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 #define DEBUG_TYPE "x86-isel"
16 #include "X86ISelLowering.h"
17 #include "Utils/X86ShuffleDecode.h"
19 #include "X86InstrBuilder.h"
20 #include "X86TargetMachine.h"
21 #include "X86TargetObjectFile.h"
22 #include "llvm/ADT/SmallSet.h"
23 #include "llvm/ADT/Statistic.h"
24 #include "llvm/ADT/StringExtras.h"
25 #include "llvm/ADT/VariadicFunction.h"
26 #include "llvm/CodeGen/IntrinsicLowering.h"
27 #include "llvm/CodeGen/MachineFrameInfo.h"
28 #include "llvm/CodeGen/MachineFunction.h"
29 #include "llvm/CodeGen/MachineInstrBuilder.h"
30 #include "llvm/CodeGen/MachineJumpTableInfo.h"
31 #include "llvm/CodeGen/MachineModuleInfo.h"
32 #include "llvm/CodeGen/MachineRegisterInfo.h"
33 #include "llvm/IR/CallingConv.h"
34 #include "llvm/IR/Constants.h"
35 #include "llvm/IR/DerivedTypes.h"
36 #include "llvm/IR/Function.h"
37 #include "llvm/IR/GlobalAlias.h"
38 #include "llvm/IR/GlobalVariable.h"
39 #include "llvm/IR/Instructions.h"
40 #include "llvm/IR/Intrinsics.h"
41 #include "llvm/IR/LLVMContext.h"
42 #include "llvm/MC/MCAsmInfo.h"
43 #include "llvm/MC/MCContext.h"
44 #include "llvm/MC/MCExpr.h"
45 #include "llvm/MC/MCSymbol.h"
46 #include "llvm/Support/CallSite.h"
47 #include "llvm/Support/Debug.h"
48 #include "llvm/Support/ErrorHandling.h"
49 #include "llvm/Support/MathExtras.h"
50 #include "llvm/Target/TargetOptions.h"
55 STATISTIC(NumTailCalls, "Number of tail calls");
57 // Forward declarations.
58 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1,
61 static SDValue ExtractSubVector(SDValue Vec, unsigned IdxVal,
62 SelectionDAG &DAG, SDLoc dl,
63 unsigned vectorWidth) {
64 assert((vectorWidth == 128 || vectorWidth == 256) &&
65 "Unsupported vector width");
66 EVT VT = Vec.getValueType();
67 EVT ElVT = VT.getVectorElementType();
68 unsigned Factor = VT.getSizeInBits()/vectorWidth;
69 EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT,
70 VT.getVectorNumElements()/Factor);
72 // Extract from UNDEF is UNDEF.
73 if (Vec.getOpcode() == ISD::UNDEF)
74 return DAG.getUNDEF(ResultVT);
76 // Extract the relevant vectorWidth bits. Generate an EXTRACT_SUBVECTOR
77 unsigned ElemsPerChunk = vectorWidth / ElVT.getSizeInBits();
79 // This is the index of the first element of the vectorWidth-bit chunk
81 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits()) / vectorWidth)
84 // If the input is a buildvector just emit a smaller one.
85 if (Vec.getOpcode() == ISD::BUILD_VECTOR)
86 return DAG.getNode(ISD::BUILD_VECTOR, dl, ResultVT,
87 Vec->op_begin()+NormalizedIdxVal, ElemsPerChunk);
89 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal);
90 SDValue Result = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec,
96 /// Generate a DAG to grab 128-bits from a vector > 128 bits. This
97 /// sets things up to match to an AVX VEXTRACTF128 / VEXTRACTI128
98 /// or AVX-512 VEXTRACTF32x4 / VEXTRACTI32x4
99 /// instructions or a simple subregister reference. Idx is an index in the
100 /// 128 bits we want. It need not be aligned to a 128-bit bounday. That makes
101 /// lowering EXTRACT_VECTOR_ELT operations easier.
102 static SDValue Extract128BitVector(SDValue Vec, unsigned IdxVal,
103 SelectionDAG &DAG, SDLoc dl) {
104 assert((Vec.getValueType().is256BitVector() ||
105 Vec.getValueType().is512BitVector()) && "Unexpected vector size!");
106 return ExtractSubVector(Vec, IdxVal, DAG, dl, 128);
109 /// Generate a DAG to grab 256-bits from a 512-bit vector.
110 static SDValue Extract256BitVector(SDValue Vec, unsigned IdxVal,
111 SelectionDAG &DAG, SDLoc dl) {
112 assert(Vec.getValueType().is512BitVector() && "Unexpected vector size!");
113 return ExtractSubVector(Vec, IdxVal, DAG, dl, 256);
116 static SDValue InsertSubVector(SDValue Result, SDValue Vec,
117 unsigned IdxVal, SelectionDAG &DAG,
118 SDLoc dl, unsigned vectorWidth) {
119 assert((vectorWidth == 128 || vectorWidth == 256) &&
120 "Unsupported vector width");
121 // Inserting UNDEF is Result
122 if (Vec.getOpcode() == ISD::UNDEF)
124 EVT VT = Vec.getValueType();
125 EVT ElVT = VT.getVectorElementType();
126 EVT ResultVT = Result.getValueType();
128 // Insert the relevant vectorWidth bits.
129 unsigned ElemsPerChunk = vectorWidth/ElVT.getSizeInBits();
131 // This is the index of the first element of the vectorWidth-bit chunk
133 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits())/vectorWidth)
136 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal);
137 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec,
140 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
141 /// sets things up to match to an AVX VINSERTF128/VINSERTI128 or
142 /// AVX-512 VINSERTF32x4/VINSERTI32x4 instructions or a
143 /// simple superregister reference. Idx is an index in the 128 bits
144 /// we want. It need not be aligned to a 128-bit bounday. That makes
145 /// lowering INSERT_VECTOR_ELT operations easier.
146 static SDValue Insert128BitVector(SDValue Result, SDValue Vec,
147 unsigned IdxVal, SelectionDAG &DAG,
149 assert(Vec.getValueType().is128BitVector() && "Unexpected vector size!");
150 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 128);
153 static SDValue Insert256BitVector(SDValue Result, SDValue Vec,
154 unsigned IdxVal, SelectionDAG &DAG,
156 assert(Vec.getValueType().is256BitVector() && "Unexpected vector size!");
157 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 256);
160 /// Concat two 128-bit vectors into a 256 bit vector using VINSERTF128
161 /// instructions. This is used because creating CONCAT_VECTOR nodes of
162 /// BUILD_VECTORS returns a larger BUILD_VECTOR while we're trying to lower
163 /// large BUILD_VECTORS.
164 static SDValue Concat128BitVectors(SDValue V1, SDValue V2, EVT VT,
165 unsigned NumElems, SelectionDAG &DAG,
167 SDValue V = Insert128BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
168 return Insert128BitVector(V, V2, NumElems/2, DAG, dl);
171 static SDValue Concat256BitVectors(SDValue V1, SDValue V2, EVT VT,
172 unsigned NumElems, SelectionDAG &DAG,
174 SDValue V = Insert256BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
175 return Insert256BitVector(V, V2, NumElems/2, DAG, dl);
178 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
179 const X86Subtarget *Subtarget = &TM.getSubtarget<X86Subtarget>();
180 bool is64Bit = Subtarget->is64Bit();
182 if (Subtarget->isTargetEnvMacho()) {
184 return new X86_64MachoTargetObjectFile();
185 return new TargetLoweringObjectFileMachO();
188 if (Subtarget->isTargetLinux())
189 return new X86LinuxTargetObjectFile();
190 if (Subtarget->isTargetELF())
191 return new TargetLoweringObjectFileELF();
192 if (Subtarget->isTargetCOFF() && !Subtarget->isTargetEnvMacho())
193 return new TargetLoweringObjectFileCOFF();
194 llvm_unreachable("unknown subtarget type");
197 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
198 : TargetLowering(TM, createTLOF(TM)) {
199 Subtarget = &TM.getSubtarget<X86Subtarget>();
200 X86ScalarSSEf64 = Subtarget->hasSSE2();
201 X86ScalarSSEf32 = Subtarget->hasSSE1();
202 TD = getDataLayout();
204 resetOperationActions();
207 void X86TargetLowering::resetOperationActions() {
208 const TargetMachine &TM = getTargetMachine();
209 static bool FirstTimeThrough = true;
211 // If none of the target options have changed, then we don't need to reset the
212 // operation actions.
213 if (!FirstTimeThrough && TO == TM.Options) return;
215 if (!FirstTimeThrough) {
216 // Reinitialize the actions.
218 FirstTimeThrough = false;
223 // Set up the TargetLowering object.
224 static const MVT IntVTs[] = { MVT::i8, MVT::i16, MVT::i32, MVT::i64 };
226 // X86 is weird, it always uses i8 for shift amounts and setcc results.
227 setBooleanContents(ZeroOrOneBooleanContent);
228 // X86-SSE is even stranger. It uses -1 or 0 for vector masks.
229 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
231 // For 64-bit since we have so many registers use the ILP scheduler, for
232 // 32-bit code use the register pressure specific scheduling.
233 // For Atom, always use ILP scheduling.
234 if (Subtarget->isAtom())
235 setSchedulingPreference(Sched::ILP);
236 else if (Subtarget->is64Bit())
237 setSchedulingPreference(Sched::ILP);
239 setSchedulingPreference(Sched::RegPressure);
240 const X86RegisterInfo *RegInfo =
241 static_cast<const X86RegisterInfo*>(TM.getRegisterInfo());
242 setStackPointerRegisterToSaveRestore(RegInfo->getStackRegister());
244 // Bypass expensive divides on Atom when compiling with O2
245 if (Subtarget->hasSlowDivide() && TM.getOptLevel() >= CodeGenOpt::Default) {
246 addBypassSlowDiv(32, 8);
247 if (Subtarget->is64Bit())
248 addBypassSlowDiv(64, 16);
251 if (Subtarget->isTargetWindows() && !Subtarget->isTargetCygMing()) {
252 // Setup Windows compiler runtime calls.
253 setLibcallName(RTLIB::SDIV_I64, "_alldiv");
254 setLibcallName(RTLIB::UDIV_I64, "_aulldiv");
255 setLibcallName(RTLIB::SREM_I64, "_allrem");
256 setLibcallName(RTLIB::UREM_I64, "_aullrem");
257 setLibcallName(RTLIB::MUL_I64, "_allmul");
258 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::X86_StdCall);
259 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::X86_StdCall);
260 setLibcallCallingConv(RTLIB::SREM_I64, CallingConv::X86_StdCall);
261 setLibcallCallingConv(RTLIB::UREM_I64, CallingConv::X86_StdCall);
262 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::X86_StdCall);
264 // The _ftol2 runtime function has an unusual calling conv, which
265 // is modeled by a special pseudo-instruction.
266 setLibcallName(RTLIB::FPTOUINT_F64_I64, 0);
267 setLibcallName(RTLIB::FPTOUINT_F32_I64, 0);
268 setLibcallName(RTLIB::FPTOUINT_F64_I32, 0);
269 setLibcallName(RTLIB::FPTOUINT_F32_I32, 0);
272 if (Subtarget->isTargetDarwin()) {
273 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
274 setUseUnderscoreSetJmp(false);
275 setUseUnderscoreLongJmp(false);
276 } else if (Subtarget->isTargetMingw()) {
277 // MS runtime is weird: it exports _setjmp, but longjmp!
278 setUseUnderscoreSetJmp(true);
279 setUseUnderscoreLongJmp(false);
281 setUseUnderscoreSetJmp(true);
282 setUseUnderscoreLongJmp(true);
285 // Set up the register classes.
286 addRegisterClass(MVT::i8, &X86::GR8RegClass);
287 addRegisterClass(MVT::i16, &X86::GR16RegClass);
288 addRegisterClass(MVT::i32, &X86::GR32RegClass);
289 if (Subtarget->is64Bit())
290 addRegisterClass(MVT::i64, &X86::GR64RegClass);
292 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
294 // We don't accept any truncstore of integer registers.
295 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
296 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
297 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
298 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
299 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
300 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
302 // SETOEQ and SETUNE require checking two conditions.
303 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
304 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
305 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
306 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
307 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
308 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
310 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
312 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
313 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
314 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
316 if (Subtarget->is64Bit()) {
317 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
318 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
319 } else if (!TM.Options.UseSoftFloat) {
320 // We have an algorithm for SSE2->double, and we turn this into a
321 // 64-bit FILD followed by conditional FADD for other targets.
322 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
323 // We have an algorithm for SSE2, and we turn this into a 64-bit
324 // FILD for other targets.
325 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
328 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
330 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
331 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
333 if (!TM.Options.UseSoftFloat) {
334 // SSE has no i16 to fp conversion, only i32
335 if (X86ScalarSSEf32) {
336 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
337 // f32 and f64 cases are Legal, f80 case is not
338 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
340 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
341 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
344 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
345 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
348 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
349 // are Legal, f80 is custom lowered.
350 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
351 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
353 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
355 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
356 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
358 if (X86ScalarSSEf32) {
359 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
360 // f32 and f64 cases are Legal, f80 case is not
361 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
363 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
364 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
367 // Handle FP_TO_UINT by promoting the destination to a larger signed
369 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
370 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
371 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
373 if (Subtarget->is64Bit()) {
374 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
375 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
376 } else if (!TM.Options.UseSoftFloat) {
377 // Since AVX is a superset of SSE3, only check for SSE here.
378 if (Subtarget->hasSSE1() && !Subtarget->hasSSE3())
379 // Expand FP_TO_UINT into a select.
380 // FIXME: We would like to use a Custom expander here eventually to do
381 // the optimal thing for SSE vs. the default expansion in the legalizer.
382 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
384 // With SSE3 we can use fisttpll to convert to a signed i64; without
385 // SSE, we're stuck with a fistpll.
386 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
389 if (isTargetFTOL()) {
390 // Use the _ftol2 runtime function, which has a pseudo-instruction
391 // to handle its weird calling convention.
392 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Custom);
395 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
396 if (!X86ScalarSSEf64) {
397 setOperationAction(ISD::BITCAST , MVT::f32 , Expand);
398 setOperationAction(ISD::BITCAST , MVT::i32 , Expand);
399 if (Subtarget->is64Bit()) {
400 setOperationAction(ISD::BITCAST , MVT::f64 , Expand);
401 // Without SSE, i64->f64 goes through memory.
402 setOperationAction(ISD::BITCAST , MVT::i64 , Expand);
406 // Scalar integer divide and remainder are lowered to use operations that
407 // produce two results, to match the available instructions. This exposes
408 // the two-result form to trivial CSE, which is able to combine x/y and x%y
409 // into a single instruction.
411 // Scalar integer multiply-high is also lowered to use two-result
412 // operations, to match the available instructions. However, plain multiply
413 // (low) operations are left as Legal, as there are single-result
414 // instructions for this in x86. Using the two-result multiply instructions
415 // when both high and low results are needed must be arranged by dagcombine.
416 for (unsigned i = 0; i != array_lengthof(IntVTs); ++i) {
418 setOperationAction(ISD::MULHS, VT, Expand);
419 setOperationAction(ISD::MULHU, VT, Expand);
420 setOperationAction(ISD::SDIV, VT, Expand);
421 setOperationAction(ISD::UDIV, VT, Expand);
422 setOperationAction(ISD::SREM, VT, Expand);
423 setOperationAction(ISD::UREM, VT, Expand);
425 // Add/Sub overflow ops with MVT::Glues are lowered to EFLAGS dependences.
426 setOperationAction(ISD::ADDC, VT, Custom);
427 setOperationAction(ISD::ADDE, VT, Custom);
428 setOperationAction(ISD::SUBC, VT, Custom);
429 setOperationAction(ISD::SUBE, VT, Custom);
432 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
433 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
434 setOperationAction(ISD::BR_CC , MVT::f32, Expand);
435 setOperationAction(ISD::BR_CC , MVT::f64, Expand);
436 setOperationAction(ISD::BR_CC , MVT::f80, Expand);
437 setOperationAction(ISD::BR_CC , MVT::i8, Expand);
438 setOperationAction(ISD::BR_CC , MVT::i16, Expand);
439 setOperationAction(ISD::BR_CC , MVT::i32, Expand);
440 setOperationAction(ISD::BR_CC , MVT::i64, Expand);
441 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
442 if (Subtarget->is64Bit())
443 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
444 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
445 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
446 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
447 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
448 setOperationAction(ISD::FREM , MVT::f32 , Expand);
449 setOperationAction(ISD::FREM , MVT::f64 , Expand);
450 setOperationAction(ISD::FREM , MVT::f80 , Expand);
451 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
453 // Promote the i8 variants and force them on up to i32 which has a shorter
455 setOperationAction(ISD::CTTZ , MVT::i8 , Promote);
456 AddPromotedToType (ISD::CTTZ , MVT::i8 , MVT::i32);
457 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i8 , Promote);
458 AddPromotedToType (ISD::CTTZ_ZERO_UNDEF , MVT::i8 , MVT::i32);
459 if (Subtarget->hasBMI()) {
460 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i16 , Expand);
461 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32 , Expand);
462 if (Subtarget->is64Bit())
463 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
465 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
466 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
467 if (Subtarget->is64Bit())
468 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
471 if (Subtarget->hasLZCNT()) {
472 // When promoting the i8 variants, force them to i32 for a shorter
474 setOperationAction(ISD::CTLZ , MVT::i8 , Promote);
475 AddPromotedToType (ISD::CTLZ , MVT::i8 , MVT::i32);
476 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Promote);
477 AddPromotedToType (ISD::CTLZ_ZERO_UNDEF, MVT::i8 , MVT::i32);
478 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Expand);
479 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Expand);
480 if (Subtarget->is64Bit())
481 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
483 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
484 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
485 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
486 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Custom);
487 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Custom);
488 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Custom);
489 if (Subtarget->is64Bit()) {
490 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
491 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Custom);
495 if (Subtarget->hasPOPCNT()) {
496 setOperationAction(ISD::CTPOP , MVT::i8 , Promote);
498 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
499 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
500 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
501 if (Subtarget->is64Bit())
502 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
505 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
506 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
508 // These should be promoted to a larger select which is supported.
509 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
510 // X86 wants to expand cmov itself.
511 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
512 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
513 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
514 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
515 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
516 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
517 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
518 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
519 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
520 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
521 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
522 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
523 if (Subtarget->is64Bit()) {
524 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
525 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
527 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
528 // NOTE: EH_SJLJ_SETJMP/_LONGJMP supported here is NOT intended to support
529 // SjLj exception handling but a light-weight setjmp/longjmp replacement to
530 // support continuation, user-level threading, and etc.. As a result, no
531 // other SjLj exception interfaces are implemented and please don't build
532 // your own exception handling based on them.
533 // LLVM/Clang supports zero-cost DWARF exception handling.
534 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
535 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
538 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
539 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
540 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
541 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
542 if (Subtarget->is64Bit())
543 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
544 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
545 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
546 if (Subtarget->is64Bit()) {
547 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
548 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
549 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
550 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
551 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
553 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
554 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
555 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
556 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
557 if (Subtarget->is64Bit()) {
558 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
559 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
560 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
563 if (Subtarget->hasSSE1())
564 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
566 setOperationAction(ISD::ATOMIC_FENCE , MVT::Other, Custom);
568 // Expand certain atomics
569 for (unsigned i = 0; i != array_lengthof(IntVTs); ++i) {
571 setOperationAction(ISD::ATOMIC_CMP_SWAP, VT, Custom);
572 setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
573 setOperationAction(ISD::ATOMIC_STORE, VT, Custom);
576 if (!Subtarget->is64Bit()) {
577 setOperationAction(ISD::ATOMIC_LOAD, MVT::i64, Custom);
578 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
579 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
580 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
581 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
582 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
583 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
584 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
585 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i64, Custom);
586 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i64, Custom);
587 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i64, Custom);
588 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i64, Custom);
591 if (Subtarget->hasCmpxchg16b()) {
592 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i128, Custom);
595 // FIXME - use subtarget debug flags
596 if (!Subtarget->isTargetDarwin() &&
597 !Subtarget->isTargetELF() &&
598 !Subtarget->isTargetCygMing()) {
599 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
602 if (Subtarget->is64Bit()) {
603 setExceptionPointerRegister(X86::RAX);
604 setExceptionSelectorRegister(X86::RDX);
606 setExceptionPointerRegister(X86::EAX);
607 setExceptionSelectorRegister(X86::EDX);
609 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
610 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
612 setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
613 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
615 setOperationAction(ISD::TRAP, MVT::Other, Legal);
616 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Legal);
618 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
619 setOperationAction(ISD::VASTART , MVT::Other, Custom);
620 setOperationAction(ISD::VAEND , MVT::Other, Expand);
621 if (Subtarget->is64Bit() && !Subtarget->isTargetWin64()) {
622 // TargetInfo::X86_64ABIBuiltinVaList
623 setOperationAction(ISD::VAARG , MVT::Other, Custom);
624 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
626 // TargetInfo::CharPtrBuiltinVaList
627 setOperationAction(ISD::VAARG , MVT::Other, Expand);
628 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
631 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
632 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
634 if (Subtarget->isTargetCOFF() && !Subtarget->isTargetEnvMacho())
635 setOperationAction(ISD::DYNAMIC_STACKALLOC, Subtarget->is64Bit() ?
636 MVT::i64 : MVT::i32, Custom);
637 else if (TM.Options.EnableSegmentedStacks)
638 setOperationAction(ISD::DYNAMIC_STACKALLOC, Subtarget->is64Bit() ?
639 MVT::i64 : MVT::i32, Custom);
641 setOperationAction(ISD::DYNAMIC_STACKALLOC, Subtarget->is64Bit() ?
642 MVT::i64 : MVT::i32, Expand);
644 if (!TM.Options.UseSoftFloat && X86ScalarSSEf64) {
645 // f32 and f64 use SSE.
646 // Set up the FP register classes.
647 addRegisterClass(MVT::f32, &X86::FR32RegClass);
648 addRegisterClass(MVT::f64, &X86::FR64RegClass);
650 // Use ANDPD to simulate FABS.
651 setOperationAction(ISD::FABS , MVT::f64, Custom);
652 setOperationAction(ISD::FABS , MVT::f32, Custom);
654 // Use XORP to simulate FNEG.
655 setOperationAction(ISD::FNEG , MVT::f64, Custom);
656 setOperationAction(ISD::FNEG , MVT::f32, Custom);
658 // Use ANDPD and ORPD to simulate FCOPYSIGN.
659 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
660 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
662 // Lower this to FGETSIGNx86 plus an AND.
663 setOperationAction(ISD::FGETSIGN, MVT::i64, Custom);
664 setOperationAction(ISD::FGETSIGN, MVT::i32, Custom);
666 // We don't support sin/cos/fmod
667 setOperationAction(ISD::FSIN , MVT::f64, Expand);
668 setOperationAction(ISD::FCOS , MVT::f64, Expand);
669 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
670 setOperationAction(ISD::FSIN , MVT::f32, Expand);
671 setOperationAction(ISD::FCOS , MVT::f32, Expand);
672 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
674 // Expand FP immediates into loads from the stack, except for the special
676 addLegalFPImmediate(APFloat(+0.0)); // xorpd
677 addLegalFPImmediate(APFloat(+0.0f)); // xorps
678 } else if (!TM.Options.UseSoftFloat && X86ScalarSSEf32) {
679 // Use SSE for f32, x87 for f64.
680 // Set up the FP register classes.
681 addRegisterClass(MVT::f32, &X86::FR32RegClass);
682 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
684 // Use ANDPS to simulate FABS.
685 setOperationAction(ISD::FABS , MVT::f32, Custom);
687 // Use XORP to simulate FNEG.
688 setOperationAction(ISD::FNEG , MVT::f32, Custom);
690 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
692 // Use ANDPS and ORPS to simulate FCOPYSIGN.
693 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
694 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
696 // We don't support sin/cos/fmod
697 setOperationAction(ISD::FSIN , MVT::f32, Expand);
698 setOperationAction(ISD::FCOS , MVT::f32, Expand);
699 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
701 // Special cases we handle for FP constants.
702 addLegalFPImmediate(APFloat(+0.0f)); // xorps
703 addLegalFPImmediate(APFloat(+0.0)); // FLD0
704 addLegalFPImmediate(APFloat(+1.0)); // FLD1
705 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
706 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
708 if (!TM.Options.UnsafeFPMath) {
709 setOperationAction(ISD::FSIN , MVT::f64, Expand);
710 setOperationAction(ISD::FCOS , MVT::f64, Expand);
711 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
713 } else if (!TM.Options.UseSoftFloat) {
714 // f32 and f64 in x87.
715 // Set up the FP register classes.
716 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
717 addRegisterClass(MVT::f32, &X86::RFP32RegClass);
719 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
720 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
721 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
722 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
724 if (!TM.Options.UnsafeFPMath) {
725 setOperationAction(ISD::FSIN , MVT::f64, Expand);
726 setOperationAction(ISD::FSIN , MVT::f32, Expand);
727 setOperationAction(ISD::FCOS , MVT::f64, Expand);
728 setOperationAction(ISD::FCOS , MVT::f32, Expand);
729 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
730 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
732 addLegalFPImmediate(APFloat(+0.0)); // FLD0
733 addLegalFPImmediate(APFloat(+1.0)); // FLD1
734 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
735 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
736 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
737 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
738 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
739 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
742 // We don't support FMA.
743 setOperationAction(ISD::FMA, MVT::f64, Expand);
744 setOperationAction(ISD::FMA, MVT::f32, Expand);
746 // Long double always uses X87.
747 if (!TM.Options.UseSoftFloat) {
748 addRegisterClass(MVT::f80, &X86::RFP80RegClass);
749 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
750 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
752 APFloat TmpFlt = APFloat::getZero(APFloat::x87DoubleExtended);
753 addLegalFPImmediate(TmpFlt); // FLD0
755 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
758 APFloat TmpFlt2(+1.0);
759 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
761 addLegalFPImmediate(TmpFlt2); // FLD1
762 TmpFlt2.changeSign();
763 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
766 if (!TM.Options.UnsafeFPMath) {
767 setOperationAction(ISD::FSIN , MVT::f80, Expand);
768 setOperationAction(ISD::FCOS , MVT::f80, Expand);
769 setOperationAction(ISD::FSINCOS, MVT::f80, Expand);
772 setOperationAction(ISD::FFLOOR, MVT::f80, Expand);
773 setOperationAction(ISD::FCEIL, MVT::f80, Expand);
774 setOperationAction(ISD::FTRUNC, MVT::f80, Expand);
775 setOperationAction(ISD::FRINT, MVT::f80, Expand);
776 setOperationAction(ISD::FNEARBYINT, MVT::f80, Expand);
777 setOperationAction(ISD::FMA, MVT::f80, Expand);
780 // Always use a library call for pow.
781 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
782 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
783 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
785 setOperationAction(ISD::FLOG, MVT::f80, Expand);
786 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
787 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
788 setOperationAction(ISD::FEXP, MVT::f80, Expand);
789 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
791 // First set operation action for all vector types to either promote
792 // (for widening) or expand (for scalarization). Then we will selectively
793 // turn on ones that can be effectively codegen'd.
794 for (int i = MVT::FIRST_VECTOR_VALUETYPE;
795 i <= MVT::LAST_VECTOR_VALUETYPE; ++i) {
796 MVT VT = (MVT::SimpleValueType)i;
797 setOperationAction(ISD::ADD , VT, Expand);
798 setOperationAction(ISD::SUB , VT, Expand);
799 setOperationAction(ISD::FADD, VT, Expand);
800 setOperationAction(ISD::FNEG, VT, Expand);
801 setOperationAction(ISD::FSUB, VT, Expand);
802 setOperationAction(ISD::MUL , VT, Expand);
803 setOperationAction(ISD::FMUL, VT, Expand);
804 setOperationAction(ISD::SDIV, VT, Expand);
805 setOperationAction(ISD::UDIV, VT, Expand);
806 setOperationAction(ISD::FDIV, VT, Expand);
807 setOperationAction(ISD::SREM, VT, Expand);
808 setOperationAction(ISD::UREM, VT, Expand);
809 setOperationAction(ISD::LOAD, VT, Expand);
810 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
811 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT,Expand);
812 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
813 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT,Expand);
814 setOperationAction(ISD::INSERT_SUBVECTOR, VT,Expand);
815 setOperationAction(ISD::FABS, VT, Expand);
816 setOperationAction(ISD::FSIN, VT, Expand);
817 setOperationAction(ISD::FSINCOS, VT, Expand);
818 setOperationAction(ISD::FCOS, VT, Expand);
819 setOperationAction(ISD::FSINCOS, VT, Expand);
820 setOperationAction(ISD::FREM, VT, Expand);
821 setOperationAction(ISD::FMA, VT, Expand);
822 setOperationAction(ISD::FPOWI, VT, Expand);
823 setOperationAction(ISD::FSQRT, VT, Expand);
824 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
825 setOperationAction(ISD::FFLOOR, VT, Expand);
826 setOperationAction(ISD::FCEIL, VT, Expand);
827 setOperationAction(ISD::FTRUNC, VT, Expand);
828 setOperationAction(ISD::FRINT, VT, Expand);
829 setOperationAction(ISD::FNEARBYINT, VT, Expand);
830 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
831 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
832 setOperationAction(ISD::SDIVREM, VT, Expand);
833 setOperationAction(ISD::UDIVREM, VT, Expand);
834 setOperationAction(ISD::FPOW, VT, Expand);
835 setOperationAction(ISD::CTPOP, VT, Expand);
836 setOperationAction(ISD::CTTZ, VT, Expand);
837 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
838 setOperationAction(ISD::CTLZ, VT, Expand);
839 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
840 setOperationAction(ISD::SHL, VT, Expand);
841 setOperationAction(ISD::SRA, VT, Expand);
842 setOperationAction(ISD::SRL, VT, Expand);
843 setOperationAction(ISD::ROTL, VT, Expand);
844 setOperationAction(ISD::ROTR, VT, Expand);
845 setOperationAction(ISD::BSWAP, VT, Expand);
846 setOperationAction(ISD::SETCC, VT, Expand);
847 setOperationAction(ISD::FLOG, VT, Expand);
848 setOperationAction(ISD::FLOG2, VT, Expand);
849 setOperationAction(ISD::FLOG10, VT, Expand);
850 setOperationAction(ISD::FEXP, VT, Expand);
851 setOperationAction(ISD::FEXP2, VT, Expand);
852 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
853 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
854 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
855 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
856 setOperationAction(ISD::SIGN_EXTEND_INREG, VT,Expand);
857 setOperationAction(ISD::TRUNCATE, VT, Expand);
858 setOperationAction(ISD::SIGN_EXTEND, VT, Expand);
859 setOperationAction(ISD::ZERO_EXTEND, VT, Expand);
860 setOperationAction(ISD::ANY_EXTEND, VT, Expand);
861 setOperationAction(ISD::VSELECT, VT, Expand);
862 for (int InnerVT = MVT::FIRST_VECTOR_VALUETYPE;
863 InnerVT <= MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
864 setTruncStoreAction(VT,
865 (MVT::SimpleValueType)InnerVT, Expand);
866 setLoadExtAction(ISD::SEXTLOAD, VT, Expand);
867 setLoadExtAction(ISD::ZEXTLOAD, VT, Expand);
868 setLoadExtAction(ISD::EXTLOAD, VT, Expand);
871 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
872 // with -msoft-float, disable use of MMX as well.
873 if (!TM.Options.UseSoftFloat && Subtarget->hasMMX()) {
874 addRegisterClass(MVT::x86mmx, &X86::VR64RegClass);
875 // No operations on x86mmx supported, everything uses intrinsics.
878 // MMX-sized vectors (other than x86mmx) are expected to be expanded
879 // into smaller operations.
880 setOperationAction(ISD::MULHS, MVT::v8i8, Expand);
881 setOperationAction(ISD::MULHS, MVT::v4i16, Expand);
882 setOperationAction(ISD::MULHS, MVT::v2i32, Expand);
883 setOperationAction(ISD::MULHS, MVT::v1i64, Expand);
884 setOperationAction(ISD::AND, MVT::v8i8, Expand);
885 setOperationAction(ISD::AND, MVT::v4i16, Expand);
886 setOperationAction(ISD::AND, MVT::v2i32, Expand);
887 setOperationAction(ISD::AND, MVT::v1i64, Expand);
888 setOperationAction(ISD::OR, MVT::v8i8, Expand);
889 setOperationAction(ISD::OR, MVT::v4i16, Expand);
890 setOperationAction(ISD::OR, MVT::v2i32, Expand);
891 setOperationAction(ISD::OR, MVT::v1i64, Expand);
892 setOperationAction(ISD::XOR, MVT::v8i8, Expand);
893 setOperationAction(ISD::XOR, MVT::v4i16, Expand);
894 setOperationAction(ISD::XOR, MVT::v2i32, Expand);
895 setOperationAction(ISD::XOR, MVT::v1i64, Expand);
896 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Expand);
897 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Expand);
898 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2i32, Expand);
899 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Expand);
900 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v1i64, Expand);
901 setOperationAction(ISD::SELECT, MVT::v8i8, Expand);
902 setOperationAction(ISD::SELECT, MVT::v4i16, Expand);
903 setOperationAction(ISD::SELECT, MVT::v2i32, Expand);
904 setOperationAction(ISD::SELECT, MVT::v1i64, Expand);
905 setOperationAction(ISD::BITCAST, MVT::v8i8, Expand);
906 setOperationAction(ISD::BITCAST, MVT::v4i16, Expand);
907 setOperationAction(ISD::BITCAST, MVT::v2i32, Expand);
908 setOperationAction(ISD::BITCAST, MVT::v1i64, Expand);
910 if (!TM.Options.UseSoftFloat && Subtarget->hasSSE1()) {
911 addRegisterClass(MVT::v4f32, &X86::VR128RegClass);
913 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
914 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
915 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
916 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
917 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
918 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
919 setOperationAction(ISD::FABS, MVT::v4f32, Custom);
920 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
921 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
922 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
923 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
924 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
927 if (!TM.Options.UseSoftFloat && Subtarget->hasSSE2()) {
928 addRegisterClass(MVT::v2f64, &X86::VR128RegClass);
930 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
931 // registers cannot be used even for integer operations.
932 addRegisterClass(MVT::v16i8, &X86::VR128RegClass);
933 addRegisterClass(MVT::v8i16, &X86::VR128RegClass);
934 addRegisterClass(MVT::v4i32, &X86::VR128RegClass);
935 addRegisterClass(MVT::v2i64, &X86::VR128RegClass);
937 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
938 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
939 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
940 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
941 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
942 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
943 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
944 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
945 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
946 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
947 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
948 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
949 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
950 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
951 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
952 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
953 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
954 setOperationAction(ISD::FABS, MVT::v2f64, Custom);
956 setOperationAction(ISD::SETCC, MVT::v2i64, Custom);
957 setOperationAction(ISD::SETCC, MVT::v16i8, Custom);
958 setOperationAction(ISD::SETCC, MVT::v8i16, Custom);
959 setOperationAction(ISD::SETCC, MVT::v4i32, Custom);
961 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
962 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
963 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
964 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
965 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
967 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
968 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
969 MVT VT = (MVT::SimpleValueType)i;
970 // Do not attempt to custom lower non-power-of-2 vectors
971 if (!isPowerOf2_32(VT.getVectorNumElements()))
973 // Do not attempt to custom lower non-128-bit vectors
974 if (!VT.is128BitVector())
976 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
977 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
978 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
981 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
982 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
983 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
984 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
985 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
986 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
988 if (Subtarget->is64Bit()) {
989 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
990 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
993 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
994 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
995 MVT VT = (MVT::SimpleValueType)i;
997 // Do not attempt to promote non-128-bit vectors
998 if (!VT.is128BitVector())
1001 setOperationAction(ISD::AND, VT, Promote);
1002 AddPromotedToType (ISD::AND, VT, MVT::v2i64);
1003 setOperationAction(ISD::OR, VT, Promote);
1004 AddPromotedToType (ISD::OR, VT, MVT::v2i64);
1005 setOperationAction(ISD::XOR, VT, Promote);
1006 AddPromotedToType (ISD::XOR, VT, MVT::v2i64);
1007 setOperationAction(ISD::LOAD, VT, Promote);
1008 AddPromotedToType (ISD::LOAD, VT, MVT::v2i64);
1009 setOperationAction(ISD::SELECT, VT, Promote);
1010 AddPromotedToType (ISD::SELECT, VT, MVT::v2i64);
1013 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
1015 // Custom lower v2i64 and v2f64 selects.
1016 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
1017 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
1018 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
1019 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
1021 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
1022 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
1024 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Custom);
1025 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
1026 // As there is no 64-bit GPR available, we need build a special custom
1027 // sequence to convert from v2i32 to v2f32.
