 const uint32_t *RegMask = MF->getTarget()
19486 .getSubtargetImpl()
19487 ->getRegisterInfo()
19488 ->getCallPreservedMask(CallingConv::C);
19490 BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
19492 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
19493 .addExternalSymbol("__morestack_allocate_stack_space")
19494 .addRegMask(RegMask)
19495 .addReg(X86::RDI, RegState::Implicit)
19496 .addReg(X86::RAX, RegState::ImplicitDefine);
19497 } else if (Is64Bit) {
19498 BuildMI(mallocMBB, DL, TII->get(X86::MOV32rr), X86::EDI)
19500 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
19501 .addExternalSymbol("__morestack_allocate_stack_space")
19502 .addRegMask(RegMask)
19503 .addReg(X86::EDI, RegState::Implicit)
19504 .addReg(X86::EAX, RegState::ImplicitDefine);
19506 BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg)
19508 BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg);
19509 BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32))
19510 .addExternalSymbol("__morestack_allocate_stack_space")
19511 .addRegMask(RegMask)
19512 .addReg(X86::EAX, RegState::ImplicitDefine);
19516 BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg)
19519 BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg)
19520 .addReg(IsLP64 ? X86::RAX : X86::EAX);
19521 BuildMI(mallocMBB, DL, TII->get(X86::JMP_4)).addMBB(continueMBB);
19523 // Set up the CFG correctly.
19524 BB->addSuccessor(bumpMBB);
19525 BB->addSuccessor(mallocMBB);
19526 mallocMBB->addSuccessor(continueMBB);
19527 bumpMBB->addSuccessor(continueMBB);
19529 // Take care of the PHI nodes.
19530 BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI),
19531 MI->getOperand(0).getReg())
19532 .addReg(mallocPtrVReg).addMBB(mallocMBB)
19533 .addReg(bumpSPPtrVReg).addMBB(bumpMBB);
19535 // Delete the original pseudo instruction.
19536 MI->eraseFromParent();
19539 return continueMBB;
19542 MachineBasicBlock *
19543 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
19544 MachineBasicBlock *BB) const {
19545 const TargetInstrInfo *TII = BB->getParent()->getSubtarget().getInstrInfo();
19546 DebugLoc DL = MI->getDebugLoc();
19548 assert(!Subtarget->isTargetMacho());
19550 // The lowering is pretty easy: we're just emitting the call to _alloca. The
19551 // non-trivial part is impdef of ESP.
19553 if (Subtarget->isTargetWin64()) {
19554 if (Subtarget->isTargetCygMing()) {
19555 // ___chkstk(Mingw64):
19556 // Clobbers R10, R11, RAX and EFLAGS.
19558 BuildMI(*BB, MI, DL, TII->get(X86::W64ALLOCA))
19559 .addExternalSymbol("___chkstk")
19560 .addReg(X86::RAX, RegState::Implicit)
19561 .addReg(X86::RSP, RegState::Implicit)
19562 .addReg(X86::RAX, RegState::Define | RegState::Implicit)
19563 .addReg(X86::RSP, RegState::Define | RegState::Implicit)
19564 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
19566 // __chkstk(MSVCRT): does not update stack pointer.
19567 // Clobbers R10, R11 and EFLAGS.
19568 BuildMI(*BB, MI, DL, TII->get(X86::W64ALLOCA))
19569 .addExternalSymbol("__chkstk")
19570 .addReg(X86::RAX, RegState::Implicit)
19571 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
19572 // RAX has the offset to be subtracted from RSP.
19573 BuildMI(*BB, MI, DL, TII->get(X86::SUB64rr), X86::RSP)
19578 const char *StackProbeSymbol =
19579 Subtarget->isTargetKnownWindowsMSVC() ? "_chkstk" : "_alloca";
19581 BuildMI(*BB, MI, DL, TII->get(X86::CALLpcrel32))
19582 .addExternalSymbol(StackProbeSymbol)
19583 .addReg(X86::EAX, RegState::Implicit)
19584 .addReg(X86::ESP, RegState::Implicit)
19585 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
19586 .addReg(X86::ESP, RegState::Define | RegState::Implicit)
19587 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
19590 MI->eraseFromParent(); // The pseudo instruction is gone now.
19594 MachineBasicBlock *
19595 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
19596 MachineBasicBlock *BB) const {
19597 // This is pretty easy. We're taking the value that we received from
19598 // our load from the relocation, sticking it in either RDI (x86-64)
19599 // or EAX and doing an indirect call. The return value will then
19600 // be in the normal return register.
19601 MachineFunction *F = BB->getParent();
19602 const X86InstrInfo *TII =
19603 static_cast<const X86InstrInfo *>(F->getSubtarget().getInstrInfo());
19604 DebugLoc DL = MI->getDebugLoc();
19606 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
19607 assert(MI->getOperand(3).isGlobal() && "This should be a global");
19609 // Get a register mask for the lowered call.
19610 // FIXME: The 32-bit calls have non-standard calling conventions. Use a
19611 // proper register mask.
19612 const uint32_t *RegMask = F->getTarget()
19613 .getSubtargetImpl()
19614 ->getRegisterInfo()
19615 ->getCallPreservedMask(CallingConv::C);
19616 if (Subtarget->is64Bit()) {
19617 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
19618 TII->get(X86::MOV64rm), X86::RDI)
19620 .addImm(0).addReg(0)
19621 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
19622 MI->getOperand(3).getTargetFlags())
19624 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
19625 addDirectMem(MIB, X86::RDI);
19626 MIB.addReg(X86::RAX, RegState::ImplicitDefine).addRegMask(RegMask);
19627 } else if (F->getTarget().getRelocationModel() != Reloc::PIC_) {
19628 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
19629 TII->get(X86::MOV32rm), X86::EAX)
19631 .addImm(0).addReg(0)
19632 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
19633 MI->getOperand(3).getTargetFlags())
19635 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
19636 addDirectMem(MIB, X86::EAX);
19637 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
19639 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
19640 TII->get(X86::MOV32rm), X86::EAX)
19641 .addReg(TII->getGlobalBaseReg(F))
19642 .addImm(0).addReg(0)
19643 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
19644 MI->getOperand(3).getTargetFlags())
19646 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
19647 addDirectMem(MIB, X86::EAX);
19648 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
19651 MI->eraseFromParent(); // The pseudo instruction is gone now.
19655 MachineBasicBlock *
19656 X86TargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
19657 MachineBasicBlock *MBB) const {
19658 DebugLoc DL = MI->getDebugLoc();
19659 MachineFunction *MF = MBB->getParent();
19660 const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo();
19661 MachineRegisterInfo &MRI = MF->getRegInfo();
19663 const BasicBlock *BB = MBB->getBasicBlock();
19664 MachineFunction::iterator I = MBB;
19667 // Memory Reference
19668 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
19669 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
19672 unsigned MemOpndSlot = 0;
19674 unsigned CurOp = 0;
19676 DstReg = MI->getOperand(CurOp++).getReg();
19677 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
19678 assert(RC->hasType(MVT::i32) && "Invalid destination!");
19679 unsigned mainDstReg = MRI.createVirtualRegister(RC);
19680 unsigned restoreDstReg = MRI.createVirtualRegister(RC);
19682 MemOpndSlot = CurOp;
19684 MVT PVT = getPointerTy();
19685 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
19686 "Invalid Pointer Size!");
19688 // For v = setjmp(buf), we generate
19691 // buf[LabelOffset] = restoreMBB
19692 // SjLjSetup restoreMBB
19698 // v = phi(main, restore)
19703 MachineBasicBlock *thisMBB = MBB;
19704 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
19705 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
19706 MachineBasicBlock *restoreMBB = MF->CreateMachineBasicBlock(BB);
19707 MF->insert(I, mainMBB);
19708 MF->insert(I, sinkMBB);
19709 MF->push_back(restoreMBB);
19711 MachineInstrBuilder MIB;
19713 // Transfer the remainder of BB and its successor edges to sinkMBB.
19714 sinkMBB->splice(sinkMBB->begin(), MBB,
19715 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
19716 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
19719 unsigned PtrStoreOpc = 0;
19720 unsigned LabelReg = 0;
19721 const int64_t LabelOffset = 1 * PVT.getStoreSize();
19722 Reloc::Model RM = MF->getTarget().getRelocationModel();
19723 bool UseImmLabel = (MF->getTarget().getCodeModel() == CodeModel::Small) &&
19724 (RM == Reloc::Static || RM == Reloc::DynamicNoPIC);
19726 // Prepare IP either in reg or imm.
19727 if (!UseImmLabel) {
19728 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr;
19729 const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
19730 LabelReg = MRI.createVirtualRegister(PtrRC);
19731 if (Subtarget->is64Bit()) {
19732 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA64r), LabelReg)
19736 .addMBB(restoreMBB)
19739 const X86InstrInfo *XII = static_cast<const X86InstrInfo*>(TII);
19740 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA32r), LabelReg)
19741 .addReg(XII->getGlobalBaseReg(MF))
19744 .addMBB(restoreMBB, Subtarget->ClassifyBlockAddressReference())
19748 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi;
19750 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PtrStoreOpc));
19751 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
19752 if (i == X86::AddrDisp)
19753 MIB.addDisp(MI->getOperand(MemOpndSlot + i), LabelOffset);
19755 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
19758 MIB.addReg(LabelReg);
19760 MIB.addMBB(restoreMBB);
19761 MIB.setMemRefs(MMOBegin, MMOEnd);
19763 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
19764 .addMBB(restoreMBB);
19766 const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
19767 MF->getSubtarget().getRegisterInfo());
19768 MIB.addRegMask(RegInfo->getNoPreservedMask());
19769 thisMBB->addSuccessor(mainMBB);
19770 thisMBB->addSuccessor(restoreMBB);
19774 BuildMI(mainMBB, DL, TII->get(X86::MOV32r0), mainDstReg);
19775 mainMBB->addSuccessor(sinkMBB);
19778 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
19779 TII->get(X86::PHI), DstReg)
19780 .addReg(mainDstReg).addMBB(mainMBB)
19781 .addReg(restoreDstReg).addMBB(restoreMBB);
19784 BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1);
19785 BuildMI(restoreMBB, DL, TII->get(X86::JMP_4)).addMBB(sinkMBB);
19786 restoreMBB->addSuccessor(sinkMBB);
19788 MI->eraseFromParent();
19792 MachineBasicBlock *
19793 X86TargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
19794 MachineBasicBlock *MBB) const {
19795 DebugLoc DL = MI->getDebugLoc();
19796 MachineFunction *MF = MBB->getParent();
19797 const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo();
19798 MachineRegisterInfo &MRI = MF->getRegInfo();
19800 // Memory Reference
19801 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
19802 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
19804 MVT PVT = getPointerTy();
19805 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
19806 "Invalid Pointer Size!");
19808 const TargetRegisterClass *RC =
19809 (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass;
19810 unsigned Tmp = MRI.createVirtualRegister(RC);
19811 // Since FP is only updated here but NOT referenced, it's treated as GPR.
19812 const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
19813 MF->getSubtarget().getRegisterInfo());
19814 unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
19815 unsigned SP = RegInfo->getStackRegister();
19817 MachineInstrBuilder MIB;
19819 const int64_t LabelOffset = 1 * PVT.getStoreSize();
19820 const int64_t SPOffset = 2 * PVT.getStoreSize();
19822 unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm;
19823 unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r;
19826 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP);
19827 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
19828 MIB.addOperand(MI->getOperand(i));
19829 MIB.setMemRefs(MMOBegin, MMOEnd);
19831 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp);
19832 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
19833 if (i == X86::AddrDisp)
19834 MIB.addDisp(MI->getOperand(i), LabelOffset);
19836 MIB.addOperand(MI->getOperand(i));
19838 MIB.setMemRefs(MMOBegin, MMOEnd);
19840 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP);
19841 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
19842 if (i == X86::AddrDisp)
19843 MIB.addDisp(MI->getOperand(i), SPOffset);
19845 MIB.addOperand(MI->getOperand(i));
19847 MIB.setMemRefs(MMOBegin, MMOEnd);
19849 BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp);
19851 MI->eraseFromParent();
19855 // Replace 213-type (isel default) FMA3 instructions with 231-type for
19856 // accumulator loops. Writing back to the accumulator allows the coalescer
19857 // to remove extra copies in the loop.
19858 MachineBasicBlock *
19859 X86TargetLowering::emitFMA3Instr(MachineInstr *MI,
19860 MachineBasicBlock *MBB) const {
19861 MachineOperand &AddendOp = MI->getOperand(3);
19863 // Bail out early if the addend isn't a register - we can't switch these.
19864 if (!AddendOp.isReg())
19867 MachineFunction &MF = *MBB->getParent();
19868 MachineRegisterInfo &MRI = MF.getRegInfo();
19870 // Check whether the addend is defined by a PHI:
19871 assert(MRI.hasOneDef(AddendOp.getReg()) && "Multiple defs in SSA?");
19872 MachineInstr &AddendDef = *MRI.def_instr_begin(AddendOp.getReg());
19873 if (!AddendDef.isPHI())
19876 // Look for the following pattern:
19878 // %addend = phi [%entry, 0], [%loop, %result]
19880 // %result<tied1> = FMA213 %m2<tied0>, %m1, %addend
19884 // %addend = phi [%entry, 0], [%loop, %result]
19886 // %result<tied1> = FMA231 %addend<tied0>, %m1, %m2
19888 for (unsigned i = 1, e = AddendDef.getNumOperands(); i < e; i += 2) {
19889 assert(AddendDef.getOperand(i).isReg());
19890 MachineOperand PHISrcOp = AddendDef.getOperand(i);
19891 MachineInstr &PHISrcInst = *MRI.def_instr_begin(PHISrcOp.getReg());
19892 if (&PHISrcInst == MI) {
19893 // Found a matching instruction.
19894 unsigned NewFMAOpc = 0;
19895 switch (MI->getOpcode()) {
19896 case X86::VFMADDPDr213r: NewFMAOpc = X86::VFMADDPDr231r; break;
19897 case X86::VFMADDPSr213r: NewFMAOpc = X86::VFMADDPSr231r; break;
19898 case X86::VFMADDSDr213r: NewFMAOpc = X86::VFMADDSDr231r; break;
19899 case X86::VFMADDSSr213r: NewFMAOpc = X86::VFMADDSSr231r; break;
19900 case X86::VFMSUBPDr213r: NewFMAOpc = X86::VFMSUBPDr231r; break;
19901 case X86::VFMSUBPSr213r: NewFMAOpc = X86::VFMSUBPSr231r; break;
19902 case X86::VFMSUBSDr213r: NewFMAOpc = X86::VFMSUBSDr231r; break;
19903 case X86::VFMSUBSSr213r: NewFMAOpc = X86::VFMSUBSSr231r; break;
19904 case X86::VFNMADDPDr213r: NewFMAOpc = X86::VFNMADDPDr231r; break;
19905 case X86::VFNMADDPSr213r: NewFMAOpc = X86::VFNMADDPSr231r; break;
19906 case X86::VFNMADDSDr213r: NewFMAOpc = X86::VFNMADDSDr231r; break;
19907 case X86::VFNMADDSSr213r: NewFMAOpc = X86::VFNMADDSSr231r; break;
19908 case X86::VFNMSUBPDr213r: NewFMAOpc = X86::VFNMSUBPDr231r; break;
19909 case X86::VFNMSUBPSr213r: NewFMAOpc = X86::VFNMSUBPSr231r; break;
19910 case X86::VFNMSUBSDr213r: NewFMAOpc = X86::VFNMSUBSDr231r; break;
19911 case X86::VFNMSUBSSr213r: NewFMAOpc = X86::VFNMSUBSSr231r; break;
19912 case X86::VFMADDPDr213rY: NewFMAOpc = X86::VFMADDPDr231rY; break;
19913 case X86::VFMADDPSr213rY: NewFMAOpc = X86::VFMADDPSr231rY; break;
19914 case X86::VFMSUBPDr213rY: NewFMAOpc = X86::VFMSUBPDr231rY; break;
19915 case X86::VFMSUBPSr213rY: NewFMAOpc = X86::VFMSUBPSr231rY; break;
19916 case X86::VFNMADDPDr213rY: NewFMAOpc = X86::VFNMADDPDr231rY; break;
19917 case X86::VFNMADDPSr213rY: NewFMAOpc = X86::VFNMADDPSr231rY; break;
19918 case X86::VFNMSUBPDr213rY: NewFMAOpc = X86::VFNMSUBPDr231rY; break;
19919 case X86::VFNMSUBPSr213rY: NewFMAOpc = X86::VFNMSUBPSr231rY; break;
19920 default: llvm_unreachable("Unrecognized FMA variant.");
19923 const TargetInstrInfo &TII = *MF.getSubtarget().getInstrInfo();
19924 MachineInstrBuilder MIB =
19925 BuildMI(MF, MI->getDebugLoc(), TII.get(NewFMAOpc))
19926 .addOperand(MI->getOperand(0))
19927 .addOperand(MI->getOperand(3))
19928 .addOperand(MI->getOperand(2))
19929 .addOperand(MI->getOperand(1));
19930 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
19931 MI->eraseFromParent();
19938 MachineBasicBlock *
19939 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
19940 MachineBasicBlock *BB) const {
19941 switch (MI->getOpcode()) {
19942 default: llvm_unreachable("Unexpected instr type to insert");
19943 case X86::TAILJMPd64:
19944 case X86::TAILJMPr64:
19945 case X86::TAILJMPm64:
19946 llvm_unreachable("TAILJMP64 would not be touched here.");
19947 case X86::TCRETURNdi64:
19948 case X86::TCRETURNri64:
19949 case X86::TCRETURNmi64:
19951 case X86::WIN_ALLOCA:
19952 return EmitLoweredWinAlloca(MI, BB);
19953 case X86::SEG_ALLOCA_32:
19954 case X86::SEG_ALLOCA_64:
19955 return EmitLoweredSegAlloca(MI, BB);
19956 case X86::TLSCall_32:
19957 case X86::TLSCall_64:
19958 return EmitLoweredTLSCall(MI, BB);
19959 case X86::CMOV_GR8:
19960 case X86::CMOV_FR32:
19961 case X86::CMOV_FR64:
19962 case X86::CMOV_V4F32:
19963 case X86::CMOV_V2F64:
19964 case X86::CMOV_V2I64:
19965 case X86::CMOV_V8F32:
19966 case X86::CMOV_V4F64:
19967 case X86::CMOV_V4I64:
19968 case X86::CMOV_V16F32:
19969 case X86::CMOV_V8F64:
19970 case X86::CMOV_V8I64:
19971 case X86::CMOV_GR16:
19972 case X86::CMOV_GR32:
19973 case X86::CMOV_RFP32:
19974 case X86::CMOV_RFP64:
19975 case X86::CMOV_RFP80:
19976 return EmitLoweredSelect(MI, BB);
19978 case X86::FP32_TO_INT16_IN_MEM:
19979 case X86::FP32_TO_INT32_IN_MEM:
19980 case X86::FP32_TO_INT64_IN_MEM:
19981 case X86::FP64_TO_INT16_IN_MEM:
19982 case X86::FP64_TO_INT32_IN_MEM:
19983 case X86::FP64_TO_INT64_IN_MEM:
19984 case X86::FP80_TO_INT16_IN_MEM:
19985 case X86::FP80_TO_INT32_IN_MEM:
19986 case X86::FP80_TO_INT64_IN_MEM: {
19987 MachineFunction *F = BB->getParent();
19988 const TargetInstrInfo *TII = F->getSubtarget().getInstrInfo();
19989 DebugLoc DL = MI->getDebugLoc();
19991 // Change the floating point control register to use "round towards zero"
19992 // mode when truncating to an integer value.
19993 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
19994 addFrameReference(BuildMI(*BB, MI, DL,
19995 TII->get(X86::FNSTCW16m)), CWFrameIdx);
19997 // Load the old value of the high byte of the control word...
19999 F->getRegInfo().createVirtualRegister(&X86::GR16RegClass);
20000 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
20003 // Set the high part to be round to zero...
20004 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
20007 // Reload the modified control word now...
20008 addFrameReference(BuildMI(*BB, MI, DL,
20009 TII->get(X86::FLDCW16m)), CWFrameIdx);
20011 // Restore the memory image of control word to original value
20012 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
20015 // Get the X86 opcode to use.
20017 switch (MI->getOpcode()) {
20018 default: llvm_unreachable("illegal opcode!");
20019 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
20020 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
20021 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
20022 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
20023 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
20024 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
20025 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
20026 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
20027 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
20031 MachineOperand &Op = MI->getOperand(0);
20033 AM.BaseType = X86AddressMode::RegBase;
20034 AM.Base.Reg = Op.getReg();
20036 AM.BaseType = X86AddressMode::FrameIndexBase;
20037 AM.Base.FrameIndex = Op.getIndex();
20039 Op = MI->getOperand(1);
20041 AM.Scale = Op.getImm();
20042 Op = MI->getOperand(2);
20044 AM.IndexReg = Op.getImm();
20045 Op = MI->getOperand(3);
20046 if (Op.isGlobal()) {
20047 AM.GV = Op.getGlobal();
20049 AM.Disp = Op.getImm();
20051 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
20052 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
20054 // Reload the original control word now.
20055 addFrameReference(BuildMI(*BB, MI, DL,
20056 TII->get(X86::FLDCW16m)), CWFrameIdx);
20058 MI->eraseFromParent(); // The pseudo instruction is gone now.
20061 // String/text processing lowering.
20062 case X86::PCMPISTRM128REG:
20063 case X86::VPCMPISTRM128REG:
20064 case X86::PCMPISTRM128MEM:
20065 case X86::VPCMPISTRM128MEM:
20066 case X86::PCMPESTRM128REG:
20067 case X86::VPCMPESTRM128REG:
20068 case X86::PCMPESTRM128MEM:
20069 case X86::VPCMPESTRM128MEM:
20070 assert(Subtarget->hasSSE42() &&
20071 "Target must have SSE4.2 or AVX features enabled");
20072 return EmitPCMPSTRM(MI, BB, BB->getParent()->getSubtarget().getInstrInfo());
20074 // String/text processing lowering.
20075 case X86::PCMPISTRIREG:
20076 case X86::VPCMPISTRIREG:
20077 case X86::PCMPISTRIMEM:
20078 case X86::VPCMPISTRIMEM:
20079 case X86::PCMPESTRIREG:
20080 case X86::VPCMPESTRIREG:
20081 case X86::PCMPESTRIMEM:
20082 case X86::VPCMPESTRIMEM:
20083 assert(Subtarget->hasSSE42() &&
20084 "Target must have SSE4.2 or AVX features enabled");
20085 return EmitPCMPSTRI(MI, BB, BB->getParent()->getSubtarget().getInstrInfo());
20087 // Thread synchronization.