1028 if (!Subtarget->is64Bit())
1029 setOperationAction(ISD::UINT_TO_FP, MVT::v2f32, Custom);
1031 setOperationAction(ISD::FP_EXTEND, MVT::v2f32, Custom);
1032 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Custom);
1034 setLoadExtAction(ISD::EXTLOAD, MVT::v2f32, Legal);
1037 if (!TM.Options.UseSoftFloat && Subtarget->hasSSE41()) {
1038 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
1039 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
1040 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
1041 setOperationAction(ISD::FRINT, MVT::f32, Legal);
1042 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
1043 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
1044 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
1045 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
1046 setOperationAction(ISD::FRINT, MVT::f64, Legal);
1047 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
1049 setOperationAction(ISD::FFLOOR, MVT::v4f32, Legal);
1050 setOperationAction(ISD::FCEIL, MVT::v4f32, Legal);
1051 setOperationAction(ISD::FTRUNC, MVT::v4f32, Legal);
1052 setOperationAction(ISD::FRINT, MVT::v4f32, Legal);
1053 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Legal);
1054 setOperationAction(ISD::FFLOOR, MVT::v2f64, Legal);
1055 setOperationAction(ISD::FCEIL, MVT::v2f64, Legal);
1056 setOperationAction(ISD::FTRUNC, MVT::v2f64, Legal);
1057 setOperationAction(ISD::FRINT, MVT::v2f64, Legal);
1058 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Legal);
1060 // FIXME: Do we need to handle scalar-to-vector here?
1061 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
1063 setOperationAction(ISD::VSELECT, MVT::v2f64, Legal);
1064 setOperationAction(ISD::VSELECT, MVT::v2i64, Legal);
1065 setOperationAction(ISD::VSELECT, MVT::v16i8, Legal);
1066 setOperationAction(ISD::VSELECT, MVT::v4i32, Legal);
1067 setOperationAction(ISD::VSELECT, MVT::v4f32, Legal);
1069 // i8 and i16 vectors are custom , because the source register and source
1070 // source memory operand types are not the same width. f32 vectors are
1071 // custom since the immediate controlling the insert encodes additional
1073 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
1074 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
1075 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
1076 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
1078 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
1079 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
1080 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
1081 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
1083 // FIXME: these should be Legal but thats only for the case where
1084 // the index is constant. For now custom expand to deal with that.
1085 if (Subtarget->is64Bit()) {
1086 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
1087 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
1091 if (Subtarget->hasSSE2()) {
1092 setOperationAction(ISD::SRL, MVT::v8i16, Custom);
1093 setOperationAction(ISD::SRL, MVT::v16i8, Custom);
1095 setOperationAction(ISD::SHL, MVT::v8i16, Custom);
1096 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
1098 setOperationAction(ISD::SRA, MVT::v8i16, Custom);
1099 setOperationAction(ISD::SRA, MVT::v16i8, Custom);
1101 // In the customized shift lowering, the legal cases in AVX2 will be
1103 setOperationAction(ISD::SRL, MVT::v2i64, Custom);
1104 setOperationAction(ISD::SRL, MVT::v4i32, Custom);
1106 setOperationAction(ISD::SHL, MVT::v2i64, Custom);
1107 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
1109 setOperationAction(ISD::SRA, MVT::v4i32, Custom);
1111 setOperationAction(ISD::SDIV, MVT::v8i16, Custom);
1112 setOperationAction(ISD::SDIV, MVT::v4i32, Custom);
1115 if (!TM.Options.UseSoftFloat && Subtarget->hasFp256()) {
1116 addRegisterClass(MVT::v32i8, &X86::VR256RegClass);
1117 addRegisterClass(MVT::v16i16, &X86::VR256RegClass);
1118 addRegisterClass(MVT::v8i32, &X86::VR256RegClass);
1119 addRegisterClass(MVT::v8f32, &X86::VR256RegClass);
1120 addRegisterClass(MVT::v4i64, &X86::VR256RegClass);
1121 addRegisterClass(MVT::v4f64, &X86::VR256RegClass);
1123 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
1124 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
1125 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
1127 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
1128 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
1129 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
1130 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
1131 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
1132 setOperationAction(ISD::FFLOOR, MVT::v8f32, Legal);
1133 setOperationAction(ISD::FCEIL, MVT::v8f32, Legal);
1134 setOperationAction(ISD::FTRUNC, MVT::v8f32, Legal);
1135 setOperationAction(ISD::FRINT, MVT::v8f32, Legal);
1136 setOperationAction(ISD::FNEARBYINT, MVT::v8f32, Legal);
1137 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
1138 setOperationAction(ISD::FABS, MVT::v8f32, Custom);
1140 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
1141 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
1142 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
1143 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
1144 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
1145 setOperationAction(ISD::FFLOOR, MVT::v4f64, Legal);
1146 setOperationAction(ISD::FCEIL, MVT::v4f64, Legal);
1147 setOperationAction(ISD::FTRUNC, MVT::v4f64, Legal);
1148 setOperationAction(ISD::FRINT, MVT::v4f64, Legal);
1149 setOperationAction(ISD::FNEARBYINT, MVT::v4f64, Legal);
1150 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
1151 setOperationAction(ISD::FABS, MVT::v4f64, Custom);
1153 setOperationAction(ISD::TRUNCATE, MVT::v8i16, Custom);
1154 setOperationAction(ISD::TRUNCATE, MVT::v4i32, Custom);
1156 setOperationAction(ISD::FP_TO_SINT, MVT::v8i16, Custom);
1158 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1159 setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Promote);
1160 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1161 setOperationAction(ISD::FP_ROUND, MVT::v4f32, Legal);
1163 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i32, Custom);
1164 setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Custom);
1165 setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Custom);
1167 setLoadExtAction(ISD::EXTLOAD, MVT::v4f32, Legal);
1169 setOperationAction(ISD::SRL, MVT::v16i16, Custom);
1170 setOperationAction(ISD::SRL, MVT::v32i8, Custom);
1172 setOperationAction(ISD::SHL, MVT::v16i16, Custom);
1173 setOperationAction(ISD::SHL, MVT::v32i8, Custom);
1175 setOperationAction(ISD::SRA, MVT::v16i16, Custom);
1176 setOperationAction(ISD::SRA, MVT::v32i8, Custom);
1178 setOperationAction(ISD::SDIV, MVT::v16i16, Custom);
1180 setOperationAction(ISD::SETCC, MVT::v32i8, Custom);
1181 setOperationAction(ISD::SETCC, MVT::v16i16, Custom);
1182 setOperationAction(ISD::SETCC, MVT::v8i32, Custom);
1183 setOperationAction(ISD::SETCC, MVT::v4i64, Custom);
1185 setOperationAction(ISD::SELECT, MVT::v4f64, Custom);
1186 setOperationAction(ISD::SELECT, MVT::v4i64, Custom);
1187 setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
1189 setOperationAction(ISD::VSELECT, MVT::v4f64, Legal);
1190 setOperationAction(ISD::VSELECT, MVT::v4i64, Legal);
1191 setOperationAction(ISD::VSELECT, MVT::v8i32, Legal);
1192 setOperationAction(ISD::VSELECT, MVT::v8f32, Legal);
1194 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i64, Custom);
1195 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i32, Custom);
1196 setOperationAction(ISD::ZERO_EXTEND, MVT::v4i64, Custom);
1197 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i32, Custom);
1198 setOperationAction(ISD::ANY_EXTEND, MVT::v4i64, Custom);
1199 setOperationAction(ISD::ANY_EXTEND, MVT::v8i32, Custom);
1201 if (Subtarget->hasFMA() || Subtarget->hasFMA4()) {
1202 setOperationAction(ISD::FMA, MVT::v8f32, Legal);
1203 setOperationAction(ISD::FMA, MVT::v4f64, Legal);
1204 setOperationAction(ISD::FMA, MVT::v4f32, Legal);
1205 setOperationAction(ISD::FMA, MVT::v2f64, Legal);
1206 setOperationAction(ISD::FMA, MVT::f32, Legal);
1207 setOperationAction(ISD::FMA, MVT::f64, Legal);
1210 if (Subtarget->hasInt256()) {
1211 setOperationAction(ISD::ADD, MVT::v4i64, Legal);
1212 setOperationAction(ISD::ADD, MVT::v8i32, Legal);
1213 setOperationAction(ISD::ADD, MVT::v16i16, Legal);
1214 setOperationAction(ISD::ADD, MVT::v32i8, Legal);
1216 setOperationAction(ISD::SUB, MVT::v4i64, Legal);
1217 setOperationAction(ISD::SUB, MVT::v8i32, Legal);
1218 setOperationAction(ISD::SUB, MVT::v16i16, Legal);
1219 setOperationAction(ISD::SUB, MVT::v32i8, Legal);
1221 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1222 setOperationAction(ISD::MUL, MVT::v8i32, Legal);
1223 setOperationAction(ISD::MUL, MVT::v16i16, Legal);
1224 // Don't lower v32i8 because there is no 128-bit byte mul
1226 setOperationAction(ISD::VSELECT, MVT::v32i8, Legal);
1228 setOperationAction(ISD::SDIV, MVT::v8i32, Custom);
1230 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
1231 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
1232 setOperationAction(ISD::ADD, MVT::v16i16, Custom);
1233 setOperationAction(ISD::ADD, MVT::v32i8, Custom);
1235 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
1236 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
1237 setOperationAction(ISD::SUB, MVT::v16i16, Custom);
1238 setOperationAction(ISD::SUB, MVT::v32i8, Custom);
1240 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1241 setOperationAction(ISD::MUL, MVT::v8i32, Custom);
1242 setOperationAction(ISD::MUL, MVT::v16i16, Custom);
1243 // Don't lower v32i8 because there is no 128-bit byte mul
1246 // In the customized shift lowering, the legal cases in AVX2 will be
1248 setOperationAction(ISD::SRL, MVT::v4i64, Custom);
1249 setOperationAction(ISD::SRL, MVT::v8i32, Custom);
1251 setOperationAction(ISD::SHL, MVT::v4i64, Custom);
1252 setOperationAction(ISD::SHL, MVT::v8i32, Custom);
1254 setOperationAction(ISD::SRA, MVT::v8i32, Custom);
1256 // Custom lower several nodes for 256-bit types.
1257 for (int i = MVT::FIRST_VECTOR_VALUETYPE;
1258 i <= MVT::LAST_VECTOR_VALUETYPE; ++i) {
1259 MVT VT = (MVT::SimpleValueType)i;
1261 // Extract subvector is special because the value type
1262 // (result) is 128-bit but the source is 256-bit wide.
1263 if (VT.is128BitVector())
1264 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1266 // Do not attempt to custom lower other non-256-bit vectors
1267 if (!VT.is256BitVector())
1270 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1271 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1272 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1273 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1274 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1275 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1276 setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
1279 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
1280 for (int i = MVT::v32i8; i != MVT::v4i64; ++i) {
1281 MVT VT = (MVT::SimpleValueType)i;
1283 // Do not attempt to promote non-256-bit vectors
1284 if (!VT.is256BitVector())
1287 setOperationAction(ISD::AND, VT, Promote);
1288 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
1289 setOperationAction(ISD::OR, VT, Promote);
1290 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
1291 setOperationAction(ISD::XOR, VT, Promote);
1292 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
1293 setOperationAction(ISD::LOAD, VT, Promote);
1294 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
1295 setOperationAction(ISD::SELECT, VT, Promote);
1296 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
1300 if (!TM.Options.UseSoftFloat && Subtarget->hasAVX512()) {
1301 addRegisterClass(MVT::v16i32, &X86::VR512RegClass);
1302 addRegisterClass(MVT::v16f32, &X86::VR512RegClass);
1303 addRegisterClass(MVT::v8i64, &X86::VR512RegClass);
1304 addRegisterClass(MVT::v8f64, &X86::VR512RegClass);
1306 addRegisterClass(MVT::v8i1, &X86::VK8RegClass);
1307 addRegisterClass(MVT::v16i1, &X86::VK16RegClass);
1309 setLoadExtAction(ISD::EXTLOAD, MVT::v8f32, Legal);
1310 setOperationAction(ISD::LOAD, MVT::v16f32, Legal);
1311 setOperationAction(ISD::LOAD, MVT::v8f64, Legal);
1312 setOperationAction(ISD::LOAD, MVT::v8i64, Legal);
1313 setOperationAction(ISD::LOAD, MVT::v16i32, Legal);
1314 setOperationAction(ISD::LOAD, MVT::v16i1, Legal);
1316 setOperationAction(ISD::FADD, MVT::v16f32, Legal);
1317 setOperationAction(ISD::FSUB, MVT::v16f32, Legal);
1318 setOperationAction(ISD::FMUL, MVT::v16f32, Legal);
1319 setOperationAction(ISD::FDIV, MVT::v16f32, Legal);
1320 setOperationAction(ISD::FSQRT, MVT::v16f32, Legal);
1321 setOperationAction(ISD::FNEG, MVT::v16f32, Custom);
1323 setOperationAction(ISD::FADD, MVT::v8f64, Legal);
1324 setOperationAction(ISD::FSUB, MVT::v8f64, Legal);
1325 setOperationAction(ISD::FMUL, MVT::v8f64, Legal);
1326 setOperationAction(ISD::FDIV, MVT::v8f64, Legal);
1327 setOperationAction(ISD::FSQRT, MVT::v8f64, Legal);
1328 setOperationAction(ISD::FNEG, MVT::v8f64, Custom);
1329 setOperationAction(ISD::FMA, MVT::v8f64, Legal);
1330 setOperationAction(ISD::FMA, MVT::v16f32, Legal);
1331 setOperationAction(ISD::SDIV, MVT::v16i32, Custom);
1334 setOperationAction(ISD::FP_TO_SINT, MVT::v16i32, Legal);
1335 setOperationAction(ISD::FP_TO_UINT, MVT::v16i32, Legal);
1336 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1337 setOperationAction(ISD::SINT_TO_FP, MVT::v16i32, Legal);
1338 setOperationAction(ISD::UINT_TO_FP, MVT::v16i32, Legal);
1339 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1340 setOperationAction(ISD::FP_ROUND, MVT::v8f32, Legal);
1341 setOperationAction(ISD::FP_EXTEND, MVT::v8f32, Legal);
1343 setOperationAction(ISD::TRUNCATE, MVT::i1, Legal);
1344 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1345 setOperationAction(ISD::TRUNCATE, MVT::v8i32, Custom);
1346 setOperationAction(ISD::TRUNCATE, MVT::v8i1, Custom);
1347 setOperationAction(ISD::TRUNCATE, MVT::v16i1, Custom);
1348 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i32, Custom);
1349 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i64, Custom);
1350 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i32, Custom);
1351 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i64, Custom);
1352 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i8, Custom);
1353 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i16, Custom);
1354 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1356 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f64, Custom);
1357 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i64, Custom);
1358 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16f32, Custom);
1359 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i32, Custom);
1360 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i1, Custom);
1362 setOperationAction(ISD::SETCC, MVT::v16i1, Custom);
1363 setOperationAction(ISD::SETCC, MVT::v8i1, Custom);
1365 setOperationAction(ISD::MUL, MVT::v8i64, Custom);
1367 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i1, Custom);
1368 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i1, Custom);
1369 setOperationAction(ISD::SELECT, MVT::v8f64, Custom);
1370 setOperationAction(ISD::SELECT, MVT::v8i64, Custom);
1371 setOperationAction(ISD::SELECT, MVT::v16f32, Custom);
1373 setOperationAction(ISD::ADD, MVT::v8i64, Legal);
1374 setOperationAction(ISD::ADD, MVT::v16i32, Legal);
1376 setOperationAction(ISD::SUB, MVT::v8i64, Legal);
1377 setOperationAction(ISD::SUB, MVT::v16i32, Legal);
1379 setOperationAction(ISD::MUL, MVT::v16i32, Legal);
1381 setOperationAction(ISD::SRL, MVT::v8i64, Custom);
1382 setOperationAction(ISD::SRL, MVT::v16i32, Custom);
1384 setOperationAction(ISD::SHL, MVT::v8i64, Custom);
1385 setOperationAction(ISD::SHL, MVT::v16i32, Custom);
1387 setOperationAction(ISD::SRA, MVT::v8i64, Custom);
1388 setOperationAction(ISD::SRA, MVT::v16i32, Custom);
1390 setOperationAction(ISD::AND, MVT::v8i64, Legal);
1391 setOperationAction(ISD::OR, MVT::v8i64, Legal);
1392 setOperationAction(ISD::XOR, MVT::v8i64, Legal);
1393 setOperationAction(ISD::AND, MVT::v16i32, Legal);
1394 setOperationAction(ISD::OR, MVT::v16i32, Legal);
1395 setOperationAction(ISD::XOR, MVT::v16i32, Legal);
1397 // Custom lower several nodes.
1398 for (int i = MVT::FIRST_VECTOR_VALUETYPE;
1399 i <= MVT::LAST_VECTOR_VALUETYPE; ++i) {
1400 MVT VT = (MVT::SimpleValueType)i;
1402 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1403 // Extract subvector is special because the value type
1404 // (result) is 256/128-bit but the source is 512-bit wide.
1405 if (VT.is128BitVector() || VT.is256BitVector())
1406 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1408 if (VT.getVectorElementType() == MVT::i1)
1409 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
1411 // Do not attempt to custom lower other non-512-bit vectors
1412 if (!VT.is512BitVector())
1415 if ( EltSize >= 32) {
1416 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1417 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1418 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1419 setOperationAction(ISD::VSELECT, VT, Legal);
1420 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1421 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1422 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1425 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1426 MVT VT = (MVT::SimpleValueType)i;
1428 // Do not attempt to promote non-256-bit vectors
1429 if (!VT.is512BitVector())
1432 setOperationAction(ISD::SELECT, VT, Promote);
1433 AddPromotedToType (ISD::SELECT, VT, MVT::v8i64);
1437 // SIGN_EXTEND_INREGs are evaluated by the extend type. Handle the expansion
1438 // of this type with custom code.
1439 for (int VT = MVT::FIRST_VECTOR_VALUETYPE;
1440 VT != MVT::LAST_VECTOR_VALUETYPE; VT++) {
1441 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,
1445 // We want to custom lower some of our intrinsics.
1446 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1447 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
1449 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1450 // handle type legalization for these operations here.
1452 // FIXME: We really should do custom legalization for addition and
1453 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1454 // than generic legalization for 64-bit multiplication-with-overflow, though.
1455 for (unsigned i = 0, e = 3+Subtarget->is64Bit(); i != e; ++i) {
1456 // Add/Sub/Mul with overflow operations are custom lowered.
1458 setOperationAction(ISD::SADDO, VT, Custom);
1459 setOperationAction(ISD::UADDO, VT, Custom);
1460 setOperationAction(ISD::SSUBO, VT, Custom);
1461 setOperationAction(ISD::USUBO, VT, Custom);
1462 setOperationAction(ISD::SMULO, VT, Custom);
1463 setOperationAction(ISD::UMULO, VT, Custom);
1466 // There are no 8-bit 3-address imul/mul instructions
1467 setOperationAction(ISD::SMULO, MVT::i8, Expand);
1468 setOperationAction(ISD::UMULO, MVT::i8, Expand);
1470 if (!Subtarget->is64Bit()) {
1471 // These libcalls are not available in 32-bit.
1472 setLibcallName(RTLIB::SHL_I128, 0);
1473 setLibcallName(RTLIB::SRL_I128, 0);
1474 setLibcallName(RTLIB::SRA_I128, 0);
1477 // Combine sin / cos into one node or libcall if possible.
1478 if (Subtarget->hasSinCos()) {
1479 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
1480 setLibcallName(RTLIB::SINCOS_F64, "sincos");
1481 if (Subtarget->isTargetDarwin()) {
1482 // For MacOSX, we don't want to the normal expansion of a libcall to
1483 // sincos. We want to issue a libcall to __sincos_stret to avoid memory
1485 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
1486 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
1490 // We have target-specific dag combine patterns for the following nodes:
1491 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1492 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1493 setTargetDAGCombine(ISD::VSELECT);
1494 setTargetDAGCombine(ISD::SELECT);
1495 setTargetDAGCombine(ISD::SHL);
1496 setTargetDAGCombine(ISD::SRA);
1497 setTargetDAGCombine(ISD::SRL);
1498 setTargetDAGCombine(ISD::OR);
1499 setTargetDAGCombine(ISD::AND);
1500 setTargetDAGCombine(ISD::ADD);
1501 setTargetDAGCombine(ISD::FADD);
1502 setTargetDAGCombine(ISD::FSUB);
1503 setTargetDAGCombine(ISD::FMA);
1504 setTargetDAGCombine(ISD::SUB);
1505 setTargetDAGCombine(ISD::LOAD);
1506 setTargetDAGCombine(ISD::STORE);
1507 setTargetDAGCombine(ISD::ZERO_EXTEND);
1508 setTargetDAGCombine(ISD::ANY_EXTEND);
1509 setTargetDAGCombine(ISD::SIGN_EXTEND);
1510 setTargetDAGCombine(ISD::SIGN_EXTEND_INREG);
1511 setTargetDAGCombine(ISD::TRUNCATE);
1512 setTargetDAGCombine(ISD::SINT_TO_FP);
1513 setTargetDAGCombine(ISD::SETCC);
1514 if (Subtarget->is64Bit())
1515 setTargetDAGCombine(ISD::MUL);
1516 setTargetDAGCombine(ISD::XOR);
1518 computeRegisterProperties();
1520 // On Darwin, -Os means optimize for size without hurting performance,
1521 // do not reduce the limit.
1522 MaxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1523 MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 16 : 8;
1524 MaxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1525 MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1526 MaxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1527 MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1528 setPrefLoopAlignment(4); // 2^4 bytes.
1530 // Predictable cmov don't hurt on atom because it's in-order.
1531 PredictableSelectIsExpensive = !Subtarget->isAtom();
1533 setPrefFunctionAlignment(4); // 2^4 bytes.
1536 EVT X86TargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
1537 if (!VT.isVector()) return MVT::i8;
1538 return VT.changeVectorElementTypeToInteger();
1541 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1542 /// the desired ByVal argument alignment.
1543 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1546 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1547 if (VTy->getBitWidth() == 128)
1549 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1550 unsigned EltAlign = 0;
1551 getMaxByValAlign(ATy->getElementType(), EltAlign);
1552 if (EltAlign > MaxAlign)
1553 MaxAlign = EltAlign;
1554 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1555 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1556 unsigned EltAlign = 0;
1557 getMaxByValAlign(STy->getElementType(i), EltAlign);
1558 if (EltAlign > MaxAlign)
1559 MaxAlign = EltAlign;
1566 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1567 /// function arguments in the caller parameter area. For X86, aggregates
1568 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1569 /// are at 4-byte boundaries.
1570 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty) const {
1571 if (Subtarget->is64Bit()) {
1572 // Max of 8 and alignment of type.
1573 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1580 if (Subtarget->hasSSE1())
1581 getMaxByValAlign(Ty, Align);
1585 /// getOptimalMemOpType - Returns the target specific optimal type for load
1586 /// and store operations as a result of memset, memcpy, and memmove
1587 /// lowering. If DstAlign is zero that means it's safe to destination
1588 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1589 /// means there isn't a need to check it against alignment requirement,
1590 /// probably because the source does not need to be loaded. If 'IsMemset' is
1591 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
1592 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
1593 /// source is constant so it does not need to be loaded.
1594 /// It returns EVT::Other if the type should be determined using generic
1595 /// target-independent logic.
1597 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1598 unsigned DstAlign, unsigned SrcAlign,
1599 bool IsMemset, bool ZeroMemset,
1601 MachineFunction &MF) const {
1602 const Function *F = MF.getFunction();
1603 if ((!IsMemset || ZeroMemset) &&
1604 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
1605 Attribute::NoImplicitFloat)) {
1607 (Subtarget->isUnalignedMemAccessFast() ||
1608 ((DstAlign == 0 || DstAlign >= 16) &&
1609 (SrcAlign == 0 || SrcAlign >= 16)))) {
1611 if (Subtarget->hasInt256())
1613 if (Subtarget->hasFp256())
1616 if (Subtarget->hasSSE2())
1618 if (Subtarget->hasSSE1())
1620 } else if (!MemcpyStrSrc && Size >= 8 &&
1621 !Subtarget->is64Bit() &&
1622 Subtarget->hasSSE2()) {
1623 // Do not use f64 to lower memcpy if source is string constant. It's
1624 // better to use i32 to avoid the loads.
1628 if (Subtarget->is64Bit() && Size >= 8)
1633 bool X86TargetLowering::isSafeMemOpType(MVT VT) const {
1635 return X86ScalarSSEf32;
1636 else if (VT == MVT::f64)
1637 return X86ScalarSSEf64;
1642 X86TargetLowering::allowsUnalignedMemoryAccesses(EVT VT, bool *Fast) const {
1644 *Fast = Subtarget->isUnalignedMemAccessFast();
1648 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1649 /// current function. The returned value is a member of the
1650 /// MachineJumpTableInfo::JTEntryKind enum.
1651 unsigned X86TargetLowering::getJumpTableEncoding() const {
1652 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1654 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1655 Subtarget->isPICStyleGOT())
1656 return MachineJumpTableInfo::EK_Custom32;
1658 // Otherwise, use the normal jump table encoding heuristics.
1659 return TargetLowering::getJumpTableEncoding();
1663 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1664 const MachineBasicBlock *MBB,
1665 unsigned uid,MCContext &Ctx) const{
1666 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1667 Subtarget->isPICStyleGOT());
1668 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1670 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1671 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1674 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1676 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1677 SelectionDAG &DAG) const {
1678 if (!Subtarget->is64Bit())
1679 // This doesn't have SDLoc associated with it, but is not really the
1680 // same as a Register.
1681 return DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), getPointerTy());
1685 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1686 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1688 const MCExpr *X86TargetLowering::
1689 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1690 MCContext &Ctx) const {
1691 // X86-64 uses RIP relative addressing based on the jump table label.
1692 if (Subtarget->isPICStyleRIPRel())
1693 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1695 // Otherwise, the reference is relative to the PIC base.
1696 return MCSymbolRefExpr::Create(MF->getPICBaseSymbol(), Ctx);
1699 // FIXME: Why this routine is here? Move to RegInfo!
1700 std::pair<const TargetRegisterClass*, uint8_t>
1701 X86TargetLowering::findRepresentativeClass(MVT VT) const{
1702 const TargetRegisterClass *RRC = 0;
1704 switch (VT.SimpleTy) {
1706 return TargetLowering::findRepresentativeClass(VT);
1707 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
1708 RRC = Subtarget->is64Bit() ?
1709 (const TargetRegisterClass*)&X86::GR64RegClass :
1710 (const TargetRegisterClass*)&X86::GR32RegClass;
1713 RRC = &X86::VR64RegClass;
1715 case MVT::f32: case MVT::f64:
1716 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1717 case MVT::v4f32: case MVT::v2f64:
1718 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
1720 RRC = &X86::VR128RegClass;
1723 return std::make_pair(RRC, Cost);
1726 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1727 unsigned &Offset) const {
1728 if (!Subtarget->isTargetLinux())
1731 if (Subtarget->is64Bit()) {
1732 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1734 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1746 //===----------------------------------------------------------------------===//
1747 // Return Value Calling Convention Implementation
1748 //===----------------------------------------------------------------------===//
1750 #include "X86GenCallingConv.inc"
1753 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv,
1754 MachineFunction &MF, bool isVarArg,
1755 const SmallVectorImpl<ISD::OutputArg> &Outs,
1756 LLVMContext &Context) const {
1757 SmallVector<CCValAssign, 16> RVLocs;
1758 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
1760 return CCInfo.CheckReturn(Outs, RetCC_X86);
1764 X86TargetLowering::LowerReturn(SDValue Chain,
1765 CallingConv::ID CallConv, bool isVarArg,
1766 const SmallVectorImpl<ISD::OutputArg> &Outs,
1767 const SmallVectorImpl<SDValue> &OutVals,
1768 SDLoc dl, SelectionDAG &DAG) const {
1769 MachineFunction &MF = DAG.getMachineFunction();
1770 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1772 SmallVector<CCValAssign, 16> RVLocs;
1773 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
1774 RVLocs, *DAG.getContext());
1775 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1778 SmallVector<SDValue, 6> RetOps;
1779 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1780 // Operand #1 = Bytes To Pop
1781 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
1784 // Copy the result values into the output registers.
1785 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1786 CCValAssign &VA = RVLocs[i];
1787 assert(VA.isRegLoc() && "Can only return in registers!");
1788 SDValue ValToCopy = OutVals[i];
1789 EVT ValVT = ValToCopy.getValueType();
1791 // Promote values to the appropriate types
1792 if (VA.getLocInfo() == CCValAssign::SExt)
1793 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
1794 else if (VA.getLocInfo() == CCValAssign::ZExt)
1795 ValToCopy = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), ValToCopy);
1796 else if (VA.getLocInfo() == CCValAssign::AExt)
1797 ValToCopy = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), ValToCopy);
1798 else if (VA.getLocInfo() == CCValAssign::BCvt)
1799 ValToCopy = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), ValToCopy);
1801 // If this is x86-64, and we disabled SSE, we can't return FP values,
1802 // or SSE or MMX vectors.
1803 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
1804 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
1805 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
1806 report_fatal_error("SSE register return with SSE disabled");
1808 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
1809 // llvm-gcc has never done it right and no one has noticed, so this
1810 // should be OK for now.
1811 if (ValVT == MVT::f64 &&
1812 (Subtarget->is64Bit() && !Subtarget->hasSSE2()))
1813 report_fatal_error("SSE2 register return with SSE2 disabled");
1815 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1816 // the RET instruction and handled by the FP Stackifier.
1817 if (VA.getLocReg() == X86::ST0 ||
1818 VA.getLocReg() == X86::ST1) {
1819 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1820 // change the value to the FP stack register class.
1821 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1822 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1823 RetOps.push_back(ValToCopy);
1824 // Don't emit a copytoreg.
1828 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1829 // which is returned in RAX / RDX.
1830 if (Subtarget->is64Bit()) {
1831 if (ValVT == MVT::x86mmx) {
1832 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1833 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ValToCopy);
1834 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
1836 // If we don't have SSE2 available, convert to v4f32 so the generated
1837 // register is legal.
1838 if (!Subtarget->hasSSE2())
1839 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32,ValToCopy);
1844 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1845 Flag = Chain.getValue(1);
1846 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
1849 // The x86-64 ABIs require that for returning structs by value we copy
1850 // the sret argument into %rax/%eax (depending on ABI) for the return.
1851 // Win32 requires us to put the sret argument to %eax as well.
1852 // We saved the argument into a virtual register in the entry block,
1853 // so now we copy the value out and into %rax/%eax.
1854 if (DAG.getMachineFunction().getFunction()->hasStructRetAttr() &&
1855 (Subtarget->is64Bit() || Subtarget->isTargetWindows())) {
1856 MachineFunction &MF = DAG.getMachineFunction();
1857 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1858 unsigned Reg = FuncInfo->getSRetReturnReg();
1860 "SRetReturnReg should have been set in LowerFormalArguments().");
1861 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1864 = (Subtarget->is64Bit() && !Subtarget->isTarget64BitILP32()) ?
1865 X86::RAX : X86::EAX;
1866 Chain = DAG.getCopyToReg(Chain, dl, RetValReg, Val, Flag);
1867 Flag = Chain.getValue(1);
1869 // RAX/EAX now acts like a return value.
1870 RetOps.push_back(DAG.getRegister(RetValReg, getPointerTy()));
1873 RetOps[0] = Chain; // Update chain.
1875 // Add the flag if we have it.
1877 RetOps.push_back(Flag);
1879 return DAG.getNode(X86ISD::RET_FLAG, dl,
1880 MVT::Other, &RetOps[0], RetOps.size());
1883 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
1884 if (N->getNumValues() != 1)
1886 if (!N->hasNUsesOfValue(1, 0))
1889 SDValue TCChain = Chain;
1890 SDNode *Copy = *N->use_begin();
1891 if (Copy->getOpcode() == ISD::CopyToReg) {
1892 // If the copy has a glue operand, we conservatively assume it isn't safe to
1893 // perform a tail call.
1894 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
1896 TCChain = Copy->getOperand(0);
1897 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
1900 bool HasRet = false;
1901 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
1903 if (UI->getOpcode() != X86ISD::RET_FLAG)
1916 X86TargetLowering::getTypeForExtArgOrReturn(MVT VT,
1917 ISD::NodeType ExtendKind) const {
1919 // TODO: Is this also valid on 32-bit?
1920 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
1921 ReturnMVT = MVT::i8;
1923 ReturnMVT = MVT::i32;
1925 MVT MinVT = getRegisterType(ReturnMVT);
1926 return VT.bitsLT(MinVT) ? MinVT : VT;
1929 /// LowerCallResult - Lower the result values of a call into the
1930 /// appropriate copies out of appropriate physical registers.
1933 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1934 CallingConv::ID CallConv, bool isVarArg,
1935 const SmallVectorImpl<ISD::InputArg> &Ins,
1936 SDLoc dl, SelectionDAG &DAG,
1937 SmallVectorImpl<SDValue> &InVals) const {
1939 // Assign locations to each value returned by this call.
1940 SmallVector<CCValAssign, 16> RVLocs;
1941 bool Is64Bit = Subtarget->is64Bit();
1942 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1943 getTargetMachine(), RVLocs, *DAG.getContext());
1944 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1946 // Copy all of the result registers out of their specified physreg.
1947 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
1948 CCValAssign &VA = RVLocs[i];
1949 EVT CopyVT = VA.getValVT();
1951 // If this is x86-64, and we disabled SSE, we can't return FP values
1952 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1953 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1954 report_fatal_error("SSE register return with SSE disabled");
1959 // If this is a call to a function that returns an fp value on the floating
1960 // point stack, we must guarantee the value is popped from the stack, so
1961 // a CopyFromReg is not good enough - the copy instruction may be eliminated
1962 // if the return value is not used. We use the FpPOP_RETVAL instruction
1964 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1) {
1965 // If we prefer to use the value in xmm registers, copy it out as f80 and
1966 // use a truncate to move it from fp stack reg to xmm reg.
1967 if (isScalarFPTypeInSSEReg(VA.getValVT())) CopyVT = MVT::f80;
1968 SDValue Ops[] = { Chain, InFlag };
1969 Chain = SDValue(DAG.getMachineNode(X86::FpPOP_RETVAL, dl, CopyVT,
1970 MVT::Other, MVT::Glue, Ops), 1);
1971 Val = Chain.getValue(0);
1973 // Round the f80 to the right size, which also moves it to the appropriate
1975 if (CopyVT != VA.getValVT())
1976 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1977 // This truncation won't change the value.
1978 DAG.getIntPtrConstant(1));
1980 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1981 CopyVT, InFlag).getValue(1);
1982 Val = Chain.getValue(0);
1984 InFlag = Chain.getValue(2);
1985 InVals.push_back(Val);
1991 //===----------------------------------------------------------------------===//
1992 // C & StdCall & Fast Calling Convention implementation
1993 //===----------------------------------------------------------------------===//
1994 // StdCall calling convention seems to be standard for many Windows' API
1995 // routines and around. It differs from C calling convention just a little:
1996 // callee should clean up the stack, not caller. Symbols should be also
1997 // decorated in some fancy way :) It doesn't support any vector arguments.
1998 // For info on fast calling convention see Fast Calling Convention (tail call)
1999 // implementation LowerX86_32FastCCCallTo.
2001 /// CallIsStructReturn - Determines whether a call uses struct return
2003 enum StructReturnType {
2008 static StructReturnType
2009 callIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
2011 return NotStructReturn;
2013 const ISD::ArgFlagsTy &Flags = Outs[0].Flags;
2014 if (!Flags.isSRet())
2015 return NotStructReturn;
2016 if (Flags.isInReg())
2017 return RegStructReturn;
2018 return StackStructReturn;
2021 /// ArgsAreStructReturn - Determines whether a function uses struct
2022 /// return semantics.
2023 static StructReturnType
2024 argsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
2026 return NotStructReturn;
2028 const ISD::ArgFlagsTy &Flags = Ins[0].Flags;
2029 if (!Flags.isSRet())
2030 return NotStructReturn;
2031 if (Flags.isInReg())
2032 return RegStructReturn;
2033 return StackStructReturn;
2036 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
2037 /// by "Src" to address "Dst" with size and alignment information specified by
2038 /// the specific parameter attribute. The copy will be passed as a byval
2039 /// function parameter.
2041 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2042 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2044 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
2046 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2047 /*isVolatile*/false, /*AlwaysInline=*/true,
2048 MachinePointerInfo(), MachinePointerInfo());
2051 /// IsTailCallConvention - Return true if the calling convention is one that
2052 /// supports tail call optimization.
2053 static bool IsTailCallConvention(CallingConv::ID CC) {
2054 return (CC == CallingConv::Fast || CC == CallingConv::GHC ||
2055 CC == CallingConv::HiPE);
2058 /// \brief Return true if the calling convention is a C calling convention.
2059 static bool IsCCallConvention(CallingConv::ID CC) {
2060 return (CC == CallingConv::C || CC == CallingConv::X86_64_Win64 ||
2061 CC == CallingConv::X86_64_SysV);
2064 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2065 if (!CI->isTailCall() || getTargetMachine().Options.DisableTailCalls)
2069 CallingConv::ID CalleeCC = CS.getCallingConv();
2070 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
2076 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
2077 /// a tailcall target by changing its ABI.