20089 return EmitMonitor(MI, BB, BB->getParent()->getSubtarget().getInstrInfo(),
20094 return EmitXBegin(MI, BB, BB->getParent()->getSubtarget().getInstrInfo());
20096 case X86::VASTART_SAVE_XMM_REGS:
20097 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
20099 case X86::VAARG_64:
20100 return EmitVAARG64WithCustomInserter(MI, BB);
20102 case X86::EH_SjLj_SetJmp32:
20103 case X86::EH_SjLj_SetJmp64:
20104 return emitEHSjLjSetJmp(MI, BB);
20106 case X86::EH_SjLj_LongJmp32:
20107 case X86::EH_SjLj_LongJmp64:
20108 return emitEHSjLjLongJmp(MI, BB);
20110 case TargetOpcode::STACKMAP:
20111 case TargetOpcode::PATCHPOINT:
20112 return emitPatchPoint(MI, BB);
20114 case X86::VFMADDPDr213r:
20115 case X86::VFMADDPSr213r:
20116 case X86::VFMADDSDr213r:
20117 case X86::VFMADDSSr213r:
20118 case X86::VFMSUBPDr213r:
20119 case X86::VFMSUBPSr213r:
20120 case X86::VFMSUBSDr213r:
20121 case X86::VFMSUBSSr213r:
20122 case X86::VFNMADDPDr213r:
20123 case X86::VFNMADDPSr213r:
20124 case X86::VFNMADDSDr213r:
20125 case X86::VFNMADDSSr213r:
20126 case X86::VFNMSUBPDr213r:
20127 case X86::VFNMSUBPSr213r:
20128 case X86::VFNMSUBSDr213r:
20129 case X86::VFNMSUBSSr213r:
20130 case X86::VFMADDPDr213rY:
20131 case X86::VFMADDPSr213rY:
20132 case X86::VFMSUBPDr213rY:
20133 case X86::VFMSUBPSr213rY:
20134 case X86::VFNMADDPDr213rY:
20135 case X86::VFNMADDPSr213rY:
20136 case X86::VFNMSUBPDr213rY:
20137 case X86::VFNMSUBPSr213rY:
20138 return emitFMA3Instr(MI, BB);
20142 //===----------------------------------------------------------------------===//
20143 // X86 Optimization Hooks
20144 //===----------------------------------------------------------------------===//
20146 void X86TargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
20149 const SelectionDAG &DAG,
20150 unsigned Depth) const {
20151 unsigned BitWidth = KnownZero.getBitWidth();
20152 unsigned Opc = Op.getOpcode();
20153 assert((Opc >= ISD::BUILTIN_OP_END ||
20154 Opc == ISD::INTRINSIC_WO_CHAIN ||
20155 Opc == ISD::INTRINSIC_W_CHAIN ||
20156 Opc == ISD::INTRINSIC_VOID) &&
20157 "Should use MaskedValueIsZero if you don't know whether Op"
20158 " is a target node!");
20160 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
20174 // These nodes' second result is a boolean.
20175 if (Op.getResNo() == 0)
20178 case X86ISD::SETCC:
20179 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
20181 case ISD::INTRINSIC_WO_CHAIN: {
20182 unsigned IntId = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
20183 unsigned NumLoBits = 0;
20186 case Intrinsic::x86_sse_movmsk_ps:
20187 case Intrinsic::x86_avx_movmsk_ps_256:
20188 case Intrinsic::x86_sse2_movmsk_pd:
20189 case Intrinsic::x86_avx_movmsk_pd_256:
20190 case Intrinsic::x86_mmx_pmovmskb:
20191 case Intrinsic::x86_sse2_pmovmskb_128:
20192 case Intrinsic::x86_avx2_pmovmskb: {
20193 // High bits of movmskp{s|d}, pmovmskb are known zero.
20195 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
20196 case Intrinsic::x86_sse_movmsk_ps: NumLoBits = 4; break;
20197 case Intrinsic::x86_avx_movmsk_ps_256: NumLoBits = 8; break;
20198 case Intrinsic::x86_sse2_movmsk_pd: NumLoBits = 2; break;
20199 case Intrinsic::x86_avx_movmsk_pd_256: NumLoBits = 4; break;
20200 case Intrinsic::x86_mmx_pmovmskb: NumLoBits = 8; break;
20201 case Intrinsic::x86_sse2_pmovmskb_128: NumLoBits = 16; break;
20202 case Intrinsic::x86_avx2_pmovmskb: NumLoBits = 32; break;
20204 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - NumLoBits);
20213 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(
20215 const SelectionDAG &,
20216 unsigned Depth) const {
20217 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
20218 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
20219 return Op.getValueType().getScalarType().getSizeInBits();
20225 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
20226 /// node is a GlobalAddress + offset.
20227 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
20228 const GlobalValue* &GA,
20229 int64_t &Offset) const {
20230 if (N->getOpcode() == X86ISD::Wrapper) {
20231 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
20232 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
20233 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
20237 return TargetLowering::isGAPlusOffset(N, GA, Offset);
20240 /// isShuffleHigh128VectorInsertLow - Checks whether the shuffle node is the
20241 /// same as extracting the high 128-bit part of 256-bit vector and then
20242 /// inserting the result into the low part of a new 256-bit vector
20243 static bool isShuffleHigh128VectorInsertLow(ShuffleVectorSDNode *SVOp) {
20244 EVT VT = SVOp->getValueType(0);
20245 unsigned NumElems = VT.getVectorNumElements();
20247 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
20248 for (unsigned i = 0, j = NumElems/2; i != NumElems/2; ++i, ++j)
20249 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
20250 SVOp->getMaskElt(j) >= 0)
20256 /// isShuffleLow128VectorInsertHigh - Checks whether the shuffle node is the
20257 /// same as extracting the low 128-bit part of 256-bit vector and then
20258 /// inserting the result into the high part of a new 256-bit vector
20259 static bool isShuffleLow128VectorInsertHigh(ShuffleVectorSDNode *SVOp) {
20260 EVT VT = SVOp->getValueType(0);
20261 unsigned NumElems = VT.getVectorNumElements();
20263 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
20264 for (unsigned i = NumElems/2, j = 0; i != NumElems; ++i, ++j)
20265 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
20266 SVOp->getMaskElt(j) >= 0)
20272 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
20273 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
20274 TargetLowering::DAGCombinerInfo &DCI,
20275 const X86Subtarget* Subtarget) {
20277 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
20278 SDValue V1 = SVOp->getOperand(0);
20279 SDValue V2 = SVOp->getOperand(1);
20280 EVT VT = SVOp->getValueType(0);
20281 unsigned NumElems = VT.getVectorNumElements();
20283 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
20284 V2.getOpcode() == ISD::CONCAT_VECTORS) {
20288 // V UNDEF BUILD_VECTOR UNDEF
20290 // CONCAT_VECTOR CONCAT_VECTOR
20293 // RESULT: V + zero extended
20295 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
20296 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
20297 V1.getOperand(1).getOpcode() != ISD::UNDEF)
20300 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
20303 // To match the shuffle mask, the first half of the mask should
20304 // be exactly the first vector, and all the rest a splat with the
20305 // first element of the second one.
20306 for (unsigned i = 0; i != NumElems/2; ++i)
20307 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
20308 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
20311 // If V1 is coming from a vector load then just fold to a VZEXT_LOAD.
20312 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(V1.getOperand(0))) {
20313 if (Ld->hasNUsesOfValue(1, 0)) {
20314 SDVTList Tys = DAG.getVTList(MVT::v4i64, MVT::Other);
20315 SDValue Ops[] = { Ld->getChain(), Ld->getBasePtr() };
20317 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops,
20319 Ld->getPointerInfo(),
20320 Ld->getAlignment(),
20321 false/*isVolatile*/, true/*ReadMem*/,
20322 false/*WriteMem*/);
20324 // Make sure the newly-created LOAD is in the same position as Ld in
20325 // terms of dependency. We create a TokenFactor for Ld and ResNode,
20326 // and update uses of Ld's output chain to use the TokenFactor.
20327 if (Ld->hasAnyUseOfValue(1)) {
20328 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
20329 SDValue(Ld, 1), SDValue(ResNode.getNode(), 1));
20330 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), NewChain);
20331 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(Ld, 1),
20332 SDValue(ResNode.getNode(), 1));
20335 return DAG.getNode(ISD::BITCAST, dl, VT, ResNode);
20339 // Emit a zeroed vector and insert the desired subvector on its
20341 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
20342 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0), 0, DAG, dl);
20343 return DCI.CombineTo(N, InsV);
20346 //===--------------------------------------------------------------------===//
20347 // Combine some shuffles into subvector extracts and inserts:
20350 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
20351 if (isShuffleHigh128VectorInsertLow(SVOp)) {
20352 SDValue V = Extract128BitVector(V1, NumElems/2, DAG, dl);
20353 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, 0, DAG, dl);
20354 return DCI.CombineTo(N, InsV);
20357 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
20358 if (isShuffleLow128VectorInsertHigh(SVOp)) {
20359 SDValue V = Extract128BitVector(V1, 0, DAG, dl);
20360 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, NumElems/2, DAG, dl);
20361 return DCI.CombineTo(N, InsV);
20367 /// \brief Combine an arbitrary chain of shuffles into a single instruction if
20370 /// This is the leaf of the recursive combinine below. When we have found some
20371 /// chain of single-use x86 shuffle instructions and accumulated the combined
20372 /// shuffle mask represented by them, this will try to pattern match that mask
20373 /// into either a single instruction if there is a special purpose instruction
20374 /// for this operation, or into a PSHUFB instruction which is a fully general
20375 /// instruction but should only be used to replace chains over a certain depth.
20376 static bool combineX86ShuffleChain(SDValue Op, SDValue Root, ArrayRef<int> Mask,
20377 int Depth, bool HasPSHUFB, SelectionDAG &DAG,
20378 TargetLowering::DAGCombinerInfo &DCI,
20379 const X86Subtarget *Subtarget) {
20380 assert(!Mask.empty() && "Cannot combine an empty shuffle mask!");
20382 // Find the operand that enters the chain. Note that multiple uses are OK
20383 // here, we're not going to remove the operand we find.
20384 SDValue Input = Op.getOperand(0);
20385 while (Input.getOpcode() == ISD::BITCAST)
20386 Input = Input.getOperand(0);
20388 MVT VT = Input.getSimpleValueType();
20389 MVT RootVT = Root.getSimpleValueType();
20392 // Just remove no-op shuffle masks.
20393 if (Mask.size() == 1) {
20394 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Input),
20399 // Use the float domain if the operand type is a floating point type.
20400 bool FloatDomain = VT.isFloatingPoint();
20402 // For floating point shuffles, we don't have free copies in the shuffle
20403 // instructions or the ability to load as part of the instruction, so
20404 // canonicalize their shuffles to UNPCK or MOV variants.
20406 // Note that even with AVX we prefer the PSHUFD form of shuffle for integer
20407 // vectors because it can have a load folded into it that UNPCK cannot. This
20408 // doesn't preclude something switching to the shorter encoding post-RA.
20410 if (Mask.equals(0, 0) || Mask.equals(1, 1)) {
20411 bool Lo = Mask.equals(0, 0);
20414 // Check if we have SSE3 which will let us use MOVDDUP. That instruction
20415 // is no slower than UNPCKLPD but has the option to fold the input operand
20416 // into even an unaligned memory load.
20417 if (Lo && Subtarget->hasSSE3()) {
20418 Shuffle = X86ISD::MOVDDUP;
20419 ShuffleVT = MVT::v2f64;
20421 // We have MOVLHPS and MOVHLPS throughout SSE and they encode smaller
20422 // than the UNPCK variants.
20423 Shuffle = Lo ? X86ISD::MOVLHPS : X86ISD::MOVHLPS;
20424 ShuffleVT = MVT::v4f32;
20426 if (Depth == 1 && Root->getOpcode() == Shuffle)
20427 return false; // Nothing to do!
20428 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
20429 DCI.AddToWorklist(Op.getNode());
20430 if (Shuffle == X86ISD::MOVDDUP)
20431 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
20433 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
20434 DCI.AddToWorklist(Op.getNode());
20435 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
20439 if (Subtarget->hasSSE3() &&
20440 (Mask.equals(0, 0, 2, 2) || Mask.equals(1, 1, 3, 3))) {
20441 bool Lo = Mask.equals(0, 0, 2, 2);
20442 unsigned Shuffle = Lo ? X86ISD::MOVSLDUP : X86ISD::MOVSHDUP;
20443 MVT ShuffleVT = MVT::v4f32;
20444 if (Depth == 1 && Root->getOpcode() == Shuffle)
20445 return false; // Nothing to do!
20446 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
20447 DCI.AddToWorklist(Op.getNode());
20448 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
20449 DCI.AddToWorklist(Op.getNode());
20450 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
20454 if (Mask.equals(0, 0, 1, 1) || Mask.equals(2, 2, 3, 3)) {
20455 bool Lo = Mask.equals(0, 0, 1, 1);
20456 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
20457 MVT ShuffleVT = MVT::v4f32;
20458 if (Depth == 1 && Root->getOpcode() == Shuffle)
20459 return false; // Nothing to do!
20460 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
20461 DCI.AddToWorklist(Op.getNode());
20462 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
20463 DCI.AddToWorklist(Op.getNode());
20464 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
20470 // We always canonicalize the 8 x i16 and 16 x i8 shuffles into their UNPCK
20471 // variants as none of these have single-instruction variants that are
20472 // superior to the UNPCK formulation.
20473 if (!FloatDomain &&
20474 (Mask.equals(0, 0, 1, 1, 2, 2, 3, 3) ||
20475 Mask.equals(4, 4, 5, 5, 6, 6, 7, 7) ||
20476 Mask.equals(0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7) ||
20477 Mask.equals(8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14, 14, 15,
20479 bool Lo = Mask[0] == 0;
20480 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
20481 if (Depth == 1 && Root->getOpcode() == Shuffle)
20482 return false; // Nothing to do!
20484 switch (Mask.size()) {
20486 ShuffleVT = MVT::v8i16;
20489 ShuffleVT = MVT::v16i8;
20492 llvm_unreachable("Impossible mask size!");
20494 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
20495 DCI.AddToWorklist(Op.getNode());
20496 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
20497 DCI.AddToWorklist(Op.getNode());
20498 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
20503 // Don't try to re-form single instruction chains under any circumstances now
20504 // that we've done encoding canonicalization for them.
20508 // If we have 3 or more shuffle instructions or a chain involving PSHUFB, we
20509 // can replace them with a single PSHUFB instruction profitably. Intel's
20510 // manuals suggest only using PSHUFB if doing so replacing 5 instructions, but
20511 // in practice PSHUFB tends to be *very* fast so we're more aggressive.
20512 if ((Depth >= 3 || HasPSHUFB) && Subtarget->hasSSSE3()) {
20513 SmallVector<SDValue, 16> PSHUFBMask;
20514 assert(Mask.size() <= 16 && "Can't shuffle elements smaller than bytes!");
20515 int Ratio = 16 / Mask.size();
20516 for (unsigned i = 0; i < 16; ++i) {
20517 if (Mask[i / Ratio] == SM_SentinelUndef) {
20518 PSHUFBMask.push_back(DAG.getUNDEF(MVT::i8));
20521 int M = Mask[i / Ratio] != SM_SentinelZero
20522 ? Ratio * Mask[i / Ratio] + i % Ratio
20524 PSHUFBMask.push_back(DAG.getConstant(M, MVT::i8));
20526 Op = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Input);
20527 DCI.AddToWorklist(Op.getNode());
20528 SDValue PSHUFBMaskOp =
20529 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, PSHUFBMask);
20530 DCI.AddToWorklist(PSHUFBMaskOp.getNode());
20531 Op = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, Op, PSHUFBMaskOp);
20532 DCI.AddToWorklist(Op.getNode());
20533 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
20538 // Failed to find any combines.
20542 /// \brief Fully generic combining of x86 shuffle instructions.
20544 /// This should be the last combine run over the x86 shuffle instructions. Once
20545 /// they have been fully optimized, this will recursively consider all chains
20546 /// of single-use shuffle instructions, build a generic model of the cumulative
20547 /// shuffle operation, and check for simpler instructions which implement this
20548 /// operation. We use this primarily for two purposes:
20550 /// 1) Collapse generic shuffles to specialized single instructions when
20551 /// equivalent. In most cases, this is just an encoding size win, but
20552 /// sometimes we will collapse multiple generic shuffles into a single
20553 /// special-purpose shuffle.
20554 /// 2) Look for sequences of shuffle instructions with 3 or more total
20555 /// instructions, and replace them with the slightly more expensive SSSE3
20556 /// PSHUFB instruction if available. We do this as the last combining step
20557 /// to ensure we avoid using PSHUFB if we can implement the shuffle with
20558 /// a suitable short sequence of other instructions. The PHUFB will either
20559 /// use a register or have to read from memory and so is slightly (but only
20560 /// slightly) more expensive than the other shuffle instructions.
20562 /// Because this is inherently a quadratic operation (for each shuffle in
20563 /// a chain, we recurse up the chain), the depth is limited to 8 instructions.
20564 /// This should never be an issue in practice as the shuffle lowering doesn't
20565 /// produce sequences of more than 8 instructions.
20567 /// FIXME: We will currently miss some cases where the redundant shuffling
20568 /// would simplify under the threshold for PSHUFB formation because of
20569 /// combine-ordering. To fix this, we should do the redundant instruction
20570 /// combining in this recursive walk.
20571 static bool combineX86ShufflesRecursively(SDValue Op, SDValue Root,
20572 ArrayRef<int> RootMask,
20573 int Depth, bool HasPSHUFB,
20575 TargetLowering::DAGCombinerInfo &DCI,
20576 const X86Subtarget *Subtarget) {
20577 // Bound the depth of our recursive combine because this is ultimately
20578 // quadratic in nature.
20582 // Directly rip through bitcasts to find the underlying operand.
20583 while (Op.getOpcode() == ISD::BITCAST && Op.getOperand(0).hasOneUse())
20584 Op = Op.getOperand(0);
20586 MVT VT = Op.getSimpleValueType();
20587 if (!VT.isVector())
20588 return false; // Bail if we hit a non-vector.
20589 // FIXME: This routine should be taught about 256-bit shuffles, or a 256-bit
20590 // version should be added.
20591 if (VT.getSizeInBits() != 128)
20594 assert(Root.getSimpleValueType().isVector() &&
20595 "Shuffles operate on vector types!");
20596 assert(VT.getSizeInBits() == Root.getSimpleValueType().getSizeInBits() &&
20597 "Can only combine shuffles of the same vector register size.");
20599 if (!isTargetShuffle(Op.getOpcode()))
20601 SmallVector<int, 16> OpMask;
20603 bool HaveMask = getTargetShuffleMask(Op.getNode(), VT, OpMask, IsUnary);
20604 // We only can combine unary shuffles which we can decode the mask for.
20605 if (!HaveMask || !IsUnary)
20608 assert(VT.getVectorNumElements() == OpMask.size() &&
20609 "Different mask size from vector size!");
20610 assert(((RootMask.size() > OpMask.size() &&
20611 RootMask.size() % OpMask.size() == 0) ||
20612 (OpMask.size() > RootMask.size() &&
20613 OpMask.size() % RootMask.size() == 0) ||
20614 OpMask.size() == RootMask.size()) &&
20615 "The smaller number of elements must divide the larger.");
20616 int RootRatio = std::max<int>(1, OpMask.size() / RootMask.size());
20617 int OpRatio = std::max<int>(1, RootMask.size() / OpMask.size());
20618 assert(((RootRatio == 1 && OpRatio == 1) ||
20619 (RootRatio == 1) != (OpRatio == 1)) &&
20620 "Must not have a ratio for both incoming and op masks!");
20622 SmallVector<int, 16> Mask;
20623 Mask.reserve(std::max(OpMask.size(), RootMask.size()));
20625 // Merge this shuffle operation's mask into our accumulated mask. Note that
20626 // this shuffle's mask will be the first applied to the input, followed by the
20627 // root mask to get us all the way to the root value arrangement. The reason
20628 // for this order is that we are recursing up the operation chain.
20629 for (int i = 0, e = std::max(OpMask.size(), RootMask.size()); i < e; ++i) {
20630 int RootIdx = i / RootRatio;
20631 if (RootMask[RootIdx] < 0) {
20632 // This is a zero or undef lane, we're done.
20633 Mask.push_back(RootMask[RootIdx]);
20637 int RootMaskedIdx = RootMask[RootIdx] * RootRatio + i % RootRatio;
20638 int OpIdx = RootMaskedIdx / OpRatio;
20639 if (OpMask[OpIdx] < 0) {
20640 // The incoming lanes are zero or undef, it doesn't matter which ones we
20642 Mask.push_back(OpMask[OpIdx]);
20646 // Ok, we have non-zero lanes, map them through.
20647 Mask.push_back(OpMask[OpIdx] * OpRatio +
20648 RootMaskedIdx % OpRatio);
20651 // See if we can recurse into the operand to combine more things.
20652 switch (Op.getOpcode()) {
20653 case X86ISD::PSHUFB:
20655 case X86ISD::PSHUFD:
20656 case X86ISD::PSHUFHW:
20657 case X86ISD::PSHUFLW:
20658 if (Op.getOperand(0).hasOneUse() &&
20659 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
20660 HasPSHUFB, DAG, DCI, Subtarget))
20664 case X86ISD::UNPCKL:
20665 case X86ISD::UNPCKH:
20666 assert(Op.getOperand(0) == Op.getOperand(1) && "We only combine unary shuffles!");
20667 // We can't check for single use, we have to check that this shuffle is the only user.
20668 if (Op->isOnlyUserOf(Op.getOperand(0).getNode()) &&
20669 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
20670 HasPSHUFB, DAG, DCI, Subtarget))
20675 // Minor canonicalization of the accumulated shuffle mask to make it easier
20676 // to match below. All this does is detect masks with squential pairs of
20677 // elements, and shrink them to the half-width mask. It does this in a loop
20678 // so it will reduce the size of the mask to the minimal width mask which
20679 // performs an equivalent shuffle.
20680 while (Mask.size() > 1 && canWidenShuffleElements(Mask)) {
20681 for (int i = 0, e = Mask.size() / 2; i < e; ++i)
20682 Mask[i] = Mask[2 * i] / 2;
20683 Mask.resize(Mask.size() / 2);
20686 return combineX86ShuffleChain(Op, Root, Mask, Depth, HasPSHUFB, DAG, DCI,
20690 /// \brief Get the PSHUF-style mask from PSHUF node.
20692 /// This is a very minor wrapper around getTargetShuffleMask to easy forming v4
20693 /// PSHUF-style masks that can be reused with such instructions.
20694 static SmallVector<int, 4> getPSHUFShuffleMask(SDValue N) {
20695 SmallVector<int, 4> Mask;
20697 bool HaveMask = getTargetShuffleMask(N.getNode(), N.getSimpleValueType(), Mask, IsUnary);
20701 switch (N.getOpcode()) {
20702 case X86ISD::PSHUFD:
20704 case X86ISD::PSHUFLW:
20707 case X86ISD::PSHUFHW:
20708 Mask.erase(Mask.begin(), Mask.begin() + 4);
20709 for (int &M : Mask)
20713 llvm_unreachable("No valid shuffle instruction found!");
20717 /// \brief Search for a combinable shuffle across a chain ending in pshufd.
20719 /// We walk up the chain and look for a combinable shuffle, skipping over
20720 /// shuffles that we could hoist this shuffle's transformation past without
20721 /// altering anything.
20723 combineRedundantDWordShuffle(SDValue N, MutableArrayRef<int> Mask,
20725 TargetLowering::DAGCombinerInfo &DCI) {
20726 assert(N.getOpcode() == X86ISD::PSHUFD &&
20727 "Called with something other than an x86 128-bit half shuffle!");
20730 // Walk up a single-use chain looking for a combinable shuffle. Keep a stack
20731 // of the shuffles in the chain so that we can form a fresh chain to replace
20733 SmallVector<SDValue, 8> Chain;
20734 SDValue V = N.getOperand(0);
20735 for (; V.hasOneUse(); V = V.getOperand(0)) {
20736 switch (V.getOpcode()) {
20738 return SDValue(); // Nothing combined!
20741 // Skip bitcasts as we always know the type for the target specific
20745 case X86ISD::PSHUFD:
20746 // Found another dword shuffle.
20749 case X86ISD::PSHUFLW:
20750 // Check that the low words (being shuffled) are the identity in the
20751 // dword shuffle, and the high words are self-contained.
20752 if (Mask[0] != 0 || Mask[1] != 1 ||
20753 !(Mask[2] >= 2 && Mask[2] < 4 && Mask[3] >= 2 && Mask[3] < 4))
20756 Chain.push_back(V);
20759 case X86ISD::PSHUFHW:
20760 // Check that the high words (being shuffled) are the identity in the
20761 // dword shuffle, and the low words are self-contained.