2078 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC,
2079 bool GuaranteedTailCallOpt) {
2080 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
2084 X86TargetLowering::LowerMemArgument(SDValue Chain,
2085 CallingConv::ID CallConv,
2086 const SmallVectorImpl<ISD::InputArg> &Ins,
2087 SDLoc dl, SelectionDAG &DAG,
2088 const CCValAssign &VA,
2089 MachineFrameInfo *MFI,
2091 // Create the nodes corresponding to a load from this parameter slot.
2092 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2093 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv,
2094 getTargetMachine().Options.GuaranteedTailCallOpt);
2095 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
2098 // If value is passed by pointer we have address passed instead of the value
2100 if (VA.getLocInfo() == CCValAssign::Indirect)
2101 ValVT = VA.getLocVT();
2103 ValVT = VA.getValVT();
2105 // FIXME: For now, all byval parameter objects are marked mutable. This can be
2106 // changed with more analysis.
2107 // In case of tail call optimization mark all arguments mutable. Since they
2108 // could be overwritten by lowering of arguments in case of a tail call.
2109 if (Flags.isByVal()) {
2110 unsigned Bytes = Flags.getByValSize();
2111 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2112 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
2113 return DAG.getFrameIndex(FI, getPointerTy());
2115 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2116 VA.getLocMemOffset(), isImmutable);
2117 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2118 return DAG.getLoad(ValVT, dl, Chain, FIN,
2119 MachinePointerInfo::getFixedStack(FI),
2120 false, false, false, 0);
2125 X86TargetLowering::LowerFormalArguments(SDValue Chain,
2126 CallingConv::ID CallConv,
2128 const SmallVectorImpl<ISD::InputArg> &Ins,
2131 SmallVectorImpl<SDValue> &InVals)
2133 MachineFunction &MF = DAG.getMachineFunction();
2134 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2136 const Function* Fn = MF.getFunction();
2137 if (Fn->hasExternalLinkage() &&
2138 Subtarget->isTargetCygMing() &&
2139 Fn->getName() == "main")
2140 FuncInfo->setForceFramePointer(true);
2142 MachineFrameInfo *MFI = MF.getFrameInfo();
2143 bool Is64Bit = Subtarget->is64Bit();
2144 bool IsWindows = Subtarget->isTargetWindows();
2145 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2147 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2148 "Var args not supported with calling convention fastcc, ghc or hipe");
2150 // Assign locations to all of the incoming arguments.
2151 SmallVector<CCValAssign, 16> ArgLocs;
2152 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
2153 ArgLocs, *DAG.getContext());
2155 // Allocate shadow area for Win64
2157 CCInfo.AllocateStack(32, 8);
2159 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
2161 unsigned LastVal = ~0U;
2163 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2164 CCValAssign &VA = ArgLocs[i];
2165 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
2167 assert(VA.getValNo() != LastVal &&
2168 "Don't support value assigned to multiple locs yet");
2170 LastVal = VA.getValNo();
2172 if (VA.isRegLoc()) {
2173 EVT RegVT = VA.getLocVT();
2174 const TargetRegisterClass *RC;
2175 if (RegVT == MVT::i32)
2176 RC = &X86::GR32RegClass;
2177 else if (Is64Bit && RegVT == MVT::i64)
2178 RC = &X86::GR64RegClass;
2179 else if (RegVT == MVT::f32)
2180 RC = &X86::FR32RegClass;
2181 else if (RegVT == MVT::f64)
2182 RC = &X86::FR64RegClass;
2183 else if (RegVT.is512BitVector())
2184 RC = &X86::VR512RegClass;
2185 else if (RegVT.is256BitVector())
2186 RC = &X86::VR256RegClass;
2187 else if (RegVT.is128BitVector())
2188 RC = &X86::VR128RegClass;
2189 else if (RegVT == MVT::x86mmx)
2190 RC = &X86::VR64RegClass;
2191 else if (RegVT == MVT::v8i1)
2192 RC = &X86::VK8RegClass;
2193 else if (RegVT == MVT::v16i1)
2194 RC = &X86::VK16RegClass;
2196 llvm_unreachable("Unknown argument type!");
2198 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2199 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2201 // If this is an 8 or 16-bit value, it is really passed promoted to 32
2202 // bits. Insert an assert[sz]ext to capture this, then truncate to the
2204 if (VA.getLocInfo() == CCValAssign::SExt)
2205 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2206 DAG.getValueType(VA.getValVT()));
2207 else if (VA.getLocInfo() == CCValAssign::ZExt)
2208 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2209 DAG.getValueType(VA.getValVT()));
2210 else if (VA.getLocInfo() == CCValAssign::BCvt)
2211 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2213 if (VA.isExtInLoc()) {
2214 // Handle MMX values passed in XMM regs.
2215 if (RegVT.isVector())
2216 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(), ArgValue);
2218 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2221 assert(VA.isMemLoc());
2222 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
2225 // If value is passed via pointer - do a load.
2226 if (VA.getLocInfo() == CCValAssign::Indirect)
2227 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
2228 MachinePointerInfo(), false, false, false, 0);
2230 InVals.push_back(ArgValue);
2233 // The x86-64 ABIs require that for returning structs by value we copy
2234 // the sret argument into %rax/%eax (depending on ABI) for the return.
2235 // Win32 requires us to put the sret argument to %eax as well.
2236 // Save the argument into a virtual register so that we can access it
2237 // from the return points.
2238 if (MF.getFunction()->hasStructRetAttr() &&
2239 (Subtarget->is64Bit() || Subtarget->isTargetWindows())) {
2240 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2241 unsigned Reg = FuncInfo->getSRetReturnReg();
2243 MVT PtrTy = getPointerTy();
2244 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrTy));
2245 FuncInfo->setSRetReturnReg(Reg);
2247 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
2248 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
2251 unsigned StackSize = CCInfo.getNextStackOffset();
2252 // Align stack specially for tail calls.
2253 if (FuncIsMadeTailCallSafe(CallConv,
2254 MF.getTarget().Options.GuaranteedTailCallOpt))
2255 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
2257 // If the function takes variable number of arguments, make a frame index for
2258 // the start of the first vararg value... for expansion of llvm.va_start.
2260 if (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
2261 CallConv != CallingConv::X86_ThisCall)) {
2262 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,true));
2265 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
2267 // FIXME: We should really autogenerate these arrays
2268 static const uint16_t GPR64ArgRegsWin64[] = {
2269 X86::RCX, X86::RDX, X86::R8, X86::R9
2271 static const uint16_t GPR64ArgRegs64Bit[] = {
2272 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
2274 static const uint16_t XMMArgRegs64Bit[] = {
2275 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2276 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2278 const uint16_t *GPR64ArgRegs;
2279 unsigned NumXMMRegs = 0;
2282 // The XMM registers which might contain var arg parameters are shadowed
2283 // in their paired GPR. So we only need to save the GPR to their home
2285 TotalNumIntRegs = 4;
2286 GPR64ArgRegs = GPR64ArgRegsWin64;
2288 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
2289 GPR64ArgRegs = GPR64ArgRegs64Bit;
2291 NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs64Bit,
2294 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
2297 bool NoImplicitFloatOps = Fn->getAttributes().
2298 hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat);
2299 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
2300 "SSE register cannot be used when SSE is disabled!");
2301 assert(!(NumXMMRegs && MF.getTarget().Options.UseSoftFloat &&
2302 NoImplicitFloatOps) &&
2303 "SSE register cannot be used when SSE is disabled!");
2304 if (MF.getTarget().Options.UseSoftFloat || NoImplicitFloatOps ||
2305 !Subtarget->hasSSE1())
2306 // Kernel mode asks for SSE to be disabled, so don't push them
2308 TotalNumXMMRegs = 0;
2311 const TargetFrameLowering &TFI = *getTargetMachine().getFrameLowering();
2312 // Get to the caller-allocated home save location. Add 8 to account
2313 // for the return address.
2314 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2315 FuncInfo->setRegSaveFrameIndex(
2316 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
2317 // Fixup to set vararg frame on shadow area (4 x i64).
2319 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
2321 // For X86-64, if there are vararg parameters that are passed via
2322 // registers, then we must store them to their spots on the stack so
2323 // they may be loaded by deferencing the result of va_next.
2324 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
2325 FuncInfo->setVarArgsFPOffset(TotalNumIntRegs * 8 + NumXMMRegs * 16);
2326 FuncInfo->setRegSaveFrameIndex(
2327 MFI->CreateStackObject(TotalNumIntRegs * 8 + TotalNumXMMRegs * 16, 16,
2331 // Store the integer parameter registers.
2332 SmallVector<SDValue, 8> MemOps;
2333 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
2335 unsigned Offset = FuncInfo->getVarArgsGPOffset();
2336 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
2337 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
2338 DAG.getIntPtrConstant(Offset));
2339 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
2340 &X86::GR64RegClass);
2341 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
2343 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2344 MachinePointerInfo::getFixedStack(
2345 FuncInfo->getRegSaveFrameIndex(), Offset),
2347 MemOps.push_back(Store);
2351 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
2352 // Now store the XMM (fp + vector) parameter registers.
2353 SmallVector<SDValue, 11> SaveXMMOps;
2354 SaveXMMOps.push_back(Chain);
2356 unsigned AL = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2357 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
2358 SaveXMMOps.push_back(ALVal);
2360 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2361 FuncInfo->getRegSaveFrameIndex()));
2362 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2363 FuncInfo->getVarArgsFPOffset()));
2365 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
2366 unsigned VReg = MF.addLiveIn(XMMArgRegs64Bit[NumXMMRegs],
2367 &X86::VR128RegClass);
2368 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
2369 SaveXMMOps.push_back(Val);
2371 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
2373 &SaveXMMOps[0], SaveXMMOps.size()));
2376 if (!MemOps.empty())
2377 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2378 &MemOps[0], MemOps.size());
2382 // Some CCs need callee pop.
2383 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
2384 MF.getTarget().Options.GuaranteedTailCallOpt)) {
2385 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
2387 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
2388 // If this is an sret function, the return should pop the hidden pointer.
2389 if (!Is64Bit && !IsTailCallConvention(CallConv) && !IsWindows &&
2390 argsAreStructReturn(Ins) == StackStructReturn)
2391 FuncInfo->setBytesToPopOnReturn(4);
2395 // RegSaveFrameIndex is X86-64 only.
2396 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
2397 if (CallConv == CallingConv::X86_FastCall ||
2398 CallConv == CallingConv::X86_ThisCall)
2399 // fastcc functions can't have varargs.
2400 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
2403 FuncInfo->setArgumentStackSize(StackSize);
2409 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
2410 SDValue StackPtr, SDValue Arg,
2411 SDLoc dl, SelectionDAG &DAG,
2412 const CCValAssign &VA,
2413 ISD::ArgFlagsTy Flags) const {
2414 unsigned LocMemOffset = VA.getLocMemOffset();
2415 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
2416 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
2417 if (Flags.isByVal())
2418 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
2420 return DAG.getStore(Chain, dl, Arg, PtrOff,
2421 MachinePointerInfo::getStack(LocMemOffset),
2425 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
2426 /// optimization is performed and it is required.
2428 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
2429 SDValue &OutRetAddr, SDValue Chain,
2430 bool IsTailCall, bool Is64Bit,
2431 int FPDiff, SDLoc dl) const {
2432 // Adjust the Return address stack slot.
2433 EVT VT = getPointerTy();
2434 OutRetAddr = getReturnAddressFrameIndex(DAG);
2436 // Load the "old" Return address.
2437 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
2438 false, false, false, 0);
2439 return SDValue(OutRetAddr.getNode(), 1);
2442 /// EmitTailCallStoreRetAddr - Emit a store of the return address if tail call
2443 /// optimization is performed and it is required (FPDiff!=0).
2445 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
2446 SDValue Chain, SDValue RetAddrFrIdx, EVT PtrVT,
2447 unsigned SlotSize, int FPDiff, SDLoc dl) {
2448 // Store the return address to the appropriate stack slot.
2449 if (!FPDiff) return Chain;
2450 // Calculate the new stack slot for the return address.
2451 int NewReturnAddrFI =
2452 MF.getFrameInfo()->CreateFixedObject(SlotSize, (int64_t)FPDiff - SlotSize,
2454 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, PtrVT);
2455 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
2456 MachinePointerInfo::getFixedStack(NewReturnAddrFI),
2462 X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2463 SmallVectorImpl<SDValue> &InVals) const {
2464 SelectionDAG &DAG = CLI.DAG;
2466 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
2467 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
2468 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
2469 SDValue Chain = CLI.Chain;
2470 SDValue Callee = CLI.Callee;
2471 CallingConv::ID CallConv = CLI.CallConv;
2472 bool &isTailCall = CLI.IsTailCall;
2473 bool isVarArg = CLI.IsVarArg;
2475 MachineFunction &MF = DAG.getMachineFunction();
2476 bool Is64Bit = Subtarget->is64Bit();
2477 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2478 bool IsWindows = Subtarget->isTargetWindows();
2479 StructReturnType SR = callIsStructReturn(Outs);
2480 bool IsSibcall = false;
2482 if (MF.getTarget().Options.DisableTailCalls)
2486 // Check if it's really possible to do a tail call.
2487 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
2488 isVarArg, SR != NotStructReturn,
2489 MF.getFunction()->hasStructRetAttr(), CLI.RetTy,
2490 Outs, OutVals, Ins, DAG);
2492 // Sibcalls are automatically detected tailcalls which do not require
2494 if (!MF.getTarget().Options.GuaranteedTailCallOpt && isTailCall)
2501 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2502 "Var args not supported with calling convention fastcc, ghc or hipe");
2504 // Analyze operands of the call, assigning locations to each operand.
2505 SmallVector<CCValAssign, 16> ArgLocs;
2506 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
2507 ArgLocs, *DAG.getContext());
2509 // Allocate shadow area for Win64
2511 CCInfo.AllocateStack(32, 8);
2513 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2515 // Get a count of how many bytes are to be pushed on the stack.
2516 unsigned NumBytes = CCInfo.getNextStackOffset();
2518 // This is a sibcall. The memory operands are available in caller's
2519 // own caller's stack.
2521 else if (getTargetMachine().Options.GuaranteedTailCallOpt &&
2522 IsTailCallConvention(CallConv))
2523 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
2526 if (isTailCall && !IsSibcall) {
2527 // Lower arguments at fp - stackoffset + fpdiff.
2528 X86MachineFunctionInfo *X86Info = MF.getInfo<X86MachineFunctionInfo>();
2529 unsigned NumBytesCallerPushed = X86Info->getBytesToPopOnReturn();
2531 FPDiff = NumBytesCallerPushed - NumBytes;
2533 // Set the delta of movement of the returnaddr stackslot.
2534 // But only set if delta is greater than previous delta.
2535 if (FPDiff < X86Info->getTCReturnAddrDelta())
2536 X86Info->setTCReturnAddrDelta(FPDiff);
2540 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true),
2543 SDValue RetAddrFrIdx;
2544 // Load return address for tail calls.
2545 if (isTailCall && FPDiff)
2546 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
2547 Is64Bit, FPDiff, dl);
2549 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2550 SmallVector<SDValue, 8> MemOpChains;
2553 // Walk the register/memloc assignments, inserting copies/loads. In the case
2554 // of tail call optimization arguments are handle later.
2555 const X86RegisterInfo *RegInfo =
2556 static_cast<const X86RegisterInfo*>(getTargetMachine().getRegisterInfo());
2557 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2558 CCValAssign &VA = ArgLocs[i];
2559 EVT RegVT = VA.getLocVT();
2560 SDValue Arg = OutVals[i];
2561 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2562 bool isByVal = Flags.isByVal();
2564 // Promote the value if needed.
2565 switch (VA.getLocInfo()) {
2566 default: llvm_unreachable("Unknown loc info!");
2567 case CCValAssign::Full: break;
2568 case CCValAssign::SExt:
2569 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
2571 case CCValAssign::ZExt:
2572 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
2574 case CCValAssign::AExt:
2575 if (RegVT.is128BitVector()) {
2576 // Special case: passing MMX values in XMM registers.
2577 Arg = DAG.getNode(ISD::BITCAST, dl, MVT::i64, Arg);
2578 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
2579 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
2581 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
2583 case CCValAssign::BCvt:
2584 Arg = DAG.getNode(ISD::BITCAST, dl, RegVT, Arg);
2586 case CCValAssign::Indirect: {
2587 // Store the argument.
2588 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
2589 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
2590 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
2591 MachinePointerInfo::getFixedStack(FI),
2598 if (VA.isRegLoc()) {
2599 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2600 if (isVarArg && IsWin64) {
2601 // Win64 ABI requires argument XMM reg to be copied to the corresponding
2602 // shadow reg if callee is a varargs function.
2603 unsigned ShadowReg = 0;
2604 switch (VA.getLocReg()) {
2605 case X86::XMM0: ShadowReg = X86::RCX; break;
2606 case X86::XMM1: ShadowReg = X86::RDX; break;
2607 case X86::XMM2: ShadowReg = X86::R8; break;
2608 case X86::XMM3: ShadowReg = X86::R9; break;
2611 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
2613 } else if (!IsSibcall && (!isTailCall || isByVal)) {
2614 assert(VA.isMemLoc());
2615 if (StackPtr.getNode() == 0)
2616 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
2618 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
2619 dl, DAG, VA, Flags));
2623 if (!MemOpChains.empty())
2624 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2625 &MemOpChains[0], MemOpChains.size());
2627 if (Subtarget->isPICStyleGOT()) {
2628 // ELF / PIC requires GOT in the EBX register before function calls via PLT
2631 RegsToPass.push_back(std::make_pair(unsigned(X86::EBX),
2632 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), getPointerTy())));
2634 // If we are tail calling and generating PIC/GOT style code load the
2635 // address of the callee into ECX. The value in ecx is used as target of
2636 // the tail jump. This is done to circumvent the ebx/callee-saved problem
2637 // for tail calls on PIC/GOT architectures. Normally we would just put the
2638 // address of GOT into ebx and then call target@PLT. But for tail calls
2639 // ebx would be restored (since ebx is callee saved) before jumping to the
2642 // Note: The actual moving to ECX is done further down.
2643 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2644 if (G && !G->getGlobal()->hasHiddenVisibility() &&
2645 !G->getGlobal()->hasProtectedVisibility())
2646 Callee = LowerGlobalAddress(Callee, DAG);
2647 else if (isa<ExternalSymbolSDNode>(Callee))
2648 Callee = LowerExternalSymbol(Callee, DAG);
2652 if (Is64Bit && isVarArg && !IsWin64) {
2653 // From AMD64 ABI document:
2654 // For calls that may call functions that use varargs or stdargs
2655 // (prototype-less calls or calls to functions containing ellipsis (...) in
2656 // the declaration) %al is used as hidden argument to specify the number
2657 // of SSE registers used. The contents of %al do not need to match exactly
2658 // the number of registers, but must be an ubound on the number of SSE
2659 // registers used and is in the range 0 - 8 inclusive.
2661 // Count the number of XMM registers allocated.
2662 static const uint16_t XMMArgRegs[] = {
2663 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2664 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2666 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
2667 assert((Subtarget->hasSSE1() || !NumXMMRegs)
2668 && "SSE registers cannot be used when SSE is disabled");
2670 RegsToPass.push_back(std::make_pair(unsigned(X86::AL),
2671 DAG.getConstant(NumXMMRegs, MVT::i8)));
2674 // For tail calls lower the arguments to the 'real' stack slot.
2676 // Force all the incoming stack arguments to be loaded from the stack
2677 // before any new outgoing arguments are stored to the stack, because the
2678 // outgoing stack slots may alias the incoming argument stack slots, and
2679 // the alias isn't otherwise explicit. This is slightly more conservative
2680 // than necessary, because it means that each store effectively depends
2681 // on every argument instead of just those arguments it would clobber.
2682 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2684 SmallVector<SDValue, 8> MemOpChains2;
2687 if (getTargetMachine().Options.GuaranteedTailCallOpt) {
2688 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2689 CCValAssign &VA = ArgLocs[i];
2692 assert(VA.isMemLoc());
2693 SDValue Arg = OutVals[i];
2694 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2695 // Create frame index.
2696 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2697 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2698 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2699 FIN = DAG.getFrameIndex(FI, getPointerTy());
2701 if (Flags.isByVal()) {
2702 // Copy relative to framepointer.
2703 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2704 if (StackPtr.getNode() == 0)
2705 StackPtr = DAG.getCopyFromReg(Chain, dl,
2706 RegInfo->getStackRegister(),
2708 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2710 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2714 // Store relative to framepointer.
2715 MemOpChains2.push_back(
2716 DAG.getStore(ArgChain, dl, Arg, FIN,
2717 MachinePointerInfo::getFixedStack(FI),
2723 if (!MemOpChains2.empty())
2724 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2725 &MemOpChains2[0], MemOpChains2.size());
2727 // Store the return address to the appropriate stack slot.
2728 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx,
2729 getPointerTy(), RegInfo->getSlotSize(),
2733 // Build a sequence of copy-to-reg nodes chained together with token chain
2734 // and flag operands which copy the outgoing args into registers.
2736 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2737 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2738 RegsToPass[i].second, InFlag);
2739 InFlag = Chain.getValue(1);
2742 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2743 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2744 // In the 64-bit large code model, we have to make all calls
2745 // through a register, since the call instruction's 32-bit
2746 // pc-relative offset may not be large enough to hold the whole
2748 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2749 // If the callee is a GlobalAddress node (quite common, every direct call
2750 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2753 // We should use extra load for direct calls to dllimported functions in
2755 const GlobalValue *GV = G->getGlobal();
2756 if (!GV->hasDLLImportLinkage()) {
2757 unsigned char OpFlags = 0;
2758 bool ExtraLoad = false;
2759 unsigned WrapperKind = ISD::DELETED_NODE;
2761 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2762 // external symbols most go through the PLT in PIC mode. If the symbol
2763 // has hidden or protected visibility, or if it is static or local, then
2764 // we don't need to use the PLT - we can directly call it.
2765 if (Subtarget->isTargetELF() &&
2766 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2767 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2768 OpFlags = X86II::MO_PLT;
2769 } else if (Subtarget->isPICStyleStubAny() &&
2770 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2771 (!Subtarget->getTargetTriple().isMacOSX() ||
2772 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
2773 // PC-relative references to external symbols should go through $stub,
2774 // unless we're building with the leopard linker or later, which
2775 // automatically synthesizes these stubs.
2776 OpFlags = X86II::MO_DARWIN_STUB;
2777 } else if (Subtarget->isPICStyleRIPRel() &&
2778 isa<Function>(GV) &&
2779 cast<Function>(GV)->getAttributes().
2780 hasAttribute(AttributeSet::FunctionIndex,
2781 Attribute::NonLazyBind)) {
2782 // If the function is marked as non-lazy, generate an indirect call
2783 // which loads from the GOT directly. This avoids runtime overhead
2784 // at the cost of eager binding (and one extra byte of encoding).
2785 OpFlags = X86II::MO_GOTPCREL;
2786 WrapperKind = X86ISD::WrapperRIP;
2790 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
2791 G->getOffset(), OpFlags);
2793 // Add a wrapper if needed.
2794 if (WrapperKind != ISD::DELETED_NODE)
2795 Callee = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Callee);
2796 // Add extra indirection if needed.
2798 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Callee,
2799 MachinePointerInfo::getGOT(),
2800 false, false, false, 0);
2802 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2803 unsigned char OpFlags = 0;
2805 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
2806 // external symbols should go through the PLT.
2807 if (Subtarget->isTargetELF() &&
2808 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2809 OpFlags = X86II::MO_PLT;
2810 } else if (Subtarget->isPICStyleStubAny() &&
2811 (!Subtarget->getTargetTriple().isMacOSX() ||
2812 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
2813 // PC-relative references to external symbols should go through $stub,
2814 // unless we're building with the leopard linker or later, which
2815 // automatically synthesizes these stubs.
2816 OpFlags = X86II::MO_DARWIN_STUB;
2819 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2823 // Returns a chain & a flag for retval copy to use.
2824 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2825 SmallVector<SDValue, 8> Ops;
2827 if (!IsSibcall && isTailCall) {
2828 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2829 DAG.getIntPtrConstant(0, true), InFlag, dl);
2830 InFlag = Chain.getValue(1);
2833 Ops.push_back(Chain);
2834 Ops.push_back(Callee);
2837 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2839 // Add argument registers to the end of the list so that they are known live
2841 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2842 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2843 RegsToPass[i].second.getValueType()));
2845 // Add a register mask operand representing the call-preserved registers.
2846 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2847 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
2848 assert(Mask && "Missing call preserved mask for calling convention");
2849 Ops.push_back(DAG.getRegisterMask(Mask));
2851 if (InFlag.getNode())
2852 Ops.push_back(InFlag);
2856 //// If this is the first return lowered for this function, add the regs
2857 //// to the liveout set for the function.
2858 // This isn't right, although it's probably harmless on x86; liveouts
2859 // should be computed from returns not tail calls. Consider a void
2860 // function making a tail call to a function returning int.
2861 return DAG.getNode(X86ISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
2864 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2865 InFlag = Chain.getValue(1);
2867 // Create the CALLSEQ_END node.
2868 unsigned NumBytesForCalleeToPush;
2869 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
2870 getTargetMachine().Options.GuaranteedTailCallOpt))
2871 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2872 else if (!Is64Bit && !IsTailCallConvention(CallConv) && !IsWindows &&
2873 SR == StackStructReturn)
2874 // If this is a call to a struct-return function, the callee
2875 // pops the hidden struct pointer, so we have to push it back.
2876 // This is common for Darwin/X86, Linux & Mingw32 targets.
2877 // For MSVC Win32 targets, the caller pops the hidden struct pointer.
2878 NumBytesForCalleeToPush = 4;
2880 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2882 // Returns a flag for retval copy to use.
2884 Chain = DAG.getCALLSEQ_END(Chain,
2885 DAG.getIntPtrConstant(NumBytes, true),
2886 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2889 InFlag = Chain.getValue(1);
2892 // Handle result values, copying them out of physregs into vregs that we
2894 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2895 Ins, dl, DAG, InVals);
2898 //===----------------------------------------------------------------------===//
2899 // Fast Calling Convention (tail call) implementation
2900 //===----------------------------------------------------------------------===//
2902 // Like std call, callee cleans arguments, convention except that ECX is
2903 // reserved for storing the tail called function address. Only 2 registers are
2904 // free for argument passing (inreg). Tail call optimization is performed
2906 // * tailcallopt is enabled
2907 // * caller/callee are fastcc
2908 // On X86_64 architecture with GOT-style position independent code only local
2909 // (within module) calls are supported at the moment.
2910 // To keep the stack aligned according to platform abi the function
2911 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2912 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2913 // If a tail called function callee has more arguments than the caller the
2914 // caller needs to make sure that there is room to move the RETADDR to. This is
2915 // achieved by reserving an area the size of the argument delta right after the
2916 // original REtADDR, but before the saved framepointer or the spilled registers
2917 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2929 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2930 /// for a 16 byte align requirement.
2932 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2933 SelectionDAG& DAG) const {
2934 MachineFunction &MF = DAG.getMachineFunction();
2935 const TargetMachine &TM = MF.getTarget();
2936 const X86RegisterInfo *RegInfo =
2937 static_cast<const X86RegisterInfo*>(TM.getRegisterInfo());
2938 const TargetFrameLowering &TFI = *TM.getFrameLowering();
2939 unsigned StackAlignment = TFI.getStackAlignment();
2940 uint64_t AlignMask = StackAlignment - 1;
2941 int64_t Offset = StackSize;
2942 unsigned SlotSize = RegInfo->getSlotSize();
2943 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2944 // Number smaller than 12 so just add the difference.
2945 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2947 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2948 Offset = ((~AlignMask) & Offset) + StackAlignment +
2949 (StackAlignment-SlotSize);
2954 /// MatchingStackOffset - Return true if the given stack call argument is
2955 /// already available in the same position (relatively) of the caller's
2956 /// incoming argument stack.
2958 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2959 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2960 const X86InstrInfo *TII) {
2961 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2963 if (Arg.getOpcode() == ISD::CopyFromReg) {
2964 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2965 if (!TargetRegisterInfo::isVirtualRegister(VR))
2967 MachineInstr *Def = MRI->getVRegDef(VR);
2970 if (!Flags.isByVal()) {
2971 if (!TII->isLoadFromStackSlot(Def, FI))
2974 unsigned Opcode = Def->getOpcode();
2975 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2976 Def->getOperand(1).isFI()) {
2977 FI = Def->getOperand(1).getIndex();
2978 Bytes = Flags.getByValSize();
2982 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2983 if (Flags.isByVal())
2984 // ByVal argument is passed in as a pointer but it's now being
2985 // dereferenced. e.g.
2986 // define @foo(%struct.X* %A) {
2987 // tail call @bar(%struct.X* byval %A)
2990 SDValue Ptr = Ld->getBasePtr();
2991 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2994 FI = FINode->getIndex();
2995 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
2996 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
2997 FI = FINode->getIndex();
2998 Bytes = Flags.getByValSize();
3002 assert(FI != INT_MAX);
3003 if (!MFI->isFixedObjectIndex(FI))
3005 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
3008 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
3009 /// for tail call optimization. Targets which want to do tail call
3010 /// optimization should implement this function.
3012 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
3013 CallingConv::ID CalleeCC,
3015 bool isCalleeStructRet,
3016 bool isCallerStructRet,
3018 const SmallVectorImpl<ISD::OutputArg> &Outs,
3019 const SmallVectorImpl<SDValue> &OutVals,
3020 const SmallVectorImpl<ISD::InputArg> &Ins,
3021 SelectionDAG &DAG) const {
3022 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
3025 // If -tailcallopt is specified, make fastcc functions tail-callable.
3026 const MachineFunction &MF = DAG.getMachineFunction();
3027 const Function *CallerF = MF.getFunction();
3029 // If the function return type is x86_fp80 and the callee return type is not,
3030 // then the FP_EXTEND of the call result is not a nop. It's not safe to
3031 // perform a tailcall optimization here.
3032 if (CallerF->getReturnType()->isX86_FP80Ty() && !RetTy->isX86_FP80Ty())
3035 CallingConv::ID CallerCC = CallerF->getCallingConv();
3036 bool CCMatch = CallerCC == CalleeCC;
3037 bool IsCalleeWin64 = Subtarget->isCallingConvWin64(CalleeCC);
3038 bool IsCallerWin64 = Subtarget->isCallingConvWin64(CallerCC);
3040 if (getTargetMachine().Options.GuaranteedTailCallOpt) {
3041 if (IsTailCallConvention(CalleeCC) && CCMatch)
3046 // Look for obvious safe cases to perform tail call optimization that do not
3047 // require ABI changes. This is what gcc calls sibcall.
3049 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
3050 // emit a special epilogue.
3051 const X86RegisterInfo *RegInfo =
3052 static_cast<const X86RegisterInfo*>(getTargetMachine().getRegisterInfo());
3053 if (RegInfo->needsStackRealignment(MF))
3056 // Also avoid sibcall optimization if either caller or callee uses struct
3057 // return semantics.
3058 if (isCalleeStructRet || isCallerStructRet)
3061 // An stdcall caller is expected to clean up its arguments; the callee
3062 // isn't going to do that.
3063 if (!CCMatch && CallerCC == CallingConv::X86_StdCall)
3066 // Do not sibcall optimize vararg calls unless all arguments are passed via
3068 if (isVarArg && !Outs.empty()) {
3070 // Optimizing for varargs on Win64 is unlikely to be safe without
3071 // additional testing.
3072 if (IsCalleeWin64 || IsCallerWin64)
3075 SmallVector<CCValAssign, 16> ArgLocs;
3076 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
3077 getTargetMachine(), ArgLocs, *DAG.getContext());
3079 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3080 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
3081 if (!ArgLocs[i].isRegLoc())
3085 // If the call result is in ST0 / ST1, it needs to be popped off the x87
3086 // stack. Therefore, if it's not used by the call it is not safe to optimize
3087 // this into a sibcall.
3088 bool Unused = false;
3089 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
3096 SmallVector<CCValAssign, 16> RVLocs;
3097 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(),
3098 getTargetMachine(), RVLocs, *DAG.getContext());
3099 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
3100 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3101 CCValAssign &VA = RVLocs[i];
3102 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
3107 // If the calling conventions do not match, then we'd better make sure the
3108 // results are returned in the same way as what the caller expects.
3110 SmallVector<CCValAssign, 16> RVLocs1;
3111 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
3112 getTargetMachine(), RVLocs1, *DAG.getContext());
3113 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
3115 SmallVector<CCValAssign, 16> RVLocs2;
3116 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
3117 getTargetMachine(), RVLocs2, *DAG.getContext());
3118 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
3120 if (RVLocs1.size() != RVLocs2.size())
3122 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
3123 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
3125 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
3127 if (RVLocs1[i].isRegLoc()) {
3128 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
3131 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
3137 // If the callee takes no arguments then go on to check the results of the
3139 if (!Outs.empty()) {
3140 // Check if stack adjustment is needed. For now, do not do this if any
3141 // argument is passed on the stack.
3142 SmallVector<CCValAssign, 16> ArgLocs;
3143 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
3144 getTargetMachine(), ArgLocs, *DAG.getContext());
3146 // Allocate shadow area for Win64
3148 CCInfo.AllocateStack(32, 8);
3150 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3151 if (CCInfo.getNextStackOffset()) {
3152 MachineFunction &MF = DAG.getMachineFunction();
3153 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
3156 // Check if the arguments are already laid out in the right way as
3157 // the caller's fixed stack objects.
3158 MachineFrameInfo *MFI = MF.getFrameInfo();
3159 const MachineRegisterInfo *MRI = &MF.getRegInfo();
3160 const X86InstrInfo *TII =
3161 ((const X86TargetMachine&)getTargetMachine()).getInstrInfo();
3162 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3163 CCValAssign &VA = ArgLocs[i];
3164 SDValue Arg = OutVals[i];
3165 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3166 if (VA.getLocInfo() == CCValAssign::Indirect)
3168 if (!VA.isRegLoc()) {
3169 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
3176 // If the tailcall address may be in a register, then make sure it's
3177 // possible to register allocate for it. In 32-bit, the call address can
3178 // only target EAX, EDX, or ECX since the tail call must be scheduled after
3179 // callee-saved registers are restored. These happen to be the same
3180 // registers used to pass 'inreg' arguments so watch out for those.
3181 if (!Subtarget->is64Bit() &&
3182 ((!isa<GlobalAddressSDNode>(Callee) &&
3183 !isa<ExternalSymbolSDNode>(Callee)) ||
3184 getTargetMachine().getRelocationModel() == Reloc::PIC_)) {
3185 unsigned NumInRegs = 0;
3186 // In PIC we need an extra register to formulate the address computation
3188 unsigned MaxInRegs =
3189 (getTargetMachine().getRelocationModel() == Reloc::PIC_) ? 2 : 3;
3191 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3192 CCValAssign &VA = ArgLocs[i];
3195 unsigned Reg = VA.getLocReg();
3198 case X86::EAX: case X86::EDX: case X86::ECX:
3199 if (++NumInRegs == MaxInRegs)
3211 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
3212 const TargetLibraryInfo *libInfo) const {
3213 return X86::createFastISel(funcInfo, libInfo);
3216 //===----------------------------------------------------------------------===//
3217 // Other Lowering Hooks
3218 //===----------------------------------------------------------------------===//
3220 static bool MayFoldLoad(SDValue Op) {
3221 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
3224 static bool MayFoldIntoStore(SDValue Op) {
3225 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
3228 static bool isTargetShuffle(unsigned Opcode) {
3230 default: return false;
3231 case X86ISD::PSHUFD:
3232 case X86ISD::PSHUFHW:
3233 case X86ISD::PSHUFLW:
3235 case X86ISD::PALIGNR:
3236 case X86ISD::MOVLHPS:
3237 case X86ISD::MOVLHPD:
3238 case X86ISD::MOVHLPS:
3239 case X86ISD::MOVLPS:
3240 case X86ISD::MOVLPD:
3241 case X86ISD::MOVSHDUP:
3242 case X86ISD::MOVSLDUP:
3243 case X86ISD::MOVDDUP:
3246 case X86ISD::UNPCKL:
3247 case X86ISD::UNPCKH:
3248 case X86ISD::VPERMILP:
3249 case X86ISD::VPERM2X128:
3250 case X86ISD::VPERMI:
3255 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3256 SDValue V1, SelectionDAG &DAG) {
3258 default: llvm_unreachable("Unknown x86 shuffle node");
3259 case X86ISD::MOVSHDUP:
3260 case X86ISD::MOVSLDUP:
3261 case X86ISD::MOVDDUP:
3262 return DAG.getNode(Opc, dl, VT, V1);
3266 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3267 SDValue V1, unsigned TargetMask,
3268 SelectionDAG &DAG) {
3270 default: llvm_unreachable("Unknown x86 shuffle node");
3271 case X86ISD::PSHUFD:
3272 case X86ISD::PSHUFHW:
3273 case X86ISD::PSHUFLW:
3274 case X86ISD::VPERMILP:
3275 case X86ISD::VPERMI:
3276 return DAG.getNode(Opc, dl, VT, V1, DAG.getConstant(TargetMask, MVT::i8));
3280 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3281 SDValue V1, SDValue V2, unsigned TargetMask,
3282 SelectionDAG &DAG) {
3284 default: llvm_unreachable("Unknown x86 shuffle node");
3285 case X86ISD::PALIGNR:
3287 case X86ISD::VPERM2X128:
3288 return DAG.getNode(Opc, dl, VT, V1, V2,
3289 DAG.getConstant(TargetMask, MVT::i8));
3293 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3294 SDValue V1, SDValue V2, SelectionDAG &DAG) {
3296 default: llvm_unreachable("Unknown x86 shuffle node");
3297 case X86ISD::MOVLHPS:
3298 case X86ISD::MOVLHPD:
3299 case X86ISD::MOVHLPS:
3300 case X86ISD::MOVLPS:
3301 case X86ISD::MOVLPD:
3304 case X86ISD::UNPCKL:
3305 case X86ISD::UNPCKH:
3306 return DAG.getNode(Opc, dl, VT, V1, V2);
3310 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
3311 MachineFunction &MF = DAG.getMachineFunction();
3312 const X86RegisterInfo *RegInfo =
3313 static_cast<const X86RegisterInfo*>(getTargetMachine().getRegisterInfo());
3314 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
3315 int ReturnAddrIndex = FuncInfo->getRAIndex();
3317 if (ReturnAddrIndex == 0) {
3318 // Set up a frame object for the return address.