20762 if (Mask[2] != 2 || Mask[3] != 3 ||
20763 !(Mask[0] >= 0 && Mask[0] < 2 && Mask[1] >= 0 && Mask[1] < 2))
20766 Chain.push_back(V);
20769 case X86ISD::UNPCKL:
20770 case X86ISD::UNPCKH:
20771 // For either i8 -> i16 or i16 -> i32 unpacks, we can combine a dword
20772 // shuffle into a preceding word shuffle.
20773 if (V.getValueType() != MVT::v16i8 && V.getValueType() != MVT::v8i16)
20776 // Search for a half-shuffle which we can combine with.
20777 unsigned CombineOp =
20778 V.getOpcode() == X86ISD::UNPCKL ? X86ISD::PSHUFLW : X86ISD::PSHUFHW;
20779 if (V.getOperand(0) != V.getOperand(1) ||
20780 !V->isOnlyUserOf(V.getOperand(0).getNode()))
20782 Chain.push_back(V);
20783 V = V.getOperand(0);
20785 switch (V.getOpcode()) {
20787 return SDValue(); // Nothing to combine.
20789 case X86ISD::PSHUFLW:
20790 case X86ISD::PSHUFHW:
20791 if (V.getOpcode() == CombineOp)
20794 Chain.push_back(V);
20798 V = V.getOperand(0);
20802 } while (V.hasOneUse());
20805 // Break out of the loop if we break out of the switch.
20809 if (!V.hasOneUse())
20810 // We fell out of the loop without finding a viable combining instruction.
20813 // Merge this node's mask and our incoming mask.
20814 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
20815 for (int &M : Mask)
20817 V = DAG.getNode(V.getOpcode(), DL, V.getValueType(), V.getOperand(0),
20818 getV4X86ShuffleImm8ForMask(Mask, DAG));
20820 // Rebuild the chain around this new shuffle.
20821 while (!Chain.empty()) {
20822 SDValue W = Chain.pop_back_val();
20824 if (V.getValueType() != W.getOperand(0).getValueType())
20825 V = DAG.getNode(ISD::BITCAST, DL, W.getOperand(0).getValueType(), V);
20827 switch (W.getOpcode()) {
20829 llvm_unreachable("Only PSHUF and UNPCK instructions get here!");
20831 case X86ISD::UNPCKL:
20832 case X86ISD::UNPCKH:
20833 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, V);
20836 case X86ISD::PSHUFD:
20837 case X86ISD::PSHUFLW:
20838 case X86ISD::PSHUFHW:
20839 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, W.getOperand(1));
20843 if (V.getValueType() != N.getValueType())
20844 V = DAG.getNode(ISD::BITCAST, DL, N.getValueType(), V);
20846 // Return the new chain to replace N.
20850 /// \brief Search for a combinable shuffle across a chain ending in pshuflw or pshufhw.
20852 /// We walk up the chain, skipping shuffles of the other half and looking
20853 /// through shuffles which switch halves trying to find a shuffle of the same
20854 /// pair of dwords.
20855 static bool combineRedundantHalfShuffle(SDValue N, MutableArrayRef<int> Mask,
20857 TargetLowering::DAGCombinerInfo &DCI) {
20859 (N.getOpcode() == X86ISD::PSHUFLW || N.getOpcode() == X86ISD::PSHUFHW) &&
20860 "Called with something other than an x86 128-bit half shuffle!");
20862 unsigned CombineOpcode = N.getOpcode();
20864 // Walk up a single-use chain looking for a combinable shuffle.
20865 SDValue V = N.getOperand(0);
20866 for (; V.hasOneUse(); V = V.getOperand(0)) {
20867 switch (V.getOpcode()) {
20869 return false; // Nothing combined!
20872 // Skip bitcasts as we always know the type for the target specific
20876 case X86ISD::PSHUFLW:
20877 case X86ISD::PSHUFHW:
20878 if (V.getOpcode() == CombineOpcode)
20881 // Other-half shuffles are no-ops.
20884 // Break out of the loop if we break out of the switch.
20888 if (!V.hasOneUse())
20889 // We fell out of the loop without finding a viable combining instruction.
20892 // Combine away the bottom node as its shuffle will be accumulated into
20893 // a preceding shuffle.
20894 DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
20896 // Record the old value.
20899 // Merge this node's mask and our incoming mask (adjusted to account for all
20900 // the pshufd instructions encountered).
20901 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
20902 for (int &M : Mask)
20904 V = DAG.getNode(V.getOpcode(), DL, MVT::v8i16, V.getOperand(0),
20905 getV4X86ShuffleImm8ForMask(Mask, DAG));
20907 // Check that the shuffles didn't cancel each other out. If not, we need to
20908 // combine to the new one.
20910 // Replace the combinable shuffle with the combined one, updating all users
20911 // so that we re-evaluate the chain here.
20912 DCI.CombineTo(Old.getNode(), V, /*AddTo*/ true);
20917 /// \brief Try to combine x86 target specific shuffles.
20918 static SDValue PerformTargetShuffleCombine(SDValue N, SelectionDAG &DAG,
20919 TargetLowering::DAGCombinerInfo &DCI,
20920 const X86Subtarget *Subtarget) {
20922 MVT VT = N.getSimpleValueType();
20923 SmallVector<int, 4> Mask;
20925 switch (N.getOpcode()) {
20926 case X86ISD::PSHUFD:
20927 case X86ISD::PSHUFLW:
20928 case X86ISD::PSHUFHW:
20929 Mask = getPSHUFShuffleMask(N);
20930 assert(Mask.size() == 4);
20936 // Nuke no-op shuffles that show up after combining.
20937 if (isNoopShuffleMask(Mask))
20938 return DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
20940 // Look for simplifications involving one or two shuffle instructions.
20941 SDValue V = N.getOperand(0);
20942 switch (N.getOpcode()) {
20945 case X86ISD::PSHUFLW:
20946 case X86ISD::PSHUFHW:
20947 assert(VT == MVT::v8i16);
20950 if (combineRedundantHalfShuffle(N, Mask, DAG, DCI))
20951 return SDValue(); // We combined away this shuffle, so we're done.
20953 // See if this reduces to a PSHUFD which is no more expensive and can
20954 // combine with more operations.
20955 if (canWidenShuffleElements(Mask)) {
20956 int DMask[] = {-1, -1, -1, -1};
20957 int DOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 2;
20958 DMask[DOffset + 0] = DOffset + Mask[0] / 2;
20959 DMask[DOffset + 1] = DOffset + Mask[2] / 2;
20960 V = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V);
20961 DCI.AddToWorklist(V.getNode());
20962 V = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V,
20963 getV4X86ShuffleImm8ForMask(DMask, DAG));
20964 DCI.AddToWorklist(V.getNode());
20965 return DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V);
20968 // Look for shuffle patterns which can be implemented as a single unpack.
20969 // FIXME: This doesn't handle the location of the PSHUFD generically, and
20970 // only works when we have a PSHUFD followed by two half-shuffles.
20971 if (Mask[0] == Mask[1] && Mask[2] == Mask[3] &&
20972 (V.getOpcode() == X86ISD::PSHUFLW ||
20973 V.getOpcode() == X86ISD::PSHUFHW) &&
20974 V.getOpcode() != N.getOpcode() &&
20976 SDValue D = V.getOperand(0);
20977 while (D.getOpcode() == ISD::BITCAST && D.hasOneUse())
20978 D = D.getOperand(0);
20979 if (D.getOpcode() == X86ISD::PSHUFD && D.hasOneUse()) {
20980 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
20981 SmallVector<int, 4> DMask = getPSHUFShuffleMask(D);
20982 int NOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
20983 int VOffset = V.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
20985 for (int i = 0; i < 4; ++i) {
20986 WordMask[i + NOffset] = Mask[i] + NOffset;
20987 WordMask[i + VOffset] = VMask[i] + VOffset;
20989 // Map the word mask through the DWord mask.
20991 for (int i = 0; i < 8; ++i)
20992 MappedMask[i] = 2 * DMask[WordMask[i] / 2] + WordMask[i] % 2;
20993 const int UnpackLoMask[] = {0, 0, 1, 1, 2, 2, 3, 3};
20994 const int UnpackHiMask[] = {4, 4, 5, 5, 6, 6, 7, 7};
20995 if (std::equal(std::begin(MappedMask), std::end(MappedMask),
20996 std::begin(UnpackLoMask)) ||
20997 std::equal(std::begin(MappedMask), std::end(MappedMask),
20998 std::begin(UnpackHiMask))) {
20999 // We can replace all three shuffles with an unpack.
21000 V = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, D.getOperand(0));
21001 DCI.AddToWorklist(V.getNode());
21002 return DAG.getNode(MappedMask[0] == 0 ? X86ISD::UNPCKL
21004 DL, MVT::v8i16, V, V);
21011 case X86ISD::PSHUFD:
21012 if (SDValue NewN = combineRedundantDWordShuffle(N, Mask, DAG, DCI))
21021 /// \brief Try to combine a shuffle into a target-specific add-sub node.
21023 /// We combine this directly on the abstract vector shuffle nodes so it is
21024 /// easier to generically match. We also insert dummy vector shuffle nodes for
21025 /// the operands which explicitly discard the lanes which are unused by this
21026 /// operation to try to flow through the rest of the combiner the fact that
21027 /// they're unused.
21028 static SDValue combineShuffleToAddSub(SDNode *N, SelectionDAG &DAG) {
21030 EVT VT = N->getValueType(0);
21032 // We only handle target-independent shuffles.
21033 // FIXME: It would be easy and harmless to use the target shuffle mask
21034 // extraction tool to support more.
21035 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
21038 auto *SVN = cast<ShuffleVectorSDNode>(N);
21039 ArrayRef<int> Mask = SVN->getMask();
21040 SDValue V1 = N->getOperand(0);
21041 SDValue V2 = N->getOperand(1);
21043 // We require the first shuffle operand to be the SUB node, and the second to
21044 // be the ADD node.
21045 // FIXME: We should support the commuted patterns.
21046 if (V1->getOpcode() != ISD::FSUB || V2->getOpcode() != ISD::FADD)
21049 // If there are other uses of these operations we can't fold them.
21050 if (!V1->hasOneUse() || !V2->hasOneUse())
21053 // Ensure that both operations have the same operands. Note that we can
21054 // commute the FADD operands.
21055 SDValue LHS = V1->getOperand(0), RHS = V1->getOperand(1);
21056 if ((V2->getOperand(0) != LHS || V2->getOperand(1) != RHS) &&
21057 (V2->getOperand(0) != RHS || V2->getOperand(1) != LHS))
21060 // We're looking for blends between FADD and FSUB nodes. We insist on these
21061 // nodes being lined up in a specific expected pattern.
21062 if (!(isShuffleEquivalent(Mask, 0, 3) ||
21063 isShuffleEquivalent(Mask, 0, 5, 2, 7) ||
21064 isShuffleEquivalent(Mask, 0, 9, 2, 11, 4, 13, 6, 15)))
21067 // Only specific types are legal at this point, assert so we notice if and
21068 // when these change.
21069 assert((VT == MVT::v4f32 || VT == MVT::v2f64 || VT == MVT::v8f32 ||
21070 VT == MVT::v4f64) &&
21071 "Unknown vector type encountered!");
21073 return DAG.getNode(X86ISD::ADDSUB, DL, VT, LHS, RHS);
21076 /// PerformShuffleCombine - Performs several different shuffle combines.
21077 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
21078 TargetLowering::DAGCombinerInfo &DCI,
21079 const X86Subtarget *Subtarget) {
21081 SDValue N0 = N->getOperand(0);
21082 SDValue N1 = N->getOperand(1);
21083 EVT VT = N->getValueType(0);
21085 // Don't create instructions with illegal types after legalize types has run.
21086 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
21087 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
21090 // If we have legalized the vector types, look for blends of FADD and FSUB
21091 // nodes that we can fuse into an ADDSUB node.
21092 if (TLI.isTypeLegal(VT) && Subtarget->hasSSE3())
21093 if (SDValue AddSub = combineShuffleToAddSub(N, DAG))
21096 // Combine 256-bit vector shuffles. This is only profitable when in AVX mode
21097 if (Subtarget->hasFp256() && VT.is256BitVector() &&
21098 N->getOpcode() == ISD::VECTOR_SHUFFLE)
21099 return PerformShuffleCombine256(N, DAG, DCI, Subtarget);
21101 // During Type Legalization, when promoting illegal vector types,
21102 // the backend might introduce new shuffle dag nodes and bitcasts.
21104 // This code performs the following transformation:
21105 // fold: (shuffle (bitcast (BINOP A, B)), Undef, <Mask>) ->
21106 // (shuffle (BINOP (bitcast A), (bitcast B)), Undef, <Mask>)
21108 // We do this only if both the bitcast and the BINOP dag nodes have
21109 // one use. Also, perform this transformation only if the new binary
21110 // operation is legal. This is to avoid introducing dag nodes that
21111 // potentially need to be further expanded (or custom lowered) into a
21112 // less optimal sequence of dag nodes.
21113 if (!DCI.isBeforeLegalize() && DCI.isBeforeLegalizeOps() &&
21114 N1.getOpcode() == ISD::UNDEF && N0.hasOneUse() &&
21115 N0.getOpcode() == ISD::BITCAST) {
21116 SDValue BC0 = N0.getOperand(0);
21117 EVT SVT = BC0.getValueType();
21118 unsigned Opcode = BC0.getOpcode();
21119 unsigned NumElts = VT.getVectorNumElements();
21121 if (BC0.hasOneUse() && SVT.isVector() &&
21122 SVT.getVectorNumElements() * 2 == NumElts &&
21123 TLI.isOperationLegal(Opcode, VT)) {
21124 bool CanFold = false;
21136 unsigned SVTNumElts = SVT.getVectorNumElements();
21137 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
21138 for (unsigned i = 0, e = SVTNumElts; i != e && CanFold; ++i)
21139 CanFold = SVOp->getMaskElt(i) == (int)(i * 2);
21140 for (unsigned i = SVTNumElts, e = NumElts; i != e && CanFold; ++i)
21141 CanFold = SVOp->getMaskElt(i) < 0;
21144 SDValue BC00 = DAG.getNode(ISD::BITCAST, dl, VT, BC0.getOperand(0));
21145 SDValue BC01 = DAG.getNode(ISD::BITCAST, dl, VT, BC0.getOperand(1));
21146 SDValue NewBinOp = DAG.getNode(BC0.getOpcode(), dl, VT, BC00, BC01);
21147 return DAG.getVectorShuffle(VT, dl, NewBinOp, N1, &SVOp->getMask()[0]);
21152 // Only handle 128 wide vector from here on.
21153 if (!VT.is128BitVector())
21156 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
21157 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
21158 // consecutive, non-overlapping, and in the right order.
21159 SmallVector<SDValue, 16> Elts;
21160 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
21161 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
21163 SDValue LD = EltsFromConsecutiveLoads(VT, Elts, dl, DAG, true);
21167 if (isTargetShuffle(N->getOpcode())) {
21169 PerformTargetShuffleCombine(SDValue(N, 0), DAG, DCI, Subtarget);
21170 if (Shuffle.getNode())
21173 // Try recursively combining arbitrary sequences of x86 shuffle
21174 // instructions into higher-order shuffles. We do this after combining
21175 // specific PSHUF instruction sequences into their minimal form so that we
21176 // can evaluate how many specialized shuffle instructions are involved in
21177 // a particular chain.
21178 SmallVector<int, 1> NonceMask; // Just a placeholder.
21179 NonceMask.push_back(0);
21180 if (combineX86ShufflesRecursively(SDValue(N, 0), SDValue(N, 0), NonceMask,
21181 /*Depth*/ 1, /*HasPSHUFB*/ false, DAG,
21183 return SDValue(); // This routine will use CombineTo to replace N.
21189 /// PerformTruncateCombine - Converts truncate operation to
21190 /// a sequence of vector shuffle operations.
21191 /// It is possible when we truncate 256-bit vector to 128-bit vector
21192 static SDValue PerformTruncateCombine(SDNode *N, SelectionDAG &DAG,
21193 TargetLowering::DAGCombinerInfo &DCI,
21194 const X86Subtarget *Subtarget) {
21198 /// XFormVExtractWithShuffleIntoLoad - Check if a vector extract from a target
21199 /// specific shuffle of a load can be folded into a single element load.
21200 /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
21201 /// shuffles have been customed lowered so we need to handle those here.
21202 static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
21203 TargetLowering::DAGCombinerInfo &DCI) {
21204 if (DCI.isBeforeLegalizeOps())
21207 SDValue InVec = N->getOperand(0);
21208 SDValue EltNo = N->getOperand(1);
21210 if (!isa<ConstantSDNode>(EltNo))
21213 EVT VT = InVec.getValueType();
21215 if (InVec.getOpcode() == ISD::BITCAST) {
21216 // Don't duplicate a load with other uses.
21217 if (!InVec.hasOneUse())
21219 EVT BCVT = InVec.getOperand(0).getValueType();
21220 if (BCVT.getVectorNumElements() != VT.getVectorNumElements())
21222 InVec = InVec.getOperand(0);
21225 if (!isTargetShuffle(InVec.getOpcode()))
21228 // Don't duplicate a load with other uses.
21229 if (!InVec.hasOneUse())
21232 SmallVector<int, 16> ShuffleMask;
21234 if (!getTargetShuffleMask(InVec.getNode(), VT.getSimpleVT(), ShuffleMask,
21238 // Select the input vector, guarding against out of range extract vector.
21239 unsigned NumElems = VT.getVectorNumElements();
21240 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
21241 int Idx = (Elt > (int)NumElems) ? -1 : ShuffleMask[Elt];
21242 SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0)
21243 : InVec.getOperand(1);
21245 // If inputs to shuffle are the same for both ops, then allow 2 uses
21246 unsigned AllowedUses = InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
21248 if (LdNode.getOpcode() == ISD::BITCAST) {
21249 // Don't duplicate a load with other uses.
21250 if (!LdNode.getNode()->hasNUsesOfValue(AllowedUses, 0))
21253 AllowedUses = 1; // only allow 1 load use if we have a bitcast
21254 LdNode = LdNode.getOperand(0);
21257 if (!ISD::isNormalLoad(LdNode.getNode()))
21260 LoadSDNode *LN0 = cast<LoadSDNode>(LdNode);
21262 if (!LN0 ||!LN0->hasNUsesOfValue(AllowedUses, 0) || LN0->isVolatile())
21265 EVT EltVT = N->getValueType(0);
21266 // If there's a bitcast before the shuffle, check if the load type and
21267 // alignment is valid.
21268 unsigned Align = LN0->getAlignment();
21269 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
21270 unsigned NewAlign = TLI.getDataLayout()->getABITypeAlignment(
21271 EltVT.getTypeForEVT(*DAG.getContext()));
21273 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, EltVT))
21276 // All checks match so transform back to vector_shuffle so that DAG combiner
21277 // can finish the job
21280 // Create shuffle node taking into account the case that its a unary shuffle
21281 SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(VT) : InVec.getOperand(1);
21282 Shuffle = DAG.getVectorShuffle(InVec.getValueType(), dl,
21283 InVec.getOperand(0), Shuffle,
21285 Shuffle = DAG.getNode(ISD::BITCAST, dl, VT, Shuffle);
21286 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle,
21290 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
21291 /// generation and convert it from being a bunch of shuffles and extracts
21292 /// to a simple store and scalar loads to extract the elements.
21293 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
21294 TargetLowering::DAGCombinerInfo &DCI) {
21295 SDValue NewOp = XFormVExtractWithShuffleIntoLoad(N, DAG, DCI);
21296 if (NewOp.getNode())
21299 SDValue InputVector = N->getOperand(0);
21301 // Detect whether we are trying to convert from mmx to i32 and the bitcast
21302 // from mmx to v2i32 has a single usage.
21303 if (InputVector.getNode()->getOpcode() == llvm::ISD::BITCAST &&
21304 InputVector.getNode()->getOperand(0).getValueType() == MVT::x86mmx &&
21305 InputVector.hasOneUse() && N->getValueType(0) == MVT::i32)
21306 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
21307 N->getValueType(0),
21308 InputVector.getNode()->getOperand(0));
21310 // Only operate on vectors of 4 elements, where the alternative shuffling
21311 // gets to be more expensive.
21312 if (InputVector.getValueType() != MVT::v4i32)
21315 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
21316 // single use which is a sign-extend or zero-extend, and all elements are
21318 SmallVector<SDNode *, 4> Uses;
21319 unsigned ExtractedElements = 0;
21320 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
21321 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
21322 if (UI.getUse().getResNo() != InputVector.getResNo())
21325 SDNode *Extract = *UI;
21326 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
21329 if (Extract->getValueType(0) != MVT::i32)
21331 if (!Extract->hasOneUse())
21333 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
21334 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
21336 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
21339 // Record which element was extracted.
21340 ExtractedElements |=
21341 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
21343 Uses.push_back(Extract);
21346 // If not all the elements were used, this may not be worthwhile.
21347 if (ExtractedElements != 15)
21350 // Ok, we've now decided to do the transformation.
21351 SDLoc dl(InputVector);
21353 // Store the value to a temporary stack slot.
21354 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
21355 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
21356 MachinePointerInfo(), false, false, 0);
21358 // Replace each use (extract) with a load of the appropriate element.
21359 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
21360 UE = Uses.end(); UI != UE; ++UI) {
21361 SDNode *Extract = *UI;
21363 // cOMpute the element's address.
21364 SDValue Idx = Extract->getOperand(1);
21366 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
21367 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
21368 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
21369 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
21371 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(),
21372 StackPtr, OffsetVal);
21374 // Load the scalar.
21375 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch,
21376 ScalarAddr, MachinePointerInfo(),
21377 false, false, false, 0);
21379 // Replace the exact with the load.
21380 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
21383 // The replacement was made in place; don't return anything.
21387 /// \brief Matches a VSELECT onto min/max or return 0 if the node doesn't match.
21388 static std::pair<unsigned, bool>
21389 matchIntegerMINMAX(SDValue Cond, EVT VT, SDValue LHS, SDValue RHS,
21390 SelectionDAG &DAG, const X86Subtarget *Subtarget) {
21391 if (!VT.isVector())
21392 return std::make_pair(0, false);
21394 bool NeedSplit = false;
21395 switch (VT.getSimpleVT().SimpleTy) {
21396 default: return std::make_pair(0, false);
21400 if (!Subtarget->hasAVX2())
21402 if (!Subtarget->hasAVX())
21403 return std::make_pair(0, false);
21408 if (!Subtarget->hasSSE2())
21409 return std::make_pair(0, false);
21412 // SSE2 has only a small subset of the operations.
21413 bool hasUnsigned = Subtarget->hasSSE41() ||
21414 (Subtarget->hasSSE2() && VT == MVT::v16i8);
21415 bool hasSigned = Subtarget->hasSSE41() ||
21416 (Subtarget->hasSSE2() && VT == MVT::v8i16);
21418 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
21421 // Check for x CC y ? x : y.
21422 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
21423 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
21428 Opc = hasUnsigned ? X86ISD::UMIN : 0; break;
21431 Opc = hasUnsigned ? X86ISD::UMAX : 0; break;
21434 Opc = hasSigned ? X86ISD::SMIN : 0; break;
21437 Opc = hasSigned ? X86ISD::SMAX : 0; break;
21439 // Check for x CC y ? y : x -- a min/max with reversed arms.