3319 unsigned SlotSize = RegInfo->getSlotSize();
3320 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize,
3323 FuncInfo->setRAIndex(ReturnAddrIndex);
3326 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
3329 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
3330 bool hasSymbolicDisplacement) {
3331 // Offset should fit into 32 bit immediate field.
3332 if (!isInt<32>(Offset))
3335 // If we don't have a symbolic displacement - we don't have any extra
3337 if (!hasSymbolicDisplacement)
3340 // FIXME: Some tweaks might be needed for medium code model.
3341 if (M != CodeModel::Small && M != CodeModel::Kernel)
3344 // For small code model we assume that latest object is 16MB before end of 31
3345 // bits boundary. We may also accept pretty large negative constants knowing
3346 // that all objects are in the positive half of address space.
3347 if (M == CodeModel::Small && Offset < 16*1024*1024)
3350 // For kernel code model we know that all object resist in the negative half
3351 // of 32bits address space. We may not accept negative offsets, since they may
3352 // be just off and we may accept pretty large positive ones.
3353 if (M == CodeModel::Kernel && Offset > 0)
3359 /// isCalleePop - Determines whether the callee is required to pop its
3360 /// own arguments. Callee pop is necessary to support tail calls.
3361 bool X86::isCalleePop(CallingConv::ID CallingConv,
3362 bool is64Bit, bool IsVarArg, bool TailCallOpt) {
3366 switch (CallingConv) {
3369 case CallingConv::X86_StdCall:
3371 case CallingConv::X86_FastCall:
3373 case CallingConv::X86_ThisCall:
3375 case CallingConv::Fast:
3377 case CallingConv::GHC:
3379 case CallingConv::HiPE:
3384 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
3385 /// specific condition code, returning the condition code and the LHS/RHS of the
3386 /// comparison to make.
3387 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
3388 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
3390 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
3391 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
3392 // X > -1 -> X == 0, jump !sign.
3393 RHS = DAG.getConstant(0, RHS.getValueType());
3394 return X86::COND_NS;
3396 if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
3397 // X < 0 -> X == 0, jump on sign.
3400 if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
3402 RHS = DAG.getConstant(0, RHS.getValueType());
3403 return X86::COND_LE;
3407 switch (SetCCOpcode) {
3408 default: llvm_unreachable("Invalid integer condition!");
3409 case ISD::SETEQ: return X86::COND_E;
3410 case ISD::SETGT: return X86::COND_G;
3411 case ISD::SETGE: return X86::COND_GE;
3412 case ISD::SETLT: return X86::COND_L;
3413 case ISD::SETLE: return X86::COND_LE;
3414 case ISD::SETNE: return X86::COND_NE;
3415 case ISD::SETULT: return X86::COND_B;
3416 case ISD::SETUGT: return X86::COND_A;
3417 case ISD::SETULE: return X86::COND_BE;
3418 case ISD::SETUGE: return X86::COND_AE;
3422 // First determine if it is required or is profitable to flip the operands.
3424 // If LHS is a foldable load, but RHS is not, flip the condition.
3425 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
3426 !ISD::isNON_EXTLoad(RHS.getNode())) {
3427 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
3428 std::swap(LHS, RHS);
3431 switch (SetCCOpcode) {
3437 std::swap(LHS, RHS);
3441 // On a floating point condition, the flags are set as follows:
3443 // 0 | 0 | 0 | X > Y
3444 // 0 | 0 | 1 | X < Y
3445 // 1 | 0 | 0 | X == Y
3446 // 1 | 1 | 1 | unordered
3447 switch (SetCCOpcode) {
3448 default: llvm_unreachable("Condcode should be pre-legalized away");
3450 case ISD::SETEQ: return X86::COND_E;
3451 case ISD::SETOLT: // flipped
3453 case ISD::SETGT: return X86::COND_A;
3454 case ISD::SETOLE: // flipped
3456 case ISD::SETGE: return X86::COND_AE;
3457 case ISD::SETUGT: // flipped
3459 case ISD::SETLT: return X86::COND_B;
3460 case ISD::SETUGE: // flipped
3462 case ISD::SETLE: return X86::COND_BE;
3464 case ISD::SETNE: return X86::COND_NE;
3465 case ISD::SETUO: return X86::COND_P;
3466 case ISD::SETO: return X86::COND_NP;
3468 case ISD::SETUNE: return X86::COND_INVALID;
3472 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
3473 /// code. Current x86 isa includes the following FP cmov instructions:
3474 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
3475 static bool hasFPCMov(unsigned X86CC) {
3491 /// isFPImmLegal - Returns true if the target can instruction select the
3492 /// specified FP immediate natively. If false, the legalizer will
3493 /// materialize the FP immediate as a load from a constant pool.
3494 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3495 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
3496 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
3502 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
3503 /// the specified range (L, H].
3504 static bool isUndefOrInRange(int Val, int Low, int Hi) {
3505 return (Val < 0) || (Val >= Low && Val < Hi);
3508 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
3509 /// specified value.
3510 static bool isUndefOrEqual(int Val, int CmpVal) {
3511 return (Val < 0 || Val == CmpVal);
3514 /// isSequentialOrUndefInRange - Return true if every element in Mask, beginning
3515 /// from position Pos and ending in Pos+Size, falls within the specified
3516 /// sequential range (L, L+Pos]. or is undef.
3517 static bool isSequentialOrUndefInRange(ArrayRef<int> Mask,
3518 unsigned Pos, unsigned Size, int Low) {
3519 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
3520 if (!isUndefOrEqual(Mask[i], Low))
3525 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
3526 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
3527 /// the second operand.
3528 static bool isPSHUFDMask(ArrayRef<int> Mask, MVT VT) {
3529 if (VT == MVT::v4f32 || VT == MVT::v4i32 )
3530 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
3531 if (VT == MVT::v2f64 || VT == MVT::v2i64)
3532 return (Mask[0] < 2 && Mask[1] < 2);
3536 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
3537 /// is suitable for input to PSHUFHW.
3538 static bool isPSHUFHWMask(ArrayRef<int> Mask, MVT VT, bool HasInt256) {
3539 if (VT != MVT::v8i16 && (!HasInt256 || VT != MVT::v16i16))
3542 // Lower quadword copied in order or undef.
3543 if (!isSequentialOrUndefInRange(Mask, 0, 4, 0))
3546 // Upper quadword shuffled.
3547 for (unsigned i = 4; i != 8; ++i)
3548 if (!isUndefOrInRange(Mask[i], 4, 8))
3551 if (VT == MVT::v16i16) {
3552 // Lower quadword copied in order or undef.
3553 if (!isSequentialOrUndefInRange(Mask, 8, 4, 8))
3556 // Upper quadword shuffled.
3557 for (unsigned i = 12; i != 16; ++i)
3558 if (!isUndefOrInRange(Mask[i], 12, 16))
3565 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
3566 /// is suitable for input to PSHUFLW.
3567 static bool isPSHUFLWMask(ArrayRef<int> Mask, MVT VT, bool HasInt256) {
3568 if (VT != MVT::v8i16 && (!HasInt256 || VT != MVT::v16i16))
3571 // Upper quadword copied in order.
3572 if (!isSequentialOrUndefInRange(Mask, 4, 4, 4))
3575 // Lower quadword shuffled.
3576 for (unsigned i = 0; i != 4; ++i)
3577 if (!isUndefOrInRange(Mask[i], 0, 4))
3580 if (VT == MVT::v16i16) {
3581 // Upper quadword copied in order.
3582 if (!isSequentialOrUndefInRange(Mask, 12, 4, 12))
3585 // Lower quadword shuffled.
3586 for (unsigned i = 8; i != 12; ++i)
3587 if (!isUndefOrInRange(Mask[i], 8, 12))
3594 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
3595 /// is suitable for input to PALIGNR.
3596 static bool isPALIGNRMask(ArrayRef<int> Mask, MVT VT,
3597 const X86Subtarget *Subtarget) {
3598 if ((VT.is128BitVector() && !Subtarget->hasSSSE3()) ||
3599 (VT.is256BitVector() && !Subtarget->hasInt256()))
3602 unsigned NumElts = VT.getVectorNumElements();
3603 unsigned NumLanes = VT.is512BitVector() ? 1: VT.getSizeInBits()/128;
3604 unsigned NumLaneElts = NumElts/NumLanes;
3606 // Do not handle 64-bit element shuffles with palignr.
3607 if (NumLaneElts == 2)
3610 for (unsigned l = 0; l != NumElts; l+=NumLaneElts) {
3612 for (i = 0; i != NumLaneElts; ++i) {
3617 // Lane is all undef, go to next lane
3618 if (i == NumLaneElts)
3621 int Start = Mask[i+l];
3623 // Make sure its in this lane in one of the sources
3624 if (!isUndefOrInRange(Start, l, l+NumLaneElts) &&
3625 !isUndefOrInRange(Start, l+NumElts, l+NumElts+NumLaneElts))
3628 // If not lane 0, then we must match lane 0
3629 if (l != 0 && Mask[i] >= 0 && !isUndefOrEqual(Start, Mask[i]+l))
3632 // Correct second source to be contiguous with first source
3633 if (Start >= (int)NumElts)
3634 Start -= NumElts - NumLaneElts;
3636 // Make sure we're shifting in the right direction.
3637 if (Start <= (int)(i+l))
3642 // Check the rest of the elements to see if they are consecutive.
3643 for (++i; i != NumLaneElts; ++i) {
3644 int Idx = Mask[i+l];
3646 // Make sure its in this lane
3647 if (!isUndefOrInRange(Idx, l, l+NumLaneElts) &&
3648 !isUndefOrInRange(Idx, l+NumElts, l+NumElts+NumLaneElts))
3651 // If not lane 0, then we must match lane 0
3652 if (l != 0 && Mask[i] >= 0 && !isUndefOrEqual(Idx, Mask[i]+l))
3655 if (Idx >= (int)NumElts)
3656 Idx -= NumElts - NumLaneElts;
3658 if (!isUndefOrEqual(Idx, Start+i))
3667 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3668 /// the two vector operands have swapped position.
3669 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask,
3670 unsigned NumElems) {
3671 for (unsigned i = 0; i != NumElems; ++i) {
3675 else if (idx < (int)NumElems)
3676 Mask[i] = idx + NumElems;
3678 Mask[i] = idx - NumElems;
3682 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
3683 /// specifies a shuffle of elements that is suitable for input to 128/256-bit
3684 /// SHUFPS and SHUFPD. If Commuted is true, then it checks for sources to be
3685 /// reverse of what x86 shuffles want.
3686 static bool isSHUFPMask(ArrayRef<int> Mask, MVT VT, bool Commuted = false) {
3688 unsigned NumElems = VT.getVectorNumElements();
3689 unsigned NumLanes = VT.getSizeInBits()/128;
3690 unsigned NumLaneElems = NumElems/NumLanes;
3692 if (NumLaneElems != 2 && NumLaneElems != 4)
3695 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3696 bool symetricMaskRequired =
3697 (VT.getSizeInBits() >= 256) && (EltSize == 32);
3699 // VSHUFPSY divides the resulting vector into 4 chunks.
3700 // The sources are also splitted into 4 chunks, and each destination
3701 // chunk must come from a different source chunk.
3703 // SRC1 => X7 X6 X5 X4 X3 X2 X1 X0
3704 // SRC2 => Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y9
3706 // DST => Y7..Y4, Y7..Y4, X7..X4, X7..X4,
3707 // Y3..Y0, Y3..Y0, X3..X0, X3..X0
3709 // VSHUFPDY divides the resulting vector into 4 chunks.
3710 // The sources are also splitted into 4 chunks, and each destination
3711 // chunk must come from a different source chunk.
3713 // SRC1 => X3 X2 X1 X0
3714 // SRC2 => Y3 Y2 Y1 Y0
3716 // DST => Y3..Y2, X3..X2, Y1..Y0, X1..X0
3718 SmallVector<int, 4> MaskVal(NumLaneElems, -1);
3719 unsigned HalfLaneElems = NumLaneElems/2;
3720 for (unsigned l = 0; l != NumElems; l += NumLaneElems) {
3721 for (unsigned i = 0; i != NumLaneElems; ++i) {
3722 int Idx = Mask[i+l];
3723 unsigned RngStart = l + ((Commuted == (i<HalfLaneElems)) ? NumElems : 0);
3724 if (!isUndefOrInRange(Idx, RngStart, RngStart+NumLaneElems))
3726 // For VSHUFPSY, the mask of the second half must be the same as the
3727 // first but with the appropriate offsets. This works in the same way as
3728 // VPERMILPS works with masks.
3729 if (!symetricMaskRequired || Idx < 0)
3731 if (MaskVal[i] < 0) {
3732 MaskVal[i] = Idx - l;
3735 if ((signed)(Idx - l) != MaskVal[i])
3743 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
3744 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
3745 static bool isMOVHLPSMask(ArrayRef<int> Mask, MVT VT) {
3746 if (!VT.is128BitVector())
3749 unsigned NumElems = VT.getVectorNumElements();
3754 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
3755 return isUndefOrEqual(Mask[0], 6) &&
3756 isUndefOrEqual(Mask[1], 7) &&
3757 isUndefOrEqual(Mask[2], 2) &&
3758 isUndefOrEqual(Mask[3], 3);
3761 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
3762 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
3764 static bool isMOVHLPS_v_undef_Mask(ArrayRef<int> Mask, MVT VT) {
3765 if (!VT.is128BitVector())
3768 unsigned NumElems = VT.getVectorNumElements();
3773 return isUndefOrEqual(Mask[0], 2) &&
3774 isUndefOrEqual(Mask[1], 3) &&
3775 isUndefOrEqual(Mask[2], 2) &&
3776 isUndefOrEqual(Mask[3], 3);
3779 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
3780 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
3781 static bool isMOVLPMask(ArrayRef<int> Mask, MVT VT) {
3782 if (!VT.is128BitVector())
3785 unsigned NumElems = VT.getVectorNumElements();
3787 if (NumElems != 2 && NumElems != 4)
3790 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3791 if (!isUndefOrEqual(Mask[i], i + NumElems))
3794 for (unsigned i = NumElems/2, e = NumElems; i != e; ++i)
3795 if (!isUndefOrEqual(Mask[i], i))
3801 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
3802 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
3803 static bool isMOVLHPSMask(ArrayRef<int> Mask, MVT VT) {
3804 if (!VT.is128BitVector())
3807 unsigned NumElems = VT.getVectorNumElements();
3809 if (NumElems != 2 && NumElems != 4)
3812 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3813 if (!isUndefOrEqual(Mask[i], i))
3816 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3817 if (!isUndefOrEqual(Mask[i + e], i + NumElems))
3824 // Some special combinations that can be optimized.
3827 SDValue Compact8x32ShuffleNode(ShuffleVectorSDNode *SVOp,
3828 SelectionDAG &DAG) {
3829 MVT VT = SVOp->getSimpleValueType(0);
3832 if (VT != MVT::v8i32 && VT != MVT::v8f32)
3835 ArrayRef<int> Mask = SVOp->getMask();
3837 // These are the special masks that may be optimized.
3838 static const int MaskToOptimizeEven[] = {0, 8, 2, 10, 4, 12, 6, 14};
3839 static const int MaskToOptimizeOdd[] = {1, 9, 3, 11, 5, 13, 7, 15};
3840 bool MatchEvenMask = true;
3841 bool MatchOddMask = true;
3842 for (int i=0; i<8; ++i) {
3843 if (!isUndefOrEqual(Mask[i], MaskToOptimizeEven[i]))
3844 MatchEvenMask = false;
3845 if (!isUndefOrEqual(Mask[i], MaskToOptimizeOdd[i]))
3846 MatchOddMask = false;
3849 if (!MatchEvenMask && !MatchOddMask)
3852 SDValue UndefNode = DAG.getNode(ISD::UNDEF, dl, VT);
3854 SDValue Op0 = SVOp->getOperand(0);
3855 SDValue Op1 = SVOp->getOperand(1);
3857 if (MatchEvenMask) {
3858 // Shift the second operand right to 32 bits.
3859 static const int ShiftRightMask[] = {-1, 0, -1, 2, -1, 4, -1, 6 };
3860 Op1 = DAG.getVectorShuffle(VT, dl, Op1, UndefNode, ShiftRightMask);
3862 // Shift the first operand left to 32 bits.
3863 static const int ShiftLeftMask[] = {1, -1, 3, -1, 5, -1, 7, -1 };
3864 Op0 = DAG.getVectorShuffle(VT, dl, Op0, UndefNode, ShiftLeftMask);
3866 static const int BlendMask[] = {0, 9, 2, 11, 4, 13, 6, 15};
3867 return DAG.getVectorShuffle(VT, dl, Op0, Op1, BlendMask);
3870 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
3871 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
3872 static bool isUNPCKLMask(ArrayRef<int> Mask, MVT VT,
3873 bool HasInt256, bool V2IsSplat = false) {
3875 assert(VT.getSizeInBits() >= 128 &&
3876 "Unsupported vector type for unpckl");
3878 // AVX defines UNPCK* to operate independently on 128-bit lanes.
3880 unsigned NumOf256BitLanes;
3881 unsigned NumElts = VT.getVectorNumElements();
3882 if (VT.is256BitVector()) {
3883 if (NumElts != 4 && NumElts != 8 &&
3884 (!HasInt256 || (NumElts != 16 && NumElts != 32)))
3887 NumOf256BitLanes = 1;
3888 } else if (VT.is512BitVector()) {
3889 assert(VT.getScalarType().getSizeInBits() >= 32 &&
3890 "Unsupported vector type for unpckh");
3892 NumOf256BitLanes = 2;
3895 NumOf256BitLanes = 1;
3898 unsigned NumEltsInStride = NumElts/NumOf256BitLanes;
3899 unsigned NumLaneElts = NumEltsInStride/NumLanes;
3901 for (unsigned l256 = 0; l256 < NumOf256BitLanes; l256 += 1) {
3902 for (unsigned l = 0; l != NumEltsInStride; l += NumLaneElts) {
3903 for (unsigned i = 0, j = l; i != NumLaneElts; i += 2, ++j) {
3904 int BitI = Mask[l256*NumEltsInStride+l+i];
3905 int BitI1 = Mask[l256*NumEltsInStride+l+i+1];
3906 if (!isUndefOrEqual(BitI, j+l256*NumElts))
3908 if (V2IsSplat && !isUndefOrEqual(BitI1, NumElts))
3910 if (!isUndefOrEqual(BitI1, j+l256*NumElts+NumEltsInStride))
3918 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
3919 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
3920 static bool isUNPCKHMask(ArrayRef<int> Mask, MVT VT,
3921 bool HasInt256, bool V2IsSplat = false) {
3922 assert(VT.getSizeInBits() >= 128 &&
3923 "Unsupported vector type for unpckh");
3925 // AVX defines UNPCK* to operate independently on 128-bit lanes.
3927 unsigned NumOf256BitLanes;
3928 unsigned NumElts = VT.getVectorNumElements();
3929 if (VT.is256BitVector()) {
3930 if (NumElts != 4 && NumElts != 8 &&
3931 (!HasInt256 || (NumElts != 16 && NumElts != 32)))
3934 NumOf256BitLanes = 1;
3935 } else if (VT.is512BitVector()) {
3936 assert(VT.getScalarType().getSizeInBits() >= 32 &&
3937 "Unsupported vector type for unpckh");
3939 NumOf256BitLanes = 2;
3942 NumOf256BitLanes = 1;
3945 unsigned NumEltsInStride = NumElts/NumOf256BitLanes;
3946 unsigned NumLaneElts = NumEltsInStride/NumLanes;
3948 for (unsigned l256 = 0; l256 < NumOf256BitLanes; l256 += 1) {
3949 for (unsigned l = 0; l != NumEltsInStride; l += NumLaneElts) {
3950 for (unsigned i = 0, j = l+NumLaneElts/2; i != NumLaneElts; i += 2, ++j) {
3951 int BitI = Mask[l256*NumEltsInStride+l+i];
3952 int BitI1 = Mask[l256*NumEltsInStride+l+i+1];
3953 if (!isUndefOrEqual(BitI, j+l256*NumElts))
3955 if (V2IsSplat && !isUndefOrEqual(BitI1, NumElts))
3957 if (!isUndefOrEqual(BitI1, j+l256*NumElts+NumEltsInStride))
3965 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
3966 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
3968 static bool isUNPCKL_v_undef_Mask(ArrayRef<int> Mask, MVT VT, bool HasInt256) {
3969 unsigned NumElts = VT.getVectorNumElements();
3970 bool Is256BitVec = VT.is256BitVector();
3972 if (VT.is512BitVector())
3974 assert((VT.is128BitVector() || VT.is256BitVector()) &&
3975 "Unsupported vector type for unpckh");
3977 if (Is256BitVec && NumElts != 4 && NumElts != 8 &&
3978 (!HasInt256 || (NumElts != 16 && NumElts != 32)))
3981 // For 256-bit i64/f64, use MOVDDUPY instead, so reject the matching pattern
3982 // FIXME: Need a better way to get rid of this, there's no latency difference
3983 // between UNPCKLPD and MOVDDUP, the later should always be checked first and
3984 // the former later. We should also remove the "_undef" special mask.
3985 if (NumElts == 4 && Is256BitVec)
3988 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
3989 // independently on 128-bit lanes.
3990 unsigned NumLanes = VT.getSizeInBits()/128;
3991 unsigned NumLaneElts = NumElts/NumLanes;
3993 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
3994 for (unsigned i = 0, j = l; i != NumLaneElts; i += 2, ++j) {
3995 int BitI = Mask[l+i];
3996 int BitI1 = Mask[l+i+1];
3998 if (!isUndefOrEqual(BitI, j))
4000 if (!isUndefOrEqual(BitI1, j))
4008 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
4009 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
4011 static bool isUNPCKH_v_undef_Mask(ArrayRef<int> Mask, MVT VT, bool HasInt256) {
4012 unsigned NumElts = VT.getVectorNumElements();
4014 if (VT.is512BitVector())
4017 assert((VT.is128BitVector() || VT.is256BitVector()) &&
4018 "Unsupported vector type for unpckh");
4020 if (VT.is256BitVector() && NumElts != 4 && NumElts != 8 &&
4021 (!HasInt256 || (NumElts != 16 && NumElts != 32)))
4024 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
4025 // independently on 128-bit lanes.
4026 unsigned NumLanes = VT.getSizeInBits()/128;
4027 unsigned NumLaneElts = NumElts/NumLanes;
4029 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
4030 for (unsigned i = 0, j = l+NumLaneElts/2; i != NumLaneElts; i += 2, ++j) {
4031 int BitI = Mask[l+i];
4032 int BitI1 = Mask[l+i+1];
4033 if (!isUndefOrEqual(BitI, j))
4035 if (!isUndefOrEqual(BitI1, j))
4042 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
4043 /// specifies a shuffle of elements that is suitable for input to MOVSS,
4044 /// MOVSD, and MOVD, i.e. setting the lowest element.
4045 static bool isMOVLMask(ArrayRef<int> Mask, EVT VT) {
4046 if (VT.getVectorElementType().getSizeInBits() < 32)
4048 if (!VT.is128BitVector())
4051 unsigned NumElts = VT.getVectorNumElements();
4053 if (!isUndefOrEqual(Mask[0], NumElts))
4056 for (unsigned i = 1; i != NumElts; ++i)
4057 if (!isUndefOrEqual(Mask[i], i))
4063 /// isVPERM2X128Mask - Match 256-bit shuffles where the elements are considered
4064 /// as permutations between 128-bit chunks or halves. As an example: this
4066 /// vector_shuffle <4, 5, 6, 7, 12, 13, 14, 15>
4067 /// The first half comes from the second half of V1 and the second half from the
4068 /// the second half of V2.
4069 static bool isVPERM2X128Mask(ArrayRef<int> Mask, MVT VT, bool HasFp256) {
4070 if (!HasFp256 || !VT.is256BitVector())
4073 // The shuffle result is divided into half A and half B. In total the two
4074 // sources have 4 halves, namely: C, D, E, F. The final values of A and
4075 // B must come from C, D, E or F.
4076 unsigned HalfSize = VT.getVectorNumElements()/2;
4077 bool MatchA = false, MatchB = false;
4079 // Check if A comes from one of C, D, E, F.
4080 for (unsigned Half = 0; Half != 4; ++Half) {
4081 if (isSequentialOrUndefInRange(Mask, 0, HalfSize, Half*HalfSize)) {
4087 // Check if B comes from one of C, D, E, F.
4088 for (unsigned Half = 0; Half != 4; ++Half) {
4089 if (isSequentialOrUndefInRange(Mask, HalfSize, HalfSize, Half*HalfSize)) {
4095 return MatchA && MatchB;
4098 /// getShuffleVPERM2X128Immediate - Return the appropriate immediate to shuffle
4099 /// the specified VECTOR_MASK mask with VPERM2F128/VPERM2I128 instructions.
4100 static unsigned getShuffleVPERM2X128Immediate(ShuffleVectorSDNode *SVOp) {
4101 MVT VT = SVOp->getSimpleValueType(0);
4103 unsigned HalfSize = VT.getVectorNumElements()/2;
4105 unsigned FstHalf = 0, SndHalf = 0;
4106 for (unsigned i = 0; i < HalfSize; ++i) {
4107 if (SVOp->getMaskElt(i) > 0) {
4108 FstHalf = SVOp->getMaskElt(i)/HalfSize;
4112 for (unsigned i = HalfSize; i < HalfSize*2; ++i) {
4113 if (SVOp->getMaskElt(i) > 0) {
4114 SndHalf = SVOp->getMaskElt(i)/HalfSize;
4119 return (FstHalf | (SndHalf << 4));
4122 // Symetric in-lane mask. Each lane has 4 elements (for imm8)
4123 static bool isPermImmMask(ArrayRef<int> Mask, MVT VT, unsigned& Imm8) {
4124 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4128 unsigned NumElts = VT.getVectorNumElements();
4130 if (VT.is128BitVector() || (VT.is256BitVector() && EltSize == 64)) {
4131 for (unsigned i = 0; i != NumElts; ++i) {
4134 Imm8 |= Mask[i] << (i*2);
4139 unsigned LaneSize = 4;
4140 SmallVector<int, 4> MaskVal(LaneSize, -1);
4142 for (unsigned l = 0; l != NumElts; l += LaneSize) {
4143 for (unsigned i = 0; i != LaneSize; ++i) {
4144 if (!isUndefOrInRange(Mask[i+l], l, l+LaneSize))
4148 if (MaskVal[i] < 0) {
4149 MaskVal[i] = Mask[i+l] - l;
4150 Imm8 |= MaskVal[i] << (i*2);
4153 if (Mask[i+l] != (signed)(MaskVal[i]+l))
4160 /// isVPERMILPMask - Return true if the specified VECTOR_SHUFFLE operand
4161 /// specifies a shuffle of elements that is suitable for input to VPERMILPD*.
4162 /// Note that VPERMIL mask matching is different depending whether theunderlying
4163 /// type is 32 or 64. In the VPERMILPS the high half of the mask should point
4164 /// to the same elements of the low, but to the higher half of the source.
4165 /// In VPERMILPD the two lanes could be shuffled independently of each other
4166 /// with the same restriction that lanes can't be crossed. Also handles PSHUFDY.
4167 static bool isVPERMILPMask(ArrayRef<int> Mask, MVT VT) {
4168 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4169 if (VT.getSizeInBits() < 256 || EltSize < 32)
4171 bool symetricMaskRequired = (EltSize == 32);
4172 unsigned NumElts = VT.getVectorNumElements();
4174 unsigned NumLanes = VT.getSizeInBits()/128;
4175 unsigned LaneSize = NumElts/NumLanes;
4176 // 2 or 4 elements in one lane
4178 SmallVector<int, 4> ExpectedMaskVal(LaneSize, -1);
4179 for (unsigned l = 0; l != NumElts; l += LaneSize) {
4180 for (unsigned i = 0; i != LaneSize; ++i) {
4181 if (!isUndefOrInRange(Mask[i+l], l, l+LaneSize))
4183 if (symetricMaskRequired) {
4184 if (ExpectedMaskVal[i] < 0 && Mask[i+l] >= 0) {
4185 ExpectedMaskVal[i] = Mask[i+l] - l;
4188 if (!isUndefOrEqual(Mask[i+l], ExpectedMaskVal[i]+l))
4196 /// isCommutedMOVLMask - Returns true if the shuffle mask is except the reverse
4197 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
4198 /// element of vector 2 and the other elements to come from vector 1 in order.
4199 static bool isCommutedMOVLMask(ArrayRef<int> Mask, MVT VT,
4200 bool V2IsSplat = false, bool V2IsUndef = false) {
4201 if (!VT.is128BitVector())
4204 unsigned NumOps = VT.getVectorNumElements();
4205 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
4208 if (!isUndefOrEqual(Mask[0], 0))
4211 for (unsigned i = 1; i != NumOps; ++i)
4212 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
4213 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
4214 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
4220 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
4221 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
4222 /// Masks to match: <1, 1, 3, 3> or <1, 1, 3, 3, 5, 5, 7, 7>
4223 static bool isMOVSHDUPMask(ArrayRef<int> Mask, MVT VT,
4224 const X86Subtarget *Subtarget) {
4225 if (!Subtarget->hasSSE3())
4228 unsigned NumElems = VT.getVectorNumElements();
4230 if ((VT.is128BitVector() && NumElems != 4) ||
4231 (VT.is256BitVector() && NumElems != 8) ||
4232 (VT.is512BitVector() && NumElems != 16))
4235 // "i+1" is the value the indexed mask element must have
4236 for (unsigned i = 0; i != NumElems; i += 2)
4237 if (!isUndefOrEqual(Mask[i], i+1) ||
4238 !isUndefOrEqual(Mask[i+1], i+1))
4244 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
4245 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
4246 /// Masks to match: <0, 0, 2, 2> or <0, 0, 2, 2, 4, 4, 6, 6>
4247 static bool isMOVSLDUPMask(ArrayRef<int> Mask, MVT VT,
4248 const X86Subtarget *Subtarget) {
4249 if (!Subtarget->hasSSE3())
4252 unsigned NumElems = VT.getVectorNumElements();
4254 if ((VT.is128BitVector() && NumElems != 4) ||
4255 (VT.is256BitVector() && NumElems != 8) ||
4256 (VT.is512BitVector() && NumElems != 16))
4259 // "i" is the value the indexed mask element must have
4260 for (unsigned i = 0; i != NumElems; i += 2)
4261 if (!isUndefOrEqual(Mask[i], i) ||
4262 !isUndefOrEqual(Mask[i+1], i))
4268 /// isMOVDDUPYMask - Return true if the specified VECTOR_SHUFFLE operand
4269 /// specifies a shuffle of elements that is suitable for input to 256-bit
4270 /// version of MOVDDUP.
4271 static bool isMOVDDUPYMask(ArrayRef<int> Mask, MVT VT, bool HasFp256) {
4272 if (!HasFp256 || !VT.is256BitVector())
4275 unsigned NumElts = VT.getVectorNumElements();
4279 for (unsigned i = 0; i != NumElts/2; ++i)
4280 if (!isUndefOrEqual(Mask[i], 0))
4282 for (unsigned i = NumElts/2; i != NumElts; ++i)
4283 if (!isUndefOrEqual(Mask[i], NumElts/2))
4288 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
4289 /// specifies a shuffle of elements that is suitable for input to 128-bit
4290 /// version of MOVDDUP.
4291 static bool isMOVDDUPMask(ArrayRef<int> Mask, MVT VT) {
4292 if (!VT.is128BitVector())
4295 unsigned e = VT.getVectorNumElements() / 2;
4296 for (unsigned i = 0; i != e; ++i)
4297 if (!isUndefOrEqual(Mask[i], i))
4299 for (unsigned i = 0; i != e; ++i)
4300 if (!isUndefOrEqual(Mask[e+i], i))
4305 /// isVEXTRACTIndex - Return true if the specified
4306 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
4307 /// suitable for instruction that extract 128 or 256 bit vectors
4308 static bool isVEXTRACTIndex(SDNode *N, unsigned vecWidth) {
4309 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4310 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4313 // The index should be aligned on a vecWidth-bit boundary.
4315 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4317 MVT VT = N->getSimpleValueType(0);
4318 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4319 bool Result = (Index * ElSize) % vecWidth == 0;
4324 /// isVINSERTIndex - Return true if the specified INSERT_SUBVECTOR
4325 /// operand specifies a subvector insert that is suitable for input to
4326 /// insertion of 128 or 256-bit subvectors
4327 static bool isVINSERTIndex(SDNode *N, unsigned vecWidth) {
4328 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4329 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4331 // The index should be aligned on a vecWidth-bit boundary.
4333 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4335 MVT VT = N->getSimpleValueType(0);
4336 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4337 bool Result = (Index * ElSize) % vecWidth == 0;
4342 bool X86::isVINSERT128Index(SDNode *N) {
4343 return isVINSERTIndex(N, 128);
4346 bool X86::isVINSERT256Index(SDNode *N) {
4347 return isVINSERTIndex(N, 256);
4350 bool X86::isVEXTRACT128Index(SDNode *N) {
4351 return isVEXTRACTIndex(N, 128);
4354 bool X86::isVEXTRACT256Index(SDNode *N) {
4355 return isVEXTRACTIndex(N, 256);
4358 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
4359 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
4360 /// Handles 128-bit and 256-bit.
4361 static unsigned getShuffleSHUFImmediate(ShuffleVectorSDNode *N) {
4362 MVT VT = N->getSimpleValueType(0);
4364 assert((VT.getSizeInBits() >= 128) &&
4365 "Unsupported vector type for PSHUF/SHUFP");
4367 // Handle 128 and 256-bit vector lengths. AVX defines PSHUF/SHUFP to operate
4368 // independently on 128-bit lanes.
4369 unsigned NumElts = VT.getVectorNumElements();
4370 unsigned NumLanes = VT.getSizeInBits()/128;
4371 unsigned NumLaneElts = NumElts/NumLanes;
4373 assert((NumLaneElts == 2 || NumLaneElts == 4 || NumLaneElts == 8) &&
4374 "Only supports 2, 4 or 8 elements per lane");
4376 unsigned Shift = (NumLaneElts >= 4) ? 1 : 0;
4378 for (unsigned i = 0; i != NumElts; ++i) {
4379 int Elt = N->getMaskElt(i);
4380 if (Elt < 0) continue;
4381 Elt &= NumLaneElts - 1;
4382 unsigned ShAmt = (i << Shift) % 8;
4383 Mask |= Elt << ShAmt;
4389 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
4390 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
4391 static unsigned getShufflePSHUFHWImmediate(ShuffleVectorSDNode *N) {
4392 MVT VT = N->getSimpleValueType(0);
4394 assert((VT == MVT::v8i16 || VT == MVT::v16i16) &&
4395 "Unsupported vector type for PSHUFHW");
4397 unsigned NumElts = VT.getVectorNumElements();
4400 for (unsigned l = 0; l != NumElts; l += 8) {
4401 // 8 nodes per lane, but we only care about the last 4.
4402 for (unsigned i = 0; i < 4; ++i) {
4403 int Elt = N->getMaskElt(l+i+4);
4404 if (Elt < 0) continue;
4405 Elt &= 0x3; // only 2-bits.
4406 Mask |= Elt << (i * 2);
4413 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
4414 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
4415 static unsigned getShufflePSHUFLWImmediate(ShuffleVectorSDNode *N) {
4416 MVT VT = N->getSimpleValueType(0);
4418 assert((VT == MVT::v8i16 || VT == MVT::v16i16) &&
4419 "Unsupported vector type for PSHUFHW");
4421 unsigned NumElts = VT.getVectorNumElements();
4424 for (unsigned l = 0; l != NumElts; l += 8) {
4425 // 8 nodes per lane, but we only care about the first 4.