21440 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
21441 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
21446 Opc = hasUnsigned ? X86ISD::UMAX : 0; break;
21449 Opc = hasUnsigned ? X86ISD::UMIN : 0; break;
21452 Opc = hasSigned ? X86ISD::SMAX : 0; break;
21455 Opc = hasSigned ? X86ISD::SMIN : 0; break;
21459 return std::make_pair(Opc, NeedSplit);
21463 TransformVSELECTtoBlendVECTOR_SHUFFLE(SDNode *N, SelectionDAG &DAG,
21464 const X86Subtarget *Subtarget) {
21466 SDValue Cond = N->getOperand(0);
21467 SDValue LHS = N->getOperand(1);
21468 SDValue RHS = N->getOperand(2);
21470 if (Cond.getOpcode() == ISD::SIGN_EXTEND) {
21471 SDValue CondSrc = Cond->getOperand(0);
21472 if (CondSrc->getOpcode() == ISD::SIGN_EXTEND_INREG)
21473 Cond = CondSrc->getOperand(0);
21476 MVT VT = N->getSimpleValueType(0);
21477 MVT EltVT = VT.getVectorElementType();
21478 unsigned NumElems = VT.getVectorNumElements();
21479 // There is no blend with immediate in AVX-512.
21480 if (VT.is512BitVector())
21483 if (!Subtarget->hasSSE41() || EltVT == MVT::i8)
21485 if (!Subtarget->hasInt256() && VT == MVT::v16i16)
21488 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
21491 // A vselect where all conditions and data are constants can be optimized into
21492 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
21493 if (ISD::isBuildVectorOfConstantSDNodes(LHS.getNode()) &&
21494 ISD::isBuildVectorOfConstantSDNodes(RHS.getNode()))
21497 unsigned MaskValue = 0;
21498 if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue))
21501 SmallVector<int, 8> ShuffleMask(NumElems, -1);
21502 for (unsigned i = 0; i < NumElems; ++i) {
21503 // Be sure we emit undef where we can.
21504 if (Cond.getOperand(i)->getOpcode() == ISD::UNDEF)
21505 ShuffleMask[i] = -1;
21507 ShuffleMask[i] = i + NumElems * ((MaskValue >> i) & 1);
21510 return DAG.getVectorShuffle(VT, dl, LHS, RHS, &ShuffleMask[0]);
21513 /// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT
21515 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
21516 TargetLowering::DAGCombinerInfo &DCI,
21517 const X86Subtarget *Subtarget) {
21519 SDValue Cond = N->getOperand(0);
21520 // Get the LHS/RHS of the select.
21521 SDValue LHS = N->getOperand(1);
21522 SDValue RHS = N->getOperand(2);
21523 EVT VT = LHS.getValueType();
21524 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
21526 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
21527 // instructions match the semantics of the common C idiom x<y?x:y but not
21528 // x<=y?x:y, because of how they handle negative zero (which can be
21529 // ignored in unsafe-math mode).
21530 if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() &&
21531 VT != MVT::f80 && TLI.isTypeLegal(VT) &&
21532 (Subtarget->hasSSE2() ||
21533 (Subtarget->hasSSE1() && VT.getScalarType() == MVT::f32))) {
21534 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
21536 unsigned Opcode = 0;
21537 // Check for x CC y ? x : y.
21538 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
21539 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
21543 // Converting this to a min would handle NaNs incorrectly, and swapping
21544 // the operands would cause it to handle comparisons between positive
21545 // and negative zero incorrectly.
21546 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
21547 if (!DAG.getTarget().Options.UnsafeFPMath &&
21548 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
21550 std::swap(LHS, RHS);
21552 Opcode = X86ISD::FMIN;
21555 // Converting this to a min would handle comparisons between positive
21556 // and negative zero incorrectly.
21557 if (!DAG.getTarget().Options.UnsafeFPMath &&
21558 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
21560 Opcode = X86ISD::FMIN;
21563 // Converting this to a min would handle both negative zeros and NaNs
21564 // incorrectly, but we can swap the operands to fix both.
21565 std::swap(LHS, RHS);
21569 Opcode = X86ISD::FMIN;
21573 // Converting this to a max would handle comparisons between positive
21574 // and negative zero incorrectly.
21575 if (!DAG.getTarget().Options.UnsafeFPMath &&
21576 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
21578 Opcode = X86ISD::FMAX;
21581 // Converting this to a max would handle NaNs incorrectly, and swapping
21582 // the operands would cause it to handle comparisons between positive
21583 // and negative zero incorrectly.
21584 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
21585 if (!DAG.getTarget().Options.UnsafeFPMath &&
21586 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
21588 std::swap(LHS, RHS);
21590 Opcode = X86ISD::FMAX;
21593 // Converting this to a max would handle both negative zeros and NaNs
21594 // incorrectly, but we can swap the operands to fix both.
21595 std::swap(LHS, RHS);
21599 Opcode = X86ISD::FMAX;
21602 // Check for x CC y ? y : x -- a min/max with reversed arms.
21603 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
21604 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
21608 // Converting this to a min would handle comparisons between positive
21609 // and negative zero incorrectly, and swapping the operands would
21610 // cause it to handle NaNs incorrectly.
21611 if (!DAG.getTarget().Options.UnsafeFPMath &&
21612 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
21613 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
21615 std::swap(LHS, RHS);
21617 Opcode = X86ISD::FMIN;
21620 // Converting this to a min would handle NaNs incorrectly.
21621 if (!DAG.getTarget().Options.UnsafeFPMath &&
21622 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
21624 Opcode = X86ISD::FMIN;
21627 // Converting this to a min would handle both negative zeros and NaNs
21628 // incorrectly, but we can swap the operands to fix both.
21629 std::swap(LHS, RHS);
21633 Opcode = X86ISD::FMIN;
21637 // Converting this to a max would handle NaNs incorrectly.
21638 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
21640 Opcode = X86ISD::FMAX;
21643 // Converting this to a max would handle comparisons between positive
21644 // and negative zero incorrectly, and swapping the operands would
21645 // cause it to handle NaNs incorrectly.
21646 if (!DAG.getTarget().Options.UnsafeFPMath &&
21647 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
21648 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
21650 std::swap(LHS, RHS);
21652 Opcode = X86ISD::FMAX;
21655 // Converting this to a max would handle both negative zeros and NaNs
21656 // incorrectly, but we can swap the operands to fix both.
21657 std::swap(LHS, RHS);
21661 Opcode = X86ISD::FMAX;
21667 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
21670 EVT CondVT = Cond.getValueType();
21671 if (Subtarget->hasAVX512() && VT.isVector() && CondVT.isVector() &&
21672 CondVT.getVectorElementType() == MVT::i1) {
21673 // v16i8 (select v16i1, v16i8, v16i8) does not have a proper
21674 // lowering on KNL. In this case we convert it to
21675 // v16i8 (select v16i8, v16i8, v16i8) and use AVX instruction.
21676 // The same situation for all 128 and 256-bit vectors of i8 and i16.
21677 // Since SKX these selects have a proper lowering.
21678 EVT OpVT = LHS.getValueType();
21679 if ((OpVT.is128BitVector() || OpVT.is256BitVector()) &&
21680 (OpVT.getVectorElementType() == MVT::i8 ||
21681 OpVT.getVectorElementType() == MVT::i16) &&
21682 !(Subtarget->hasBWI() && Subtarget->hasVLX())) {
21683 Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, OpVT, Cond);
21684 DCI.AddToWorklist(Cond.getNode());
21685 return DAG.getNode(N->getOpcode(), DL, OpVT, Cond, LHS, RHS);
21688 // If this is a select between two integer constants, try to do some
21690 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
21691 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
21692 // Don't do this for crazy integer types.
21693 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
21694 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
21695 // so that TrueC (the true value) is larger than FalseC.
21696 bool NeedsCondInvert = false;
21698 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
21699 // Efficiently invertible.
21700 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
21701 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
21702 isa<ConstantSDNode>(Cond.getOperand(1))))) {
21703 NeedsCondInvert = true;
21704 std::swap(TrueC, FalseC);
21707 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
21708 if (FalseC->getAPIntValue() == 0 &&
21709 TrueC->getAPIntValue().isPowerOf2()) {
21710 if (NeedsCondInvert) // Invert the condition if needed.
21711 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
21712 DAG.getConstant(1, Cond.getValueType()));
21714 // Zero extend the condition if needed.
21715 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
21717 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
21718 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
21719 DAG.getConstant(ShAmt, MVT::i8));
21722 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
21723 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
21724 if (NeedsCondInvert) // Invert the condition if needed.
21725 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
21726 DAG.getConstant(1, Cond.getValueType()));
21728 // Zero extend the condition if needed.
21729 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
21730 FalseC->getValueType(0), Cond);
21731 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
21732 SDValue(FalseC, 0));
21735 // Optimize cases that will turn into an LEA instruction. This requires
21736 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
21737 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
21738 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
21739 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
21741 bool isFastMultiplier = false;
21743 switch ((unsigned char)Diff) {
21745 case 1: // result = add base, cond
21746 case 2: // result = lea base( , cond*2)
21747 case 3: // result = lea base(cond, cond*2)
21748 case 4: // result = lea base( , cond*4)
21749 case 5: // result = lea base(cond, cond*4)
21750 case 8: // result = lea base( , cond*8)
21751 case 9: // result = lea base(cond, cond*8)
21752 isFastMultiplier = true;
21757 if (isFastMultiplier) {
21758 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
21759 if (NeedsCondInvert) // Invert the condition if needed.
21760 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
21761 DAG.getConstant(1, Cond.getValueType()));
21763 // Zero extend the condition if needed.
21764 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
21766 // Scale the condition by the difference.
21768 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
21769 DAG.getConstant(Diff, Cond.getValueType()));
21771 // Add the base if non-zero.
21772 if (FalseC->getAPIntValue() != 0)
21773 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
21774 SDValue(FalseC, 0));
21781 // Canonicalize max and min:
21782 // (x > y) ? x : y -> (x >= y) ? x : y
21783 // (x < y) ? x : y -> (x <= y) ? x : y
21784 // This allows use of COND_S / COND_NS (see TranslateX86CC) which eliminates
21785 // the need for an extra compare
21786 // against zero. e.g.
21787 // (x - y) > 0 : (x - y) ? 0 -> (x - y) >= 0 : (x - y) ? 0
21789 // testl %edi, %edi
21791 // cmovgl %edi, %eax
21795 // cmovsl %eax, %edi
21796 if (N->getOpcode() == ISD::SELECT && Cond.getOpcode() == ISD::SETCC &&
21797 DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
21798 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
21799 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
21804 ISD::CondCode NewCC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGE;
21805 Cond = DAG.getSetCC(SDLoc(Cond), Cond.getValueType(),
21806 Cond.getOperand(0), Cond.getOperand(1), NewCC);
21807 return DAG.getNode(ISD::SELECT, DL, VT, Cond, LHS, RHS);
21812 // Early exit check
21813 if (!TLI.isTypeLegal(VT))
21816 // Match VSELECTs into subs with unsigned saturation.
21817 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
21818 // psubus is available in SSE2 and AVX2 for i8 and i16 vectors.
21819 ((Subtarget->hasSSE2() && (VT == MVT::v16i8 || VT == MVT::v8i16)) ||
21820 (Subtarget->hasAVX2() && (VT == MVT::v32i8 || VT == MVT::v16i16)))) {
21821 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
21823 // Check if one of the arms of the VSELECT is a zero vector. If it's on the
21824 // left side invert the predicate to simplify logic below.
21826 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
21828 CC = ISD::getSetCCInverse(CC, true);
21829 } else if (ISD::isBuildVectorAllZeros(RHS.getNode())) {
21833 if (Other.getNode() && Other->getNumOperands() == 2 &&
21834 DAG.isEqualTo(Other->getOperand(0), Cond.getOperand(0))) {
21835 SDValue OpLHS = Other->getOperand(0), OpRHS = Other->getOperand(1);
21836 SDValue CondRHS = Cond->getOperand(1);
21838 // Look for a general sub with unsigned saturation first.
21839 // x >= y ? x-y : 0 --> subus x, y
21840 // x > y ? x-y : 0 --> subus x, y
21841 if ((CC == ISD::SETUGE || CC == ISD::SETUGT) &&
21842 Other->getOpcode() == ISD::SUB && DAG.isEqualTo(OpRHS, CondRHS))
21843 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
21845 if (auto *OpRHSBV = dyn_cast<BuildVectorSDNode>(OpRHS))
21846 if (auto *OpRHSConst = OpRHSBV->getConstantSplatNode()) {
21847 if (auto *CondRHSBV = dyn_cast<BuildVectorSDNode>(CondRHS))
21848 if (auto *CondRHSConst = CondRHSBV->getConstantSplatNode())
21849 // If the RHS is a constant we have to reverse the const
21850 // canonicalization.
21851 // x > C-1 ? x+-C : 0 --> subus x, C
21852 if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD &&
21853 CondRHSConst->getAPIntValue() ==
21854 (-OpRHSConst->getAPIntValue() - 1))
21855 return DAG.getNode(
21856 X86ISD::SUBUS, DL, VT, OpLHS,
21857 DAG.getConstant(-OpRHSConst->getAPIntValue(), VT));
21859 // Another special case: If C was a sign bit, the sub has been
21860 // canonicalized into a xor.
21861 // FIXME: Would it be better to use computeKnownBits to determine
21862 // whether it's safe to decanonicalize the xor?
21863 // x s< 0 ? x^C : 0 --> subus x, C
21864 if (CC == ISD::SETLT && Other->getOpcode() == ISD::XOR &&
21865 ISD::isBuildVectorAllZeros(CondRHS.getNode()) &&
21866 OpRHSConst->getAPIntValue().isSignBit())
21867 // Note that we have to rebuild the RHS constant here to ensure we
21868 // don't rely on particular values of undef lanes.
21869 return DAG.getNode(
21870 X86ISD::SUBUS, DL, VT, OpLHS,
21871 DAG.getConstant(OpRHSConst->getAPIntValue(), VT));
21876 // Try to match a min/max vector operation.
21877 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC) {
21878 std::pair<unsigned, bool> ret = matchIntegerMINMAX(Cond, VT, LHS, RHS, DAG, Subtarget);
21879 unsigned Opc = ret.first;
21880 bool NeedSplit = ret.second;
21882 if (Opc && NeedSplit) {
21883 unsigned NumElems = VT.getVectorNumElements();
21884 // Extract the LHS vectors
21885 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, DL);
21886 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, DL);
21888 // Extract the RHS vectors
21889 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, DL);
21890 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, DL);
21892 // Create min/max for each subvector
21893 LHS = DAG.getNode(Opc, DL, LHS1.getValueType(), LHS1, RHS1);
21894 RHS = DAG.getNode(Opc, DL, LHS2.getValueType(), LHS2, RHS2);
21896 // Merge the result
21897 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LHS, RHS);
21899 return DAG.getNode(Opc, DL, VT, LHS, RHS);
21902 // Simplify vector selection if the selector will be produced by CMPP*/PCMP*.
21903 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
21904 // Check if SETCC has already been promoted
21905 TLI.getSetCCResultType(*DAG.getContext(), VT) == CondVT &&
21906 // Check that condition value type matches vselect operand type
21909 assert(Cond.getValueType().isVector() &&
21910 "vector select expects a vector selector!");
21912 bool TValIsAllOnes = ISD::isBuildVectorAllOnes(LHS.getNode());
21913 bool FValIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
21915 if (!TValIsAllOnes && !FValIsAllZeros) {
21916 // Try invert the condition if true value is not all 1s and false value
21918 bool TValIsAllZeros = ISD::isBuildVectorAllZeros(LHS.getNode());
21919 bool FValIsAllOnes = ISD::isBuildVectorAllOnes(RHS.getNode());
21921 if (TValIsAllZeros || FValIsAllOnes) {
21922 SDValue CC = Cond.getOperand(2);
21923 ISD::CondCode NewCC =
21924 ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
21925 Cond.getOperand(0).getValueType().isInteger());
21926 Cond = DAG.getSetCC(DL, CondVT, Cond.getOperand(0), Cond.getOperand(1), NewCC);
21927 std::swap(LHS, RHS);
21928 TValIsAllOnes = FValIsAllOnes;
21929 FValIsAllZeros = TValIsAllZeros;
21933 if (TValIsAllOnes || FValIsAllZeros) {
21936 if (TValIsAllOnes && FValIsAllZeros)
21938 else if (TValIsAllOnes)
21939 Ret = DAG.getNode(ISD::OR, DL, CondVT, Cond,
21940 DAG.getNode(ISD::BITCAST, DL, CondVT, RHS));
21941 else if (FValIsAllZeros)
21942 Ret = DAG.getNode(ISD::AND, DL, CondVT, Cond,
21943 DAG.getNode(ISD::BITCAST, DL, CondVT, LHS));
21945 return DAG.getNode(ISD::BITCAST, DL, VT, Ret);
21949 // Try to fold this VSELECT into a MOVSS/MOVSD
21950 if (N->getOpcode() == ISD::VSELECT &&
21951 Cond.getOpcode() == ISD::BUILD_VECTOR && !DCI.isBeforeLegalize()) {
21952 if (VT == MVT::v4i32 || VT == MVT::v4f32 ||
21953 (Subtarget->hasSSE2() && (VT == MVT::v2i64 || VT == MVT::v2f64))) {
21954 bool CanFold = false;
21955 unsigned NumElems = Cond.getNumOperands();
21959 if (isZero(Cond.getOperand(0))) {
21962 // fold (vselect <0,-1,-1,-1>, A, B) -> (movss A, B)
21963 // fold (vselect <0,-1> -> (movsd A, B)
21964 for (unsigned i = 1, e = NumElems; i != e && CanFold; ++i)
21965 CanFold = isAllOnes(Cond.getOperand(i));
21966 } else if (isAllOnes(Cond.getOperand(0))) {
21970 // fold (vselect <-1,0,0,0>, A, B) -> (movss B, A)
21971 // fold (vselect <-1,0> -> (movsd B, A)
21972 for (unsigned i = 1, e = NumElems; i != e && CanFold; ++i)
21973 CanFold = isZero(Cond.getOperand(i));
21977 if (VT == MVT::v4i32 || VT == MVT::v4f32)
21978 return getTargetShuffleNode(X86ISD::MOVSS, DL, VT, A, B, DAG);
21979 return getTargetShuffleNode(X86ISD::MOVSD, DL, VT, A, B, DAG);
21982 if (Subtarget->hasSSE2() && (VT == MVT::v4i32 || VT == MVT::v4f32)) {
21983 // fold (v4i32: vselect <0,0,-1,-1>, A, B) ->
21984 // (v4i32 (bitcast (movsd (v2i64 (bitcast A)),
21985 // (v2i64 (bitcast B)))))
21987 // fold (v4f32: vselect <0,0,-1,-1>, A, B) ->
21988 // (v4f32 (bitcast (movsd (v2f64 (bitcast A)),
21989 // (v2f64 (bitcast B)))))
21991 // fold (v4i32: vselect <-1,-1,0,0>, A, B) ->
21992 // (v4i32 (bitcast (movsd (v2i64 (bitcast B)),
21993 // (v2i64 (bitcast A)))))
21995 // fold (v4f32: vselect <-1,-1,0,0>, A, B) ->
21996 // (v4f32 (bitcast (movsd (v2f64 (bitcast B)),
21997 // (v2f64 (bitcast A)))))
21999 CanFold = (isZero(Cond.getOperand(0)) &&
22000 isZero(Cond.getOperand(1)) &&
22001 isAllOnes(Cond.getOperand(2)) &&
22002 isAllOnes(Cond.getOperand(3)));
22004 if (!CanFold && isAllOnes(Cond.getOperand(0)) &&
22005 isAllOnes(Cond.getOperand(1)) &&
22006 isZero(Cond.getOperand(2)) &&
22007 isZero(Cond.getOperand(3))) {
22009 std::swap(LHS, RHS);
22013 EVT NVT = (VT == MVT::v4i32) ? MVT::v2i64 : MVT::v2f64;
22014 SDValue NewA = DAG.getNode(ISD::BITCAST, DL, NVT, LHS);
22015 SDValue NewB = DAG.getNode(ISD::BITCAST, DL, NVT, RHS);
22016 SDValue Select = getTargetShuffleNode(X86ISD::MOVSD, DL, NVT, NewA,
22018 return DAG.getNode(ISD::BITCAST, DL, VT, Select);
22024 // If we know that this node is legal then we know that it is going to be
22025 // matched by one of the SSE/AVX BLEND instructions. These instructions only
22026 // depend on the highest bit in each word. Try to use SimplifyDemandedBits
22027 // to simplify previous instructions.
22028 if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() &&
22029 !DCI.isBeforeLegalize() &&
22030 // We explicitly check against v8i16 and v16i16 because, although
22031 // they're marked as Custom, they might only be legal when Cond is a
22032 // build_vector of constants. This will be taken care in a later
22034 (TLI.isOperationLegalOrCustom(ISD::VSELECT, VT) && VT != MVT::v16i16 &&
22035 VT != MVT::v8i16)) {
22036 unsigned BitWidth = Cond.getValueType().getScalarType().getSizeInBits();
22038 // Don't optimize vector selects that map to mask-registers.
22042 // Check all uses of that condition operand to check whether it will be
22043 // consumed by non-BLEND instructions, which may depend on all bits are set
22045 for (SDNode::use_iterator I = Cond->use_begin(),
22046 E = Cond->use_end(); I != E; ++I)
22047 if (I->getOpcode() != ISD::VSELECT)
22048 // TODO: Add other opcodes eventually lowered into BLEND.
22051 assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size");
22052 APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);
22054 APInt KnownZero, KnownOne;
22055 TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(),
22056 DCI.isBeforeLegalizeOps());
22057 if (TLO.ShrinkDemandedConstant(Cond, DemandedMask) ||
22058 TLI.SimplifyDemandedBits(Cond, DemandedMask, KnownZero, KnownOne, TLO))
22059 DCI.CommitTargetLoweringOpt(TLO);
22062 // We should generate an X86ISD::BLENDI from a vselect if its argument
22063 // is a sign_extend_inreg of an any_extend of a BUILD_VECTOR of
22064 // constants. This specific pattern gets generated when we split a
22065 // selector for a 512 bit vector in a machine without AVX512 (but with
22066 // 256-bit vectors), during legalization:
22068 // (vselect (sign_extend (any_extend (BUILD_VECTOR)) i1) LHS RHS)
22070 // Iff we find this pattern and the build_vectors are built from
22071 // constants, we translate the vselect into a shuffle_vector that we
22072 // know will be matched by LowerVECTOR_SHUFFLEtoBlend.
22073 if (N->getOpcode() == ISD::VSELECT && !DCI.isBeforeLegalize()) {
22074 SDValue Shuffle = TransformVSELECTtoBlendVECTOR_SHUFFLE(N, DAG, Subtarget);
22075 if (Shuffle.getNode())
22082 // Check whether a boolean test is testing a boolean value generated by
22083 // X86ISD::SETCC. If so, return the operand of that SETCC and proper condition
22086 // Simplify the following patterns:
22087 // (Op (CMP (SETCC Cond EFLAGS) 1) EQ) or
22088 // (Op (CMP (SETCC Cond EFLAGS) 0) NEQ)
22089 // to (Op EFLAGS Cond)
22091 // (Op (CMP (SETCC Cond EFLAGS) 0) EQ) or
22092 // (Op (CMP (SETCC Cond EFLAGS) 1) NEQ)
22093 // to (Op EFLAGS !Cond)
22095 // where Op could be BRCOND or CMOV.
22097 static SDValue checkBoolTestSetCCCombine(SDValue Cmp, X86::CondCode &CC) {
22098 // Quit if not CMP and SUB with its value result used.
22099 if (Cmp.getOpcode() != X86ISD::CMP &&
22100 (Cmp.getOpcode() != X86ISD::SUB || Cmp.getNode()->hasAnyUseOfValue(0)))
22103 // Quit if not used as a boolean value.
22104 if (CC != X86::COND_E && CC != X86::COND_NE)
22107 // Check CMP operands. One of them should be 0 or 1 and the other should be
22108 // an SetCC or extended from it.
22109 SDValue Op1 = Cmp.getOperand(0);
22110 SDValue Op2 = Cmp.getOperand(1);
22113 const ConstantSDNode* C = nullptr;
22114 bool needOppositeCond = (CC == X86::COND_E);
22115 bool checkAgainstTrue = false; // Is it a comparison against 1?
22117 if ((C = dyn_cast<ConstantSDNode>(Op1)))
22119 else if ((C = dyn_cast<ConstantSDNode>(Op2)))
22121 else // Quit if all operands are not constants.
22124 if (C->getZExtValue() == 1) {
22125 needOppositeCond = !needOppositeCond;
22126 checkAgainstTrue = true;
22127 } else if (C->getZExtValue() != 0)
22128 // Quit if the constant is neither 0 or 1.
22131 bool truncatedToBoolWithAnd = false;
22132 // Skip (zext $x), (trunc $x), or (and $x, 1) node.
22133 while (SetCC.getOpcode() == ISD::ZERO_EXTEND ||
22134 SetCC.getOpcode() == ISD::TRUNCATE ||
22135 SetCC.getOpcode() == ISD::AND) {
22136 if (SetCC.getOpcode() == ISD::AND) {
22138 ConstantSDNode *CS;
22139 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(0))) &&
22140 CS->getZExtValue() == 1)
22142 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(1))) &&
22143 CS->getZExtValue() == 1)
22147 SetCC = SetCC.getOperand(OpIdx);
22148 truncatedToBoolWithAnd = true;
22150 SetCC = SetCC.getOperand(0);
22153 switch (SetCC.getOpcode()) {
22154 case X86ISD::SETCC_CARRY:
22155 // Since SETCC_CARRY gives output based on R = CF ? ~0 : 0, it's unsafe to
22156 // simplify it if the result of SETCC_CARRY is not canonicalized to 0 or 1,
22157 // i.e. it's a comparison against true but the result of SETCC_CARRY is not
22158 // truncated to i1 using 'and'.
22159 if (checkAgainstTrue && !truncatedToBoolWithAnd)
22161 assert(X86::CondCode(SetCC.getConstantOperandVal(0)) == X86::COND_B &&
22162 "Invalid use of SETCC_CARRY!");
22164 case X86ISD::SETCC:
22165 // Set the condition code or opposite one if necessary.
22166 CC = X86::CondCode(SetCC.getConstantOperandVal(0));
22167 if (needOppositeCond)
22168 CC = X86::GetOppositeBranchCondition(CC);
22169 return SetCC.getOperand(1);
22170 case X86ISD::CMOV: {
22171 // Check whether false/true value has canonical one, i.e. 0 or 1.
22172 ConstantSDNode *FVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(0));
22173 ConstantSDNode *TVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(1));
22174 // Quit if true value is not a constant.
22177 // Quit if false value is not a constant.
22179 SDValue Op = SetCC.getOperand(0);
22180 // Skip 'zext' or 'trunc' node.
22181 if (Op.getOpcode() == ISD::ZERO_EXTEND ||
22182 Op.getOpcode() == ISD::TRUNCATE)
22183 Op = Op.getOperand(0);
22184 // A special case for rdrand/rdseed, where 0 is set if false cond is
22186 if ((Op.getOpcode() != X86ISD::RDRAND &&
22187 Op.getOpcode() != X86ISD::RDSEED) || Op.getResNo() != 0)
22190 // Quit if false value is not the constant 0 or 1.
22191 bool FValIsFalse = true;
22192 if (FVal && FVal->getZExtValue() != 0) {
22193 if (FVal->getZExtValue() != 1)
22195 // If FVal is 1, opposite cond is needed.
22196 needOppositeCond = !needOppositeCond;
22197 FValIsFalse = false;
22199 // Quit if TVal is not the constant opposite of FVal.
22200 if (FValIsFalse && TVal->getZExtValue() != 1)
22202 if (!FValIsFalse && TVal->getZExtValue() != 0)
22204 CC = X86::CondCode(SetCC.getConstantOperandVal(2));
22205 if (needOppositeCond)
22206 CC = X86::GetOppositeBranchCondition(CC);
22207 return SetCC.getOperand(3);
22214 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
22215 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
22216 TargetLowering::DAGCombinerInfo &DCI,
22217 const X86Subtarget *Subtarget) {
22220 // If the flag operand isn't dead, don't touch this CMOV.
22221 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
22224 SDValue FalseOp = N->getOperand(0);
22225 SDValue TrueOp = N->getOperand(1);
22226 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
22227 SDValue Cond = N->getOperand(3);
22229 if (CC == X86::COND_E || CC == X86::COND_NE) {
22230 switch (Cond.getOpcode()) {
22234 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
22235 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
22236 return (CC == X86::COND_E) ? FalseOp : TrueOp;
22242 Flags = checkBoolTestSetCCCombine(Cond, CC);
22243 if (Flags.getNode() &&
22244 // Extra check as FCMOV only supports a subset of X86 cond.
22245 (FalseOp.getValueType() != MVT::f80 || hasFPCMov(CC))) {
22246 SDValue Ops[] = { FalseOp, TrueOp,
22247 DAG.getConstant(CC, MVT::i8), Flags };
22248 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
22251 // If this is a select between two integer constants, try to do some
22252 // optimizations. Note that the operands are ordered the opposite of SELECT
22254 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
22255 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
22256 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
22257 // larger than FalseC (the false value).
22258 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
22259 CC = X86::GetOppositeBranchCondition(CC);
22260 std::swap(TrueC, FalseC);
22261 std::swap(TrueOp, FalseOp);
22264 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
22265 // This is efficient for any integer data type (including i8/i16) and
22267 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
22268 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
22269 DAG.getConstant(CC, MVT::i8), Cond);
22271 // Zero extend the condition if needed.
22272 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
22274 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
22275 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
22276 DAG.getConstant(ShAmt, MVT::i8));
22277 if (N->getNumValues() == 2) // Dead flag value?
22278 return DCI.CombineTo(N, Cond, SDValue());
22282 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
22283 // for any integer data type, including i8/i16.
22284 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
22285 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
22286 DAG.getConstant(CC, MVT::i8), Cond);
22288 // Zero extend the condition if needed.
22289 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
22290 FalseC->getValueType(0), Cond);
22291 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
22292 SDValue(FalseC, 0));
22294 if (N->getNumValues() == 2) // Dead flag value?
22295 return DCI.CombineTo(N, Cond, SDValue());
22299 // Optimize cases that will turn into an LEA instruction. This requires
22300 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
22301 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
22302 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
22303 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
22305 bool isFastMultiplier = false;
22307 switch ((unsigned char)Diff) {
22309 case 1: // result = add base, cond
22310 case 2: // result = lea base( , cond*2)
22311 case 3: // result = lea base(cond, cond*2)
22312 case 4: // result = lea base( , cond*4)
22313 case 5: // result = lea base(cond, cond*4)
22314 case 8: // result = lea base( , cond*8)
22315 case 9: // result = lea base(cond, cond*8)
22316 isFastMultiplier = true;
22321 if (isFastMultiplier) {
22322 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
22323 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
22324 DAG.getConstant(CC, MVT::i8), Cond);
22325 // Zero extend the condition if needed.
22326 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
22328 // Scale the condition by the difference.
22330 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
22331 DAG.getConstant(Diff, Cond.getValueType()));
22333 // Add the base if non-zero.
22334 if (FalseC->getAPIntValue() != 0)
22335 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
22336 SDValue(FalseC, 0));
22337 if (N->getNumValues() == 2) // Dead flag value?
22338 return DCI.CombineTo(N, Cond, SDValue());
22345 // Handle these cases:
22346 // (select (x != c), e, c) -> select (x != c), e, x),
22347 // (select (x == c), c, e) -> select (x == c), x, e)
22348 // where the c is an integer constant, and the "select" is the combination
22349 // of CMOV and CMP.
22351 // The rationale for this change is that the conditional-move from a constant
22352 // needs two instructions, however, conditional-move from a register needs
22353 // only one instruction.
22355 // CAVEAT: By replacing a constant with a symbolic value, it may obscure
22356 // some instruction-combining opportunities. This opt needs to be
22357 // postponed as late as possible.
22359 if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) {
22360 // the DCI.xxxx conditions are provided to postpone the optimization as
22361 // late as possible.
22363 ConstantSDNode *CmpAgainst = nullptr;
22364 if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) &&
22365 (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) &&
22366 !isa<ConstantSDNode>(Cond.getOperand(0))) {
22368 if (CC == X86::COND_NE &&
22369 CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) {
22370 CC = X86::GetOppositeBranchCondition(CC);
22371 std::swap(TrueOp, FalseOp);
22374 if (CC == X86::COND_E &&
22375 CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) {
22376 SDValue Ops[] = { FalseOp, Cond.getOperand(0),
22377 DAG.getConstant(CC, MVT::i8), Cond };
22378 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops);
22386 static SDValue PerformINTRINSIC_WO_CHAINCombine(SDNode *N, SelectionDAG &DAG,
22387 const X86Subtarget *Subtarget) {
22388 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
22390 default: return SDValue();
22391 // SSE/AVX/AVX2 blend intrinsics.
22392 case Intrinsic::x86_avx2_pblendvb:
22393 case Intrinsic::x86_avx2_pblendw:
22394 case Intrinsic::x86_avx2_pblendd_128:
22395 case Intrinsic::x86_avx2_pblendd_256:
22396 // Don't try to simplify this intrinsic if we don't have AVX2.
22397 if (!Subtarget->hasAVX2())
22400 case Intrinsic::x86_avx_blend_pd_256:
22401 case Intrinsic::x86_avx_blend_ps_256:
22402 case Intrinsic::x86_avx_blendv_pd_256:
22403 case Intrinsic::x86_avx_blendv_ps_256:
22404 // Don't try to simplify this intrinsic if we don't have AVX.
22405 if (!Subtarget->hasAVX())
22408 case Intrinsic::x86_sse41_pblendw:
22409 case Intrinsic::x86_sse41_blendpd:
22410 case Intrinsic::x86_sse41_blendps:
22411 case Intrinsic::x86_sse41_blendvps:
22412 case Intrinsic::x86_sse41_blendvpd:
22413 case Intrinsic::x86_sse41_pblendvb: {
22414 SDValue Op0 = N->getOperand(1);
22415 SDValue Op1 = N->getOperand(2);
22416 SDValue Mask = N->getOperand(3);
22418 // Don't try to simplify this intrinsic if we don't have SSE4.1.
22419 if (!Subtarget->hasSSE41())
22422 // fold (blend A, A, Mask) -> A
22425 // fold (blend A, B, allZeros) -> A
22426 if (ISD::isBuildVectorAllZeros(Mask.getNode()))
22428 // fold (blend A, B, allOnes) -> B
22429 if (ISD::isBuildVectorAllOnes(Mask.getNode()))
22432 // Simplify the case where the mask is a constant i32 value.
22433 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Mask)) {
22434 if (C->isNullValue())
22436 if (C->isAllOnesValue())
22443 // Packed SSE2/AVX2 arithmetic shift immediate intrinsics.
22444 case Intrinsic::x86_sse2_psrai_w:
22445 case Intrinsic::x86_sse2_psrai_d:
22446 case Intrinsic::x86_avx2_psrai_w:
22447 case Intrinsic::x86_avx2_psrai_d:
22448 case Intrinsic::x86_sse2_psra_w:
22449 case Intrinsic::x86_sse2_psra_d:
22450 case Intrinsic::x86_avx2_psra_w:
22451 case Intrinsic::x86_avx2_psra_d: {
22452 SDValue Op0 = N->getOperand(1);
22453 SDValue Op1 = N->getOperand(2);
22454 EVT VT = Op0.getValueType();
22455 assert(VT.isVector() && "Expected a vector type!");
22457 if (isa<BuildVectorSDNode>(Op1))
22458 Op1 = Op1.getOperand(0);
22460 if (!isa<ConstantSDNode>(Op1))
22463 EVT SVT = VT.getVectorElementType();
22464 unsigned SVTBits = SVT.getSizeInBits();
22466 ConstantSDNode *CND = cast<ConstantSDNode>(Op1);
22467 const APInt &C = APInt(SVTBits, CND->getAPIntValue().getZExtValue());
22468 uint64_t ShAmt = C.getZExtValue();
22470 // Don't try to convert this shift into a ISD::SRA if the shift
22471 // count is bigger than or equal to the element size.
22472 if (ShAmt >= SVTBits)
22475 // Trivial case: if the shift count is zero, then fold this
22476 // into the first operand.
22480 // Replace this packed shift intrinsic with a target independent
22482 SDValue Splat = DAG.getConstant(C, VT);
22483 return DAG.getNode(ISD::SRA, SDLoc(N), VT, Op0, Splat);
22488 /// PerformMulCombine - Optimize a single multiply with constant into two
22489 /// in order to implement it with two cheaper instructions, e.g.
22490 /// LEA + SHL, LEA + LEA.
22491 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
22492 TargetLowering::DAGCombinerInfo &DCI) {
22493 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
22496 EVT VT = N->getValueType(0);
22497 if (VT != MVT::i64)
22500 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
22503 uint64_t MulAmt = C->getZExtValue();
22504 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
22507 uint64_t MulAmt1 = 0;
22508 uint64_t MulAmt2 = 0;
22509 if ((MulAmt % 9) == 0) {
22511 MulAmt2 = MulAmt / 9;
22512 } else if ((MulAmt % 5) == 0) {
22514 MulAmt2 = MulAmt / 5;
22515 } else if ((MulAmt % 3) == 0) {
22517 MulAmt2 = MulAmt / 3;
22520 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
22523 if (isPowerOf2_64(MulAmt2) &&
22524 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
22525 // If second multiplifer is pow2, issue it first. We want the multiply by
22526 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
22528 std::swap(MulAmt1, MulAmt2);
22531 if (isPowerOf2_64(MulAmt1))
22532 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
22533 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
22535 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
22536 DAG.getConstant(MulAmt1, VT));
22538 if (isPowerOf2_64(MulAmt2))
22539 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
22540 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
22542 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
22543 DAG.getConstant(MulAmt2, VT));
22545 // Do not add new nodes to DAG combiner worklist.
22546 DCI.CombineTo(N, NewMul, false);
22551 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
22552 SDValue N0 = N->getOperand(0);
22553 SDValue N1 = N->getOperand(1);
22554 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
22555 EVT VT = N0.getValueType();
22557 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
22558 // since the result of setcc_c is all zero's or all ones.
22559 if (VT.isInteger() && !VT.isVector() &&
22560 N1C && N0.getOpcode() == ISD::AND &&
22561 N0.getOperand(1).getOpcode() == ISD::Constant) {
22562 SDValue N00 = N0.getOperand(0);
22563 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
22564 ((N00.getOpcode() == ISD::ANY_EXTEND ||
22565 N00.getOpcode() == ISD::ZERO_EXTEND) &&
22566 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
22567 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
22568 APInt ShAmt = N1C->getAPIntValue();
22569 Mask = Mask.shl(ShAmt);
22571 return DAG.getNode(ISD::AND, SDLoc(N), VT,
22572 N00, DAG.getConstant(Mask, VT));
22576 // Hardware support for vector shifts is sparse which makes us scalarize the
22577 // vector operations in many cases. Also, on sandybridge ADD is faster than
22579 // (shl V, 1) -> add V,V
22580 if (auto *N1BV = dyn_cast<BuildVectorSDNode>(N1))
22581 if (auto *N1SplatC = N1BV->getConstantSplatNode()) {
22582 assert(N0.getValueType().isVector() && "Invalid vector shift type");
22583 // We shift all of the values by one. In many cases we do not have
22584 // hardware support for this operation. This is better expressed as an ADD
22586 if (N1SplatC->getZExtValue() == 1)
22587 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N0);
22593 /// \brief Returns a vector of 0s if the node in input is a vector logical
22594 /// shift by a constant amount which is known to be bigger than or equal
22595 /// to the vector element size in bits.
22596 static SDValue performShiftToAllZeros(SDNode *N, SelectionDAG &DAG,
22597 const X86Subtarget *Subtarget) {
22598 EVT VT = N->getValueType(0);
22600 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16 &&
22601 (!Subtarget->hasInt256() ||
22602 (VT != MVT::v4i64 && VT != MVT::v8i32 && VT != MVT::v16i16)))
22605 SDValue Amt = N->getOperand(1);
22607 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Amt))
22608 if (auto *AmtSplat = AmtBV->getConstantSplatNode()) {
22609 APInt ShiftAmt = AmtSplat->getAPIntValue();
22610 unsigned MaxAmount = VT.getVectorElementType().getSizeInBits();
22612 // SSE2/AVX2 logical shifts always return a vector of 0s
22613 // if the shift amount is bigger than or equal to
22614 // the element size. The constant shift amount will be
22615 // encoded as a 8-bit immediate.
22616 if (ShiftAmt.trunc(8).uge(MaxAmount))
22617 return getZeroVector(VT, Subtarget, DAG, DL);
22623 /// PerformShiftCombine - Combine shifts.
22624 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
22625 TargetLowering::DAGCombinerInfo &DCI,
22626 const X86Subtarget *Subtarget) {
22627 if (N->getOpcode() == ISD::SHL) {
22628 SDValue V = PerformSHLCombine(N, DAG);
22629 if (V.getNode()) return V;
22632 if (N->getOpcode() != ISD::SRA) {
22633 // Try to fold this logical shift into a zero vector.
22634 SDValue V = performShiftToAllZeros(N, DAG, Subtarget);
22635 if (V.getNode()) return V;
22641 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
22642 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
22643 // and friends. Likewise for OR -> CMPNEQSS.
22644 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
22645 TargetLowering::DAGCombinerInfo &DCI,
22646 const X86Subtarget *Subtarget) {
22649 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
22650 // we're requiring SSE2 for both.
22651 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
22652 SDValue N0 = N->getOperand(0);
22653 SDValue N1 = N->getOperand(1);
22654 SDValue CMP0 = N0->getOperand(1);
22655 SDValue CMP1 = N1->getOperand(1);
22658 // The SETCCs should both refer to the same CMP.
22659 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
22662 SDValue CMP00 = CMP0->getOperand(0);
22663 SDValue CMP01 = CMP0->getOperand(1);
22664 EVT VT = CMP00.getValueType();
22666 if (VT == MVT::f32 || VT == MVT::f64) {
22667 bool ExpectingFlags = false;
22668 // Check for any users that want flags:
22669 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
22670 !ExpectingFlags && UI != UE; ++UI)
22671 switch (UI->getOpcode()) {
22676 ExpectingFlags = true;
22678 case ISD::CopyToReg:
22679 case ISD::SIGN_EXTEND:
22680 case ISD::ZERO_EXTEND:
22681 case ISD::ANY_EXTEND:
22685 if (!ExpectingFlags) {
22686 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
22687 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
22689 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
22690 X86::CondCode tmp = cc0;
22695 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
22696 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
22697 // FIXME: need symbolic constants for these magic numbers.
22698 // See X86ATTInstPrinter.cpp:printSSECC().
22699 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
22700 if (Subtarget->hasAVX512()) {
22701 SDValue FSetCC = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CMP00,
22702 CMP01, DAG.getConstant(x86cc, MVT::i8));
22703 if (N->getValueType(0) != MVT::i1)
22704 return DAG.getNode(ISD::ZERO_EXTEND, DL, N->getValueType(0),
22708 SDValue OnesOrZeroesF = DAG.getNode(X86ISD::FSETCC, DL,
22709 CMP00.getValueType(), CMP00, CMP01,
22710 DAG.getConstant(x86cc, MVT::i8));
22712 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
22713 MVT IntVT = is64BitFP ? MVT::i64 : MVT::i32;
22715 if (is64BitFP && !Subtarget->is64Bit()) {
22716 // On a 32-bit target, we cannot bitcast the 64-bit float to a
22717 // 64-bit integer, since that's not a legal type. Since
22718 // OnesOrZeroesF is all ones of all zeroes, we don't need all the
22719 // bits, but can do this little dance to extract the lowest 32 bits
22720 // and work with those going forward.
22721 SDValue Vector64 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64,
22723 SDValue Vector32 = DAG.getNode(ISD::BITCAST, DL, MVT::v4f32,
22725 OnesOrZeroesF = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32,
22726 Vector32, DAG.getIntPtrConstant(0));
22730 SDValue OnesOrZeroesI = DAG.getNode(ISD::BITCAST, DL, IntVT, OnesOrZeroesF);
22731 SDValue ANDed = DAG.getNode(ISD::AND, DL, IntVT, OnesOrZeroesI,
22732 DAG.getConstant(1, IntVT));
22733 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8, ANDed);
22734 return OneBitOfTruth;
22742 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
22743 /// so it can be folded inside ANDNP.
22744 static bool CanFoldXORWithAllOnes(const SDNode *N) {
22745 EVT VT = N->getValueType(0);
22747 // Match direct AllOnes for 128 and 256-bit vectors
22748 if (ISD::isBuildVectorAllOnes(N))
22751 // Look through a bit convert.
22752 if (N->getOpcode() == ISD::BITCAST)
22753 N = N->getOperand(0).getNode();
22755 // Sometimes the operand may come from a insert_subvector building a 256-bit
22757 if (VT.is256BitVector() &&
22758 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
22759 SDValue V1 = N->getOperand(0);
22760 SDValue V2 = N->getOperand(1);
22762 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
22763 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
22764 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
22765 ISD::isBuildVectorAllOnes(V2.getNode()))
22772 // On AVX/AVX2 the type v8i1 is legalized to v8i16, which is an XMM sized
22773 // register. In most cases we actually compare or select YMM-sized registers
22774 // and mixing the two types creates horrible code. This method optimizes
22775 // some of the transition sequences.
22776 static SDValue WidenMaskArithmetic(SDNode *N, SelectionDAG &DAG,
22777 TargetLowering::DAGCombinerInfo &DCI,
22778 const X86Subtarget *Subtarget) {
22779 EVT VT = N->getValueType(0);
22780 if (!VT.is256BitVector())
22783 assert((N->getOpcode() == ISD::ANY_EXTEND ||
22784 N->getOpcode() == ISD::ZERO_EXTEND ||
22785 N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node");
22787 SDValue Narrow = N->getOperand(0);
22788 EVT NarrowVT = Narrow->getValueType(0);
22789 if (!NarrowVT.is128BitVector())
22792 if (Narrow->getOpcode() != ISD::XOR &&
22793 Narrow->getOpcode() != ISD::AND &&
22794 Narrow->getOpcode() != ISD::OR)
22797 SDValue N0 = Narrow->getOperand(0);
22798 SDValue N1 = Narrow->getOperand(1);
22801 // The Left side has to be a trunc.
22802 if (N0.getOpcode() != ISD::TRUNCATE)
22805 // The type of the truncated inputs.
22806 EVT WideVT = N0->getOperand(0)->getValueType(0);
22810 // The right side has to be a 'trunc' or a constant vector.
22811 bool RHSTrunc = N1.getOpcode() == ISD::TRUNCATE;
22812 ConstantSDNode *RHSConstSplat = nullptr;
22813 if (auto *RHSBV = dyn_cast<BuildVectorSDNode>(N1))
22814 RHSConstSplat = RHSBV->getConstantSplatNode();
22815 if (!RHSTrunc && !RHSConstSplat)
22818 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22820 if (!TLI.isOperationLegalOrPromote(Narrow->getOpcode(), WideVT))
22823 // Set N0 and N1 to hold the inputs to the new wide operation.
22824 N0 = N0->getOperand(0);
22825 if (RHSConstSplat) {
22826 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT.getScalarType(),
22827 SDValue(RHSConstSplat, 0));
22828 SmallVector<SDValue, 8> C(WideVT.getVectorNumElements(), N1);
22829 N1 = DAG.getNode(ISD::BUILD_VECTOR, DL, WideVT, C);
22830 } else if (RHSTrunc) {
22831 N1 = N1->getOperand(0);
22834 // Generate the wide operation.