4426 for (unsigned i = 0; i < 4; ++i) {
4427 int Elt = N->getMaskElt(l+i);
4428 if (Elt < 0) continue;
4429 Elt &= 0x3; // only 2-bits
4430 Mask |= Elt << (i * 2);
4437 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
4438 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
4439 static unsigned getShufflePALIGNRImmediate(ShuffleVectorSDNode *SVOp) {
4440 MVT VT = SVOp->getSimpleValueType(0);
4441 unsigned EltSize = VT.is512BitVector() ? 1 :
4442 VT.getVectorElementType().getSizeInBits() >> 3;
4444 unsigned NumElts = VT.getVectorNumElements();
4445 unsigned NumLanes = VT.is512BitVector() ? 1 : VT.getSizeInBits()/128;
4446 unsigned NumLaneElts = NumElts/NumLanes;
4450 for (i = 0; i != NumElts; ++i) {
4451 Val = SVOp->getMaskElt(i);
4455 if (Val >= (int)NumElts)
4456 Val -= NumElts - NumLaneElts;
4458 assert(Val - i > 0 && "PALIGNR imm should be positive");
4459 return (Val - i) * EltSize;
4462 static unsigned getExtractVEXTRACTImmediate(SDNode *N, unsigned vecWidth) {
4463 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4464 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4465 llvm_unreachable("Illegal extract subvector for VEXTRACT");
4468 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4470 MVT VecVT = N->getOperand(0).getSimpleValueType();
4471 MVT ElVT = VecVT.getVectorElementType();
4473 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4474 return Index / NumElemsPerChunk;
4477 static unsigned getInsertVINSERTImmediate(SDNode *N, unsigned vecWidth) {
4478 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4479 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4480 llvm_unreachable("Illegal insert subvector for VINSERT");
4483 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4485 MVT VecVT = N->getSimpleValueType(0);
4486 MVT ElVT = VecVT.getVectorElementType();
4488 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4489 return Index / NumElemsPerChunk;
4492 /// getExtractVEXTRACT128Immediate - Return the appropriate immediate
4493 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF128
4494 /// and VINSERTI128 instructions.
4495 unsigned X86::getExtractVEXTRACT128Immediate(SDNode *N) {
4496 return getExtractVEXTRACTImmediate(N, 128);
4499 /// getExtractVEXTRACT256Immediate - Return the appropriate immediate
4500 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF64x4
4501 /// and VINSERTI64x4 instructions.
4502 unsigned X86::getExtractVEXTRACT256Immediate(SDNode *N) {
4503 return getExtractVEXTRACTImmediate(N, 256);
4506 /// getInsertVINSERT128Immediate - Return the appropriate immediate
4507 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF128
4508 /// and VINSERTI128 instructions.
4509 unsigned X86::getInsertVINSERT128Immediate(SDNode *N) {
4510 return getInsertVINSERTImmediate(N, 128);
4513 /// getInsertVINSERT256Immediate - Return the appropriate immediate
4514 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF46x4
4515 /// and VINSERTI64x4 instructions.
4516 unsigned X86::getInsertVINSERT256Immediate(SDNode *N) {
4517 return getInsertVINSERTImmediate(N, 256);
4520 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
4522 bool X86::isZeroNode(SDValue Elt) {
4523 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Elt))
4524 return CN->isNullValue();
4525 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Elt))
4526 return CFP->getValueAPF().isPosZero();
4530 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
4531 /// their permute mask.
4532 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
4533 SelectionDAG &DAG) {
4534 MVT VT = SVOp->getSimpleValueType(0);
4535 unsigned NumElems = VT.getVectorNumElements();
4536 SmallVector<int, 8> MaskVec;
4538 for (unsigned i = 0; i != NumElems; ++i) {
4539 int Idx = SVOp->getMaskElt(i);
4541 if (Idx < (int)NumElems)
4546 MaskVec.push_back(Idx);
4548 return DAG.getVectorShuffle(VT, SDLoc(SVOp), SVOp->getOperand(1),
4549 SVOp->getOperand(0), &MaskVec[0]);
4552 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
4553 /// match movhlps. The lower half elements should come from upper half of
4554 /// V1 (and in order), and the upper half elements should come from the upper
4555 /// half of V2 (and in order).
4556 static bool ShouldXformToMOVHLPS(ArrayRef<int> Mask, MVT VT) {
4557 if (!VT.is128BitVector())
4559 if (VT.getVectorNumElements() != 4)
4561 for (unsigned i = 0, e = 2; i != e; ++i)
4562 if (!isUndefOrEqual(Mask[i], i+2))
4564 for (unsigned i = 2; i != 4; ++i)
4565 if (!isUndefOrEqual(Mask[i], i+4))
4570 /// isScalarLoadToVector - Returns true if the node is a scalar load that
4571 /// is promoted to a vector. It also returns the LoadSDNode by reference if
4573 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
4574 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
4576 N = N->getOperand(0).getNode();
4577 if (!ISD::isNON_EXTLoad(N))
4580 *LD = cast<LoadSDNode>(N);
4584 // Test whether the given value is a vector value which will be legalized
4586 static bool WillBeConstantPoolLoad(SDNode *N) {
4587 if (N->getOpcode() != ISD::BUILD_VECTOR)
4590 // Check for any non-constant elements.
4591 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
4592 switch (N->getOperand(i).getNode()->getOpcode()) {
4594 case ISD::ConstantFP:
4601 // Vectors of all-zeros and all-ones are materialized with special
4602 // instructions rather than being loaded.
4603 return !ISD::isBuildVectorAllZeros(N) &&
4604 !ISD::isBuildVectorAllOnes(N);
4607 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
4608 /// match movlp{s|d}. The lower half elements should come from lower half of
4609 /// V1 (and in order), and the upper half elements should come from the upper
4610 /// half of V2 (and in order). And since V1 will become the source of the
4611 /// MOVLP, it must be either a vector load or a scalar load to vector.
4612 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
4613 ArrayRef<int> Mask, MVT VT) {
4614 if (!VT.is128BitVector())
4617 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
4619 // Is V2 is a vector load, don't do this transformation. We will try to use
4620 // load folding shufps op.
4621 if (ISD::isNON_EXTLoad(V2) || WillBeConstantPoolLoad(V2))
4624 unsigned NumElems = VT.getVectorNumElements();
4626 if (NumElems != 2 && NumElems != 4)
4628 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
4629 if (!isUndefOrEqual(Mask[i], i))
4631 for (unsigned i = NumElems/2, e = NumElems; i != e; ++i)
4632 if (!isUndefOrEqual(Mask[i], i+NumElems))
4637 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
4639 static bool isSplatVector(SDNode *N) {
4640 if (N->getOpcode() != ISD::BUILD_VECTOR)
4643 SDValue SplatValue = N->getOperand(0);
4644 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
4645 if (N->getOperand(i) != SplatValue)
4650 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
4651 /// to an zero vector.
4652 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
4653 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
4654 SDValue V1 = N->getOperand(0);
4655 SDValue V2 = N->getOperand(1);
4656 unsigned NumElems = N->getValueType(0).getVectorNumElements();
4657 for (unsigned i = 0; i != NumElems; ++i) {
4658 int Idx = N->getMaskElt(i);
4659 if (Idx >= (int)NumElems) {
4660 unsigned Opc = V2.getOpcode();
4661 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
4663 if (Opc != ISD::BUILD_VECTOR ||
4664 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
4666 } else if (Idx >= 0) {
4667 unsigned Opc = V1.getOpcode();
4668 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
4670 if (Opc != ISD::BUILD_VECTOR ||
4671 !X86::isZeroNode(V1.getOperand(Idx)))
4678 /// getZeroVector - Returns a vector of specified type with all zero elements.
4680 static SDValue getZeroVector(EVT VT, const X86Subtarget *Subtarget,
4681 SelectionDAG &DAG, SDLoc dl) {
4682 assert(VT.isVector() && "Expected a vector type");
4684 // Always build SSE zero vectors as <4 x i32> bitcasted
4685 // to their dest type. This ensures they get CSE'd.
4687 if (VT.is128BitVector()) { // SSE
4688 if (Subtarget->hasSSE2()) { // SSE2
4689 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
4690 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4692 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
4693 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
4695 } else if (VT.is256BitVector()) { // AVX
4696 if (Subtarget->hasInt256()) { // AVX2
4697 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
4698 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4699 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops,
4700 array_lengthof(Ops));
4702 // 256-bit logic and arithmetic instructions in AVX are all
4703 // floating-point, no support for integer ops. Emit fp zeroed vectors.
4704 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
4705 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4706 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops,
4707 array_lengthof(Ops));
4709 } else if (VT.is512BitVector()) { // AVX-512
4710 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
4711 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
4712 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4713 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops, 16);
4715 llvm_unreachable("Unexpected vector type");
4717 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
4720 /// getOnesVector - Returns a vector of specified type with all bits set.
4721 /// Always build ones vectors as <4 x i32> or <8 x i32>. For 256-bit types with
4722 /// no AVX2 supprt, use two <4 x i32> inserted in a <8 x i32> appropriately.
4723 /// Then bitcast to their original type, ensuring they get CSE'd.
4724 static SDValue getOnesVector(MVT VT, bool HasInt256, SelectionDAG &DAG,
4726 assert(VT.isVector() && "Expected a vector type");
4728 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
4730 if (VT.is256BitVector()) {
4731 if (HasInt256) { // AVX2
4732 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4733 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops,
4734 array_lengthof(Ops));
4736 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4737 Vec = Concat128BitVectors(Vec, Vec, MVT::v8i32, 8, DAG, dl);
4739 } else if (VT.is128BitVector()) {
4740 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4742 llvm_unreachable("Unexpected vector type");
4744 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
4747 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
4748 /// that point to V2 points to its first element.
4749 static void NormalizeMask(SmallVectorImpl<int> &Mask, unsigned NumElems) {
4750 for (unsigned i = 0; i != NumElems; ++i) {
4751 if (Mask[i] > (int)NumElems) {
4757 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
4758 /// operation of specified width.
4759 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1,
4761 unsigned NumElems = VT.getVectorNumElements();
4762 SmallVector<int, 8> Mask;
4763 Mask.push_back(NumElems);
4764 for (unsigned i = 1; i != NumElems; ++i)
4766 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4769 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
4770 static SDValue getUnpackl(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4772 unsigned NumElems = VT.getVectorNumElements();
4773 SmallVector<int, 8> Mask;
4774 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
4776 Mask.push_back(i + NumElems);
4778 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4781 /// getUnpackh - Returns a vector_shuffle node for an unpackh operation.
4782 static SDValue getUnpackh(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4784 unsigned NumElems = VT.getVectorNumElements();
4785 SmallVector<int, 8> Mask;
4786 for (unsigned i = 0, Half = NumElems/2; i != Half; ++i) {
4787 Mask.push_back(i + Half);
4788 Mask.push_back(i + NumElems + Half);
4790 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4793 // PromoteSplati8i16 - All i16 and i8 vector types can't be used directly by
4794 // a generic shuffle instruction because the target has no such instructions.
4795 // Generate shuffles which repeat i16 and i8 several times until they can be
4796 // represented by v4f32 and then be manipulated by target suported shuffles.
4797 static SDValue PromoteSplati8i16(SDValue V, SelectionDAG &DAG, int &EltNo) {
4798 MVT VT = V.getSimpleValueType();
4799 int NumElems = VT.getVectorNumElements();
4802 while (NumElems > 4) {
4803 if (EltNo < NumElems/2) {
4804 V = getUnpackl(DAG, dl, VT, V, V);
4806 V = getUnpackh(DAG, dl, VT, V, V);
4807 EltNo -= NumElems/2;
4814 /// getLegalSplat - Generate a legal splat with supported x86 shuffles
4815 static SDValue getLegalSplat(SelectionDAG &DAG, SDValue V, int EltNo) {
4816 MVT VT = V.getSimpleValueType();
4819 if (VT.is128BitVector()) {
4820 V = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, V);
4821 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
4822 V = DAG.getVectorShuffle(MVT::v4f32, dl, V, DAG.getUNDEF(MVT::v4f32),
4824 } else if (VT.is256BitVector()) {
4825 // To use VPERMILPS to splat scalars, the second half of indicies must
4826 // refer to the higher part, which is a duplication of the lower one,
4827 // because VPERMILPS can only handle in-lane permutations.
4828 int SplatMask[8] = { EltNo, EltNo, EltNo, EltNo,
4829 EltNo+4, EltNo+4, EltNo+4, EltNo+4 };
4831 V = DAG.getNode(ISD::BITCAST, dl, MVT::v8f32, V);
4832 V = DAG.getVectorShuffle(MVT::v8f32, dl, V, DAG.getUNDEF(MVT::v8f32),
4835 llvm_unreachable("Vector size not supported");
4837 return DAG.getNode(ISD::BITCAST, dl, VT, V);
4840 /// PromoteSplat - Splat is promoted to target supported vector shuffles.
4841 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG) {
4842 MVT SrcVT = SV->getSimpleValueType(0);
4843 SDValue V1 = SV->getOperand(0);
4846 int EltNo = SV->getSplatIndex();
4847 int NumElems = SrcVT.getVectorNumElements();
4848 bool Is256BitVec = SrcVT.is256BitVector();
4850 assert(((SrcVT.is128BitVector() && NumElems > 4) || Is256BitVec) &&
4851 "Unknown how to promote splat for type");
4853 // Extract the 128-bit part containing the splat element and update
4854 // the splat element index when it refers to the higher register.
4856 V1 = Extract128BitVector(V1, EltNo, DAG, dl);
4857 if (EltNo >= NumElems/2)
4858 EltNo -= NumElems/2;
4861 // All i16 and i8 vector types can't be used directly by a generic shuffle
4862 // instruction because the target has no such instruction. Generate shuffles
4863 // which repeat i16 and i8 several times until they fit in i32, and then can
4864 // be manipulated by target suported shuffles.
4865 MVT EltVT = SrcVT.getVectorElementType();
4866 if (EltVT == MVT::i8 || EltVT == MVT::i16)
4867 V1 = PromoteSplati8i16(V1, DAG, EltNo);
4869 // Recreate the 256-bit vector and place the same 128-bit vector
4870 // into the low and high part. This is necessary because we want
4871 // to use VPERM* to shuffle the vectors
4873 V1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, SrcVT, V1, V1);
4876 return getLegalSplat(DAG, V1, EltNo);
4879 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
4880 /// vector of zero or undef vector. This produces a shuffle where the low
4881 /// element of V2 is swizzled into the zero/undef vector, landing at element
4882 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
4883 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
4885 const X86Subtarget *Subtarget,
4886 SelectionDAG &DAG) {
4887 MVT VT = V2.getSimpleValueType();
4889 ? getZeroVector(VT, Subtarget, DAG, SDLoc(V2)) : DAG.getUNDEF(VT);
4890 unsigned NumElems = VT.getVectorNumElements();
4891 SmallVector<int, 16> MaskVec;
4892 for (unsigned i = 0; i != NumElems; ++i)
4893 // If this is the insertion idx, put the low elt of V2 here.
4894 MaskVec.push_back(i == Idx ? NumElems : i);
4895 return DAG.getVectorShuffle(VT, SDLoc(V2), V1, V2, &MaskVec[0]);
4898 /// getTargetShuffleMask - Calculates the shuffle mask corresponding to the
4899 /// target specific opcode. Returns true if the Mask could be calculated.
4900 /// Sets IsUnary to true if only uses one source.
4901 static bool getTargetShuffleMask(SDNode *N, MVT VT,
4902 SmallVectorImpl<int> &Mask, bool &IsUnary) {
4903 unsigned NumElems = VT.getVectorNumElements();
4907 switch(N->getOpcode()) {
4909 ImmN = N->getOperand(N->getNumOperands()-1);
4910 DecodeSHUFPMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4912 case X86ISD::UNPCKH:
4913 DecodeUNPCKHMask(VT, Mask);
4915 case X86ISD::UNPCKL:
4916 DecodeUNPCKLMask(VT, Mask);
4918 case X86ISD::MOVHLPS:
4919 DecodeMOVHLPSMask(NumElems, Mask);
4921 case X86ISD::MOVLHPS:
4922 DecodeMOVLHPSMask(NumElems, Mask);
4924 case X86ISD::PALIGNR:
4925 ImmN = N->getOperand(N->getNumOperands()-1);
4926 DecodePALIGNRMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4928 case X86ISD::PSHUFD:
4929 case X86ISD::VPERMILP:
4930 ImmN = N->getOperand(N->getNumOperands()-1);
4931 DecodePSHUFMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4934 case X86ISD::PSHUFHW:
4935 ImmN = N->getOperand(N->getNumOperands()-1);
4936 DecodePSHUFHWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4939 case X86ISD::PSHUFLW:
4940 ImmN = N->getOperand(N->getNumOperands()-1);
4941 DecodePSHUFLWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4944 case X86ISD::VPERMI:
4945 ImmN = N->getOperand(N->getNumOperands()-1);
4946 DecodeVPERMMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4950 case X86ISD::MOVSD: {
4951 // The index 0 always comes from the first element of the second source,
4952 // this is why MOVSS and MOVSD are used in the first place. The other
4953 // elements come from the other positions of the first source vector
4954 Mask.push_back(NumElems);
4955 for (unsigned i = 1; i != NumElems; ++i) {
4960 case X86ISD::VPERM2X128:
4961 ImmN = N->getOperand(N->getNumOperands()-1);
4962 DecodeVPERM2X128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4963 if (Mask.empty()) return false;
4965 case X86ISD::MOVDDUP:
4966 case X86ISD::MOVLHPD:
4967 case X86ISD::MOVLPD:
4968 case X86ISD::MOVLPS:
4969 case X86ISD::MOVSHDUP:
4970 case X86ISD::MOVSLDUP:
4971 // Not yet implemented
4973 default: llvm_unreachable("unknown target shuffle node");
4979 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
4980 /// element of the result of the vector shuffle.
4981 static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG,
4984 return SDValue(); // Limit search depth.
4986 SDValue V = SDValue(N, 0);
4987 EVT VT = V.getValueType();
4988 unsigned Opcode = V.getOpcode();
4990 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
4991 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
4992 int Elt = SV->getMaskElt(Index);
4995 return DAG.getUNDEF(VT.getVectorElementType());
4997 unsigned NumElems = VT.getVectorNumElements();
4998 SDValue NewV = (Elt < (int)NumElems) ? SV->getOperand(0)
4999 : SV->getOperand(1);
5000 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG, Depth+1);
5003 // Recurse into target specific vector shuffles to find scalars.
5004 if (isTargetShuffle(Opcode)) {
5005 MVT ShufVT = V.getSimpleValueType();
5006 unsigned NumElems = ShufVT.getVectorNumElements();
5007 SmallVector<int, 16> ShuffleMask;
5010 if (!getTargetShuffleMask(N, ShufVT, ShuffleMask, IsUnary))
5013 int Elt = ShuffleMask[Index];
5015 return DAG.getUNDEF(ShufVT.getVectorElementType());
5017 SDValue NewV = (Elt < (int)NumElems) ? N->getOperand(0)
5019 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG,
5023 // Actual nodes that may contain scalar elements
5024 if (Opcode == ISD::BITCAST) {
5025 V = V.getOperand(0);
5026 EVT SrcVT = V.getValueType();
5027 unsigned NumElems = VT.getVectorNumElements();
5029 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
5033 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
5034 return (Index == 0) ? V.getOperand(0)
5035 : DAG.getUNDEF(VT.getVectorElementType());
5037 if (V.getOpcode() == ISD::BUILD_VECTOR)
5038 return V.getOperand(Index);
5043 /// getNumOfConsecutiveZeros - Return the number of elements of a vector
5044 /// shuffle operation which come from a consecutively from a zero. The
5045 /// search can start in two different directions, from left or right.
5046 /// We count undefs as zeros until PreferredNum is reached.
5047 static unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp,
5048 unsigned NumElems, bool ZerosFromLeft,
5050 unsigned PreferredNum = -1U) {
5051 unsigned NumZeros = 0;
5052 for (unsigned i = 0; i != NumElems; ++i) {
5053 unsigned Index = ZerosFromLeft ? i : NumElems - i - 1;
5054 SDValue Elt = getShuffleScalarElt(SVOp, Index, DAG, 0);
5058 if (X86::isZeroNode(Elt))
5060 else if (Elt.getOpcode() == ISD::UNDEF) // Undef as zero up to PreferredNum.
5061 NumZeros = std::min(NumZeros + 1, PreferredNum);
5069 /// isShuffleMaskConsecutive - Check if the shuffle mask indicies [MaskI, MaskE)
5070 /// correspond consecutively to elements from one of the vector operands,
5071 /// starting from its index OpIdx. Also tell OpNum which source vector operand.
5073 bool isShuffleMaskConsecutive(ShuffleVectorSDNode *SVOp,
5074 unsigned MaskI, unsigned MaskE, unsigned OpIdx,
5075 unsigned NumElems, unsigned &OpNum) {
5076 bool SeenV1 = false;
5077 bool SeenV2 = false;
5079 for (unsigned i = MaskI; i != MaskE; ++i, ++OpIdx) {
5080 int Idx = SVOp->getMaskElt(i);
5081 // Ignore undef indicies
5085 if (Idx < (int)NumElems)
5090 // Only accept consecutive elements from the same vector
5091 if ((Idx % NumElems != OpIdx) || (SeenV1 && SeenV2))
5095 OpNum = SeenV1 ? 0 : 1;
5099 /// isVectorShiftRight - Returns true if the shuffle can be implemented as a
5100 /// logical left shift of a vector.
5101 static bool isVectorShiftRight(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
5102 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
5104 SVOp->getSimpleValueType(0).getVectorNumElements();
5105 unsigned NumZeros = getNumOfConsecutiveZeros(
5106 SVOp, NumElems, false /* check zeros from right */, DAG,
5107 SVOp->getMaskElt(0));
5113 // Considering the elements in the mask that are not consecutive zeros,
5114 // check if they consecutively come from only one of the source vectors.
5116 // V1 = {X, A, B, C} 0
5118 // vector_shuffle V1, V2 <1, 2, 3, X>
5120 if (!isShuffleMaskConsecutive(SVOp,
5121 0, // Mask Start Index
5122 NumElems-NumZeros, // Mask End Index(exclusive)
5123 NumZeros, // Where to start looking in the src vector
5124 NumElems, // Number of elements in vector
5125 OpSrc)) // Which source operand ?
5130 ShVal = SVOp->getOperand(OpSrc);
5134 /// isVectorShiftLeft - Returns true if the shuffle can be implemented as a
5135 /// logical left shift of a vector.
5136 static bool isVectorShiftLeft(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
5137 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
5139 SVOp->getSimpleValueType(0).getVectorNumElements();
5140 unsigned NumZeros = getNumOfConsecutiveZeros(
5141 SVOp, NumElems, true /* check zeros from left */, DAG,
5142 NumElems - SVOp->getMaskElt(NumElems - 1) - 1);
5148 // Considering the elements in the mask that are not consecutive zeros,
5149 // check if they consecutively come from only one of the source vectors.
5151 // 0 { A, B, X, X } = V2
5153 // vector_shuffle V1, V2 <X, X, 4, 5>
5155 if (!isShuffleMaskConsecutive(SVOp,
5156 NumZeros, // Mask Start Index
5157 NumElems, // Mask End Index(exclusive)
5158 0, // Where to start looking in the src vector
5159 NumElems, // Number of elements in vector
5160 OpSrc)) // Which source operand ?
5165 ShVal = SVOp->getOperand(OpSrc);
5169 /// isVectorShift - Returns true if the shuffle can be implemented as a
5170 /// logical left or right shift of a vector.
5171 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
5172 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
5173 // Although the logic below support any bitwidth size, there are no
5174 // shift instructions which handle more than 128-bit vectors.
5175 if (!SVOp->getSimpleValueType(0).is128BitVector())
5178 if (isVectorShiftLeft(SVOp, DAG, isLeft, ShVal, ShAmt) ||
5179 isVectorShiftRight(SVOp, DAG, isLeft, ShVal, ShAmt))
5185 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
5187 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
5188 unsigned NumNonZero, unsigned NumZero,
5190 const X86Subtarget* Subtarget,
5191 const TargetLowering &TLI) {
5198 for (unsigned i = 0; i < 16; ++i) {
5199 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
5200 if (ThisIsNonZero && First) {
5202 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
5204 V = DAG.getUNDEF(MVT::v8i16);
5209 SDValue ThisElt(0, 0), LastElt(0, 0);
5210 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
5211 if (LastIsNonZero) {
5212 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
5213 MVT::i16, Op.getOperand(i-1));
5215 if (ThisIsNonZero) {
5216 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
5217 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
5218 ThisElt, DAG.getConstant(8, MVT::i8));
5220 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
5224 if (ThisElt.getNode())
5225 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
5226 DAG.getIntPtrConstant(i/2));
5230 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V);
5233 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
5235 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
5236 unsigned NumNonZero, unsigned NumZero,
5238 const X86Subtarget* Subtarget,
5239 const TargetLowering &TLI) {
5246 for (unsigned i = 0; i < 8; ++i) {
5247 bool isNonZero = (NonZeros & (1 << i)) != 0;
5251 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
5253 V = DAG.getUNDEF(MVT::v8i16);
5256 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
5257 MVT::v8i16, V, Op.getOperand(i),
5258 DAG.getIntPtrConstant(i));
5265 /// getVShift - Return a vector logical shift node.
5267 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
5268 unsigned NumBits, SelectionDAG &DAG,
5269 const TargetLowering &TLI, SDLoc dl) {
5270 assert(VT.is128BitVector() && "Unknown type for VShift");
5271 EVT ShVT = MVT::v2i64;
5272 unsigned Opc = isLeft ? X86ISD::VSHLDQ : X86ISD::VSRLDQ;
5273 SrcOp = DAG.getNode(ISD::BITCAST, dl, ShVT, SrcOp);
5274 return DAG.getNode(ISD::BITCAST, dl, VT,
5275 DAG.getNode(Opc, dl, ShVT, SrcOp,
5276 DAG.getConstant(NumBits,
5277 TLI.getScalarShiftAmountTy(SrcOp.getValueType()))));
5281 LowerAsSplatVectorLoad(SDValue SrcOp, MVT VT, SDLoc dl, SelectionDAG &DAG) {
5283 // Check if the scalar load can be widened into a vector load. And if
5284 // the address is "base + cst" see if the cst can be "absorbed" into
5285 // the shuffle mask.
5286 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
5287 SDValue Ptr = LD->getBasePtr();
5288 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
5290 EVT PVT = LD->getValueType(0);
5291 if (PVT != MVT::i32 && PVT != MVT::f32)
5296 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
5297 FI = FINode->getIndex();
5299 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
5300 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
5301 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
5302 Offset = Ptr.getConstantOperandVal(1);
5303 Ptr = Ptr.getOperand(0);
5308 // FIXME: 256-bit vector instructions don't require a strict alignment,
5309 // improve this code to support it better.
5310 unsigned RequiredAlign = VT.getSizeInBits()/8;
5311 SDValue Chain = LD->getChain();
5312 // Make sure the stack object alignment is at least 16 or 32.
5313 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5314 if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
5315 if (MFI->isFixedObjectIndex(FI)) {
5316 // Can't change the alignment. FIXME: It's possible to compute
5317 // the exact stack offset and reference FI + adjust offset instead.
5318 // If someone *really* cares about this. That's the way to implement it.
5321 MFI->setObjectAlignment(FI, RequiredAlign);
5325 // (Offset % 16 or 32) must be multiple of 4. Then address is then
5326 // Ptr + (Offset & ~15).
5329 if ((Offset % RequiredAlign) & 3)
5331 int64_t StartOffset = Offset & ~(RequiredAlign-1);
5333 Ptr = DAG.getNode(ISD::ADD, SDLoc(Ptr), Ptr.getValueType(),
5334 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
5336 int EltNo = (Offset - StartOffset) >> 2;
5337 unsigned NumElems = VT.getVectorNumElements();
5339 EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
5340 SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
5341 LD->getPointerInfo().getWithOffset(StartOffset),
5342 false, false, false, 0);
5344 SmallVector<int, 8> Mask;
5345 for (unsigned i = 0; i != NumElems; ++i)
5346 Mask.push_back(EltNo);
5348 return DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), &Mask[0]);
5354 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
5355 /// vector of type 'VT', see if the elements can be replaced by a single large
5356 /// load which has the same value as a build_vector whose operands are 'elts'.
5358 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
5360 /// FIXME: we'd also like to handle the case where the last elements are zero
5361 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
5362 /// There's even a handy isZeroNode for that purpose.
5363 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
5364 SDLoc &DL, SelectionDAG &DAG) {
5365 EVT EltVT = VT.getVectorElementType();
5366 unsigned NumElems = Elts.size();
5368 LoadSDNode *LDBase = NULL;
5369 unsigned LastLoadedElt = -1U;
5371 // For each element in the initializer, see if we've found a load or an undef.
5372 // If we don't find an initial load element, or later load elements are
5373 // non-consecutive, bail out.
5374 for (unsigned i = 0; i < NumElems; ++i) {
5375 SDValue Elt = Elts[i];
5377 if (!Elt.getNode() ||
5378 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
5381 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
5383 LDBase = cast<LoadSDNode>(Elt.getNode());
5387 if (Elt.getOpcode() == ISD::UNDEF)
5390 LoadSDNode *LD = cast<LoadSDNode>(Elt);
5391 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
5396 // If we have found an entire vector of loads and undefs, then return a large
5397 // load of the entire vector width starting at the base pointer. If we found
5398 // consecutive loads for the low half, generate a vzext_load node.
5399 if (LastLoadedElt == NumElems - 1) {
5400 SDValue NewLd = SDValue();
5401 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
5402 NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
5403 LDBase->getPointerInfo(),
5404 LDBase->isVolatile(), LDBase->isNonTemporal(),
5405 LDBase->isInvariant(), 0);
5406 NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
5407 LDBase->getPointerInfo(),
5408 LDBase->isVolatile(), LDBase->isNonTemporal(),
5409 LDBase->isInvariant(), LDBase->getAlignment());
5411 if (LDBase->hasAnyUseOfValue(1)) {
5412 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5414 SDValue(NewLd.getNode(), 1));
5415 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5416 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5417 SDValue(NewLd.getNode(), 1));
5422 if (NumElems == 4 && LastLoadedElt == 1 &&
5423 DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
5424 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
5425 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
5427 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys, Ops,
5428 array_lengthof(Ops), MVT::i64,
5429 LDBase->getPointerInfo(),
5430 LDBase->getAlignment(),
5431 false/*isVolatile*/, true/*ReadMem*/,
5434 // Make sure the newly-created LOAD is in the same position as LDBase in
5435 // terms of dependency. We create a TokenFactor for LDBase and ResNode, and
5436 // update uses of LDBase's output chain to use the TokenFactor.
5437 if (LDBase->hasAnyUseOfValue(1)) {
5438 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5439 SDValue(LDBase, 1), SDValue(ResNode.getNode(), 1));
5440 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5441 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5442 SDValue(ResNode.getNode(), 1));
5445 return DAG.getNode(ISD::BITCAST, DL, VT, ResNode);
5450 /// LowerVectorBroadcast - Attempt to use the vbroadcast instruction
5451 /// to generate a splat value for the following cases:
5452 /// 1. A splat BUILD_VECTOR which uses a single scalar load, or a constant.
5453 /// 2. A splat shuffle which uses a scalar_to_vector node which comes from
5454 /// a scalar load, or a constant.
5455 /// The VBROADCAST node is returned when a pattern is found,
5456 /// or SDValue() otherwise.
5457 static SDValue LowerVectorBroadcast(SDValue Op, const X86Subtarget* Subtarget,
5458 SelectionDAG &DAG) {
5459 if (!Subtarget->hasFp256())
5462 MVT VT = Op.getSimpleValueType();
5465 assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&
5466 "Unsupported vector type for broadcast.");
5471 switch (Op.getOpcode()) {
5473 // Unknown pattern found.
5476 case ISD::BUILD_VECTOR: {
5477 // The BUILD_VECTOR node must be a splat.
5478 if (!isSplatVector(Op.getNode()))
5481 Ld = Op.getOperand(0);
5482 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5483 Ld.getOpcode() == ISD::ConstantFP);
5485 // The suspected load node has several users. Make sure that all
5486 // of its users are from the BUILD_VECTOR node.
5487 // Constants may have multiple users.
5488 if (!ConstSplatVal && !Ld->hasNUsesOfValue(VT.getVectorNumElements(), 0))
5493 case ISD::VECTOR_SHUFFLE: {
5494 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5496 // Shuffles must have a splat mask where the first element is
5498 if ((!SVOp->isSplat()) || SVOp->getMaskElt(0) != 0)
5501 SDValue Sc = Op.getOperand(0);
5502 if (Sc.getOpcode() != ISD::SCALAR_TO_VECTOR &&
5503 Sc.getOpcode() != ISD::BUILD_VECTOR) {
5505 if (!Subtarget->hasInt256())
5508 // Use the register form of the broadcast instruction available on AVX2.
5509 if (VT.getSizeInBits() >= 256)
5510 Sc = Extract128BitVector(Sc, 0, DAG, dl);
5511 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Sc);
5514 Ld = Sc.getOperand(0);
5515 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5516 Ld.getOpcode() == ISD::ConstantFP);
5518 // The scalar_to_vector node and the suspected
5519 // load node must have exactly one user.
5520 // Constants may have multiple users.
5522 // AVX-512 has register version of the broadcast
5523 bool hasRegVer = Subtarget->hasAVX512() && VT.is512BitVector() &&
5524 Ld.getValueType().getSizeInBits() >= 32;
5525 if (!ConstSplatVal && ((!Sc.hasOneUse() || !Ld.hasOneUse()) &&
5532 bool IsGE256 = (VT.getSizeInBits() >= 256);
5534 // Handle the broadcasting a single constant scalar from the constant pool
5535 // into a vector. On Sandybridge it is still better to load a constant vector
5536 // from the constant pool and not to broadcast it from a scalar.
5537 if (ConstSplatVal && Subtarget->hasInt256()) {
5538 EVT CVT = Ld.getValueType();
5539 assert(!CVT.isVector() && "Must not broadcast a vector type");
5540 unsigned ScalarSize = CVT.getSizeInBits();
5542 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64)) {
5543 const Constant *C = 0;
5544 if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld))
5545 C = CI->getConstantIntValue();
5546 else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld))
5547 C = CF->getConstantFPValue();
5549 assert(C && "Invalid constant type");
5551 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5552 SDValue CP = DAG.getConstantPool(C, TLI.getPointerTy());
5553 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
5554 Ld = DAG.getLoad(CVT, dl, DAG.getEntryNode(), CP,
5555 MachinePointerInfo::getConstantPool(),
5556 false, false, false, Alignment);
5558 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5562 bool IsLoad = ISD::isNormalLoad(Ld.getNode());
5563 unsigned ScalarSize = Ld.getValueType().getSizeInBits();
5565 // Handle AVX2 in-register broadcasts.
5566 if (!IsLoad && Subtarget->hasInt256() &&
5567 (ScalarSize == 32 || (IsGE256 && ScalarSize == 64)))
5568 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5570 // The scalar source must be a normal load.
5574 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64))
5575 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5577 // The integer check is needed for the 64-bit into 128-bit so it doesn't match
5578 // double since there is no vbroadcastsd xmm
5579 if (Subtarget->hasInt256() && Ld.getValueType().isInteger()) {
5580 if (ScalarSize == 8 || ScalarSize == 16 || ScalarSize == 64)
5581 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5584 // Unsupported broadcast.
5588 static SDValue buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) {
5589 MVT VT = Op.getSimpleValueType();
5591 // Skip if insert_vec_elt is not supported.
5592 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5593 if (!TLI.isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT))
5597 unsigned NumElems = Op.getNumOperands();
5601 SmallVector<unsigned, 4> InsertIndices;
5602 SmallVector<int, 8> Mask(NumElems, -1);
5604 for (unsigned i = 0; i != NumElems; ++i) {
5605 unsigned Opc = Op.getOperand(i).getOpcode();
5607 if (Opc == ISD::UNDEF)
5610 if (Opc != ISD::EXTRACT_VECTOR_ELT) {
5611 // Quit if more than 1 elements need inserting.
5612 if (InsertIndices.size() > 1)
5615 InsertIndices.push_back(i);
5619 SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0);
5620 SDValue ExtIdx = Op.getOperand(i).getOperand(1);
5622 // Quit if extracted from vector of different type.
5623 if (ExtractedFromVec.getValueType() != VT)
5626 // Quit if non-constant index.
5627 if (!isa<ConstantSDNode>(ExtIdx))
5630 if (VecIn1.getNode() == 0)
5631 VecIn1 = ExtractedFromVec;
5632 else if (VecIn1 != ExtractedFromVec) {
5633 if (VecIn2.getNode() == 0)
5634 VecIn2 = ExtractedFromVec;
5635 else if (VecIn2 != ExtractedFromVec)
5636 // Quit if more than 2 vectors to shuffle
5640 unsigned Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue();
5642 if (ExtractedFromVec == VecIn1)
5644 else if (ExtractedFromVec == VecIn2)
5645 Mask[i] = Idx + NumElems;
5648 if (VecIn1.getNode() == 0)
5651 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
5652 SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, &Mask[0]);
5653 for (unsigned i = 0, e = InsertIndices.size(); i != e; ++i) {
5654 unsigned Idx = InsertIndices[i];
5655 NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx),
5656 DAG.getIntPtrConstant(Idx));
5662 // Lower BUILD_VECTOR operation for v8i1 and v16i1 types.