22835 SDValue Op = DAG.getNode(Narrow->getOpcode(), DL, WideVT, N0, N1);
22836 unsigned Opcode = N->getOpcode();
22838 case ISD::ANY_EXTEND:
22840 case ISD::ZERO_EXTEND: {
22841 unsigned InBits = NarrowVT.getScalarType().getSizeInBits();
22842 APInt Mask = APInt::getAllOnesValue(InBits);
22843 Mask = Mask.zext(VT.getScalarType().getSizeInBits());
22844 return DAG.getNode(ISD::AND, DL, VT,
22845 Op, DAG.getConstant(Mask, VT));
22847 case ISD::SIGN_EXTEND:
22848 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT,
22849 Op, DAG.getValueType(NarrowVT));
22851 llvm_unreachable("Unexpected opcode");
22855 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
22856 TargetLowering::DAGCombinerInfo &DCI,
22857 const X86Subtarget *Subtarget) {
22858 EVT VT = N->getValueType(0);
22859 if (DCI.isBeforeLegalizeOps())
22862 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
22866 // Create BEXTR instructions
22867 // BEXTR is ((X >> imm) & (2**size-1))
22868 if (VT == MVT::i32 || VT == MVT::i64) {
22869 SDValue N0 = N->getOperand(0);
22870 SDValue N1 = N->getOperand(1);
22873 // Check for BEXTR.
22874 if ((Subtarget->hasBMI() || Subtarget->hasTBM()) &&
22875 (N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::SRL)) {
22876 ConstantSDNode *MaskNode = dyn_cast<ConstantSDNode>(N1);
22877 ConstantSDNode *ShiftNode = dyn_cast<ConstantSDNode>(N0.getOperand(1));
22878 if (MaskNode && ShiftNode) {
22879 uint64_t Mask = MaskNode->getZExtValue();
22880 uint64_t Shift = ShiftNode->getZExtValue();
22881 if (isMask_64(Mask)) {
22882 uint64_t MaskSize = CountPopulation_64(Mask);
22883 if (Shift + MaskSize <= VT.getSizeInBits())
22884 return DAG.getNode(X86ISD::BEXTR, DL, VT, N0.getOperand(0),
22885 DAG.getConstant(Shift | (MaskSize << 8), VT));
22893 // Want to form ANDNP nodes:
22894 // 1) In the hopes of then easily combining them with OR and AND nodes
22895 // to form PBLEND/PSIGN.
22896 // 2) To match ANDN packed intrinsics
22897 if (VT != MVT::v2i64 && VT != MVT::v4i64)
22900 SDValue N0 = N->getOperand(0);
22901 SDValue N1 = N->getOperand(1);
22904 // Check LHS for vnot
22905 if (N0.getOpcode() == ISD::XOR &&
22906 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
22907 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
22908 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
22910 // Check RHS for vnot
22911 if (N1.getOpcode() == ISD::XOR &&
22912 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
22913 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
22914 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
22919 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
22920 TargetLowering::DAGCombinerInfo &DCI,
22921 const X86Subtarget *Subtarget) {
22922 if (DCI.isBeforeLegalizeOps())
22925 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
22929 SDValue N0 = N->getOperand(0);
22930 SDValue N1 = N->getOperand(1);
22931 EVT VT = N->getValueType(0);
22933 // look for psign/blend
22934 if (VT == MVT::v2i64 || VT == MVT::v4i64) {
22935 if (!Subtarget->hasSSSE3() ||
22936 (VT == MVT::v4i64 && !Subtarget->hasInt256()))
22939 // Canonicalize pandn to RHS
22940 if (N0.getOpcode() == X86ISD::ANDNP)
22942 // or (and (m, y), (pandn m, x))
22943 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
22944 SDValue Mask = N1.getOperand(0);
22945 SDValue X = N1.getOperand(1);
22947 if (N0.getOperand(0) == Mask)
22948 Y = N0.getOperand(1);
22949 if (N0.getOperand(1) == Mask)
22950 Y = N0.getOperand(0);
22952 // Check to see if the mask appeared in both the AND and ANDNP and
22956 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
22957 // Look through mask bitcast.
22958 if (Mask.getOpcode() == ISD::BITCAST)
22959 Mask = Mask.getOperand(0);
22960 if (X.getOpcode() == ISD::BITCAST)
22961 X = X.getOperand(0);
22962 if (Y.getOpcode() == ISD::BITCAST)
22963 Y = Y.getOperand(0);
22965 EVT MaskVT = Mask.getValueType();
22967 // Validate that the Mask operand is a vector sra node.
22968 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
22969 // there is no psrai.b
22970 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
22971 unsigned SraAmt = ~0;
22972 if (Mask.getOpcode() == ISD::SRA) {
22973 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Mask.getOperand(1)))
22974 if (auto *AmtConst = AmtBV->getConstantSplatNode())
22975 SraAmt = AmtConst->getZExtValue();
22976 } else if (Mask.getOpcode() == X86ISD::VSRAI) {
22977 SDValue SraC = Mask.getOperand(1);
22978 SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
22980 if ((SraAmt + 1) != EltBits)
22985 // Now we know we at least have a plendvb with the mask val. See if
22986 // we can form a psignb/w/d.
22987 // psign = x.type == y.type == mask.type && y = sub(0, x);
22988 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
22989 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
22990 X.getValueType() == MaskVT && Y.getValueType() == MaskVT) {
22991 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
22992 "Unsupported VT for PSIGN");
22993 Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask.getOperand(0));
22994 return DAG.getNode(ISD::BITCAST, DL, VT, Mask);
22996 // PBLENDVB only available on SSE 4.1
22997 if (!Subtarget->hasSSE41())
23000 EVT BlendVT = (VT == MVT::v4i64) ? MVT::v32i8 : MVT::v16i8;
23002 X = DAG.getNode(ISD::BITCAST, DL, BlendVT, X);
23003 Y = DAG.getNode(ISD::BITCAST, DL, BlendVT, Y);
23004 Mask = DAG.getNode(ISD::BITCAST, DL, BlendVT, Mask);
23005 Mask = DAG.getNode(ISD::VSELECT, DL, BlendVT, Mask, Y, X);
23006 return DAG.getNode(ISD::BITCAST, DL, VT, Mask);
23010 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
23013 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
23014 MachineFunction &MF = DAG.getMachineFunction();
23015 bool OptForSize = MF.getFunction()->getAttributes().
23016 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize);
23018 // SHLD/SHRD instructions have lower register pressure, but on some
23019 // platforms they have higher latency than the equivalent
23020 // series of shifts/or that would otherwise be generated.
23021 // Don't fold (or (x << c) | (y >> (64 - c))) if SHLD/SHRD instructions
23022 // have higher latencies and we are not optimizing for size.
23023 if (!OptForSize && Subtarget->isSHLDSlow())
23026 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
23028 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
23030 if (!N0.hasOneUse() || !N1.hasOneUse())
23033 SDValue ShAmt0 = N0.getOperand(1);
23034 if (ShAmt0.getValueType() != MVT::i8)
23036 SDValue ShAmt1 = N1.getOperand(1);
23037 if (ShAmt1.getValueType() != MVT::i8)
23039 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
23040 ShAmt0 = ShAmt0.getOperand(0);
23041 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
23042 ShAmt1 = ShAmt1.getOperand(0);
23045 unsigned Opc = X86ISD::SHLD;
23046 SDValue Op0 = N0.getOperand(0);
23047 SDValue Op1 = N1.getOperand(0);
23048 if (ShAmt0.getOpcode() == ISD::SUB) {
23049 Opc = X86ISD::SHRD;
23050 std::swap(Op0, Op1);
23051 std::swap(ShAmt0, ShAmt1);
23054 unsigned Bits = VT.getSizeInBits();
23055 if (ShAmt1.getOpcode() == ISD::SUB) {
23056 SDValue Sum = ShAmt1.getOperand(0);
23057 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
23058 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
23059 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
23060 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
23061 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
23062 return DAG.getNode(Opc, DL, VT,
23064 DAG.getNode(ISD::TRUNCATE, DL,
23067 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
23068 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
23070 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
23071 return DAG.getNode(Opc, DL, VT,
23072 N0.getOperand(0), N1.getOperand(0),
23073 DAG.getNode(ISD::TRUNCATE, DL,
23080 // Generate NEG and CMOV for integer abs.
23081 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
23082 EVT VT = N->getValueType(0);
23084 // Since X86 does not have CMOV for 8-bit integer, we don't convert
23085 // 8-bit integer abs to NEG and CMOV.
23086 if (VT.isInteger() && VT.getSizeInBits() == 8)
23089 SDValue N0 = N->getOperand(0);
23090 SDValue N1 = N->getOperand(1);
23093 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
23094 // and change it to SUB and CMOV.
23095 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
23096 N0.getOpcode() == ISD::ADD &&
23097 N0.getOperand(1) == N1 &&
23098 N1.getOpcode() == ISD::SRA &&
23099 N1.getOperand(0) == N0.getOperand(0))
23100 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
23101 if (Y1C->getAPIntValue() == VT.getSizeInBits()-1) {
23102 // Generate SUB & CMOV.
23103 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32),
23104 DAG.getConstant(0, VT), N0.getOperand(0));
23106 SDValue Ops[] = { N0.getOperand(0), Neg,
23107 DAG.getConstant(X86::COND_GE, MVT::i8),
23108 SDValue(Neg.getNode(), 1) };
23109 return DAG.getNode(X86ISD::CMOV, DL, DAG.getVTList(VT, MVT::Glue), Ops);
23114 // PerformXorCombine - Attempts to turn XOR nodes into BLSMSK nodes
23115 static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
23116 TargetLowering::DAGCombinerInfo &DCI,
23117 const X86Subtarget *Subtarget) {
23118 if (DCI.isBeforeLegalizeOps())
23121 if (Subtarget->hasCMov()) {
23122 SDValue RV = performIntegerAbsCombine(N, DAG);
23130 /// PerformLOADCombine - Do target-specific dag combines on LOAD nodes.
23131 static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
23132 TargetLowering::DAGCombinerInfo &DCI,
23133 const X86Subtarget *Subtarget) {
23134 LoadSDNode *Ld = cast<LoadSDNode>(N);
23135 EVT RegVT = Ld->getValueType(0);
23136 EVT MemVT = Ld->getMemoryVT();
23138 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23140 // On Sandybridge unaligned 256bit loads are inefficient.
23141 ISD::LoadExtType Ext = Ld->getExtensionType();
23142 unsigned Alignment = Ld->getAlignment();
23143 bool IsAligned = Alignment == 0 || Alignment >= MemVT.getSizeInBits()/8;
23144 if (RegVT.is256BitVector() && !Subtarget->hasInt256() &&
23145 !DCI.isBeforeLegalizeOps() && !IsAligned && Ext == ISD::NON_EXTLOAD) {
23146 unsigned NumElems = RegVT.getVectorNumElements();
23150 SDValue Ptr = Ld->getBasePtr();
23151 SDValue Increment = DAG.getConstant(16, TLI.getPointerTy());
23153 EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
23155 SDValue Load1 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
23156 Ld->getPointerInfo(), Ld->isVolatile(),
23157 Ld->isNonTemporal(), Ld->isInvariant(),
23159 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
23160 SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
23161 Ld->getPointerInfo(), Ld->isVolatile(),
23162 Ld->isNonTemporal(), Ld->isInvariant(),
23163 std::min(16U, Alignment));
23164 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
23166 Load2.getValue(1));
23168 SDValue NewVec = DAG.getUNDEF(RegVT);
23169 NewVec = Insert128BitVector(NewVec, Load1, 0, DAG, dl);
23170 NewVec = Insert128BitVector(NewVec, Load2, NumElems/2, DAG, dl);
23171 return DCI.CombineTo(N, NewVec, TF, true);
23177 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
23178 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
23179 const X86Subtarget *Subtarget) {
23180 StoreSDNode *St = cast<StoreSDNode>(N);
23181 EVT VT = St->getValue().getValueType();
23182 EVT StVT = St->getMemoryVT();
23184 SDValue StoredVal = St->getOperand(1);
23185 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23187 // If we are saving a concatenation of two XMM registers, perform two stores.
23188 // On Sandy Bridge, 256-bit memory operations are executed by two
23189 // 128-bit ports. However, on Haswell it is better to issue a single 256-bit
23190 // memory operation.
23191 unsigned Alignment = St->getAlignment();
23192 bool IsAligned = Alignment == 0 || Alignment >= VT.getSizeInBits()/8;
23193 if (VT.is256BitVector() && !Subtarget->hasInt256() &&
23194 StVT == VT && !IsAligned) {
23195 unsigned NumElems = VT.getVectorNumElements();
23199 SDValue Value0 = Extract128BitVector(StoredVal, 0, DAG, dl);
23200 SDValue Value1 = Extract128BitVector(StoredVal, NumElems/2, DAG, dl);
23202 SDValue Stride = DAG.getConstant(16, TLI.getPointerTy());
23203 SDValue Ptr0 = St->getBasePtr();
23204 SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
23206 SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
23207 St->getPointerInfo(), St->isVolatile(),
23208 St->isNonTemporal(), Alignment);
23209 SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
23210 St->getPointerInfo(), St->isVolatile(),
23211 St->isNonTemporal(),
23212 std::min(16U, Alignment));
23213 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
23216 // Optimize trunc store (of multiple scalars) to shuffle and store.
23217 // First, pack all of the elements in one place. Next, store to memory
23218 // in fewer chunks.
23219 if (St->isTruncatingStore() && VT.isVector()) {
23220 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23221 unsigned NumElems = VT.getVectorNumElements();
23222 assert(StVT != VT && "Cannot truncate to the same type");
23223 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
23224 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
23226 // From, To sizes and ElemCount must be pow of two
23227 if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
23228 // We are going to use the original vector elt for storing.
23229 // Accumulated smaller vector elements must be a multiple of the store size.
23230 if (0 != (NumElems * FromSz) % ToSz) return SDValue();
23232 unsigned SizeRatio = FromSz / ToSz;
23234 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
23236 // Create a type on which we perform the shuffle
23237 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
23238 StVT.getScalarType(), NumElems*SizeRatio);
23240 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
23242 SDValue WideVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, St->getValue());
23243 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
23244 for (unsigned i = 0; i != NumElems; ++i)
23245 ShuffleVec[i] = i * SizeRatio;
23247 // Can't shuffle using an illegal type.
23248 if (!TLI.isTypeLegal(WideVecVT))
23251 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
23252 DAG.getUNDEF(WideVecVT),
23254 // At this point all of the data is stored at the bottom of the
23255 // register. We now need to save it to mem.
23257 // Find the largest store unit
23258 MVT StoreType = MVT::i8;
23259 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
23260 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
23261 MVT Tp = (MVT::SimpleValueType)tp;
23262 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToSz)
23266 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
23267 if (TLI.isTypeLegal(MVT::f64) && StoreType.getSizeInBits() < 64 &&
23268 (64 <= NumElems * ToSz))
23269 StoreType = MVT::f64;
23271 // Bitcast the original vector into a vector of store-size units
23272 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
23273 StoreType, VT.getSizeInBits()/StoreType.getSizeInBits());
23274 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
23275 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, dl, StoreVecVT, Shuff);
23276 SmallVector<SDValue, 8> Chains;
23277 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
23278 TLI.getPointerTy());
23279 SDValue Ptr = St->getBasePtr();
23281 // Perform one or more big stores into memory.
23282 for (unsigned i=0, e=(ToSz*NumElems)/StoreType.getSizeInBits(); i!=e; ++i) {
23283 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
23284 StoreType, ShuffWide,
23285 DAG.getIntPtrConstant(i));
23286 SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr,
23287 St->getPointerInfo(), St->isVolatile(),
23288 St->isNonTemporal(), St->getAlignment());
23289 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
23290 Chains.push_back(Ch);
23293 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
23296 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
23297 // the FP state in cases where an emms may be missing.
23298 // A preferable solution to the general problem is to figure out the right
23299 // places to insert EMMS. This qualifies as a quick hack.
23301 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
23302 if (VT.getSizeInBits() != 64)
23305 const Function *F = DAG.getMachineFunction().getFunction();
23306 bool NoImplicitFloatOps = F->getAttributes().
23307 hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat);
23308 bool F64IsLegal = !DAG.getTarget().Options.UseSoftFloat && !NoImplicitFloatOps
23309 && Subtarget->hasSSE2();
23310 if ((VT.isVector() ||
23311 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
23312 isa<LoadSDNode>(St->getValue()) &&
23313 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
23314 St->getChain().hasOneUse() && !St->isVolatile()) {
23315 SDNode* LdVal = St->getValue().getNode();
23316 LoadSDNode *Ld = nullptr;
23317 int TokenFactorIndex = -1;
23318 SmallVector<SDValue, 8> Ops;
23319 SDNode* ChainVal = St->getChain().getNode();
23320 // Must be a store of a load. We currently handle two cases: the load
23321 // is a direct child, and it's under an intervening TokenFactor. It is
23322 // possible to dig deeper under nested TokenFactors.
23323 if (ChainVal == LdVal)
23324 Ld = cast<LoadSDNode>(St->getChain());
23325 else if (St->getValue().hasOneUse() &&
23326 ChainVal->getOpcode() == ISD::TokenFactor) {
23327 for (unsigned i = 0, e = ChainVal->getNumOperands(); i != e; ++i) {
23328 if (ChainVal->getOperand(i).getNode() == LdVal) {
23329 TokenFactorIndex = i;
23330 Ld = cast<LoadSDNode>(St->getValue());
23332 Ops.push_back(ChainVal->getOperand(i));
23336 if (!Ld || !ISD::isNormalLoad(Ld))
23339 // If this is not the MMX case, i.e. we are just turning i64 load/store
23340 // into f64 load/store, avoid the transformation if there are multiple
23341 // uses of the loaded value.
23342 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
23347 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
23348 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
23350 if (Subtarget->is64Bit() || F64IsLegal) {
23351 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
23352 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
23353 Ld->getPointerInfo(), Ld->isVolatile(),
23354 Ld->isNonTemporal(), Ld->isInvariant(),
23355 Ld->getAlignment());
23356 SDValue NewChain = NewLd.getValue(1);
23357 if (TokenFactorIndex != -1) {
23358 Ops.push_back(NewChain);
23359 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
23361 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
23362 St->getPointerInfo(),
23363 St->isVolatile(), St->isNonTemporal(),
23364 St->getAlignment());
23367 // Otherwise, lower to two pairs of 32-bit loads / stores.
23368 SDValue LoAddr = Ld->getBasePtr();
23369 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
23370 DAG.getConstant(4, MVT::i32));
23372 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
23373 Ld->getPointerInfo(),
23374 Ld->isVolatile(), Ld->isNonTemporal(),
23375 Ld->isInvariant(), Ld->getAlignment());
23376 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
23377 Ld->getPointerInfo().getWithOffset(4),
23378 Ld->isVolatile(), Ld->isNonTemporal(),
23380 MinAlign(Ld->getAlignment(), 4));
23382 SDValue NewChain = LoLd.getValue(1);
23383 if (TokenFactorIndex != -1) {
23384 Ops.push_back(LoLd);
23385 Ops.push_back(HiLd);
23386 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
23389 LoAddr = St->getBasePtr();
23390 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
23391 DAG.getConstant(4, MVT::i32));
23393 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
23394 St->getPointerInfo(),
23395 St->isVolatile(), St->isNonTemporal(),
23396 St->getAlignment());
23397 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
23398 St->getPointerInfo().getWithOffset(4),
23400 St->isNonTemporal(),
23401 MinAlign(St->getAlignment(), 4));
23402 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
23407 /// isHorizontalBinOp - Return 'true' if this vector operation is "horizontal"
23408 /// and return the operands for the horizontal operation in LHS and RHS. A
23409 /// horizontal operation performs the binary operation on successive elements
23410 /// of its first operand, then on successive elements of its second operand,
23411 /// returning the resulting values in a vector. For example, if
23412 /// A = < float a0, float a1, float a2, float a3 >
23414 /// B = < float b0, float b1, float b2, float b3 >
23415 /// then the result of doing a horizontal operation on A and B is
23416 /// A horizontal-op B = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >.
23417 /// In short, LHS and RHS are inspected to see if LHS op RHS is of the form
23418 /// A horizontal-op B, for some already available A and B, and if so then LHS is
23419 /// set to A, RHS to B, and the routine returns 'true'.
23420 /// Note that the binary operation should have the property that if one of the
23421 /// operands is UNDEF then the result is UNDEF.
23422 static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool IsCommutative) {
23423 // Look for the following pattern: if
23424 // A = < float a0, float a1, float a2, float a3 >
23425 // B = < float b0, float b1, float b2, float b3 >
23427 // LHS = VECTOR_SHUFFLE A, B, <0, 2, 4, 6>
23428 // RHS = VECTOR_SHUFFLE A, B, <1, 3, 5, 7>
23429 // then LHS op RHS = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >
23430 // which is A horizontal-op B.
23432 // At least one of the operands should be a vector shuffle.
23433 if (LHS.getOpcode() != ISD::VECTOR_SHUFFLE &&
23434 RHS.getOpcode() != ISD::VECTOR_SHUFFLE)
23437 MVT VT = LHS.getSimpleValueType();
23439 assert((VT.is128BitVector() || VT.is256BitVector()) &&
23440 "Unsupported vector type for horizontal add/sub");
23442 // Handle 128 and 256-bit vector lengths. AVX defines horizontal add/sub to
23443 // operate independently on 128-bit lanes.
23444 unsigned NumElts = VT.getVectorNumElements();
23445 unsigned NumLanes = VT.getSizeInBits()/128;
23446 unsigned NumLaneElts = NumElts / NumLanes;
23447 assert((NumLaneElts % 2 == 0) &&
23448 "Vector type should have an even number of elements in each lane");
23449 unsigned HalfLaneElts = NumLaneElts/2;
23451 // View LHS in the form
23452 // LHS = VECTOR_SHUFFLE A, B, LMask
23453 // If LHS is not a shuffle then pretend it is the shuffle
23454 // LHS = VECTOR_SHUFFLE LHS, undef, <0, 1, ..., N-1>
23455 // NOTE: in what follows a default initialized SDValue represents an UNDEF of
23458 SmallVector<int, 16> LMask(NumElts);
23459 if (LHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
23460 if (LHS.getOperand(0).getOpcode() != ISD::UNDEF)
23461 A = LHS.getOperand(0);
23462 if (LHS.getOperand(1).getOpcode() != ISD::UNDEF)
23463 B = LHS.getOperand(1);
23464 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(LHS.getNode())->getMask();
23465 std::copy(Mask.begin(), Mask.end(), LMask.begin());
23467 if (LHS.getOpcode() != ISD::UNDEF)
23469 for (unsigned i = 0; i != NumElts; ++i)
23473 // Likewise, view RHS in the form
23474 // RHS = VECTOR_SHUFFLE C, D, RMask
23476 SmallVector<int, 16> RMask(NumElts);
23477 if (RHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
23478 if (RHS.getOperand(0).getOpcode() != ISD::UNDEF)
23479 C = RHS.getOperand(0);
23480 if (RHS.getOperand(1).getOpcode() != ISD::UNDEF)
23481 D = RHS.getOperand(1);
23482 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(RHS.getNode())->getMask();
23483 std::copy(Mask.begin(), Mask.end(), RMask.begin());
23485 if (RHS.getOpcode() != ISD::UNDEF)
23487 for (unsigned i = 0; i != NumElts; ++i)
23491 // Check that the shuffles are both shuffling the same vectors.
23492 if (!(A == C && B == D) && !(A == D && B == C))
23495 // If everything is UNDEF then bail out: it would be better to fold to UNDEF.
23496 if (!A.getNode() && !B.getNode())
23499 // If A and B occur in reverse order in RHS, then "swap" them (which means
23500 // rewriting the mask).
23502 CommuteVectorShuffleMask(RMask, NumElts);
23504 // At this point LHS and RHS are equivalent to
23505 // LHS = VECTOR_SHUFFLE A, B, LMask
23506 // RHS = VECTOR_SHUFFLE A, B, RMask
23507 // Check that the masks correspond to performing a horizontal operation.