5664 X86TargetLowering::LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const {
5666 MVT VT = Op.getSimpleValueType();
5667 assert((VT.getVectorElementType() == MVT::i1) && (VT.getSizeInBits() <= 16) &&
5668 "Unexpected type in LowerBUILD_VECTORvXi1!");
5671 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5672 SDValue Cst = DAG.getTargetConstant(0, MVT::i1);
5673 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
5674 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
5675 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT,
5676 Ops, VT.getVectorNumElements());
5679 if (ISD::isBuildVectorAllOnes(Op.getNode())) {
5680 SDValue Cst = DAG.getTargetConstant(1, MVT::i1);
5681 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
5682 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
5683 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT,
5684 Ops, VT.getVectorNumElements());
5687 bool AllContants = true;
5688 uint64_t Immediate = 0;
5689 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5690 SDValue In = Op.getOperand(idx);
5691 if (In.getOpcode() == ISD::UNDEF)
5693 if (!isa<ConstantSDNode>(In)) {
5694 AllContants = false;
5697 if (cast<ConstantSDNode>(In)->getZExtValue())
5698 Immediate |= (1ULL << idx);
5702 SDValue FullMask = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1,
5703 DAG.getConstant(Immediate, MVT::i16));
5704 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, FullMask,
5705 DAG.getIntPtrConstant(0));
5708 // Splat vector (with undefs)
5709 SDValue In = Op.getOperand(0);
5710 for (unsigned i = 1, e = Op.getNumOperands(); i != e; ++i) {
5711 if (Op.getOperand(i) != In && Op.getOperand(i).getOpcode() != ISD::UNDEF)
5712 llvm_unreachable("Unsupported predicate operation");
5715 SDValue EFLAGS, X86CC;
5716 if (In.getOpcode() == ISD::SETCC) {
5717 SDValue Op0 = In.getOperand(0);
5718 SDValue Op1 = In.getOperand(1);
5719 ISD::CondCode CC = cast<CondCodeSDNode>(In.getOperand(2))->get();
5720 bool isFP = Op1.getValueType().isFloatingPoint();
5721 unsigned X86CCVal = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
5723 assert(X86CCVal != X86::COND_INVALID && "Unsupported predicate operation");
5725 X86CC = DAG.getConstant(X86CCVal, MVT::i8);
5726 EFLAGS = EmitCmp(Op0, Op1, X86CCVal, DAG);
5727 EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
5728 } else if (In.getOpcode() == X86ISD::SETCC) {
5729 X86CC = In.getOperand(0);
5730 EFLAGS = In.getOperand(1);
5739 // res = allOnes ### CMOVNE -1, %res
5742 MVT InVT = In.getSimpleValueType();
5743 SDValue Bit1 = DAG.getNode(ISD::AND, dl, InVT, In, DAG.getConstant(1, InVT));
5744 EFLAGS = EmitTest(Bit1, X86::COND_NE, DAG);
5745 X86CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5748 if (VT == MVT::v16i1) {
5749 SDValue Cst1 = DAG.getConstant(-1, MVT::i16);
5750 SDValue Cst0 = DAG.getConstant(0, MVT::i16);
5751 SDValue CmovOp = DAG.getNode(X86ISD::CMOV, dl, MVT::i16,
5752 Cst0, Cst1, X86CC, EFLAGS);
5753 return DAG.getNode(ISD::BITCAST, dl, VT, CmovOp);
5756 if (VT == MVT::v8i1) {
5757 SDValue Cst1 = DAG.getConstant(-1, MVT::i32);
5758 SDValue Cst0 = DAG.getConstant(0, MVT::i32);
5759 SDValue CmovOp = DAG.getNode(X86ISD::CMOV, dl, MVT::i32,
5760 Cst0, Cst1, X86CC, EFLAGS);
5761 CmovOp = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, CmovOp);
5762 return DAG.getNode(ISD::BITCAST, dl, VT, CmovOp);
5764 llvm_unreachable("Unsupported predicate operation");
5768 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5771 MVT VT = Op.getSimpleValueType();
5772 MVT ExtVT = VT.getVectorElementType();
5773 unsigned NumElems = Op.getNumOperands();
5775 // Generate vectors for predicate vectors.
5776 if (VT.getScalarType() == MVT::i1 && Subtarget->hasAVX512())
5777 return LowerBUILD_VECTORvXi1(Op, DAG);
5779 // Vectors containing all zeros can be matched by pxor and xorps later
5780 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5781 // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
5782 // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
5783 if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32)
5786 return getZeroVector(VT, Subtarget, DAG, dl);
5789 // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
5790 // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use
5791 // vpcmpeqd on 256-bit vectors.
5792 if (Subtarget->hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) {
5793 if (VT == MVT::v4i32 || (VT == MVT::v8i32 && Subtarget->hasInt256()))
5796 if (!VT.is512BitVector())
5797 return getOnesVector(VT, Subtarget->hasInt256(), DAG, dl);
5800 SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG);
5801 if (Broadcast.getNode())
5804 unsigned EVTBits = ExtVT.getSizeInBits();
5806 unsigned NumZero = 0;
5807 unsigned NumNonZero = 0;
5808 unsigned NonZeros = 0;
5809 bool IsAllConstants = true;
5810 SmallSet<SDValue, 8> Values;
5811 for (unsigned i = 0; i < NumElems; ++i) {
5812 SDValue Elt = Op.getOperand(i);
5813 if (Elt.getOpcode() == ISD::UNDEF)
5816 if (Elt.getOpcode() != ISD::Constant &&
5817 Elt.getOpcode() != ISD::ConstantFP)
5818 IsAllConstants = false;
5819 if (X86::isZeroNode(Elt))
5822 NonZeros |= (1 << i);
5827 // All undef vector. Return an UNDEF. All zero vectors were handled above.
5828 if (NumNonZero == 0)
5829 return DAG.getUNDEF(VT);
5831 // Special case for single non-zero, non-undef, element.
5832 if (NumNonZero == 1) {
5833 unsigned Idx = countTrailingZeros(NonZeros);
5834 SDValue Item = Op.getOperand(Idx);
5836 // If this is an insertion of an i64 value on x86-32, and if the top bits of
5837 // the value are obviously zero, truncate the value to i32 and do the
5838 // insertion that way. Only do this if the value is non-constant or if the
5839 // value is a constant being inserted into element 0. It is cheaper to do
5840 // a constant pool load than it is to do a movd + shuffle.
5841 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
5842 (!IsAllConstants || Idx == 0)) {
5843 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
5845 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
5846 EVT VecVT = MVT::v4i32;
5847 unsigned VecElts = 4;
5849 // Truncate the value (which may itself be a constant) to i32, and
5850 // convert it to a vector with movd (S2V+shuffle to zero extend).
5851 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
5852 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
5853 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
5855 // Now we have our 32-bit value zero extended in the low element of
5856 // a vector. If Idx != 0, swizzle it into place.
5858 SmallVector<int, 4> Mask;
5859 Mask.push_back(Idx);
5860 for (unsigned i = 1; i != VecElts; ++i)
5862 Item = DAG.getVectorShuffle(VecVT, dl, Item, DAG.getUNDEF(VecVT),
5865 return DAG.getNode(ISD::BITCAST, dl, VT, Item);
5869 // If we have a constant or non-constant insertion into the low element of
5870 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
5871 // the rest of the elements. This will be matched as movd/movq/movss/movsd
5872 // depending on what the source datatype is.
5875 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5877 if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
5878 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
5879 if (VT.is256BitVector() || VT.is512BitVector()) {
5880 SDValue ZeroVec = getZeroVector(VT, Subtarget, DAG, dl);
5881 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, ZeroVec,
5882 Item, DAG.getIntPtrConstant(0));
5884 assert(VT.is128BitVector() && "Expected an SSE value type!");
5885 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5886 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
5887 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
5890 if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
5891 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
5892 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
5893 if (VT.is256BitVector()) {
5894 SDValue ZeroVec = getZeroVector(MVT::v8i32, Subtarget, DAG, dl);
5895 Item = Insert128BitVector(ZeroVec, Item, 0, DAG, dl);
5897 assert(VT.is128BitVector() && "Expected an SSE value type!");
5898 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
5900 return DAG.getNode(ISD::BITCAST, dl, VT, Item);
5904 // Is it a vector logical left shift?
5905 if (NumElems == 2 && Idx == 1 &&
5906 X86::isZeroNode(Op.getOperand(0)) &&
5907 !X86::isZeroNode(Op.getOperand(1))) {
5908 unsigned NumBits = VT.getSizeInBits();
5909 return getVShift(true, VT,
5910 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5911 VT, Op.getOperand(1)),
5912 NumBits/2, DAG, *this, dl);
5915 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
5918 // Otherwise, if this is a vector with i32 or f32 elements, and the element
5919 // is a non-constant being inserted into an element other than the low one,
5920 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
5921 // movd/movss) to move this into the low element, then shuffle it into
5923 if (EVTBits == 32) {
5924 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5926 // Turn it into a shuffle of zero and zero-extended scalar to vector.
5927 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0, Subtarget, DAG);
5928 SmallVector<int, 8> MaskVec;
5929 for (unsigned i = 0; i != NumElems; ++i)
5930 MaskVec.push_back(i == Idx ? 0 : 1);
5931 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
5935 // Splat is obviously ok. Let legalizer expand it to a shuffle.
5936 if (Values.size() == 1) {
5937 if (EVTBits == 32) {
5938 // Instead of a shuffle like this:
5939 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
5940 // Check if it's possible to issue this instead.
5941 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
5942 unsigned Idx = countTrailingZeros(NonZeros);
5943 SDValue Item = Op.getOperand(Idx);
5944 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
5945 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
5950 // A vector full of immediates; various special cases are already
5951 // handled, so this is best done with a single constant-pool load.
5955 // For AVX-length vectors, build the individual 128-bit pieces and use
5956 // shuffles to put them in place.
5957 if (VT.is256BitVector()) {
5958 SmallVector<SDValue, 32> V;
5959 for (unsigned i = 0; i != NumElems; ++i)
5960 V.push_back(Op.getOperand(i));
5962 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
5964 // Build both the lower and upper subvector.
5965 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT, &V[0], NumElems/2);
5966 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT, &V[NumElems / 2],
5969 // Recreate the wider vector with the lower and upper part.
5970 return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
5973 // Let legalizer expand 2-wide build_vectors.
5974 if (EVTBits == 64) {
5975 if (NumNonZero == 1) {
5976 // One half is zero or undef.
5977 unsigned Idx = countTrailingZeros(NonZeros);
5978 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
5979 Op.getOperand(Idx));
5980 return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG);
5985 // If element VT is < 32 bits, convert it to inserts into a zero vector.
5986 if (EVTBits == 8 && NumElems == 16) {
5987 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
5989 if (V.getNode()) return V;
5992 if (EVTBits == 16 && NumElems == 8) {
5993 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
5995 if (V.getNode()) return V;
5998 // If element VT is == 32 bits, turn it into a number of shuffles.
5999 SmallVector<SDValue, 8> V(NumElems);
6000 if (NumElems == 4 && NumZero > 0) {
6001 for (unsigned i = 0; i < 4; ++i) {
6002 bool isZero = !(NonZeros & (1 << i));
6004 V[i] = getZeroVector(VT, Subtarget, DAG, dl);
6006 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6009 for (unsigned i = 0; i < 2; ++i) {
6010 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
6013 V[i] = V[i*2]; // Must be a zero vector.
6016 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
6019 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
6022 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
6027 bool Reverse1 = (NonZeros & 0x3) == 2;
6028 bool Reverse2 = ((NonZeros & (0x3 << 2)) >> 2) == 2;
6032 static_cast<int>(Reverse2 ? NumElems+1 : NumElems),
6033 static_cast<int>(Reverse2 ? NumElems : NumElems+1)
6035 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
6038 if (Values.size() > 1 && VT.is128BitVector()) {
6039 // Check for a build vector of consecutive loads.
6040 for (unsigned i = 0; i < NumElems; ++i)
6041 V[i] = Op.getOperand(i);
6043 // Check for elements which are consecutive loads.
6044 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
6048 // Check for a build vector from mostly shuffle plus few inserting.
6049 SDValue Sh = buildFromShuffleMostly(Op, DAG);
6053 // For SSE 4.1, use insertps to put the high elements into the low element.
6054 if (getSubtarget()->hasSSE41()) {
6056 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
6057 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
6059 Result = DAG.getUNDEF(VT);
6061 for (unsigned i = 1; i < NumElems; ++i) {
6062 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
6063 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
6064 Op.getOperand(i), DAG.getIntPtrConstant(i));
6069 // Otherwise, expand into a number of unpckl*, start by extending each of
6070 // our (non-undef) elements to the full vector width with the element in the
6071 // bottom slot of the vector (which generates no code for SSE).
6072 for (unsigned i = 0; i < NumElems; ++i) {
6073 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
6074 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6076 V[i] = DAG.getUNDEF(VT);
6079 // Next, we iteratively mix elements, e.g. for v4f32:
6080 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
6081 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
6082 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
6083 unsigned EltStride = NumElems >> 1;
6084 while (EltStride != 0) {
6085 for (unsigned i = 0; i < EltStride; ++i) {
6086 // If V[i+EltStride] is undef and this is the first round of mixing,
6087 // then it is safe to just drop this shuffle: V[i] is already in the
6088 // right place, the one element (since it's the first round) being
6089 // inserted as undef can be dropped. This isn't safe for successive
6090 // rounds because they will permute elements within both vectors.
6091 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
6092 EltStride == NumElems/2)
6095 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
6104 // LowerAVXCONCAT_VECTORS - 256-bit AVX can use the vinsertf128 instruction
6105 // to create 256-bit vectors from two other 128-bit ones.
6106 static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
6108 MVT ResVT = Op.getSimpleValueType();
6110 assert((ResVT.is256BitVector() ||
6111 ResVT.is512BitVector()) && "Value type must be 256-/512-bit wide");
6113 SDValue V1 = Op.getOperand(0);
6114 SDValue V2 = Op.getOperand(1);
6115 unsigned NumElems = ResVT.getVectorNumElements();
6116 if(ResVT.is256BitVector())
6117 return Concat128BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6119 return Concat256BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6122 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
6123 assert(Op.getNumOperands() == 2);
6125 // AVX/AVX-512 can use the vinsertf128 instruction to create 256-bit vectors
6126 // from two other 128-bit ones.
6127 return LowerAVXCONCAT_VECTORS(Op, DAG);
6130 // Try to lower a shuffle node into a simple blend instruction.
6132 LowerVECTOR_SHUFFLEtoBlend(ShuffleVectorSDNode *SVOp,
6133 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
6134 SDValue V1 = SVOp->getOperand(0);
6135 SDValue V2 = SVOp->getOperand(1);
6137 MVT VT = SVOp->getSimpleValueType(0);
6138 MVT EltVT = VT.getVectorElementType();
6139 unsigned NumElems = VT.getVectorNumElements();
6141 if (!Subtarget->hasSSE41() || EltVT == MVT::i8)
6143 if (!Subtarget->hasInt256() && VT == MVT::v16i16)
6146 // Check the mask for BLEND and build the value.
6147 unsigned MaskValue = 0;
6148 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
6149 unsigned NumLanes = (NumElems-1)/8 + 1;
6150 unsigned NumElemsInLane = NumElems / NumLanes;
6152 // Blend for v16i16 should be symetric for the both lanes.
6153 for (unsigned i = 0; i < NumElemsInLane; ++i) {
6155 int SndLaneEltIdx = (NumLanes == 2) ?
6156 SVOp->getMaskElt(i + NumElemsInLane) : -1;
6157 int EltIdx = SVOp->getMaskElt(i);
6159 if ((EltIdx < 0 || EltIdx == (int)i) &&
6160 (SndLaneEltIdx < 0 || SndLaneEltIdx == (int)(i + NumElemsInLane)))
6163 if (((unsigned)EltIdx == (i + NumElems)) &&
6164 (SndLaneEltIdx < 0 ||
6165 (unsigned)SndLaneEltIdx == i + NumElems + NumElemsInLane))
6166 MaskValue |= (1<<i);
6171 // Convert i32 vectors to floating point if it is not AVX2.
6172 // AVX2 introduced VPBLENDD instruction for 128 and 256-bit vectors.
6174 if (EltVT == MVT::i64 || (EltVT == MVT::i32 && !Subtarget->hasInt256())) {
6175 BlendVT = MVT::getVectorVT(MVT::getFloatingPointVT(EltVT.getSizeInBits()),
6177 V1 = DAG.getNode(ISD::BITCAST, dl, VT, V1);
6178 V2 = DAG.getNode(ISD::BITCAST, dl, VT, V2);
6181 SDValue Ret = DAG.getNode(X86ISD::BLENDI, dl, BlendVT, V1, V2,
6182 DAG.getConstant(MaskValue, MVT::i32));
6183 return DAG.getNode(ISD::BITCAST, dl, VT, Ret);
6186 // v8i16 shuffles - Prefer shuffles in the following order:
6187 // 1. [all] pshuflw, pshufhw, optional move
6188 // 2. [ssse3] 1 x pshufb
6189 // 3. [ssse3] 2 x pshufb + 1 x por
6190 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
6192 LowerVECTOR_SHUFFLEv8i16(SDValue Op, const X86Subtarget *Subtarget,
6193 SelectionDAG &DAG) {
6194 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
6195 SDValue V1 = SVOp->getOperand(0);
6196 SDValue V2 = SVOp->getOperand(1);
6198 SmallVector<int, 8> MaskVals;
6200 // Determine if more than 1 of the words in each of the low and high quadwords
6201 // of the result come from the same quadword of one of the two inputs. Undef
6202 // mask values count as coming from any quadword, for better codegen.
6203 unsigned LoQuad[] = { 0, 0, 0, 0 };
6204 unsigned HiQuad[] = { 0, 0, 0, 0 };
6205 std::bitset<4> InputQuads;
6206 for (unsigned i = 0; i < 8; ++i) {
6207 unsigned *Quad = i < 4 ? LoQuad : HiQuad;
6208 int EltIdx = SVOp->getMaskElt(i);
6209 MaskVals.push_back(EltIdx);
6218 InputQuads.set(EltIdx / 4);
6221 int BestLoQuad = -1;
6222 unsigned MaxQuad = 1;
6223 for (unsigned i = 0; i < 4; ++i) {
6224 if (LoQuad[i] > MaxQuad) {
6226 MaxQuad = LoQuad[i];
6230 int BestHiQuad = -1;
6232 for (unsigned i = 0; i < 4; ++i) {
6233 if (HiQuad[i] > MaxQuad) {
6235 MaxQuad = HiQuad[i];
6239 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
6240 // of the two input vectors, shuffle them into one input vector so only a
6241 // single pshufb instruction is necessary. If There are more than 2 input
6242 // quads, disable the next transformation since it does not help SSSE3.
6243 bool V1Used = InputQuads[0] || InputQuads[1];
6244 bool V2Used = InputQuads[2] || InputQuads[3];
6245 if (Subtarget->hasSSSE3()) {
6246 if (InputQuads.count() == 2 && V1Used && V2Used) {
6247 BestLoQuad = InputQuads[0] ? 0 : 1;
6248 BestHiQuad = InputQuads[2] ? 2 : 3;
6250 if (InputQuads.count() > 2) {
6256 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
6257 // the shuffle mask. If a quad is scored as -1, that means that it contains
6258 // words from all 4 input quadwords.
6260 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
6262 BestLoQuad < 0 ? 0 : BestLoQuad,
6263 BestHiQuad < 0 ? 1 : BestHiQuad
6265 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
6266 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V1),
6267 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V2), &MaskV[0]);
6268 NewV = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, NewV);
6270 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
6271 // source words for the shuffle, to aid later transformations.
6272 bool AllWordsInNewV = true;
6273 bool InOrder[2] = { true, true };
6274 for (unsigned i = 0; i != 8; ++i) {
6275 int idx = MaskVals[i];
6277 InOrder[i/4] = false;
6278 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
6280 AllWordsInNewV = false;
6284 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
6285 if (AllWordsInNewV) {
6286 for (int i = 0; i != 8; ++i) {
6287 int idx = MaskVals[i];
6290 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
6291 if ((idx != i) && idx < 4)
6293 if ((idx != i) && idx > 3)
6302 // If we've eliminated the use of V2, and the new mask is a pshuflw or
6303 // pshufhw, that's as cheap as it gets. Return the new shuffle.
6304 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
6305 unsigned Opc = pshufhw ? X86ISD::PSHUFHW : X86ISD::PSHUFLW;
6306 unsigned TargetMask = 0;
6307 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
6308 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
6309 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(NewV.getNode());
6310 TargetMask = pshufhw ? getShufflePSHUFHWImmediate(SVOp):
6311 getShufflePSHUFLWImmediate(SVOp);
6312 V1 = NewV.getOperand(0);
6313 return getTargetShuffleNode(Opc, dl, MVT::v8i16, V1, TargetMask, DAG);
6317 // Promote splats to a larger type which usually leads to more efficient code.
6318 // FIXME: Is this true if pshufb is available?
6319 if (SVOp->isSplat())
6320 return PromoteSplat(SVOp, DAG);
6322 // If we have SSSE3, and all words of the result are from 1 input vector,
6323 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
6324 // is present, fall back to case 4.
6325 if (Subtarget->hasSSSE3()) {
6326 SmallVector<SDValue,16> pshufbMask;
6328 // If we have elements from both input vectors, set the high bit of the
6329 // shuffle mask element to zero out elements that come from V2 in the V1
6330 // mask, and elements that come from V1 in the V2 mask, so that the two
6331 // results can be OR'd together.
6332 bool TwoInputs = V1Used && V2Used;
6333 for (unsigned i = 0; i != 8; ++i) {
6334 int EltIdx = MaskVals[i] * 2;
6335 int Idx0 = (TwoInputs && (EltIdx >= 16)) ? 0x80 : EltIdx;
6336 int Idx1 = (TwoInputs && (EltIdx >= 16)) ? 0x80 : EltIdx+1;
6337 pshufbMask.push_back(DAG.getConstant(Idx0, MVT::i8));
6338 pshufbMask.push_back(DAG.getConstant(Idx1, MVT::i8));
6340 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V1);
6341 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
6342 DAG.getNode(ISD::BUILD_VECTOR, dl,
6343 MVT::v16i8, &pshufbMask[0], 16));
6345 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
6347 // Calculate the shuffle mask for the second input, shuffle it, and
6348 // OR it with the first shuffled input.
6350 for (unsigned i = 0; i != 8; ++i) {
6351 int EltIdx = MaskVals[i] * 2;
6352 int Idx0 = (EltIdx < 16) ? 0x80 : EltIdx - 16;
6353 int Idx1 = (EltIdx < 16) ? 0x80 : EltIdx - 15;
6354 pshufbMask.push_back(DAG.getConstant(Idx0, MVT::i8));
6355 pshufbMask.push_back(DAG.getConstant(Idx1, MVT::i8));
6357 V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V2);
6358 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
6359 DAG.getNode(ISD::BUILD_VECTOR, dl,
6360 MVT::v16i8, &pshufbMask[0], 16));
6361 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
6362 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
6365 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
6366 // and update MaskVals with new element order.
6367 std::bitset<8> InOrder;
6368 if (BestLoQuad >= 0) {
6369 int MaskV[] = { -1, -1, -1, -1, 4, 5, 6, 7 };
6370 for (int i = 0; i != 4; ++i) {
6371 int idx = MaskVals[i];
6374 } else if ((idx / 4) == BestLoQuad) {
6379 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
6382 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3()) {
6383 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(NewV.getNode());
6384 NewV = getTargetShuffleNode(X86ISD::PSHUFLW, dl, MVT::v8i16,
6386 getShufflePSHUFLWImmediate(SVOp), DAG);
6390 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
6391 // and update MaskVals with the new element order.
6392 if (BestHiQuad >= 0) {
6393 int MaskV[] = { 0, 1, 2, 3, -1, -1, -1, -1 };
6394 for (unsigned i = 4; i != 8; ++i) {
6395 int idx = MaskVals[i];
6398 } else if ((idx / 4) == BestHiQuad) {
6399 MaskV[i] = (idx & 3) + 4;
6403 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
6406 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3()) {
6407 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(NewV.getNode());
6408 NewV = getTargetShuffleNode(X86ISD::PSHUFHW, dl, MVT::v8i16,
6410 getShufflePSHUFHWImmediate(SVOp), DAG);
6414 // In case BestHi & BestLo were both -1, which means each quadword has a word
6415 // from each of the four input quadwords, calculate the InOrder bitvector now
6416 // before falling through to the insert/extract cleanup.
6417 if (BestLoQuad == -1 && BestHiQuad == -1) {
6419 for (int i = 0; i != 8; ++i)
6420 if (MaskVals[i] < 0 || MaskVals[i] == i)
6424 // The other elements are put in the right place using pextrw and pinsrw.
6425 for (unsigned i = 0; i != 8; ++i) {
6428 int EltIdx = MaskVals[i];
6431 SDValue ExtOp = (EltIdx < 8) ?
6432 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
6433 DAG.getIntPtrConstant(EltIdx)) :
6434 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
6435 DAG.getIntPtrConstant(EltIdx - 8));
6436 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
6437 DAG.getIntPtrConstant(i));
6442 // v16i8 shuffles - Prefer shuffles in the following order:
6443 // 1. [ssse3] 1 x pshufb
6444 // 2. [ssse3] 2 x pshufb + 1 x por
6445 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
6446 static SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
6447 const X86Subtarget* Subtarget,
6448 SelectionDAG &DAG) {
6449 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6450 SDValue V1 = SVOp->getOperand(0);
6451 SDValue V2 = SVOp->getOperand(1);
6453 ArrayRef<int> MaskVals = SVOp->getMask();
6455 // Promote splats to a larger type which usually leads to more efficient code.
6456 // FIXME: Is this true if pshufb is available?
6457 if (SVOp->isSplat())
6458 return PromoteSplat(SVOp, DAG);
6460 // If we have SSSE3, case 1 is generated when all result bytes come from
6461 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
6462 // present, fall back to case 3.
6464 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
6465 if (Subtarget->hasSSSE3()) {
6466 SmallVector<SDValue,16> pshufbMask;
6468 // If all result elements are from one input vector, then only translate
6469 // undef mask values to 0x80 (zero out result) in the pshufb mask.
6471 // Otherwise, we have elements from both input vectors, and must zero out
6472 // elements that come from V2 in the first mask, and V1 in the second mask
6473 // so that we can OR them together.
6474 for (unsigned i = 0; i != 16; ++i) {
6475 int EltIdx = MaskVals[i];
6476 if (EltIdx < 0 || EltIdx >= 16)
6478 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
6480 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
6481 DAG.getNode(ISD::BUILD_VECTOR, dl,
6482 MVT::v16i8, &pshufbMask[0], 16));
6484 // As PSHUFB will zero elements with negative indices, it's safe to ignore
6485 // the 2nd operand if it's undefined or zero.
6486 if (V2.getOpcode() == ISD::UNDEF ||
6487 ISD::isBuildVectorAllZeros(V2.getNode()))
6490 // Calculate the shuffle mask for the second input, shuffle it, and
6491 // OR it with the first shuffled input.
6493 for (unsigned i = 0; i != 16; ++i) {
6494 int EltIdx = MaskVals[i];
6495 EltIdx = (EltIdx < 16) ? 0x80 : EltIdx - 16;
6496 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
6498 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
6499 DAG.getNode(ISD::BUILD_VECTOR, dl,
6500 MVT::v16i8, &pshufbMask[0], 16));
6501 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
6504 // No SSSE3 - Calculate in place words and then fix all out of place words
6505 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
6506 // the 16 different words that comprise the two doublequadword input vectors.
6507 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
6508 V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V2);
6510 for (int i = 0; i != 8; ++i) {
6511 int Elt0 = MaskVals[i*2];
6512 int Elt1 = MaskVals[i*2+1];
6514 // This word of the result is all undef, skip it.
6515 if (Elt0 < 0 && Elt1 < 0)
6518 // This word of the result is already in the correct place, skip it.
6519 if ((Elt0 == i*2) && (Elt1 == i*2+1))
6522 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
6523 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
6526 // If Elt0 and Elt1 are defined, are consecutive, and can be load
6527 // using a single extract together, load it and store it.
6528 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
6529 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
6530 DAG.getIntPtrConstant(Elt1 / 2));
6531 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
6532 DAG.getIntPtrConstant(i));
6536 // If Elt1 is defined, extract it from the appropriate source. If the
6537 // source byte is not also odd, shift the extracted word left 8 bits
6538 // otherwise clear the bottom 8 bits if we need to do an or.
6540 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
6541 DAG.getIntPtrConstant(Elt1 / 2));
6542 if ((Elt1 & 1) == 0)
6543 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
6545 TLI.getShiftAmountTy(InsElt.getValueType())));
6547 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
6548 DAG.getConstant(0xFF00, MVT::i16));
6550 // If Elt0 is defined, extract it from the appropriate source. If the
6551 // source byte is not also even, shift the extracted word right 8 bits. If
6552 // Elt1 was also defined, OR the extracted values together before
6553 // inserting them in the result.
6555 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
6556 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
6557 if ((Elt0 & 1) != 0)
6558 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
6560 TLI.getShiftAmountTy(InsElt0.getValueType())));
6562 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
6563 DAG.getConstant(0x00FF, MVT::i16));
6564 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
6567 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
6568 DAG.getIntPtrConstant(i));
6570 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, NewV);
6573 // v32i8 shuffles - Translate to VPSHUFB if possible.
6575 SDValue LowerVECTOR_SHUFFLEv32i8(ShuffleVectorSDNode *SVOp,
6576 const X86Subtarget *Subtarget,
6577 SelectionDAG &DAG) {
6578 MVT VT = SVOp->getSimpleValueType(0);
6579 SDValue V1 = SVOp->getOperand(0);
6580 SDValue V2 = SVOp->getOperand(1);
6582 SmallVector<int, 32> MaskVals(SVOp->getMask().begin(), SVOp->getMask().end());
6584 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
6585 bool V1IsAllZero = ISD::isBuildVectorAllZeros(V1.getNode());
6586 bool V2IsAllZero = ISD::isBuildVectorAllZeros(V2.getNode());
6588 // VPSHUFB may be generated if
6589 // (1) one of input vector is undefined or zeroinitializer.
6590 // The mask value 0x80 puts 0 in the corresponding slot of the vector.
6591 // And (2) the mask indexes don't cross the 128-bit lane.
6592 if (VT != MVT::v32i8 || !Subtarget->hasInt256() ||
6593 (!V2IsUndef && !V2IsAllZero && !V1IsAllZero))
6596 if (V1IsAllZero && !V2IsAllZero) {
6597 CommuteVectorShuffleMask(MaskVals, 32);
6600 SmallVector<SDValue, 32> pshufbMask;
6601 for (unsigned i = 0; i != 32; i++) {
6602 int EltIdx = MaskVals[i];
6603 if (EltIdx < 0 || EltIdx >= 32)
6606 if ((EltIdx >= 16 && i < 16) || (EltIdx < 16 && i >= 16))
6607 // Cross lane is not allowed.
6611 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
6613 return DAG.getNode(X86ISD::PSHUFB, dl, MVT::v32i8, V1,
6614 DAG.getNode(ISD::BUILD_VECTOR, dl,
6615 MVT::v32i8, &pshufbMask[0], 32));
6618 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
6619 /// ones, or rewriting v4i32 / v4f32 as 2 wide ones if possible. This can be
6620 /// done when every pair / quad of shuffle mask elements point to elements in
6621 /// the right sequence. e.g.
6622 /// vector_shuffle X, Y, <2, 3, | 10, 11, | 0, 1, | 14, 15>
6624 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
6625 SelectionDAG &DAG) {
6626 MVT VT = SVOp->getSimpleValueType(0);
6628 unsigned NumElems = VT.getVectorNumElements();
6631 switch (VT.SimpleTy) {
6632 default: llvm_unreachable("Unexpected!");
6633 case MVT::v4f32: NewVT = MVT::v2f64; Scale = 2; break;
6634 case MVT::v4i32: NewVT = MVT::v2i64; Scale = 2; break;
6635 case MVT::v8i16: NewVT = MVT::v4i32; Scale = 2; break;
6636 case MVT::v16i8: NewVT = MVT::v4i32; Scale = 4; break;
6637 case MVT::v16i16: NewVT = MVT::v8i32; Scale = 2; break;
6638 case MVT::v32i8: NewVT = MVT::v8i32; Scale = 4; break;
6641 SmallVector<int, 8> MaskVec;
6642 for (unsigned i = 0; i != NumElems; i += Scale) {
6644 for (unsigned j = 0; j != Scale; ++j) {
6645 int EltIdx = SVOp->getMaskElt(i+j);
6649 StartIdx = (EltIdx / Scale);
6650 if (EltIdx != (int)(StartIdx*Scale + j))
6653 MaskVec.push_back(StartIdx);
6656 SDValue V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, SVOp->getOperand(0));
6657 SDValue V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, SVOp->getOperand(1));
6658 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
6661 /// getVZextMovL - Return a zero-extending vector move low node.
6663 static SDValue getVZextMovL(MVT VT, MVT OpVT,
6664 SDValue SrcOp, SelectionDAG &DAG,
6665 const X86Subtarget *Subtarget, SDLoc dl) {
6666 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
6667 LoadSDNode *LD = NULL;
6668 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
6669 LD = dyn_cast<LoadSDNode>(SrcOp);
6671 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
6673 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
6674 if ((ExtVT != MVT::i64 || Subtarget->is64Bit()) &&
6675 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
6676 SrcOp.getOperand(0).getOpcode() == ISD::BITCAST &&
6677 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
6679 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
6680 return DAG.getNode(ISD::BITCAST, dl, VT,
6681 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
6682 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6690 return DAG.getNode(ISD::BITCAST, dl, VT,
6691 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
6692 DAG.getNode(ISD::BITCAST, dl,
6696 /// LowerVECTOR_SHUFFLE_256 - Handle all 256-bit wide vectors shuffles
6697 /// which could not be matched by any known target speficic shuffle
6699 LowerVECTOR_SHUFFLE_256(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
6701 SDValue NewOp = Compact8x32ShuffleNode(SVOp, DAG);
6702 if (NewOp.getNode())
6705 MVT VT = SVOp->getSimpleValueType(0);
6707 unsigned NumElems = VT.getVectorNumElements();
6708 unsigned NumLaneElems = NumElems / 2;
6711 MVT EltVT = VT.getVectorElementType();
6712 MVT NVT = MVT::getVectorVT(EltVT, NumLaneElems);
6715 SmallVector<int, 16> Mask;
6716 for (unsigned l = 0; l < 2; ++l) {
6717 // Build a shuffle mask for the output, discovering on the fly which
6718 // input vectors to use as shuffle operands (recorded in InputUsed).
6719 // If building a suitable shuffle vector proves too hard, then bail
6720 // out with UseBuildVector set.
6721 bool UseBuildVector = false;
6722 int InputUsed[2] = { -1, -1 }; // Not yet discovered.
6723 unsigned LaneStart = l * NumLaneElems;
6724 for (unsigned i = 0; i != NumLaneElems; ++i) {
6725 // The mask element. This indexes into the input.
6726 int Idx = SVOp->getMaskElt(i+LaneStart);
6728 // the mask element does not index into any input vector.
6733 // The input vector this mask element indexes into.
6734 int Input = Idx / NumLaneElems;
6736 // Turn the index into an offset from the start of the input vector.
6737 Idx -= Input * NumLaneElems;
6739 // Find or create a shuffle vector operand to hold this input.
6741 for (OpNo = 0; OpNo < array_lengthof(InputUsed); ++OpNo) {
6742 if (InputUsed[OpNo] == Input)
6743 // This input vector is already an operand.
6745 if (InputUsed[OpNo] < 0) {
6746 // Create a new operand for this input vector.
6747 InputUsed[OpNo] = Input;
6752 if (OpNo >= array_lengthof(InputUsed)) {
6753 // More than two input vectors used! Give up on trying to create a
6754 // shuffle vector. Insert all elements into a BUILD_VECTOR instead.
6755 UseBuildVector = true;
6759 // Add the mask index for the new shuffle vector.
6760 Mask.push_back(Idx + OpNo * NumLaneElems);
6763 if (UseBuildVector) {
6764 SmallVector<SDValue, 16> SVOps;
6765 for (unsigned i = 0; i != NumLaneElems; ++i) {
6766 // The mask element. This indexes into the input.
6767 int Idx = SVOp->getMaskElt(i+LaneStart);
6769 SVOps.push_back(DAG.getUNDEF(EltVT));
6773 // The input vector this mask element indexes into.
6774 int Input = Idx / NumElems;
6776 // Turn the index into an offset from the start of the input vector.
6777 Idx -= Input * NumElems;
6779 // Extract the vector element by hand.
6780 SVOps.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
6781 SVOp->getOperand(Input),
6782 DAG.getIntPtrConstant(Idx)));
6785 // Construct the output using a BUILD_VECTOR.