23508 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
23509 for (unsigned i = 0; i != NumLaneElts; ++i) {
23510 int LIdx = LMask[i+l], RIdx = RMask[i+l];
23512 // Ignore any UNDEF components.
23513 if (LIdx < 0 || RIdx < 0 ||
23514 (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) ||
23515 (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts)))
23518 // Check that successive elements are being operated on. If not, this is
23519 // not a horizontal operation.
23520 unsigned Src = (i/HalfLaneElts); // each lane is split between srcs
23521 int Index = 2*(i%HalfLaneElts) + NumElts*Src + l;
23522 if (!(LIdx == Index && RIdx == Index + 1) &&
23523 !(IsCommutative && LIdx == Index + 1 && RIdx == Index))
23528 LHS = A.getNode() ? A : B; // If A is 'UNDEF', use B for it.
23529 RHS = B.getNode() ? B : A; // If B is 'UNDEF', use A for it.
23533 /// PerformFADDCombine - Do target-specific dag combines on floating point adds.
23534 static SDValue PerformFADDCombine(SDNode *N, SelectionDAG &DAG,
23535 const X86Subtarget *Subtarget) {
23536 EVT VT = N->getValueType(0);
23537 SDValue LHS = N->getOperand(0);
23538 SDValue RHS = N->getOperand(1);
23540 // Try to synthesize horizontal adds from adds of shuffles.
23541 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
23542 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
23543 isHorizontalBinOp(LHS, RHS, true))
23544 return DAG.getNode(X86ISD::FHADD, SDLoc(N), VT, LHS, RHS);
23548 /// PerformFSUBCombine - Do target-specific dag combines on floating point subs.
23549 static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG,
23550 const X86Subtarget *Subtarget) {
23551 EVT VT = N->getValueType(0);
23552 SDValue LHS = N->getOperand(0);
23553 SDValue RHS = N->getOperand(1);
23555 // Try to synthesize horizontal subs from subs of shuffles.
23556 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
23557 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
23558 isHorizontalBinOp(LHS, RHS, false))
23559 return DAG.getNode(X86ISD::FHSUB, SDLoc(N), VT, LHS, RHS);
23563 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
23564 /// X86ISD::FXOR nodes.
23565 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
23566 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
23567 // F[X]OR(0.0, x) -> x
23568 // F[X]OR(x, 0.0) -> x
23569 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
23570 if (C->getValueAPF().isPosZero())
23571 return N->getOperand(1);
23572 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
23573 if (C->getValueAPF().isPosZero())
23574 return N->getOperand(0);
23578 /// PerformFMinFMaxCombine - Do target-specific dag combines on X86ISD::FMIN and
23579 /// X86ISD::FMAX nodes.
23580 static SDValue PerformFMinFMaxCombine(SDNode *N, SelectionDAG &DAG) {
23581 assert(N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD::FMAX);
23583 // Only perform optimizations if UnsafeMath is used.
23584 if (!DAG.getTarget().Options.UnsafeFPMath)
23587 // If we run in unsafe-math mode, then convert the FMAX and FMIN nodes
23588 // into FMINC and FMAXC, which are Commutative operations.
23589 unsigned NewOp = 0;
23590 switch (N->getOpcode()) {
23591 default: llvm_unreachable("unknown opcode");
23592 case X86ISD::FMIN: NewOp = X86ISD::FMINC; break;
23593 case X86ISD::FMAX: NewOp = X86ISD::FMAXC; break;
23596 return DAG.getNode(NewOp, SDLoc(N), N->getValueType(0),
23597 N->getOperand(0), N->getOperand(1));
23600 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
23601 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
23602 // FAND(0.0, x) -> 0.0
23603 // FAND(x, 0.0) -> 0.0
23604 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
23605 if (C->getValueAPF().isPosZero())
23606 return N->getOperand(0);
23607 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
23608 if (C->getValueAPF().isPosZero())
23609 return N->getOperand(1);
23613 /// PerformFANDNCombine - Do target-specific dag combines on X86ISD::FANDN nodes
23614 static SDValue PerformFANDNCombine(SDNode *N, SelectionDAG &DAG) {
23615 // FANDN(x, 0.0) -> 0.0
23616 // FANDN(0.0, x) -> x
23617 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
23618 if (C->getValueAPF().isPosZero())
23619 return N->getOperand(1);
23620 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
23621 if (C->getValueAPF().isPosZero())
23622 return N->getOperand(1);
23626 static SDValue PerformBTCombine(SDNode *N,
23628 TargetLowering::DAGCombinerInfo &DCI) {
23629 // BT ignores high bits in the bit index operand.
23630 SDValue Op1 = N->getOperand(1);
23631 if (Op1.hasOneUse()) {
23632 unsigned BitWidth = Op1.getValueSizeInBits();
23633 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
23634 APInt KnownZero, KnownOne;
23635 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
23636 !DCI.isBeforeLegalizeOps());
23637 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23638 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
23639 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
23640 DCI.CommitTargetLoweringOpt(TLO);
23645 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
23646 SDValue Op = N->getOperand(0);
23647 if (Op.getOpcode() == ISD::BITCAST)
23648 Op = Op.getOperand(0);
23649 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
23650 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
23651 VT.getVectorElementType().getSizeInBits() ==
23652 OpVT.getVectorElementType().getSizeInBits()) {
23653 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
23658 static SDValue PerformSIGN_EXTEND_INREGCombine(SDNode *N, SelectionDAG &DAG,
23659 const X86Subtarget *Subtarget) {
23660 EVT VT = N->getValueType(0);
23661 if (!VT.isVector())
23664 SDValue N0 = N->getOperand(0);
23665 SDValue N1 = N->getOperand(1);
23666 EVT ExtraVT = cast<VTSDNode>(N1)->getVT();
23669 // The SIGN_EXTEND_INREG to v4i64 is expensive operation on the
23670 // both SSE and AVX2 since there is no sign-extended shift right
23671 // operation on a vector with 64-bit elements.
23672 //(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) ->
23673 // (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT)))
23674 if (VT == MVT::v4i64 && (N0.getOpcode() == ISD::ANY_EXTEND ||
23675 N0.getOpcode() == ISD::SIGN_EXTEND)) {
23676 SDValue N00 = N0.getOperand(0);
23678 // EXTLOAD has a better solution on AVX2,
23679 // it may be replaced with X86ISD::VSEXT node.
23680 if (N00.getOpcode() == ISD::LOAD && Subtarget->hasInt256())
23681 if (!ISD::isNormalLoad(N00.getNode()))
23684 if (N00.getValueType() == MVT::v4i32 && ExtraVT.getSizeInBits() < 128) {
23685 SDValue Tmp = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32,
23687 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i64, Tmp);
23693 static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
23694 TargetLowering::DAGCombinerInfo &DCI,
23695 const X86Subtarget *Subtarget) {
23696 if (!DCI.isBeforeLegalizeOps())
23699 if (!Subtarget->hasFp256())
23702 EVT VT = N->getValueType(0);
23703 if (VT.isVector() && VT.getSizeInBits() == 256) {
23704 SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget);
23712 static SDValue PerformFMACombine(SDNode *N, SelectionDAG &DAG,
23713 const X86Subtarget* Subtarget) {
23715 EVT VT = N->getValueType(0);
23717 // Let legalize expand this if it isn't a legal type yet.
23718 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
23721 EVT ScalarVT = VT.getScalarType();
23722 if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) ||
23723 (!Subtarget->hasFMA() && !Subtarget->hasFMA4()))
23726 SDValue A = N->getOperand(0);
23727 SDValue B = N->getOperand(1);
23728 SDValue C = N->getOperand(2);
23730 bool NegA = (A.getOpcode() == ISD::FNEG);
23731 bool NegB = (B.getOpcode() == ISD::FNEG);
23732 bool NegC = (C.getOpcode() == ISD::FNEG);
23734 // Negative multiplication when NegA xor NegB
23735 bool NegMul = (NegA != NegB);
23737 A = A.getOperand(0);
23739 B = B.getOperand(0);
23741 C = C.getOperand(0);
23745 Opcode = (!NegC) ? X86ISD::FMADD : X86ISD::FMSUB;
23747 Opcode = (!NegC) ? X86ISD::FNMADD : X86ISD::FNMSUB;
23749 return DAG.getNode(Opcode, dl, VT, A, B, C);
23752 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG,
23753 TargetLowering::DAGCombinerInfo &DCI,
23754 const X86Subtarget *Subtarget) {
23755 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
23756 // (and (i32 x86isd::setcc_carry), 1)
23757 // This eliminates the zext. This transformation is necessary because
23758 // ISD::SETCC is always legalized to i8.
23760 SDValue N0 = N->getOperand(0);
23761 EVT VT = N->getValueType(0);
23763 if (N0.getOpcode() == ISD::AND &&
23765 N0.getOperand(0).hasOneUse()) {
23766 SDValue N00 = N0.getOperand(0);
23767 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
23768 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
23769 if (!C || C->getZExtValue() != 1)
23771 return DAG.getNode(ISD::AND, dl, VT,
23772 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
23773 N00.getOperand(0), N00.getOperand(1)),
23774 DAG.getConstant(1, VT));
23778 if (N0.getOpcode() == ISD::TRUNCATE &&
23780 N0.getOperand(0).hasOneUse()) {
23781 SDValue N00 = N0.getOperand(0);
23782 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
23783 return DAG.getNode(ISD::AND, dl, VT,
23784 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
23785 N00.getOperand(0), N00.getOperand(1)),
23786 DAG.getConstant(1, VT));
23789 if (VT.is256BitVector()) {
23790 SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget);
23798 // Optimize x == -y --> x+y == 0
23799 // x != -y --> x+y != 0
23800 static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG,
23801 const X86Subtarget* Subtarget) {
23802 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
23803 SDValue LHS = N->getOperand(0);
23804 SDValue RHS = N->getOperand(1);
23805 EVT VT = N->getValueType(0);
23808 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB)
23809 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(LHS.getOperand(0)))
23810 if (C->getAPIntValue() == 0 && LHS.hasOneUse()) {
23811 SDValue addV = DAG.getNode(ISD::ADD, SDLoc(N),
23812 LHS.getValueType(), RHS, LHS.getOperand(1));
23813 return DAG.getSetCC(SDLoc(N), N->getValueType(0),
23814 addV, DAG.getConstant(0, addV.getValueType()), CC);
23816 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && RHS.getOpcode() == ISD::SUB)
23817 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS.getOperand(0)))
23818 if (C->getAPIntValue() == 0 && RHS.hasOneUse()) {
23819 SDValue addV = DAG.getNode(ISD::ADD, SDLoc(N),
23820 RHS.getValueType(), LHS, RHS.getOperand(1));
23821 return DAG.getSetCC(SDLoc(N), N->getValueType(0),
23822 addV, DAG.getConstant(0, addV.getValueType()), CC);
23825 if (VT.getScalarType() == MVT::i1) {
23826 bool IsSEXT0 = (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
23827 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
23828 bool IsVZero0 = ISD::isBuildVectorAllZeros(LHS.getNode());
23829 if (!IsSEXT0 && !IsVZero0)
23831 bool IsSEXT1 = (RHS.getOpcode() == ISD::SIGN_EXTEND) &&
23832 (RHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
23833 bool IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
23835 if (!IsSEXT1 && !IsVZero1)
23838 if (IsSEXT0 && IsVZero1) {
23839 assert(VT == LHS.getOperand(0).getValueType() && "Uexpected operand type");
23840 if (CC == ISD::SETEQ)
23841 return DAG.getNOT(DL, LHS.getOperand(0), VT);
23842 return LHS.getOperand(0);
23844 if (IsSEXT1 && IsVZero0) {
23845 assert(VT == RHS.getOperand(0).getValueType() && "Uexpected operand type");
23846 if (CC == ISD::SETEQ)
23847 return DAG.getNOT(DL, RHS.getOperand(0), VT);
23848 return RHS.getOperand(0);
23855 static SDValue PerformINSERTPSCombine(SDNode *N, SelectionDAG &DAG,
23856 const X86Subtarget *Subtarget) {
23858 MVT VT = N->getOperand(1)->getSimpleValueType(0);
23859 assert((VT == MVT::v4f32 || VT == MVT::v4i32) &&
23860 "X86insertps is only defined for v4x32");
23862 SDValue Ld = N->getOperand(1);
23863 if (MayFoldLoad(Ld)) {
23864 // Extract the countS bits from the immediate so we can get the proper
23865 // address when narrowing the vector load to a specific element.
23866 // When the second source op is a memory address, interps doesn't use
23867 // countS and just gets an f32 from that address.
23868 unsigned DestIndex =
23869 cast<ConstantSDNode>(N->getOperand(2))->getZExtValue() >> 6;
23870 Ld = NarrowVectorLoadToElement(cast<LoadSDNode>(Ld), DestIndex, DAG);
23874 // Create this as a scalar to vector to match the instruction pattern.
23875 SDValue LoadScalarToVector = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Ld);
23876 // countS bits are ignored when loading from memory on insertps, which
23877 // means we don't need to explicitly set them to 0.
23878 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N->getOperand(0),
23879 LoadScalarToVector, N->getOperand(2));
23882 // Helper function of PerformSETCCCombine. It is to materialize "setb reg"
23883 // as "sbb reg,reg", since it can be extended without zext and produces
23884 // an all-ones bit which is more useful than 0/1 in some cases.
23885 static SDValue MaterializeSETB(SDLoc DL, SDValue EFLAGS, SelectionDAG &DAG,
23888 return DAG.getNode(ISD::AND, DL, VT,
23889 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
23890 DAG.getConstant(X86::COND_B, MVT::i8), EFLAGS),
23891 DAG.getConstant(1, VT));
23892 assert (VT == MVT::i1 && "Unexpected type for SECCC node");
23893 return DAG.getNode(ISD::TRUNCATE, DL, MVT::i1,
23894 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
23895 DAG.getConstant(X86::COND_B, MVT::i8), EFLAGS));
23898 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
23899 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG,
23900 TargetLowering::DAGCombinerInfo &DCI,
23901 const X86Subtarget *Subtarget) {
23903 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0));
23904 SDValue EFLAGS = N->getOperand(1);
23906 if (CC == X86::COND_A) {
23907 // Try to convert COND_A into COND_B in an attempt to facilitate
23908 // materializing "setb reg".
23910 // Do not flip "e > c", where "c" is a constant, because Cmp instruction
23911 // cannot take an immediate as its first operand.
23913 if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() &&
23914 EFLAGS.getValueType().isInteger() &&
23915 !isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
23916 SDValue NewSub = DAG.getNode(X86ISD::SUB, SDLoc(EFLAGS),
23917 EFLAGS.getNode()->getVTList(),
23918 EFLAGS.getOperand(1), EFLAGS.getOperand(0));
23919 SDValue NewEFLAGS = SDValue(NewSub.getNode(), EFLAGS.getResNo());
23920 return MaterializeSETB(DL, NewEFLAGS, DAG, N->getSimpleValueType(0));
23924 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
23925 // a zext and produces an all-ones bit which is more useful than 0/1 in some
23927 if (CC == X86::COND_B)
23928 return MaterializeSETB(DL, EFLAGS, DAG, N->getSimpleValueType(0));
23932 Flags = checkBoolTestSetCCCombine(EFLAGS, CC);
23933 if (Flags.getNode()) {
23934 SDValue Cond = DAG.getConstant(CC, MVT::i8);
23935 return DAG.getNode(X86ISD::SETCC, DL, N->getVTList(), Cond, Flags);
23941 // Optimize branch condition evaluation.
23943 static SDValue PerformBrCondCombine(SDNode *N, SelectionDAG &DAG,
23944 TargetLowering::DAGCombinerInfo &DCI,
23945 const X86Subtarget *Subtarget) {
23947 SDValue Chain = N->getOperand(0);
23948 SDValue Dest = N->getOperand(1);
23949 SDValue EFLAGS = N->getOperand(3);
23950 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(2));
23954 Flags = checkBoolTestSetCCCombine(EFLAGS, CC);
23955 if (Flags.getNode()) {
23956 SDValue Cond = DAG.getConstant(CC, MVT::i8);
23957 return DAG.getNode(X86ISD::BRCOND, DL, N->getVTList(), Chain, Dest, Cond,
23964 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
23965 SelectionDAG &DAG) {
23966 // Take advantage of vector comparisons producing 0 or -1 in each lane to
23967 // optimize away operation when it's from a constant.
23969 // The general transformation is:
23970 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
23971 // AND(VECTOR_CMP(x,y), constant2)
23972 // constant2 = UNARYOP(constant)
23974 // Early exit if this isn't a vector operation, the operand of the
23975 // unary operation isn't a bitwise AND, or if the sizes of the operations
23976 // aren't the same.
23977 EVT VT = N->getValueType(0);
23978 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
23979 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
23980 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
23983 // Now check that the other operand of the AND is a constant. We could
23984 // make the transformation for non-constant splats as well, but it's unclear
23985 // that would be a benefit as it would not eliminate any operations, just
23986 // perform one more step in scalar code before moving to the vector unit.
23987 if (BuildVectorSDNode *BV =
23988 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
23989 // Bail out if the vector isn't a constant.
23990 if (!BV->isConstant())
23993 // Everything checks out. Build up the new and improved node.
23995 EVT IntVT = BV->getValueType(0);
23996 // Create a new constant of the appropriate type for the transformed
23998 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
23999 // The AND node needs bitcasts to/from an integer vector type around it.
24000 SDValue MaskConst = DAG.getNode(ISD::BITCAST, DL, IntVT, SourceConst);
24001 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
24002 N->getOperand(0)->getOperand(0), MaskConst);
24003 SDValue Res = DAG.getNode(ISD::BITCAST, DL, VT, NewAnd);
24010 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
24011 const X86TargetLowering *XTLI) {
24012 // First try to optimize away the conversion entirely when it's
24013 // conditionally from a constant. Vectors only.
24014 SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG);
24015 if (Res != SDValue())
24018 // Now move on to more general possibilities.
24019 SDValue Op0 = N->getOperand(0);
24020 EVT InVT = Op0->getValueType(0);
24022 // SINT_TO_FP(v4i8) -> SINT_TO_FP(SEXT(v4i8 to v4i32))
24023 if (InVT == MVT::v8i8 || InVT == MVT::v4i8) {
24025 MVT DstVT = InVT == MVT::v4i8 ? MVT::v4i32 : MVT::v8i32;
24026 SDValue P = DAG.getNode(ISD::SIGN_EXTEND, dl, DstVT, Op0);
24027 return DAG.getNode(ISD::SINT_TO_FP, dl, N->getValueType(0), P);
24030 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
24031 // a 32-bit target where SSE doesn't support i64->FP operations.
24032 if (Op0.getOpcode() == ISD::LOAD) {
24033 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
24034 EVT VT = Ld->getValueType(0);
24035 if (!Ld->isVolatile() && !N->getValueType(0).isVector() &&
24036 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
24037 !XTLI->getSubtarget()->is64Bit() &&
24039 SDValue FILDChain = XTLI->BuildFILD(SDValue(N, 0), Ld->getValueType(0),
24040 Ld->getChain(), Op0, DAG);
24041 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
24048 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
24049 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
24050 X86TargetLowering::DAGCombinerInfo &DCI) {
24051 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
24052 // the result is either zero or one (depending on the input carry bit).
24053 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
24054 if (X86::isZeroNode(N->getOperand(0)) &&
24055 X86::isZeroNode(N->getOperand(1)) &&
24056 // We don't have a good way to replace an EFLAGS use, so only do this when
24058 SDValue(N, 1).use_empty()) {
24060 EVT VT = N->getValueType(0);
24061 SDValue CarryOut = DAG.getConstant(0, N->getValueType(1));
24062 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
24063 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
24064 DAG.getConstant(X86::COND_B,MVT::i8),
24066 DAG.getConstant(1, VT));
24067 return DCI.CombineTo(N, Res1, CarryOut);
24073 // fold (add Y, (sete X, 0)) -> adc 0, Y
24074 // (add Y, (setne X, 0)) -> sbb -1, Y
24075 // (sub (sete X, 0), Y) -> sbb 0, Y
24076 // (sub (setne X, 0), Y) -> adc -1, Y
24077 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
24080 // Look through ZExts.
24081 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
24082 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
24085 SDValue SetCC = Ext.getOperand(0);
24086 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
24089 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
24090 if (CC != X86::COND_E && CC != X86::COND_NE)
24093 SDValue Cmp = SetCC.getOperand(1);
24094 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
24095 !X86::isZeroNode(Cmp.getOperand(1)) ||
24096 !Cmp.getOperand(0).getValueType().isInteger())
24099 SDValue CmpOp0 = Cmp.getOperand(0);
24100 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
24101 DAG.getConstant(1, CmpOp0.getValueType()));
24103 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
24104 if (CC == X86::COND_NE)
24105 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
24106 DL, OtherVal.getValueType(), OtherVal,
24107 DAG.getConstant(-1ULL, OtherVal.getValueType()), NewCmp);
24108 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
24109 DL, OtherVal.getValueType(), OtherVal,
24110 DAG.getConstant(0, OtherVal.getValueType()), NewCmp);
24113 /// PerformADDCombine - Do target-specific dag combines on integer adds.
24114 static SDValue PerformAddCombine(SDNode *N, SelectionDAG &DAG,
24115 const X86Subtarget *Subtarget) {
24116 EVT VT = N->getValueType(0);
24117 SDValue Op0 = N->getOperand(0);
24118 SDValue Op1 = N->getOperand(1);
24120 // Try to synthesize horizontal adds from adds of shuffles.
24121 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
24122 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
24123 isHorizontalBinOp(Op0, Op1, true))
24124 return DAG.getNode(X86ISD::HADD, SDLoc(N), VT, Op0, Op1);
24126 return OptimizeConditionalInDecrement(N, DAG);
24129 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG,
24130 const X86Subtarget *Subtarget) {
24131 SDValue Op0 = N->getOperand(0);
24132 SDValue Op1 = N->getOperand(1);
24134 // X86 can't encode an immediate LHS of a sub. See if we can push the
24135 // negation into a preceding instruction.
24136 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
24137 // If the RHS of the sub is a XOR with one use and a constant, invert the
24138 // immediate. Then add one to the LHS of the sub so we can turn
24139 // X-Y -> X+~Y+1, saving one register.
24140 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
24141 isa<ConstantSDNode>(Op1.getOperand(1))) {
24142 APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue();
24143 EVT VT = Op0.getValueType();
24144 SDValue NewXor = DAG.getNode(ISD::XOR, SDLoc(Op1), VT,
24146 DAG.getConstant(~XorC, VT));
24147 return DAG.getNode(ISD::ADD, SDLoc(N), VT, NewXor,
24148 DAG.getConstant(C->getAPIntValue()+1, VT));
24152 // Try to synthesize horizontal adds from adds of shuffles.