6786 Output[l] = DAG.getNode(ISD::BUILD_VECTOR, dl, NVT, &SVOps[0],
6788 } else if (InputUsed[0] < 0) {
6789 // No input vectors were used! The result is undefined.
6790 Output[l] = DAG.getUNDEF(NVT);
6792 SDValue Op0 = Extract128BitVector(SVOp->getOperand(InputUsed[0] / 2),
6793 (InputUsed[0] % 2) * NumLaneElems,
6795 // If only one input was used, use an undefined vector for the other.
6796 SDValue Op1 = (InputUsed[1] < 0) ? DAG.getUNDEF(NVT) :
6797 Extract128BitVector(SVOp->getOperand(InputUsed[1] / 2),
6798 (InputUsed[1] % 2) * NumLaneElems, DAG, dl);
6799 // At least one input vector was used. Create a new shuffle vector.
6800 Output[l] = DAG.getVectorShuffle(NVT, dl, Op0, Op1, &Mask[0]);
6806 // Concatenate the result back
6807 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Output[0], Output[1]);
6810 /// LowerVECTOR_SHUFFLE_128v4 - Handle all 128-bit wide vectors with
6811 /// 4 elements, and match them with several different shuffle types.
6813 LowerVECTOR_SHUFFLE_128v4(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
6814 SDValue V1 = SVOp->getOperand(0);
6815 SDValue V2 = SVOp->getOperand(1);
6817 MVT VT = SVOp->getSimpleValueType(0);
6819 assert(VT.is128BitVector() && "Unsupported vector size");
6821 std::pair<int, int> Locs[4];
6822 int Mask1[] = { -1, -1, -1, -1 };
6823 SmallVector<int, 8> PermMask(SVOp->getMask().begin(), SVOp->getMask().end());
6827 for (unsigned i = 0; i != 4; ++i) {
6828 int Idx = PermMask[i];
6830 Locs[i] = std::make_pair(-1, -1);
6832 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
6834 Locs[i] = std::make_pair(0, NumLo);
6838 Locs[i] = std::make_pair(1, NumHi);
6840 Mask1[2+NumHi] = Idx;
6846 if (NumLo <= 2 && NumHi <= 2) {
6847 // If no more than two elements come from either vector. This can be
6848 // implemented with two shuffles. First shuffle gather the elements.
6849 // The second shuffle, which takes the first shuffle as both of its
6850 // vector operands, put the elements into the right order.
6851 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
6853 int Mask2[] = { -1, -1, -1, -1 };
6855 for (unsigned i = 0; i != 4; ++i)
6856 if (Locs[i].first != -1) {
6857 unsigned Idx = (i < 2) ? 0 : 4;
6858 Idx += Locs[i].first * 2 + Locs[i].second;
6862 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
6865 if (NumLo == 3 || NumHi == 3) {
6866 // Otherwise, we must have three elements from one vector, call it X, and
6867 // one element from the other, call it Y. First, use a shufps to build an
6868 // intermediate vector with the one element from Y and the element from X
6869 // that will be in the same half in the final destination (the indexes don't
6870 // matter). Then, use a shufps to build the final vector, taking the half
6871 // containing the element from Y from the intermediate, and the other half
6874 // Normalize it so the 3 elements come from V1.
6875 CommuteVectorShuffleMask(PermMask, 4);
6879 // Find the element from V2.
6881 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
6882 int Val = PermMask[HiIndex];
6889 Mask1[0] = PermMask[HiIndex];
6891 Mask1[2] = PermMask[HiIndex^1];
6893 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
6896 Mask1[0] = PermMask[0];
6897 Mask1[1] = PermMask[1];
6898 Mask1[2] = HiIndex & 1 ? 6 : 4;
6899 Mask1[3] = HiIndex & 1 ? 4 : 6;
6900 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
6903 Mask1[0] = HiIndex & 1 ? 2 : 0;
6904 Mask1[1] = HiIndex & 1 ? 0 : 2;
6905 Mask1[2] = PermMask[2];
6906 Mask1[3] = PermMask[3];
6911 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
6914 // Break it into (shuffle shuffle_hi, shuffle_lo).
6915 int LoMask[] = { -1, -1, -1, -1 };
6916 int HiMask[] = { -1, -1, -1, -1 };
6918 int *MaskPtr = LoMask;
6919 unsigned MaskIdx = 0;
6922 for (unsigned i = 0; i != 4; ++i) {
6929 int Idx = PermMask[i];
6931 Locs[i] = std::make_pair(-1, -1);
6932 } else if (Idx < 4) {
6933 Locs[i] = std::make_pair(MaskIdx, LoIdx);
6934 MaskPtr[LoIdx] = Idx;
6937 Locs[i] = std::make_pair(MaskIdx, HiIdx);
6938 MaskPtr[HiIdx] = Idx;
6943 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
6944 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
6945 int MaskOps[] = { -1, -1, -1, -1 };
6946 for (unsigned i = 0; i != 4; ++i)
6947 if (Locs[i].first != -1)
6948 MaskOps[i] = Locs[i].first * 4 + Locs[i].second;
6949 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
6952 static bool MayFoldVectorLoad(SDValue V) {
6953 while (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
6954 V = V.getOperand(0);
6956 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
6957 V = V.getOperand(0);
6958 if (V.hasOneUse() && V.getOpcode() == ISD::BUILD_VECTOR &&
6959 V.getNumOperands() == 2 && V.getOperand(1).getOpcode() == ISD::UNDEF)
6960 // BUILD_VECTOR (load), undef
6961 V = V.getOperand(0);
6963 return MayFoldLoad(V);
6967 SDValue getMOVDDup(SDValue &Op, SDLoc &dl, SDValue V1, SelectionDAG &DAG) {
6968 MVT VT = Op.getSimpleValueType();
6970 // Canonizalize to v2f64.
6971 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, V1);
6972 return DAG.getNode(ISD::BITCAST, dl, VT,
6973 getTargetShuffleNode(X86ISD::MOVDDUP, dl, MVT::v2f64,
6978 SDValue getMOVLowToHigh(SDValue &Op, SDLoc &dl, SelectionDAG &DAG,
6980 SDValue V1 = Op.getOperand(0);
6981 SDValue V2 = Op.getOperand(1);
6982 MVT VT = Op.getSimpleValueType();
6984 assert(VT != MVT::v2i64 && "unsupported shuffle type");
6986 if (HasSSE2 && VT == MVT::v2f64)
6987 return getTargetShuffleNode(X86ISD::MOVLHPD, dl, VT, V1, V2, DAG);
6989 // v4f32 or v4i32: canonizalized to v4f32 (which is legal for SSE1)
6990 return DAG.getNode(ISD::BITCAST, dl, VT,
6991 getTargetShuffleNode(X86ISD::MOVLHPS, dl, MVT::v4f32,
6992 DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, V1),
6993 DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, V2), DAG));
6997 SDValue getMOVHighToLow(SDValue &Op, SDLoc &dl, SelectionDAG &DAG) {
6998 SDValue V1 = Op.getOperand(0);
6999 SDValue V2 = Op.getOperand(1);
7000 MVT VT = Op.getSimpleValueType();
7002 assert((VT == MVT::v4i32 || VT == MVT::v4f32) &&
7003 "unsupported shuffle type");
7005 if (V2.getOpcode() == ISD::UNDEF)
7009 return getTargetShuffleNode(X86ISD::MOVHLPS, dl, VT, V1, V2, DAG);
7013 SDValue getMOVLP(SDValue &Op, SDLoc &dl, SelectionDAG &DAG, bool HasSSE2) {
7014 SDValue V1 = Op.getOperand(0);
7015 SDValue V2 = Op.getOperand(1);
7016 MVT VT = Op.getSimpleValueType();
7017 unsigned NumElems = VT.getVectorNumElements();
7019 // Use MOVLPS and MOVLPD in case V1 or V2 are loads. During isel, the second
7020 // operand of these instructions is only memory, so check if there's a
7021 // potencial load folding here, otherwise use SHUFPS or MOVSD to match the
7023 bool CanFoldLoad = false;
7025 // Trivial case, when V2 comes from a load.
7026 if (MayFoldVectorLoad(V2))
7029 // When V1 is a load, it can be folded later into a store in isel, example:
7030 // (store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)), addr:$src1)
7032 // (MOVLPSmr addr:$src1, VR128:$src2)
7033 // So, recognize this potential and also use MOVLPS or MOVLPD
7034 else if (MayFoldVectorLoad(V1) && MayFoldIntoStore(Op))
7037 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7039 if (HasSSE2 && NumElems == 2)
7040 return getTargetShuffleNode(X86ISD::MOVLPD, dl, VT, V1, V2, DAG);
7043 // If we don't care about the second element, proceed to use movss.
7044 if (SVOp->getMaskElt(1) != -1)
7045 return getTargetShuffleNode(X86ISD::MOVLPS, dl, VT, V1, V2, DAG);
7048 // movl and movlp will both match v2i64, but v2i64 is never matched by
7049 // movl earlier because we make it strict to avoid messing with the movlp load
7050 // folding logic (see the code above getMOVLP call). Match it here then,
7051 // this is horrible, but will stay like this until we move all shuffle
7052 // matching to x86 specific nodes. Note that for the 1st condition all
7053 // types are matched with movsd.
7055 // FIXME: isMOVLMask should be checked and matched before getMOVLP,
7056 // as to remove this logic from here, as much as possible
7057 if (NumElems == 2 || !isMOVLMask(SVOp->getMask(), VT))
7058 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
7059 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
7062 assert(VT != MVT::v4i32 && "unsupported shuffle type");
7064 // Invert the operand order and use SHUFPS to match it.
7065 return getTargetShuffleNode(X86ISD::SHUFP, dl, VT, V2, V1,
7066 getShuffleSHUFImmediate(SVOp), DAG);
7069 // Reduce a vector shuffle to zext.
7070 static SDValue LowerVectorIntExtend(SDValue Op, const X86Subtarget *Subtarget,
7071 SelectionDAG &DAG) {
7072 // PMOVZX is only available from SSE41.
7073 if (!Subtarget->hasSSE41())
7076 MVT VT = Op.getSimpleValueType();
7078 // Only AVX2 support 256-bit vector integer extending.
7079 if (!Subtarget->hasInt256() && VT.is256BitVector())
7082 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7084 SDValue V1 = Op.getOperand(0);
7085 SDValue V2 = Op.getOperand(1);
7086 unsigned NumElems = VT.getVectorNumElements();
7088 // Extending is an unary operation and the element type of the source vector
7089 // won't be equal to or larger than i64.
7090 if (V2.getOpcode() != ISD::UNDEF || !VT.isInteger() ||
7091 VT.getVectorElementType() == MVT::i64)
7094 // Find the expansion ratio, e.g. expanding from i8 to i32 has a ratio of 4.
7095 unsigned Shift = 1; // Start from 2, i.e. 1 << 1.
7096 while ((1U << Shift) < NumElems) {
7097 if (SVOp->getMaskElt(1U << Shift) == 1)
7100 // The maximal ratio is 8, i.e. from i8 to i64.
7105 // Check the shuffle mask.
7106 unsigned Mask = (1U << Shift) - 1;
7107 for (unsigned i = 0; i != NumElems; ++i) {
7108 int EltIdx = SVOp->getMaskElt(i);
7109 if ((i & Mask) != 0 && EltIdx != -1)
7111 if ((i & Mask) == 0 && (unsigned)EltIdx != (i >> Shift))
7115 unsigned NBits = VT.getVectorElementType().getSizeInBits() << Shift;
7116 MVT NeVT = MVT::getIntegerVT(NBits);
7117 MVT NVT = MVT::getVectorVT(NeVT, NumElems >> Shift);
7119 if (!DAG.getTargetLoweringInfo().isTypeLegal(NVT))
7122 // Simplify the operand as it's prepared to be fed into shuffle.
7123 unsigned SignificantBits = NVT.getSizeInBits() >> Shift;
7124 if (V1.getOpcode() == ISD::BITCAST &&
7125 V1.getOperand(0).getOpcode() == ISD::SCALAR_TO_VECTOR &&
7126 V1.getOperand(0).getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
7127 V1.getOperand(0).getOperand(0)
7128 .getSimpleValueType().getSizeInBits() == SignificantBits) {
7129 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast x)
7130 SDValue V = V1.getOperand(0).getOperand(0).getOperand(0);
7131 ConstantSDNode *CIdx =
7132 dyn_cast<ConstantSDNode>(V1.getOperand(0).getOperand(0).getOperand(1));
7133 // If it's foldable, i.e. normal load with single use, we will let code
7134 // selection to fold it. Otherwise, we will short the conversion sequence.
7135 if (CIdx && CIdx->getZExtValue() == 0 &&
7136 (!ISD::isNormalLoad(V.getNode()) || !V.hasOneUse())) {
7137 MVT FullVT = V.getSimpleValueType();
7138 MVT V1VT = V1.getSimpleValueType();
7139 if (FullVT.getSizeInBits() > V1VT.getSizeInBits()) {
7140 // The "ext_vec_elt" node is wider than the result node.
7141 // In this case we should extract subvector from V.
7142 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast (extract_subvector x)).
7143 unsigned Ratio = FullVT.getSizeInBits() / V1VT.getSizeInBits();
7144 MVT SubVecVT = MVT::getVectorVT(FullVT.getVectorElementType(),
7145 FullVT.getVectorNumElements()/Ratio);
7146 V = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVecVT, V,
7147 DAG.getIntPtrConstant(0));
7149 V1 = DAG.getNode(ISD::BITCAST, DL, V1VT, V);
7153 return DAG.getNode(ISD::BITCAST, DL, VT,
7154 DAG.getNode(X86ISD::VZEXT, DL, NVT, V1));
7158 NormalizeVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
7159 SelectionDAG &DAG) {
7160 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7161 MVT VT = Op.getSimpleValueType();
7163 SDValue V1 = Op.getOperand(0);
7164 SDValue V2 = Op.getOperand(1);
7166 if (isZeroShuffle(SVOp))
7167 return getZeroVector(VT, Subtarget, DAG, dl);
7169 // Handle splat operations
7170 if (SVOp->isSplat()) {
7171 // Use vbroadcast whenever the splat comes from a foldable load
7172 SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG);
7173 if (Broadcast.getNode())
7177 // Check integer expanding shuffles.
7178 SDValue NewOp = LowerVectorIntExtend(Op, Subtarget, DAG);
7179 if (NewOp.getNode())
7182 // If the shuffle can be profitably rewritten as a narrower shuffle, then
7184 if (VT == MVT::v8i16 || VT == MVT::v16i8 ||
7185 VT == MVT::v16i16 || VT == MVT::v32i8) {
7186 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG);
7187 if (NewOp.getNode())
7188 return DAG.getNode(ISD::BITCAST, dl, VT, NewOp);
7189 } else if ((VT == MVT::v4i32 ||
7190 (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
7191 // FIXME: Figure out a cleaner way to do this.
7192 // Try to make use of movq to zero out the top part.
7193 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
7194 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG);
7195 if (NewOp.getNode()) {
7196 MVT NewVT = NewOp.getSimpleValueType();
7197 if (isCommutedMOVLMask(cast<ShuffleVectorSDNode>(NewOp)->getMask(),
7198 NewVT, true, false))
7199 return getVZextMovL(VT, NewVT, NewOp.getOperand(0),
7200 DAG, Subtarget, dl);
7202 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
7203 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG);
7204 if (NewOp.getNode()) {
7205 MVT NewVT = NewOp.getSimpleValueType();
7206 if (isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)->getMask(), NewVT))
7207 return getVZextMovL(VT, NewVT, NewOp.getOperand(1),
7208 DAG, Subtarget, dl);
7216 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
7217 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7218 SDValue V1 = Op.getOperand(0);
7219 SDValue V2 = Op.getOperand(1);
7220 MVT VT = Op.getSimpleValueType();
7222 unsigned NumElems = VT.getVectorNumElements();
7223 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
7224 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
7225 bool V1IsSplat = false;
7226 bool V2IsSplat = false;
7227 bool HasSSE2 = Subtarget->hasSSE2();
7228 bool HasFp256 = Subtarget->hasFp256();
7229 bool HasInt256 = Subtarget->hasInt256();
7230 MachineFunction &MF = DAG.getMachineFunction();
7231 bool OptForSize = MF.getFunction()->getAttributes().
7232 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize);
7234 assert(VT.getSizeInBits() != 64 && "Can't lower MMX shuffles");
7236 if (V1IsUndef && V2IsUndef)
7237 return DAG.getUNDEF(VT);
7239 assert(!V1IsUndef && "Op 1 of shuffle should not be undef");
7241 // Vector shuffle lowering takes 3 steps:
7243 // 1) Normalize the input vectors. Here splats, zeroed vectors, profitable
7244 // narrowing and commutation of operands should be handled.
7245 // 2) Matching of shuffles with known shuffle masks to x86 target specific
7247 // 3) Rewriting of unmatched masks into new generic shuffle operations,
7248 // so the shuffle can be broken into other shuffles and the legalizer can
7249 // try the lowering again.
7251 // The general idea is that no vector_shuffle operation should be left to
7252 // be matched during isel, all of them must be converted to a target specific
7255 // Normalize the input vectors. Here splats, zeroed vectors, profitable
7256 // narrowing and commutation of operands should be handled. The actual code
7257 // doesn't include all of those, work in progress...
7258 SDValue NewOp = NormalizeVectorShuffle(Op, Subtarget, DAG);
7259 if (NewOp.getNode())
7262 SmallVector<int, 8> M(SVOp->getMask().begin(), SVOp->getMask().end());
7264 // NOTE: isPSHUFDMask can also match both masks below (unpckl_undef and
7265 // unpckh_undef). Only use pshufd if speed is more important than size.
7266 if (OptForSize && isUNPCKL_v_undef_Mask(M, VT, HasInt256))
7267 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V1, DAG);
7268 if (OptForSize && isUNPCKH_v_undef_Mask(M, VT, HasInt256))
7269 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V1, DAG);
7271 if (isMOVDDUPMask(M, VT) && Subtarget->hasSSE3() &&
7272 V2IsUndef && MayFoldVectorLoad(V1))
7273 return getMOVDDup(Op, dl, V1, DAG);
7275 if (isMOVHLPS_v_undef_Mask(M, VT))
7276 return getMOVHighToLow(Op, dl, DAG);
7278 // Use to match splats
7279 if (HasSSE2 && isUNPCKHMask(M, VT, HasInt256) && V2IsUndef &&
7280 (VT == MVT::v2f64 || VT == MVT::v2i64))
7281 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V1, DAG);
7283 if (isPSHUFDMask(M, VT)) {
7284 // The actual implementation will match the mask in the if above and then
7285 // during isel it can match several different instructions, not only pshufd
7286 // as its name says, sad but true, emulate the behavior for now...
7287 if (isMOVDDUPMask(M, VT) && ((VT == MVT::v4f32 || VT == MVT::v2i64)))
7288 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V1, DAG);
7290 unsigned TargetMask = getShuffleSHUFImmediate(SVOp);
7292 if (HasSSE2 && (VT == MVT::v4f32 || VT == MVT::v4i32))
7293 return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1, TargetMask, DAG);
7295 if (HasFp256 && (VT == MVT::v4f32 || VT == MVT::v2f64))
7296 return getTargetShuffleNode(X86ISD::VPERMILP, dl, VT, V1, TargetMask,
7299 return getTargetShuffleNode(X86ISD::SHUFP, dl, VT, V1, V1,
7303 if (isPALIGNRMask(M, VT, Subtarget))
7304 return getTargetShuffleNode(X86ISD::PALIGNR, dl, VT, V1, V2,
7305 getShufflePALIGNRImmediate(SVOp),
7308 // Check if this can be converted into a logical shift.
7309 bool isLeft = false;
7312 bool isShift = HasSSE2 && isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
7313 if (isShift && ShVal.hasOneUse()) {
7314 // If the shifted value has multiple uses, it may be cheaper to use
7315 // v_set0 + movlhps or movhlps, etc.
7316 MVT EltVT = VT.getVectorElementType();
7317 ShAmt *= EltVT.getSizeInBits();
7318 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
7321 if (isMOVLMask(M, VT)) {
7322 if (ISD::isBuildVectorAllZeros(V1.getNode()))
7323 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
7324 if (!isMOVLPMask(M, VT)) {
7325 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
7326 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
7328 if (VT == MVT::v4i32 || VT == MVT::v4f32)
7329 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
7333 // FIXME: fold these into legal mask.
7334 if (isMOVLHPSMask(M, VT) && !isUNPCKLMask(M, VT, HasInt256))
7335 return getMOVLowToHigh(Op, dl, DAG, HasSSE2);
7337 if (isMOVHLPSMask(M, VT))
7338 return getMOVHighToLow(Op, dl, DAG);
7340 if (V2IsUndef && isMOVSHDUPMask(M, VT, Subtarget))
7341 return getTargetShuffleNode(X86ISD::MOVSHDUP, dl, VT, V1, DAG);
7343 if (V2IsUndef && isMOVSLDUPMask(M, VT, Subtarget))
7344 return getTargetShuffleNode(X86ISD::MOVSLDUP, dl, VT, V1, DAG);
7346 if (isMOVLPMask(M, VT))
7347 return getMOVLP(Op, dl, DAG, HasSSE2);
7349 if (ShouldXformToMOVHLPS(M, VT) ||
7350 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), M, VT))
7351 return CommuteVectorShuffle(SVOp, DAG);
7354 // No better options. Use a vshldq / vsrldq.
7355 MVT EltVT = VT.getVectorElementType();
7356 ShAmt *= EltVT.getSizeInBits();
7357 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
7360 bool Commuted = false;
7361 // FIXME: This should also accept a bitcast of a splat? Be careful, not
7362 // 1,1,1,1 -> v8i16 though.
7363 V1IsSplat = isSplatVector(V1.getNode());
7364 V2IsSplat = isSplatVector(V2.getNode());
7366 // Canonicalize the splat or undef, if present, to be on the RHS.
7367 if (!V2IsUndef && V1IsSplat && !V2IsSplat) {
7368 CommuteVectorShuffleMask(M, NumElems);
7370 std::swap(V1IsSplat, V2IsSplat);
7374 if (isCommutedMOVLMask(M, VT, V2IsSplat, V2IsUndef)) {
7375 // Shuffling low element of v1 into undef, just return v1.
7378 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
7379 // the instruction selector will not match, so get a canonical MOVL with
7380 // swapped operands to undo the commute.
7381 return getMOVL(DAG, dl, VT, V2, V1);
7384 if (isUNPCKLMask(M, VT, HasInt256))
7385 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V2, DAG);
7387 if (isUNPCKHMask(M, VT, HasInt256))
7388 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V2, DAG);
7391 // Normalize mask so all entries that point to V2 points to its first
7392 // element then try to match unpck{h|l} again. If match, return a
7393 // new vector_shuffle with the corrected mask.p
7394 SmallVector<int, 8> NewMask(M.begin(), M.end());
7395 NormalizeMask(NewMask, NumElems);
7396 if (isUNPCKLMask(NewMask, VT, HasInt256, true))
7397 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V2, DAG);
7398 if (isUNPCKHMask(NewMask, VT, HasInt256, true))
7399 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V2, DAG);
7403 // Commute is back and try unpck* again.
7404 // FIXME: this seems wrong.
7405 CommuteVectorShuffleMask(M, NumElems);
7407 std::swap(V1IsSplat, V2IsSplat);
7410 if (isUNPCKLMask(M, VT, HasInt256))
7411 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V2, DAG);
7413 if (isUNPCKHMask(M, VT, HasInt256))
7414 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V2, DAG);
7417 // Normalize the node to match x86 shuffle ops if needed
7418 if (!V2IsUndef && (isSHUFPMask(M, VT, /* Commuted */ true)))
7419 return CommuteVectorShuffle(SVOp, DAG);
7421 // The checks below are all present in isShuffleMaskLegal, but they are
7422 // inlined here right now to enable us to directly emit target specific
7423 // nodes, and remove one by one until they don't return Op anymore.
7425 if (ShuffleVectorSDNode::isSplatMask(&M[0], VT) &&
7426 SVOp->getSplatIndex() == 0 && V2IsUndef) {
7427 if (VT == MVT::v2f64 || VT == MVT::v2i64)
7428 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V1, DAG);
7431 if (isPSHUFHWMask(M, VT, HasInt256))
7432 return getTargetShuffleNode(X86ISD::PSHUFHW, dl, VT, V1,
7433 getShufflePSHUFHWImmediate(SVOp),
7436 if (isPSHUFLWMask(M, VT, HasInt256))
7437 return getTargetShuffleNode(X86ISD::PSHUFLW, dl, VT, V1,
7438 getShufflePSHUFLWImmediate(SVOp),
7441 if (isSHUFPMask(M, VT))
7442 return getTargetShuffleNode(X86ISD::SHUFP, dl, VT, V1, V2,
7443 getShuffleSHUFImmediate(SVOp), DAG);
7445 if (isUNPCKL_v_undef_Mask(M, VT, HasInt256))
7446 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V1, DAG);
7447 if (isUNPCKH_v_undef_Mask(M, VT, HasInt256))
7448 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V1, DAG);
7450 //===--------------------------------------------------------------------===//
7451 // Generate target specific nodes for 128 or 256-bit shuffles only
7452 // supported in the AVX instruction set.
7455 // Handle VMOVDDUPY permutations
7456 if (V2IsUndef && isMOVDDUPYMask(M, VT, HasFp256))
7457 return getTargetShuffleNode(X86ISD::MOVDDUP, dl, VT, V1, DAG);
7459 // Handle VPERMILPS/D* permutations
7460 if (isVPERMILPMask(M, VT)) {
7461 if ((HasInt256 && VT == MVT::v8i32) || VT == MVT::v16i32)
7462 return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1,
7463 getShuffleSHUFImmediate(SVOp), DAG);
7464 return getTargetShuffleNode(X86ISD::VPERMILP, dl, VT, V1,
7465 getShuffleSHUFImmediate(SVOp), DAG);
7468 // Handle VPERM2F128/VPERM2I128 permutations
7469 if (isVPERM2X128Mask(M, VT, HasFp256))
7470 return getTargetShuffleNode(X86ISD::VPERM2X128, dl, VT, V1,
7471 V2, getShuffleVPERM2X128Immediate(SVOp), DAG);
7473 SDValue BlendOp = LowerVECTOR_SHUFFLEtoBlend(SVOp, Subtarget, DAG);
7474 if (BlendOp.getNode())
7478 if (V2IsUndef && HasInt256 && isPermImmMask(M, VT, Imm8))
7479 return getTargetShuffleNode(X86ISD::VPERMI, dl, VT, V1, Imm8, DAG);
7481 if ((V2IsUndef && HasInt256 && VT.is256BitVector() && NumElems == 8) ||
7482 VT.is512BitVector()) {
7483 MVT MaskEltVT = MVT::getIntegerVT(VT.getVectorElementType().getSizeInBits());
7484 MVT MaskVectorVT = MVT::getVectorVT(MaskEltVT, NumElems);
7485 SmallVector<SDValue, 16> permclMask;
7486 for (unsigned i = 0; i != NumElems; ++i) {
7487 permclMask.push_back(DAG.getConstant((M[i]>=0) ? M[i] : 0, MaskEltVT));
7490 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVectorVT,
7491 &permclMask[0], NumElems);
7493 // Bitcast is for VPERMPS since mask is v8i32 but node takes v8f32
7494 return DAG.getNode(X86ISD::VPERMV, dl, VT,
7495 DAG.getNode(ISD::BITCAST, dl, VT, Mask), V1);
7496 return DAG.getNode(X86ISD::VPERMV3, dl, VT,
7497 DAG.getNode(ISD::BITCAST, dl, VT, Mask), V1, V2);
7500 //===--------------------------------------------------------------------===//
7501 // Since no target specific shuffle was selected for this generic one,
7502 // lower it into other known shuffles. FIXME: this isn't true yet, but
7503 // this is the plan.
7506 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
7507 if (VT == MVT::v8i16) {
7508 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(Op, Subtarget, DAG);
7509 if (NewOp.getNode())
7513 if (VT == MVT::v16i8) {
7514 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, Subtarget, DAG);
7515 if (NewOp.getNode())
7519 if (VT == MVT::v32i8) {
7520 SDValue NewOp = LowerVECTOR_SHUFFLEv32i8(SVOp, Subtarget, DAG);
7521 if (NewOp.getNode())
7525 // Handle all 128-bit wide vectors with 4 elements, and match them with
7526 // several different shuffle types.
7527 if (NumElems == 4 && VT.is128BitVector())
7528 return LowerVECTOR_SHUFFLE_128v4(SVOp, DAG);
7530 // Handle general 256-bit shuffles
7531 if (VT.is256BitVector())
7532 return LowerVECTOR_SHUFFLE_256(SVOp, DAG);
7537 static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
7538 MVT VT = Op.getSimpleValueType();
7541 if (!Op.getOperand(0).getSimpleValueType().is128BitVector())
7544 if (VT.getSizeInBits() == 8) {
7545 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
7546 Op.getOperand(0), Op.getOperand(1));
7547 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
7548 DAG.getValueType(VT));
7549 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
7552 if (VT.getSizeInBits() == 16) {
7553 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
7554 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
7556 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
7557 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
7558 DAG.getNode(ISD::BITCAST, dl,
7562 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
7563 Op.getOperand(0), Op.getOperand(1));
7564 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
7565 DAG.getValueType(VT));
7566 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
7569 if (VT == MVT::f32) {
7570 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
7571 // the result back to FR32 register. It's only worth matching if the
7572 // result has a single use which is a store or a bitcast to i32. And in
7573 // the case of a store, it's not worth it if the index is a constant 0,
7574 // because a MOVSSmr can be used instead, which is smaller and faster.
7575 if (!Op.hasOneUse())
7577 SDNode *User = *Op.getNode()->use_begin();
7578 if ((User->getOpcode() != ISD::STORE ||
7579 (isa<ConstantSDNode>(Op.getOperand(1)) &&
7580 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
7581 (User->getOpcode() != ISD::BITCAST ||
7582 User->getValueType(0) != MVT::i32))
7584 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
7585 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32,
7588 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, Extract);
7591 if (VT == MVT::i32 || VT == MVT::i64) {
7592 // ExtractPS/pextrq works with constant index.
7593 if (isa<ConstantSDNode>(Op.getOperand(1)))
7600 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
7601 SelectionDAG &DAG) const {
7603 if (!isa<ConstantSDNode>(Op.getOperand(1)))
7606 SDValue Vec = Op.getOperand(0);
7607 MVT VecVT = Vec.getSimpleValueType();
7609 // If this is a 256-bit vector result, first extract the 128-bit vector and
7610 // then extract the element from the 128-bit vector.
7611 if (VecVT.is256BitVector() || VecVT.is512BitVector()) {
7612 SDValue Idx = Op.getOperand(1);
7613 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
7615 // Get the 128-bit vector.
7616 Vec = Extract128BitVector(Vec, IdxVal, DAG, dl);
7617 MVT EltVT = VecVT.getVectorElementType();
7619 unsigned ElemsPerChunk = 128 / EltVT.getSizeInBits();
7621 //if (IdxVal >= NumElems/2)
7622 // IdxVal -= NumElems/2;
7623 IdxVal -= (IdxVal/ElemsPerChunk)*ElemsPerChunk;
7624 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
7625 DAG.getConstant(IdxVal, MVT::i32));
7628 assert(VecVT.is128BitVector() && "Unexpected vector length");
7630 if (Subtarget->hasSSE41()) {
7631 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
7636 MVT VT = Op.getSimpleValueType();
7637 // TODO: handle v16i8.
7638 if (VT.getSizeInBits() == 16) {
7639 SDValue Vec = Op.getOperand(0);
7640 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
7642 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
7643 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
7644 DAG.getNode(ISD::BITCAST, dl,
7647 // Transform it so it match pextrw which produces a 32-bit result.
7648 MVT EltVT = MVT::i32;
7649 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
7650 Op.getOperand(0), Op.getOperand(1));
7651 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
7652 DAG.getValueType(VT));
7653 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
7656 if (VT.getSizeInBits() == 32) {
7657 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
7661 // SHUFPS the element to the lowest double word, then movss.
7662 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
7663 MVT VVT = Op.getOperand(0).getSimpleValueType();
7664 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
7665 DAG.getUNDEF(VVT), Mask);
7666 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
7667 DAG.getIntPtrConstant(0));
7670 if (VT.getSizeInBits() == 64) {
7671 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
7672 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
7673 // to match extract_elt for f64.
7674 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
7678 // UNPCKHPD the element to the lowest double word, then movsd.
7679 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
7680 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
7681 int Mask[2] = { 1, -1 };
7682 MVT VVT = Op.getOperand(0).getSimpleValueType();
7683 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
7684 DAG.getUNDEF(VVT), Mask);
7685 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
7686 DAG.getIntPtrConstant(0));
7692 static SDValue LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
7693 MVT VT = Op.getSimpleValueType();
7694 MVT EltVT = VT.getVectorElementType();
7697 SDValue N0 = Op.getOperand(0);
7698 SDValue N1 = Op.getOperand(1);
7699 SDValue N2 = Op.getOperand(2);
7701 if (!VT.is128BitVector())
7704 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
7705 isa<ConstantSDNode>(N2)) {
7707 if (VT == MVT::v8i16)
7708 Opc = X86ISD::PINSRW;
7709 else if (VT == MVT::v16i8)
7710 Opc = X86ISD::PINSRB;
7712 Opc = X86ISD::PINSRB;
7714 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
7716 if (N1.getValueType() != MVT::i32)
7717 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
7718 if (N2.getValueType() != MVT::i32)
7719 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
7720 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
7723 if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
7724 // Bits [7:6] of the constant are the source select. This will always be
7725 // zero here. The DAG Combiner may combine an extract_elt index into these
7726 // bits. For example (insert (extract, 3), 2) could be matched by putting
7727 // the '3' into bits [7:6] of X86ISD::INSERTPS.
7728 // Bits [5:4] of the constant are the destination select. This is the
7729 // value of the incoming immediate.
7730 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
7731 // combine either bitwise AND or insert of float 0.0 to set these bits.
7732 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
7733 // Create this as a scalar to vector..
7734 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
7735 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
7738 if ((EltVT == MVT::i32 || EltVT == MVT::i64) && isa<ConstantSDNode>(N2)) {
7739 // PINSR* works with constant index.
7746 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
7747 MVT VT = Op.getSimpleValueType();
7748 MVT EltVT = VT.getVectorElementType();
7751 SDValue N0 = Op.getOperand(0);
7752 SDValue N1 = Op.getOperand(1);
7753 SDValue N2 = Op.getOperand(2);
7755 // If this is a 256-bit vector result, first extract the 128-bit vector,
7756 // insert the element into the extracted half and then place it back.
7757 if (VT.is256BitVector() || VT.is512BitVector()) {
7758 if (!isa<ConstantSDNode>(N2))
7761 // Get the desired 128-bit vector half.
7762 unsigned IdxVal = cast<ConstantSDNode>(N2)->getZExtValue();
7763 SDValue V = Extract128BitVector(N0, IdxVal, DAG, dl);
7765 // Insert the element into the desired half.
7766 unsigned NumEltsIn128 = 128/EltVT.getSizeInBits();
7767 unsigned IdxIn128 = IdxVal - (IdxVal/NumEltsIn128) * NumEltsIn128;
7769 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1,
7770 DAG.getConstant(IdxIn128, MVT::i32));
7772 // Insert the changed part back to the 256-bit vector
7773 return Insert128BitVector(N0, V, IdxVal, DAG, dl);
7776 if (Subtarget->hasSSE41())
7777 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
7779 if (EltVT == MVT::i8)
7782 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
7783 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
7784 // as its second argument.
7785 if (N1.getValueType() != MVT::i32)
7786 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
7787 if (N2.getValueType() != MVT::i32)
7788 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
7789 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
7794 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
7796 MVT OpVT = Op.getSimpleValueType();
7798 // If this is a 256-bit vector result, first insert into a 128-bit
7799 // vector and then insert into the 256-bit vector.
7800 if (!OpVT.is128BitVector()) {
7801 // Insert into a 128-bit vector.
7802 unsigned SizeFactor = OpVT.getSizeInBits()/128;
7803 MVT VT128 = MVT::getVectorVT(OpVT.getVectorElementType(),
7804 OpVT.getVectorNumElements() / SizeFactor);
7806 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
7808 // Insert the 128-bit vector.
7809 return Insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl);
7812 if (OpVT == MVT::v1i64 &&
7813 Op.getOperand(0).getValueType() == MVT::i64)
7814 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
7816 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
7817 assert(OpVT.is128BitVector() && "Expected an SSE type!");
7818 return DAG.getNode(ISD::BITCAST, dl, OpVT,
7819 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,AnyExt));
7822 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
7823 // a simple subregister reference or explicit instructions to grab
7824 // upper bits of a vector.