24153 EVT VT = N->getValueType(0);
24154 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
24155 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
24156 isHorizontalBinOp(Op0, Op1, true))
24157 return DAG.getNode(X86ISD::HSUB, SDLoc(N), VT, Op0, Op1);
24159 return OptimizeConditionalInDecrement(N, DAG);
24162 /// performVZEXTCombine - Performs build vector combines
24163 static SDValue performVZEXTCombine(SDNode *N, SelectionDAG &DAG,
24164 TargetLowering::DAGCombinerInfo &DCI,
24165 const X86Subtarget *Subtarget) {
24166 // (vzext (bitcast (vzext (x)) -> (vzext x)
24167 SDValue In = N->getOperand(0);
24168 while (In.getOpcode() == ISD::BITCAST)
24169 In = In.getOperand(0);
24171 if (In.getOpcode() != X86ISD::VZEXT)
24174 return DAG.getNode(X86ISD::VZEXT, SDLoc(N), N->getValueType(0),
24178 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
24179 DAGCombinerInfo &DCI) const {
24180 SelectionDAG &DAG = DCI.DAG;
24181 switch (N->getOpcode()) {
24183 case ISD::EXTRACT_VECTOR_ELT:
24184 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, DCI);
24186 case ISD::SELECT: return PerformSELECTCombine(N, DAG, DCI, Subtarget);
24187 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI, Subtarget);
24188 case ISD::ADD: return PerformAddCombine(N, DAG, Subtarget);
24189 case ISD::SUB: return PerformSubCombine(N, DAG, Subtarget);
24190 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
24191 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
24194 case ISD::SRL: return PerformShiftCombine(N, DAG, DCI, Subtarget);
24195 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
24196 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
24197 case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget);
24198 case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget);
24199 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
24200 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, this);
24201 case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
24202 case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
24204 case X86ISD::FOR: return PerformFORCombine(N, DAG);
24206 case X86ISD::FMAX: return PerformFMinFMaxCombine(N, DAG);
24207 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
24208 case X86ISD::FANDN: return PerformFANDNCombine(N, DAG);
24209 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
24210 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
24211 case ISD::ANY_EXTEND:
24212 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG, DCI, Subtarget);
24213 case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
24214 case ISD::SIGN_EXTEND_INREG:
24215 return PerformSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
24216 case ISD::TRUNCATE: return PerformTruncateCombine(N, DAG,DCI,Subtarget);
24217 case ISD::SETCC: return PerformISDSETCCCombine(N, DAG, Subtarget);
24218 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
24219 case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
24220 case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);
24221 case X86ISD::SHUFP: // Handle all target specific shuffles
24222 case X86ISD::PALIGNR:
24223 case X86ISD::UNPCKH:
24224 case X86ISD::UNPCKL:
24225 case X86ISD::MOVHLPS:
24226 case X86ISD::MOVLHPS:
24227 case X86ISD::PSHUFB:
24228 case X86ISD::PSHUFD:
24229 case X86ISD::PSHUFHW:
24230 case X86ISD::PSHUFLW:
24231 case X86ISD::MOVSS:
24232 case X86ISD::MOVSD:
24233 case X86ISD::VPERMILPI:
24234 case X86ISD::VPERM2X128:
24235 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
24236 case ISD::FMA: return PerformFMACombine(N, DAG, Subtarget);
24237 case ISD::INTRINSIC_WO_CHAIN:
24238 return PerformINTRINSIC_WO_CHAINCombine(N, DAG, Subtarget);
24239 case X86ISD::INSERTPS:
24240 return PerformINSERTPSCombine(N, DAG, Subtarget);
24241 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DAG, Subtarget);
24247 /// isTypeDesirableForOp - Return true if the target has native support for
24248 /// the specified value type and it is 'desirable' to use the type for the
24249 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
24250 /// instruction encodings are longer and some i16 instructions are slow.
24251 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
24252 if (!isTypeLegal(VT))
24254 if (VT != MVT::i16)
24261 case ISD::SIGN_EXTEND:
24262 case ISD::ZERO_EXTEND:
24263 case ISD::ANY_EXTEND:
24276 /// IsDesirableToPromoteOp - This method query the target whether it is
24277 /// beneficial for dag combiner to promote the specified node. If true, it
24278 /// should return the desired promotion type by reference.
24279 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
24280 EVT VT = Op.getValueType();
24281 if (VT != MVT::i16)
24284 bool Promote = false;
24285 bool Commute = false;
24286 switch (Op.getOpcode()) {
24289 LoadSDNode *LD = cast<LoadSDNode>(Op);
24290 // If the non-extending load has a single use and it's not live out, then it
24291 // might be folded.
24292 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
24293 Op.hasOneUse()*/) {
24294 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
24295 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
24296 // The only case where we'd want to promote LOAD (rather then it being
24297 // promoted as an operand is when it's only use is liveout.
24298 if (UI->getOpcode() != ISD::CopyToReg)
24305 case ISD::SIGN_EXTEND:
24306 case ISD::ZERO_EXTEND:
24307 case ISD::ANY_EXTEND:
24312 SDValue N0 = Op.getOperand(0);
24313 // Look out for (store (shl (load), x)).
24314 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
24327 SDValue N0 = Op.getOperand(0);
24328 SDValue N1 = Op.getOperand(1);
24329 if (!Commute && MayFoldLoad(N1))
24331 // Avoid disabling potential load folding opportunities.
24332 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
24334 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
24344 //===----------------------------------------------------------------------===//
24345 // X86 Inline Assembly Support
24346 //===----------------------------------------------------------------------===//
24349 // Helper to match a string separated by whitespace.
24350 bool matchAsmImpl(StringRef s, ArrayRef<const StringRef *> args) {
24351 s = s.substr(s.find_first_not_of(" \t")); // Skip leading whitespace.
24353 for (unsigned i = 0, e = args.size(); i != e; ++i) {
24354 StringRef piece(*args[i]);
24355 if (!s.startswith(piece)) // Check if the piece matches.
24358 s = s.substr(piece.size());
24359 StringRef::size_type pos = s.find_first_not_of(" \t");
24360 if (pos == 0) // We matched a prefix.
24368 const VariadicFunction1<bool, StringRef, StringRef, matchAsmImpl> matchAsm={};
24371 static bool clobbersFlagRegisters(const SmallVector<StringRef, 4> &AsmPieces) {
24373 if (AsmPieces.size() == 3 || AsmPieces.size() == 4) {
24374 if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{cc}") &&
24375 std::count(AsmPieces.begin(), AsmPieces.end(), "~{flags}") &&
24376 std::count(AsmPieces.begin(), AsmPieces.end(), "~{fpsr}")) {
24378 if (AsmPieces.size() == 3)
24380 else if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{dirflag}"))
24387 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
24388 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
24390 std::string AsmStr = IA->getAsmString();
24392 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
24393 if (!Ty || Ty->getBitWidth() % 16 != 0)
24396 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
24397 SmallVector<StringRef, 4> AsmPieces;
24398 SplitString(AsmStr, AsmPieces, ";\n");
24400 switch (AsmPieces.size()) {
24401 default: return false;
24403 // FIXME: this should verify that we are targeting a 486 or better. If not,
24404 // we will turn this bswap into something that will be lowered to logical
24405 // ops instead of emitting the bswap asm. For now, we don't support 486 or
24406 // lower so don't worry about this.
24408 if (matchAsm(AsmPieces[0], "bswap", "$0") ||
24409 matchAsm(AsmPieces[0], "bswapl", "$0") ||
24410 matchAsm(AsmPieces[0], "bswapq", "$0") ||
24411 matchAsm(AsmPieces[0], "bswap", "${0:q}") ||
24412 matchAsm(AsmPieces[0], "bswapl", "${0:q}") ||
24413 matchAsm(AsmPieces[0], "bswapq", "${0:q}")) {
24414 // No need to check constraints, nothing other than the equivalent of
24415 // "=r,0" would be valid here.
24416 return IntrinsicLowering::LowerToByteSwap(CI);
24419 // rorw $$8, ${0:w} --> llvm.bswap.i16
24420 if (CI->getType()->isIntegerTy(16) &&
24421 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
24422 (matchAsm(AsmPieces[0], "rorw", "$$8,", "${0:w}") ||
24423 matchAsm(AsmPieces[0], "rolw", "$$8,", "${0:w}"))) {
24425 const std::string &ConstraintsStr = IA->getConstraintString();
24426 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
24427 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
24428 if (clobbersFlagRegisters(AsmPieces))
24429 return IntrinsicLowering::LowerToByteSwap(CI);
24433 if (CI->getType()->isIntegerTy(32) &&
24434 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
24435 matchAsm(AsmPieces[0], "rorw", "$$8,", "${0:w}") &&
24436 matchAsm(AsmPieces[1], "rorl", "$$16,", "$0") &&
24437 matchAsm(AsmPieces[2], "rorw", "$$8,", "${0:w}")) {
24439 const std::string &ConstraintsStr = IA->getConstraintString();
24440 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
24441 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
24442 if (clobbersFlagRegisters(AsmPieces))
24443 return IntrinsicLowering::LowerToByteSwap(CI);
24446 if (CI->getType()->isIntegerTy(64)) {
24447 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
24448 if (Constraints.size() >= 2 &&
24449 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
24450 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
24451 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
24452 if (matchAsm(AsmPieces[0], "bswap", "%eax") &&
24453 matchAsm(AsmPieces[1], "bswap", "%edx") &&
24454 matchAsm(AsmPieces[2], "xchgl", "%eax,", "%edx"))
24455 return IntrinsicLowering::LowerToByteSwap(CI);
24463 /// getConstraintType - Given a constraint letter, return the type of
24464 /// constraint it is for this target.
24465 X86TargetLowering::ConstraintType
24466 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
24467 if (Constraint.size() == 1) {
24468 switch (Constraint[0]) {
24479 return C_RegisterClass;
24503 return TargetLowering::getConstraintType(Constraint);
24506 /// Examine constraint type and operand type and determine a weight value.
24507 /// This object must already have been set up with the operand type
24508 /// and the current alternative constraint selected.
24509 TargetLowering::ConstraintWeight
24510 X86TargetLowering::getSingleConstraintMatchWeight(
24511 AsmOperandInfo &info, const char *constraint) const {
24512 ConstraintWeight weight = CW_Invalid;
24513 Value *CallOperandVal = info.CallOperandVal;
24514 // If we don't have a value, we can't do a match,
24515 // but allow it at the lowest weight.
24516 if (!CallOperandVal)
24518 Type *type = CallOperandVal->getType();
24519 // Look at the constraint type.
24520 switch (*constraint) {
24522 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
24533 if (CallOperandVal->getType()->isIntegerTy())
24534 weight = CW_SpecificReg;
24539 if (type->isFloatingPointTy())
24540 weight = CW_SpecificReg;
24543 if (type->isX86_MMXTy() && Subtarget->hasMMX())
24544 weight = CW_SpecificReg;
24548 if (((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasSSE1()) ||
24549 ((type->getPrimitiveSizeInBits() == 256) && Subtarget->hasFp256()))
24550 weight = CW_Register;
24553 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
24554 if (C->getZExtValue() <= 31)
24555 weight = CW_Constant;
24559 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
24560 if (C->getZExtValue() <= 63)
24561 weight = CW_Constant;
24565 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
24566 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
24567 weight = CW_Constant;
24571 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
24572 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
24573 weight = CW_Constant;
24577 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
24578 if (C->getZExtValue() <= 3)
24579 weight = CW_Constant;
24583 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
24584 if (C->getZExtValue() <= 0xff)
24585 weight = CW_Constant;
24590 if (dyn_cast<ConstantFP>(CallOperandVal)) {
24591 weight = CW_Constant;
24595 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
24596 if ((C->getSExtValue() >= -0x80000000LL) &&
24597 (C->getSExtValue() <= 0x7fffffffLL))
24598 weight = CW_Constant;
24602 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
24603 if (C->getZExtValue() <= 0xffffffff)
24604 weight = CW_Constant;
24611 /// LowerXConstraint - try to replace an X constraint, which matches anything,
24612 /// with another that has more specific requirements based on the type of the
24613 /// corresponding operand.
24614 const char *X86TargetLowering::
24615 LowerXConstraint(EVT ConstraintVT) const {
24616 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
24617 // 'f' like normal targets.
24618 if (ConstraintVT.isFloatingPoint()) {
24619 if (Subtarget->hasSSE2())
24621 if (Subtarget->hasSSE1())
24625 return TargetLowering::LowerXConstraint(ConstraintVT);
24628 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
24629 /// vector. If it is invalid, don't add anything to Ops.
24630 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
24631 std::string &Constraint,
24632 std::vector<SDValue>&Ops,
24633 SelectionDAG &DAG) const {
24636 // Only support length 1 constraints for now.
24637 if (Constraint.length() > 1) return;
24639 char ConstraintLetter = Constraint[0];
24640 switch (ConstraintLetter) {
24643 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24644 if (C->getZExtValue() <= 31) {
24645 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
24651 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24652 if (C->getZExtValue() <= 63) {
24653 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
24659 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24660 if (isInt<8>(C->getSExtValue())) {
24661 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
24667 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24668 if (C->getZExtValue() <= 255) {
24669 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
24675 // 32-bit signed value
24676 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24677 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
24678 C->getSExtValue())) {
24679 // Widen to 64 bits here to get it sign extended.
24680 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
24683 // FIXME gcc accepts some relocatable values here too, but only in certain
24684 // memory models; it's complicated.
24689 // 32-bit unsigned value
24690 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24691 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
24692 C->getZExtValue())) {
24693 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
24697 // FIXME gcc accepts some relocatable values here too, but only in certain
24698 // memory models; it's complicated.
24702 // Literal immediates are always ok.
24703 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
24704 // Widen to 64 bits here to get it sign extended.
24705 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
24709 // In any sort of PIC mode addresses need to be computed at runtime by
24710 // adding in a register or some sort of table lookup. These can't
24711 // be used as immediates.
24712 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
24715 // If we are in non-pic codegen mode, we allow the address of a global (with
24716 // an optional displacement) to be used with 'i'.
24717 GlobalAddressSDNode *GA = nullptr;
24718 int64_t Offset = 0;
24720 // Match either (GA), (GA+C), (GA+C1+C2), etc.
24722 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
24723 Offset += GA->getOffset();
24725 } else if (Op.getOpcode() == ISD::ADD) {
24726 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
24727 Offset += C->getZExtValue();
24728 Op = Op.getOperand(0);
24731 } else if (Op.getOpcode() == ISD::SUB) {
24732 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
24733 Offset += -C->getZExtValue();
24734 Op = Op.getOperand(0);
24739 // Otherwise, this isn't something we can handle, reject it.
24743 const GlobalValue *GV = GA->getGlobal();
24744 // If we require an extra load to get this address, as in PIC mode, we
24745 // can't accept it.
24746 if (isGlobalStubReference(
24747 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget())))
24750 Result = DAG.getTargetGlobalAddress(GV, SDLoc(Op),
24751 GA->getValueType(0), Offset);
24756 if (Result.getNode()) {
24757 Ops.push_back(Result);
24760 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
24763 std::pair<unsigned, const TargetRegisterClass*>
24764 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
24766 // First, see if this is a constraint that directly corresponds to an LLVM
24768 if (Constraint.size() == 1) {
24769 // GCC Constraint Letters
24770 switch (Constraint[0]) {
24772 // TODO: Slight differences here in allocation order and leaving
24773 // RIP in the class. Do they matter any more here than they do
24774 // in the normal allocation?
24775 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
24776 if (Subtarget->is64Bit()) {
24777 if (VT == MVT::i32 || VT == MVT::f32)
24778 return std::make_pair(0U, &X86::GR32RegClass);
24779 if (VT == MVT::i16)
24780 return std::make_pair(0U, &X86::GR16RegClass);
24781 if (VT == MVT::i8 || VT == MVT::i1)
24782 return std::make_pair(0U, &X86::GR8RegClass);
24783 if (VT == MVT::i64 || VT == MVT::f64)
24784 return std::make_pair(0U, &X86::GR64RegClass);
24787 // 32-bit fallthrough
24788 case 'Q': // Q_REGS
24789 if (VT == MVT::i32 || VT == MVT::f32)
24790 return std::make_pair(0U, &X86::GR32_ABCDRegClass);
24791 if (VT == MVT::i16)
24792 return std::make_pair(0U, &X86::GR16_ABCDRegClass);
24793 if (VT == MVT::i8 || VT == MVT::i1)
24794 return std::make_pair(0U, &X86::GR8_ABCD_LRegClass);
24795 if (VT == MVT::i64)
24796 return std::make_pair(0U, &X86::GR64_ABCDRegClass);
24798 case 'r': // GENERAL_REGS
24799 case 'l': // INDEX_REGS
24800 if (VT == MVT::i8 || VT == MVT::i1)
24801 return std::make_pair(0U, &X86::GR8RegClass);
24802 if (VT == MVT::i16)
24803 return std::make_pair(0U, &X86::GR16RegClass);
24804 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
24805 return std::make_pair(0U, &X86::GR32RegClass);
24806 return std::make_pair(0U, &X86::GR64RegClass);
24807 case 'R': // LEGACY_REGS
24808 if (VT == MVT::i8 || VT == MVT::i1)
24809 return std::make_pair(0U, &X86::GR8_NOREXRegClass);
24810 if (VT == MVT::i16)
24811 return std::make_pair(0U, &X86::GR16_NOREXRegClass);
24812 if (VT == MVT::i32 || !Subtarget->is64Bit())
24813 return std::make_pair(0U, &X86::GR32_NOREXRegClass);
24814 return std::make_pair(0U, &X86::GR64_NOREXRegClass);
24815 case 'f': // FP Stack registers.
24816 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
24817 // value to the correct fpstack register class.
24818 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
24819 return std::make_pair(0U, &X86::RFP32RegClass);
24820 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
24821 return std::make_pair(0U, &X86::RFP64RegClass);
24822 return std::make_pair(0U, &X86::RFP80RegClass);
24823 case 'y': // MMX_REGS if MMX allowed.
24824 if (!Subtarget->hasMMX()) break;
24825 return std::make_pair(0U, &X86::VR64RegClass);
24826 case 'Y': // SSE_REGS if SSE2 allowed
24827 if (!Subtarget->hasSSE2()) break;
24829 case 'x': // SSE_REGS if SSE1 allowed or AVX_REGS if AVX allowed
24830 if (!Subtarget->hasSSE1()) break;
24832 switch (VT.SimpleTy) {
24834 // Scalar SSE types.
24837 return std::make_pair(0U, &X86::FR32RegClass);
24840 return std::make_pair(0U, &X86::FR64RegClass);
24848 return std::make_pair(0U, &X86::VR128RegClass);
24856 return std::make_pair(0U, &X86::VR256RegClass);
24861 return std::make_pair(0U, &X86::VR512RegClass);
24867 // Use the default implementation in TargetLowering to convert the register
24868 // constraint into a member of a register class.
24869 std::pair<unsigned, const TargetRegisterClass*> Res;
24870 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
24872 // Not found as a standard register?
24874 // Map st(0) -> st(7) -> ST0
24875 if (Constraint.size() == 7 && Constraint[0] == '{' &&
24876 tolower(Constraint[1]) == 's' &&
24877 tolower(Constraint[2]) == 't' &&
24878 Constraint[3] == '(' &&
24879 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
24880 Constraint[5] == ')' &&
24881 Constraint[6] == '}') {
24883 Res.first = X86::FP0+Constraint[4]-'0';
24884 Res.second = &X86::RFP80RegClass;
24888 // GCC allows "st(0)" to be called just plain "st".
24889 if (StringRef("{st}").equals_lower(Constraint)) {
24890 Res.first = X86::FP0;
24891 Res.second = &X86::RFP80RegClass;
24896 if (StringRef("{flags}").equals_lower(Constraint)) {
24897 Res.first = X86::EFLAGS;
24898 Res.second = &X86::CCRRegClass;
24902 // 'A' means EAX + EDX.
24903 if (Constraint == "A") {
24904 Res.first = X86::EAX;
24905 Res.second = &X86::GR32_ADRegClass;
24911 // Otherwise, check to see if this is a register class of the wrong value
24912 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
24913 // turn into {ax},{dx}.
24914 if (Res.second->hasType(VT))
24915 return Res; // Correct type already, nothing to do.
24917 // All of the single-register GCC register classes map their values onto
24918 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
24919 // really want an 8-bit or 32-bit register, map to the appropriate register
24920 // class and return the appropriate register.
24921 if (Res.second == &X86::GR16RegClass) {
24922 if (VT == MVT::i8 || VT == MVT::i1) {
24923 unsigned DestReg = 0;
24924 switch (Res.first) {
24926 case X86::AX: DestReg = X86::AL; break;
24927 case X86::DX: DestReg = X86::DL; break;
24928 case X86::CX: DestReg = X86::CL; break;
24929 case X86::BX: DestReg = X86::BL; break;
24932 Res.first = DestReg;
24933 Res.second = &X86::GR8RegClass;
24935 } else if (VT == MVT::i32 || VT == MVT::f32) {
24936 unsigned DestReg = 0;
24937 switch (Res.first) {
24939 case X86::AX: DestReg = X86::EAX; break;
24940 case X86::DX: DestReg = X86::EDX; break;
24941 case X86::CX: DestReg = X86::ECX; break;
24942 case X86::BX: DestReg = X86::EBX; break;
24943 case X86::SI: DestReg = X86::ESI; break;
24944 case X86::DI: DestReg = X86::EDI; break;
24945 case X86::BP: DestReg = X86::EBP; break;
24946 case X86::SP: DestReg = X86::ESP; break;
24949 Res.first = DestReg;
24950 Res.second = &X86::GR32RegClass;
24952 } else if (VT == MVT::i64 || VT == MVT::f64) {
24953 unsigned DestReg = 0;
24954 switch (Res.first) {
24956 case X86::AX: DestReg = X86::RAX; break;
24957 case X86::DX: DestReg = X86::RDX; break;
24958 case X86::CX: DestReg = X86::RCX; break;
24959 case X86::BX: DestReg = X86::RBX; break;
24960 case X86::SI: DestReg = X86::RSI; break;
24961 case X86::DI: DestReg = X86::RDI; break;
24962 case X86::BP: DestReg = X86::RBP; break;
24963 case X86::SP: DestReg = X86::RSP; break;
24966 Res.first = DestReg;
24967 Res.second = &X86::GR64RegClass;
24970 } else if (Res.second == &X86::FR32RegClass ||
24971 Res.second == &X86::FR64RegClass ||
24972 Res.second == &X86::VR128RegClass ||
24973 Res.second == &X86::VR256RegClass ||
24974 Res.second == &X86::FR32XRegClass ||
24975 Res.second == &X86::FR64XRegClass ||
24976 Res.second == &X86::VR128XRegClass ||
24977 Res.second == &X86::VR256XRegClass ||
24978 Res.second == &X86::VR512RegClass) {
24979 // Handle references to XMM physical registers that got mapped into the
24980 // wrong class. This can happen with constraints like {xmm0} where the
24981 // target independent register mapper will just pick the first match it can
24982 // find, ignoring the required type.
24984 if (VT == MVT::f32 || VT == MVT::i32)
24985 Res.second = &X86::FR32RegClass;
24986 else if (VT == MVT::f64 || VT == MVT::i64)
24987 Res.second = &X86::FR64RegClass;
24988 else if (X86::VR128RegClass.hasType(VT))
24989 Res.second = &X86::VR128RegClass;
24990 else if (X86::VR256RegClass.hasType(VT))
24991 Res.second = &X86::VR256RegClass;
24992 else if (X86::VR512RegClass.hasType(VT))
24993 Res.second = &X86::VR512RegClass;
24999 int X86TargetLowering::getScalingFactorCost(const AddrMode &AM,
25001 // Scaling factors are not free at all.
25002 // An indexed folded instruction, i.e., inst (reg1, reg2, scale),
25003 // will take 2 allocations in the out of order engine instead of 1
25004 // for plain addressing mode, i.e. inst (reg1).
25006 // vaddps (%rsi,%drx), %ymm0, %ymm1
25007 // Requires two allocations (one for the load, one for the computation)
25009 // vaddps (%rsi), %ymm0, %ymm1
25010 // Requires just 1 allocation, i.e., freeing allocations for other operations
25011 // and having less micro operations to execute.
25013 // For some X86 architectures, this is even worse because for instance for
25014 // stores, the complex addressing mode forces the instruction to use the
25015 // "load" ports instead of the dedicated "store" port.
25016 // E.g., on Haswell:
25017 // vmovaps %ymm1, (%r8, %rdi) can use port 2 or 3.
25018 // vmovaps %ymm1, (%r8) can use port 2, 3, or 7.
25019 if (isLegalAddressingMode(AM, Ty))
25020 // Scale represents reg2 * scale, thus account for 1
25021 // as soon as we use a second register.
25022 return AM.Scale != 0;
25026 bool X86TargetLowering::isTargetFTOL() const {
25027 return Subtarget->isTargetKnownWindowsMSVC() && !Subtarget->is64Bit();