7825 static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
7826 SelectionDAG &DAG) {
7828 SDValue In = Op.getOperand(0);
7829 SDValue Idx = Op.getOperand(1);
7830 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
7831 MVT ResVT = Op.getSimpleValueType();
7832 MVT InVT = In.getSimpleValueType();
7834 if (Subtarget->hasFp256()) {
7835 if (ResVT.is128BitVector() &&
7836 (InVT.is256BitVector() || InVT.is512BitVector()) &&
7837 isa<ConstantSDNode>(Idx)) {
7838 return Extract128BitVector(In, IdxVal, DAG, dl);
7840 if (ResVT.is256BitVector() && InVT.is512BitVector() &&
7841 isa<ConstantSDNode>(Idx)) {
7842 return Extract256BitVector(In, IdxVal, DAG, dl);
7848 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
7849 // simple superregister reference or explicit instructions to insert
7850 // the upper bits of a vector.
7851 static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
7852 SelectionDAG &DAG) {
7853 if (Subtarget->hasFp256()) {
7854 SDLoc dl(Op.getNode());
7855 SDValue Vec = Op.getNode()->getOperand(0);
7856 SDValue SubVec = Op.getNode()->getOperand(1);
7857 SDValue Idx = Op.getNode()->getOperand(2);
7859 if ((Op.getNode()->getSimpleValueType(0).is256BitVector() ||
7860 Op.getNode()->getSimpleValueType(0).is512BitVector()) &&
7861 SubVec.getNode()->getSimpleValueType(0).is128BitVector() &&
7862 isa<ConstantSDNode>(Idx)) {
7863 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
7864 return Insert128BitVector(Vec, SubVec, IdxVal, DAG, dl);
7867 if (Op.getNode()->getSimpleValueType(0).is512BitVector() &&
7868 SubVec.getNode()->getSimpleValueType(0).is256BitVector() &&
7869 isa<ConstantSDNode>(Idx)) {
7870 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
7871 return Insert256BitVector(Vec, SubVec, IdxVal, DAG, dl);
7877 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
7878 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
7879 // one of the above mentioned nodes. It has to be wrapped because otherwise
7880 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
7881 // be used to form addressing mode. These wrapped nodes will be selected
7884 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
7885 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
7887 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
7889 unsigned char OpFlag = 0;
7890 unsigned WrapperKind = X86ISD::Wrapper;
7891 CodeModel::Model M = getTargetMachine().getCodeModel();
7893 if (Subtarget->isPICStyleRIPRel() &&
7894 (M == CodeModel::Small || M == CodeModel::Kernel))
7895 WrapperKind = X86ISD::WrapperRIP;
7896 else if (Subtarget->isPICStyleGOT())
7897 OpFlag = X86II::MO_GOTOFF;
7898 else if (Subtarget->isPICStyleStubPIC())
7899 OpFlag = X86II::MO_PIC_BASE_OFFSET;
7901 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
7903 CP->getOffset(), OpFlag);
7905 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
7906 // With PIC, the address is actually $g + Offset.
7908 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
7909 DAG.getNode(X86ISD::GlobalBaseReg,
7910 SDLoc(), getPointerTy()),
7917 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
7918 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
7920 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
7922 unsigned char OpFlag = 0;
7923 unsigned WrapperKind = X86ISD::Wrapper;
7924 CodeModel::Model M = getTargetMachine().getCodeModel();
7926 if (Subtarget->isPICStyleRIPRel() &&
7927 (M == CodeModel::Small || M == CodeModel::Kernel))
7928 WrapperKind = X86ISD::WrapperRIP;
7929 else if (Subtarget->isPICStyleGOT())
7930 OpFlag = X86II::MO_GOTOFF;
7931 else if (Subtarget->isPICStyleStubPIC())
7932 OpFlag = X86II::MO_PIC_BASE_OFFSET;
7934 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
7937 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
7939 // With PIC, the address is actually $g + Offset.
7941 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
7942 DAG.getNode(X86ISD::GlobalBaseReg,
7943 SDLoc(), getPointerTy()),
7950 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
7951 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
7953 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
7955 unsigned char OpFlag = 0;
7956 unsigned WrapperKind = X86ISD::Wrapper;
7957 CodeModel::Model M = getTargetMachine().getCodeModel();
7959 if (Subtarget->isPICStyleRIPRel() &&
7960 (M == CodeModel::Small || M == CodeModel::Kernel)) {
7961 if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF())
7962 OpFlag = X86II::MO_GOTPCREL;
7963 WrapperKind = X86ISD::WrapperRIP;
7964 } else if (Subtarget->isPICStyleGOT()) {
7965 OpFlag = X86II::MO_GOT;
7966 } else if (Subtarget->isPICStyleStubPIC()) {
7967 OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE;
7968 } else if (Subtarget->isPICStyleStubNoDynamic()) {
7969 OpFlag = X86II::MO_DARWIN_NONLAZY;
7972 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
7975 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
7977 // With PIC, the address is actually $g + Offset.
7978 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
7979 !Subtarget->is64Bit()) {
7980 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
7981 DAG.getNode(X86ISD::GlobalBaseReg,
7982 SDLoc(), getPointerTy()),
7986 // For symbols that require a load from a stub to get the address, emit the
7988 if (isGlobalStubReference(OpFlag))
7989 Result = DAG.getLoad(getPointerTy(), DL, DAG.getEntryNode(), Result,
7990 MachinePointerInfo::getGOT(), false, false, false, 0);
7996 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
7997 // Create the TargetBlockAddressAddress node.
7998 unsigned char OpFlags =
7999 Subtarget->ClassifyBlockAddressReference();
8000 CodeModel::Model M = getTargetMachine().getCodeModel();
8001 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
8002 int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset();
8004 SDValue Result = DAG.getTargetBlockAddress(BA, getPointerTy(), Offset,
8007 if (Subtarget->isPICStyleRIPRel() &&
8008 (M == CodeModel::Small || M == CodeModel::Kernel))
8009 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
8011 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
8013 // With PIC, the address is actually $g + Offset.
8014 if (isGlobalRelativeToPICBase(OpFlags)) {
8015 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
8016 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
8024 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
8025 int64_t Offset, SelectionDAG &DAG) const {
8026 // Create the TargetGlobalAddress node, folding in the constant
8027 // offset if it is legal.
8028 unsigned char OpFlags =
8029 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
8030 CodeModel::Model M = getTargetMachine().getCodeModel();
8032 if (OpFlags == X86II::MO_NO_FLAG &&
8033 X86::isOffsetSuitableForCodeModel(Offset, M)) {
8034 // A direct static reference to a global.
8035 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
8038 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
8041 if (Subtarget->isPICStyleRIPRel() &&
8042 (M == CodeModel::Small || M == CodeModel::Kernel))
8043 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
8045 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
8047 // With PIC, the address is actually $g + Offset.
8048 if (isGlobalRelativeToPICBase(OpFlags)) {
8049 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
8050 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
8054 // For globals that require a load from a stub to get the address, emit the
8056 if (isGlobalStubReference(OpFlags))
8057 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
8058 MachinePointerInfo::getGOT(), false, false, false, 0);
8060 // If there was a non-zero offset that we didn't fold, create an explicit
8063 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
8064 DAG.getConstant(Offset, getPointerTy()));
8070 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
8071 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
8072 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
8073 return LowerGlobalAddress(GV, SDLoc(Op), Offset, DAG);
8077 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
8078 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
8079 unsigned char OperandFlags, bool LocalDynamic = false) {
8080 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8081 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
8083 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
8084 GA->getValueType(0),
8088 X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR
8092 SDValue Ops[] = { Chain, TGA, *InFlag };
8093 Chain = DAG.getNode(CallType, dl, NodeTys, Ops, array_lengthof(Ops));
8095 SDValue Ops[] = { Chain, TGA };
8096 Chain = DAG.getNode(CallType, dl, NodeTys, Ops, array_lengthof(Ops));
8099 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
8100 MFI->setAdjustsStack(true);
8102 SDValue Flag = Chain.getValue(1);
8103 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
8106 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
8108 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
8111 SDLoc dl(GA); // ? function entry point might be better
8112 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
8113 DAG.getNode(X86ISD::GlobalBaseReg,
8114 SDLoc(), PtrVT), InFlag);
8115 InFlag = Chain.getValue(1);
8117 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
8120 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
8122 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
8124 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
8125 X86::RAX, X86II::MO_TLSGD);
8128 static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA,
8134 // Get the start address of the TLS block for this module.
8135 X86MachineFunctionInfo* MFI = DAG.getMachineFunction()
8136 .getInfo<X86MachineFunctionInfo>();
8137 MFI->incNumLocalDynamicTLSAccesses();
8141 Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT, X86::RAX,
8142 X86II::MO_TLSLD, /*LocalDynamic=*/true);
8145 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
8146 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), InFlag);
8147 InFlag = Chain.getValue(1);
8148 Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX,
8149 X86II::MO_TLSLDM, /*LocalDynamic=*/true);
8152 // Note: the CleanupLocalDynamicTLSPass will remove redundant computations
8156 unsigned char OperandFlags = X86II::MO_DTPOFF;
8157 unsigned WrapperKind = X86ISD::Wrapper;
8158 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
8159 GA->getValueType(0),
8160 GA->getOffset(), OperandFlags);
8161 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
8163 // Add x@dtpoff with the base.
8164 return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base);
8167 // Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model.
8168 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
8169 const EVT PtrVT, TLSModel::Model model,
8170 bool is64Bit, bool isPIC) {
8173 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
8174 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
8175 is64Bit ? 257 : 256));
8177 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
8178 DAG.getIntPtrConstant(0),
8179 MachinePointerInfo(Ptr),
8180 false, false, false, 0);
8182 unsigned char OperandFlags = 0;
8183 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
8185 unsigned WrapperKind = X86ISD::Wrapper;
8186 if (model == TLSModel::LocalExec) {
8187 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
8188 } else if (model == TLSModel::InitialExec) {
8190 OperandFlags = X86II::MO_GOTTPOFF;
8191 WrapperKind = X86ISD::WrapperRIP;
8193 OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF;
8196 llvm_unreachable("Unexpected model");
8199 // emit "addl x@ntpoff,%eax" (local exec)
8200 // or "addl x@indntpoff,%eax" (initial exec)
8201 // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic)
8202 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
8203 GA->getValueType(0),
8204 GA->getOffset(), OperandFlags);
8205 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
8207 if (model == TLSModel::InitialExec) {
8208 if (isPIC && !is64Bit) {
8209 Offset = DAG.getNode(ISD::ADD, dl, PtrVT,
8210 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
8214 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
8215 MachinePointerInfo::getGOT(), false, false, false,
8219 // The address of the thread local variable is the add of the thread
8220 // pointer with the offset of the variable.
8221 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
8225 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
8227 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
8228 const GlobalValue *GV = GA->getGlobal();
8230 if (Subtarget->isTargetELF()) {
8231 TLSModel::Model model = getTargetMachine().getTLSModel(GV);
8234 case TLSModel::GeneralDynamic:
8235 if (Subtarget->is64Bit())
8236 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
8237 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
8238 case TLSModel::LocalDynamic:
8239 return LowerToTLSLocalDynamicModel(GA, DAG, getPointerTy(),
8240 Subtarget->is64Bit());
8241 case TLSModel::InitialExec:
8242 case TLSModel::LocalExec:
8243 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
8244 Subtarget->is64Bit(),
8245 getTargetMachine().getRelocationModel() == Reloc::PIC_);
8247 llvm_unreachable("Unknown TLS model.");
8250 if (Subtarget->isTargetDarwin()) {
8251 // Darwin only has one model of TLS. Lower to that.
8252 unsigned char OpFlag = 0;
8253 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
8254 X86ISD::WrapperRIP : X86ISD::Wrapper;
8256 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
8258 bool PIC32 = (getTargetMachine().getRelocationModel() == Reloc::PIC_) &&
8259 !Subtarget->is64Bit();
8261 OpFlag = X86II::MO_TLVP_PIC_BASE;
8263 OpFlag = X86II::MO_TLVP;
8265 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
8266 GA->getValueType(0),
8267 GA->getOffset(), OpFlag);
8268 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
8270 // With PIC32, the address is actually $g + Offset.
8272 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
8273 DAG.getNode(X86ISD::GlobalBaseReg,
8274 SDLoc(), getPointerTy()),
8277 // Lowering the machine isd will make sure everything is in the right
8279 SDValue Chain = DAG.getEntryNode();
8280 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
8281 SDValue Args[] = { Chain, Offset };
8282 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args, 2);
8284 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
8285 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8286 MFI->setAdjustsStack(true);
8288 // And our return value (tls address) is in the standard call return value
8290 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
8291 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy(),
8295 if (Subtarget->isTargetWindows() || Subtarget->isTargetMingw()) {
8296 // Just use the implicit TLS architecture
8297 // Need to generate someting similar to:
8298 // mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage
8300 // mov ecx, dword [rel _tls_index]: Load index (from C runtime)
8301 // mov rcx, qword [rdx+rcx*8]
8302 // mov eax, .tls$:tlsvar
8303 // [rax+rcx] contains the address
8304 // Windows 64bit: gs:0x58
8305 // Windows 32bit: fs:__tls_array
8307 // If GV is an alias then use the aliasee for determining
8308 // thread-localness.
8309 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
8310 GV = GA->resolveAliasedGlobal(false);
8312 SDValue Chain = DAG.getEntryNode();
8314 // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or
8315 // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly
8316 // use its literal value of 0x2C.
8317 Value *Ptr = Constant::getNullValue(Subtarget->is64Bit()
8318 ? Type::getInt8PtrTy(*DAG.getContext(),
8320 : Type::getInt32PtrTy(*DAG.getContext(),
8323 SDValue TlsArray = Subtarget->is64Bit() ? DAG.getIntPtrConstant(0x58) :
8324 (Subtarget->isTargetMingw() ? DAG.getIntPtrConstant(0x2C) :
8325 DAG.getExternalSymbol("_tls_array", getPointerTy()));
8327 SDValue ThreadPointer = DAG.getLoad(getPointerTy(), dl, Chain, TlsArray,
8328 MachinePointerInfo(Ptr),
8329 false, false, false, 0);
8331 // Load the _tls_index variable
8332 SDValue IDX = DAG.getExternalSymbol("_tls_index", getPointerTy());
8333 if (Subtarget->is64Bit())
8334 IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, getPointerTy(), Chain,
8335 IDX, MachinePointerInfo(), MVT::i32,
8338 IDX = DAG.getLoad(getPointerTy(), dl, Chain, IDX, MachinePointerInfo(),
8339 false, false, false, 0);
8341 SDValue Scale = DAG.getConstant(Log2_64_Ceil(TD->getPointerSize()),
8343 IDX = DAG.getNode(ISD::SHL, dl, getPointerTy(), IDX, Scale);
8345 SDValue res = DAG.getNode(ISD::ADD, dl, getPointerTy(), ThreadPointer, IDX);
8346 res = DAG.getLoad(getPointerTy(), dl, Chain, res, MachinePointerInfo(),
8347 false, false, false, 0);
8349 // Get the offset of start of .tls section
8350 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
8351 GA->getValueType(0),
8352 GA->getOffset(), X86II::MO_SECREL);
8353 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), TGA);
8355 // The address of the thread local variable is the add of the thread
8356 // pointer with the offset of the variable.
8357 return DAG.getNode(ISD::ADD, dl, getPointerTy(), res, Offset);
8360 llvm_unreachable("TLS not implemented for this target.");
8363 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values
8364 /// and take a 2 x i32 value to shift plus a shift amount.
8365 SDValue X86TargetLowering::LowerShiftParts(SDValue Op, SelectionDAG &DAG) const{
8366 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
8367 EVT VT = Op.getValueType();
8368 unsigned VTBits = VT.getSizeInBits();
8370 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
8371 SDValue ShOpLo = Op.getOperand(0);
8372 SDValue ShOpHi = Op.getOperand(1);
8373 SDValue ShAmt = Op.getOperand(2);
8374 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
8375 DAG.getConstant(VTBits - 1, MVT::i8))
8376 : DAG.getConstant(0, VT);
8379 if (Op.getOpcode() == ISD::SHL_PARTS) {
8380 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
8381 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
8383 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
8384 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
8387 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
8388 DAG.getConstant(VTBits, MVT::i8));
8389 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
8390 AndNode, DAG.getConstant(0, MVT::i8));
8393 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
8394 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
8395 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
8397 if (Op.getOpcode() == ISD::SHL_PARTS) {
8398 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
8399 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
8401 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
8402 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
8405 SDValue Ops[2] = { Lo, Hi };
8406 return DAG.getMergeValues(Ops, array_lengthof(Ops), dl);
8409 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
8410 SelectionDAG &DAG) const {
8411 EVT SrcVT = Op.getOperand(0).getValueType();
8413 if (SrcVT.isVector())
8416 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
8417 "Unknown SINT_TO_FP to lower!");
8419 // These are really Legal; return the operand so the caller accepts it as
8421 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
8423 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
8424 Subtarget->is64Bit()) {
8429 unsigned Size = SrcVT.getSizeInBits()/8;
8430 MachineFunction &MF = DAG.getMachineFunction();
8431 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
8432 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
8433 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
8435 MachinePointerInfo::getFixedStack(SSFI),
8437 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
8440 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
8442 SelectionDAG &DAG) const {
8446 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
8448 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
8450 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
8452 unsigned ByteSize = SrcVT.getSizeInBits()/8;
8454 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
8455 MachineMemOperand *MMO;
8457 int SSFI = FI->getIndex();
8459 DAG.getMachineFunction()
8460 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
8461 MachineMemOperand::MOLoad, ByteSize, ByteSize);
8463 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
8464 StackSlot = StackSlot.getOperand(1);
8466 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
8467 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
8469 Tys, Ops, array_lengthof(Ops),
8473 Chain = Result.getValue(1);
8474 SDValue InFlag = Result.getValue(2);
8476 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
8477 // shouldn't be necessary except that RFP cannot be live across
8478 // multiple blocks. When stackifier is fixed, they can be uncoupled.
8479 MachineFunction &MF = DAG.getMachineFunction();
8480 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
8481 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
8482 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
8483 Tys = DAG.getVTList(MVT::Other);
8485 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
8487 MachineMemOperand *MMO =
8488 DAG.getMachineFunction()
8489 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
8490 MachineMemOperand::MOStore, SSFISize, SSFISize);
8492 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
8493 Ops, array_lengthof(Ops),
8494 Op.getValueType(), MMO);
8495 Result = DAG.getLoad(Op.getValueType(), DL, Chain, StackSlot,
8496 MachinePointerInfo::getFixedStack(SSFI),
8497 false, false, false, 0);
8503 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
8504 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
8505 SelectionDAG &DAG) const {
8506 // This algorithm is not obvious. Here it is what we're trying to output:
8509 punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U }
8510 subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 }
8514 pshufd $0x4e, %xmm0, %xmm1
8520 LLVMContext *Context = DAG.getContext();
8522 // Build some magic constants.
8523 static const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 };
8524 Constant *C0 = ConstantDataVector::get(*Context, CV0);
8525 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
8527 SmallVector<Constant*,2> CV1;
8529 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
8530 APInt(64, 0x4330000000000000ULL))));
8532 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
8533 APInt(64, 0x4530000000000000ULL))));
8534 Constant *C1 = ConstantVector::get(CV1);
8535 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
8537 // Load the 64-bit value into an XMM register.
8538 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
8540 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
8541 MachinePointerInfo::getConstantPool(),
8542 false, false, false, 16);
8543 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32,
8544 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, XR1),
8547 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
8548 MachinePointerInfo::getConstantPool(),
8549 false, false, false, 16);
8550 SDValue XR2F = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Unpck1);
8551 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
8554 if (Subtarget->hasSSE3()) {
8555 // FIXME: The 'haddpd' instruction may be slower than 'movhlps + addsd'.
8556 Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub);
8558 SDValue S2F = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Sub);
8559 SDValue Shuffle = getTargetShuffleNode(X86ISD::PSHUFD, dl, MVT::v4i32,
8561 Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64,
8562 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Shuffle),
8566 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result,
8567 DAG.getIntPtrConstant(0));
8570 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
8571 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
8572 SelectionDAG &DAG) const {
8574 // FP constant to bias correct the final result.
8575 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
8578 // Load the 32-bit value into an XMM register.
8579 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
8582 // Zero out the upper parts of the register.
8583 Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG);
8585 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
8586 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Load),
8587 DAG.getIntPtrConstant(0));
8589 // Or the load with the bias.
8590 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
8591 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
8592 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
8594 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
8595 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
8596 MVT::v2f64, Bias)));
8597 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
8598 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or),
8599 DAG.getIntPtrConstant(0));
8601 // Subtract the bias.
8602 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
8604 // Handle final rounding.
8605 EVT DestVT = Op.getValueType();
8607 if (DestVT.bitsLT(MVT::f64))
8608 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
8609 DAG.getIntPtrConstant(0));
8610 if (DestVT.bitsGT(MVT::f64))
8611 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
8613 // Handle final rounding.
8617 SDValue X86TargetLowering::lowerUINT_TO_FP_vec(SDValue Op,
8618 SelectionDAG &DAG) const {
8619 SDValue N0 = Op.getOperand(0);
8620 EVT SVT = N0.getValueType();
8623 assert((SVT == MVT::v4i8 || SVT == MVT::v4i16 ||
8624 SVT == MVT::v8i8 || SVT == MVT::v8i16) &&
8625 "Custom UINT_TO_FP is not supported!");
8627 EVT NVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
8628 SVT.getVectorNumElements());
8629 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
8630 DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0));
8633 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
8634 SelectionDAG &DAG) const {
8635 SDValue N0 = Op.getOperand(0);
8638 if (Op.getValueType().isVector())
8639 return lowerUINT_TO_FP_vec(Op, DAG);
8641 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
8642 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
8643 // the optimization here.
8644 if (DAG.SignBitIsZero(N0))
8645 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
8647 EVT SrcVT = N0.getValueType();
8648 EVT DstVT = Op.getValueType();
8649 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
8650 return LowerUINT_TO_FP_i64(Op, DAG);
8651 if (SrcVT == MVT::i32 && X86ScalarSSEf64)
8652 return LowerUINT_TO_FP_i32(Op, DAG);
8653 if (Subtarget->is64Bit() && SrcVT == MVT::i64 && DstVT == MVT::f32)
8656 // Make a 64-bit buffer, and use it to build an FILD.
8657 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
8658 if (SrcVT == MVT::i32) {
8659 SDValue WordOff = DAG.getConstant(4, getPointerTy());
8660 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
8661 getPointerTy(), StackSlot, WordOff);
8662 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
8663 StackSlot, MachinePointerInfo(),
8665 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
8666 OffsetSlot, MachinePointerInfo(),
8668 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
8672 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
8673 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
8674 StackSlot, MachinePointerInfo(),
8676 // For i64 source, we need to add the appropriate power of 2 if the input
8677 // was negative. This is the same as the optimization in
8678 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
8679 // we must be careful to do the computation in x87 extended precision, not
8680 // in SSE. (The generic code can't know it's OK to do this, or how to.)
8681 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
8682 MachineMemOperand *MMO =
8683 DAG.getMachineFunction()
8684 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
8685 MachineMemOperand::MOLoad, 8, 8);
8687 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
8688 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
8689 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops,
8690 array_lengthof(Ops), MVT::i64, MMO);
8692 APInt FF(32, 0x5F800000ULL);
8694 // Check whether the sign bit is set.
8695 SDValue SignSet = DAG.getSetCC(dl,
8696 getSetCCResultType(*DAG.getContext(), MVT::i64),
8697 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
8700 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
8701 SDValue FudgePtr = DAG.getConstantPool(
8702 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
8705 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
8706 SDValue Zero = DAG.getIntPtrConstant(0);
8707 SDValue Four = DAG.getIntPtrConstant(4);
8708 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
8710 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
8712 // Load the value out, extending it from f32 to f80.
8713 // FIXME: Avoid the extend by constructing the right constant pool?
8714 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
8715 FudgePtr, MachinePointerInfo::getConstantPool(),
8716 MVT::f32, false, false, 4);
8717 // Extend everything to 80 bits to force it to be done on x87.
8718 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
8719 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
8722 std::pair<SDValue,SDValue>
8723 X86TargetLowering:: FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
8724 bool IsSigned, bool IsReplace) const {
8727 EVT DstTy = Op.getValueType();
8729 if (!IsSigned && !isIntegerTypeFTOL(DstTy)) {
8730 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
8734 assert(DstTy.getSimpleVT() <= MVT::i64 &&
8735 DstTy.getSimpleVT() >= MVT::i16 &&
8736 "Unknown FP_TO_INT to lower!");
8738 // These are really Legal.
8739 if (DstTy == MVT::i32 &&
8740 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
8741 return std::make_pair(SDValue(), SDValue());
8742 if (Subtarget->is64Bit() &&
8743 DstTy == MVT::i64 &&
8744 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
8745 return std::make_pair(SDValue(), SDValue());
8747 // We lower FP->int64 either into FISTP64 followed by a load from a temporary
8748 // stack slot, or into the FTOL runtime function.
8749 MachineFunction &MF = DAG.getMachineFunction();
8750 unsigned MemSize = DstTy.getSizeInBits()/8;
8751 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
8752 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
8755 if (!IsSigned && isIntegerTypeFTOL(DstTy))
8756 Opc = X86ISD::WIN_FTOL;
8758 switch (DstTy.getSimpleVT().SimpleTy) {
8759 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
8760 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
8761 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
8762 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
8765 SDValue Chain = DAG.getEntryNode();
8766 SDValue Value = Op.getOperand(0);
8767 EVT TheVT = Op.getOperand(0).getValueType();
8768 // FIXME This causes a redundant load/store if the SSE-class value is already
8769 // in memory, such as if it is on the callstack.
8770 if (isScalarFPTypeInSSEReg(TheVT)) {
8771 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
8772 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
8773 MachinePointerInfo::getFixedStack(SSFI),
8775 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
8777 Chain, StackSlot, DAG.getValueType(TheVT)
8780 MachineMemOperand *MMO =
8781 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
8782 MachineMemOperand::MOLoad, MemSize, MemSize);
8783 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops,
8784 array_lengthof(Ops), DstTy, MMO);
8785 Chain = Value.getValue(1);
8786 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
8787 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
8790 MachineMemOperand *MMO =
8791 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
8792 MachineMemOperand::MOStore, MemSize, MemSize);
8794 if (Opc != X86ISD::WIN_FTOL) {
8795 // Build the FP_TO_INT*_IN_MEM
8796 SDValue Ops[] = { Chain, Value, StackSlot };
8797 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
8798 Ops, array_lengthof(Ops), DstTy,
8800 return std::make_pair(FIST, StackSlot);
8802 SDValue ftol = DAG.getNode(X86ISD::WIN_FTOL, DL,
8803 DAG.getVTList(MVT::Other, MVT::Glue),
8805 SDValue eax = DAG.getCopyFromReg(ftol, DL, X86::EAX,
8806 MVT::i32, ftol.getValue(1));
8807 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), DL, X86::EDX,
8808 MVT::i32, eax.getValue(2));
8809 SDValue Ops[] = { eax, edx };
8810 SDValue pair = IsReplace
8811 ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops, array_lengthof(Ops))
8812 : DAG.getMergeValues(Ops, array_lengthof(Ops), DL);
8813 return std::make_pair(pair, SDValue());
8817 static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
8818 const X86Subtarget *Subtarget) {
8819 MVT VT = Op->getSimpleValueType(0);
8820 SDValue In = Op->getOperand(0);
8821 MVT InVT = In.getSimpleValueType();
8824 // Optimize vectors in AVX mode:
8827 // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32.
8828 // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32.
8829 // Concat upper and lower parts.
8832 // Use vpunpckldq for 4 lower elements v4i32 -> v2i64.
8833 // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64.
8834 // Concat upper and lower parts.
8837 if (((VT != MVT::v8i32) || (InVT != MVT::v8i16)) &&
8838 ((VT != MVT::v4i64) || (InVT != MVT::v4i32)))
8841 if (Subtarget->hasInt256())
8842 return DAG.getNode(X86ISD::VZEXT_MOVL, dl, VT, In);
8844 SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl);
8845 SDValue Undef = DAG.getUNDEF(InVT);
8846 bool NeedZero = Op.getOpcode() == ISD::ZERO_EXTEND;
8847 SDValue OpLo = getUnpackl(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
8848 SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
8850 MVT HVT = MVT::getVectorVT(VT.getVectorElementType(),
8851 VT.getVectorNumElements()/2);
8853 OpLo = DAG.getNode(ISD::BITCAST, dl, HVT, OpLo);
8854 OpHi = DAG.getNode(ISD::BITCAST, dl, HVT, OpHi);
8856 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
8859 static SDValue LowerZERO_EXTEND_AVX512(SDValue Op,
8860 SelectionDAG &DAG) {
8861 MVT VT = Op->getValueType(0).getSimpleVT();
8862 SDValue In = Op->getOperand(0);
8863 MVT InVT = In.getValueType().getSimpleVT();
8865 unsigned int NumElts = VT.getVectorNumElements();
8866 if (NumElts != 8 && NumElts != 16)
8869 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1)
8870 return DAG.getNode(X86ISD::VZEXT, DL, VT, In);
8872 EVT ExtVT = (NumElts == 8)? MVT::v8i64 : MVT::v16i32;
8873 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8874 // Now we have only mask extension
8875 assert(InVT.getVectorElementType() == MVT::i1);
8876 SDValue Cst = DAG.getTargetConstant(1, ExtVT.getScalarType());
8877 const Constant *C = (dyn_cast<ConstantSDNode>(Cst))->getConstantIntValue();
8878 SDValue CP = DAG.getConstantPool(C, TLI.getPointerTy());
8879 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
8880 SDValue Ld = DAG.getLoad(Cst.getValueType(), DL, DAG.getEntryNode(), CP,
8881 MachinePointerInfo::getConstantPool(),
8882 false, false, false, Alignment);
8884 SDValue Brcst = DAG.getNode(X86ISD::VBROADCASTM, DL, ExtVT, In, Ld);
8885 if (VT.is512BitVector())
8887 return DAG.getNode(X86ISD::VTRUNC, DL, VT, Brcst);
8890 static SDValue LowerANY_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
8891 SelectionDAG &DAG) {
8892 if (Subtarget->hasFp256()) {
8893 SDValue Res = LowerAVXExtend(Op, DAG, Subtarget);
8901 static SDValue LowerZERO_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
8902 SelectionDAG &DAG) {
8904 MVT VT = Op.getSimpleValueType();
8905 SDValue In = Op.getOperand(0);
8906 MVT SVT = In.getSimpleValueType();
8908 if (VT.is512BitVector() || SVT.getVectorElementType() == MVT::i1)
8909 return LowerZERO_EXTEND_AVX512(Op, DAG);
8911 if (Subtarget->hasFp256()) {
8912 SDValue Res = LowerAVXExtend(Op, DAG, Subtarget);
8917 if (!VT.is256BitVector() || !SVT.is128BitVector() ||
8918 VT.getVectorNumElements() != SVT.getVectorNumElements())
8921 assert(Subtarget->hasFp256() && "256-bit vector is observed without AVX!");
8923 // AVX2 has better support of integer extending.
8924 if (Subtarget->hasInt256())
8925 return DAG.getNode(X86ISD::VZEXT, DL, VT, In);
8927 SDValue Lo = DAG.getNode(X86ISD::VZEXT, DL, MVT::v4i32, In);
8928 static const int Mask[] = {4, 5, 6, 7, -1, -1, -1, -1};
8929 SDValue Hi = DAG.getNode(X86ISD::VZEXT, DL, MVT::v4i32,
8930 DAG.getVectorShuffle(MVT::v8i16, DL, In,
8931 DAG.getUNDEF(MVT::v8i16),
8934 return DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v8i32, Lo, Hi);
8937 SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
8939 MVT VT = Op.getSimpleValueType();
8940 SDValue In = Op.getOperand(0);
8941 MVT InVT = In.getSimpleValueType();
8942 assert(VT.getVectorNumElements() == InVT.getVectorNumElements() &&
8943 "Invalid TRUNCATE operation");
8945 if (InVT.is512BitVector() || VT.getVectorElementType() == MVT::i1) {
8946 if (VT.getVectorElementType().getSizeInBits() >=8)
8947 return DAG.getNode(X86ISD::VTRUNC, DL, VT, In);
8949 assert(VT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
8950 unsigned NumElts = InVT.getVectorNumElements();
8951 assert ((NumElts == 8 || NumElts == 16) && "Unexpected vector type");
8952 if (InVT.getSizeInBits() < 512) {
8953 MVT ExtVT = (NumElts == 16)? MVT::v16i32 : MVT::v8i64;
8954 In = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, In);
8957 SDValue Cst = DAG.getTargetConstant(1, InVT.getVectorElementType());
8958 const Constant *C = (dyn_cast<ConstantSDNode>(Cst))->getConstantIntValue();
8959 SDValue CP = DAG.getConstantPool(C, getPointerTy());
8960 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
8961 SDValue Ld = DAG.getLoad(Cst.getValueType(), DL, DAG.getEntryNode(), CP,
8962 MachinePointerInfo::getConstantPool(),
8963 false, false, false, Alignment);
8964 SDValue OneV = DAG.getNode(X86ISD::VBROADCAST, DL, InVT, Ld);
8965 SDValue And = DAG.getNode(ISD::AND, DL, InVT, OneV, In);
8966 return DAG.getNode(X86ISD::TESTM, DL, VT, And, And);
8969 if ((VT == MVT::v4i32) && (InVT == MVT::v4i64)) {
8970 // On AVX2, v4i64 -> v4i32 becomes VPERMD.
8971 if (Subtarget->hasInt256()) {
8972 static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1};
8973 In = DAG.getNode(ISD::BITCAST, DL, MVT::v8i32, In);
8974 In = DAG.getVectorShuffle(MVT::v8i32, DL, In, DAG.getUNDEF(MVT::v8i32),
8976 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, In,
8977 DAG.getIntPtrConstant(0));
8980 // On AVX, v4i64 -> v4i32 becomes a sequence that uses PSHUFD and MOVLHPS.
8981 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
8982 DAG.getIntPtrConstant(0));
8983 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
8984 DAG.getIntPtrConstant(2));
8986 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpLo);
8987 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpHi);
8990 static const int ShufMask1[] = {0, 2, 0, 0};
8991 SDValue Undef = DAG.getUNDEF(VT);
8992 OpLo = DAG.getVectorShuffle(VT, DL, OpLo, Undef, ShufMask1);
8993 OpHi = DAG.getVectorShuffle(VT, DL, OpHi, Undef, ShufMask1);
8995 // The MOVLHPS mask:
8996 static const int ShufMask2[] = {0, 1, 4, 5};
8997 return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask2);
9000 if ((VT == MVT::v8i16) && (InVT == MVT::v8i32)) {
9001 // On AVX2, v8i32 -> v8i16 becomed PSHUFB.
9002 if (Subtarget->hasInt256()) {
9003 In = DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, In);
9005 SmallVector<SDValue,32> pshufbMask;
9006 for (unsigned i = 0; i < 2; ++i) {
9007 pshufbMask.push_back(DAG.getConstant(0x0, MVT::i8));
9008 pshufbMask.push_back(DAG.getConstant(0x1, MVT::i8));
9009 pshufbMask.push_back(DAG.getConstant(0x4, MVT::i8));
9010 pshufbMask.push_back(DAG.getConstant(0x5, MVT::i8));
9011 pshufbMask.push_back(DAG.getConstant(0x8, MVT::i8));
9012 pshufbMask.push_back(DAG.getConstant(0x9, MVT::i8));
9013 pshufbMask.push_back(DAG.getConstant(0xc, MVT::i8));
9014 pshufbMask.push_back(DAG.getConstant(0xd, MVT::i8));
9015 for (unsigned j = 0; j < 8; ++j)
9016 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
9018 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8,
9019 &pshufbMask[0], 32);
9020 In = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8, In, BV);
9021 In = DAG.getNode(ISD::BITCAST, DL, MVT::v4i64, In);
9023 static const int ShufMask[] = {0, 2, -1, -1};
9024 In = DAG.getVectorShuffle(MVT::v4i64, DL, In, DAG.getUNDEF(MVT::v4i64),
9026 In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
9027 DAG.getIntPtrConstant(0));
9028 return DAG.getNode(ISD::BITCAST, DL, VT, In);
9031 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
9032 DAG.getIntPtrConstant(0));
9034 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
9035 DAG.getIntPtrConstant(4));
9037 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, OpLo);
9038 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, OpHi);
9041 static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13,
9042 -1, -1, -1, -1, -1, -1, -1, -1};
9044 SDValue Undef = DAG.getUNDEF(MVT::v16i8);
9045 OpLo = DAG.getVectorShuffle(MVT::v16i8, DL, OpLo, Undef, ShufMask1);
9046 OpHi = DAG.getVectorShuffle(MVT::v16i8, DL, OpHi, Undef, ShufMask1);
9048 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpLo);
9049 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpHi);
9051 // The MOVLHPS Mask:
9052 static const int ShufMask2[] = {0, 1, 4, 5};
9053 SDValue res = DAG.getVectorShuffle(MVT::v4i32, DL, OpLo, OpHi, ShufMask2);
9054 return DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, res);
9057 // Handle truncation of V256 to V128 using shuffles.
9058 if (!VT.is128BitVector() || !InVT.is256BitVector())
9061 assert(Subtarget->hasFp256() && "256-bit vector without AVX!");
9063 unsigned NumElems = VT.getVectorNumElements();
9064 EVT NVT = EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(),
9067 SmallVector<int, 16> MaskVec(NumElems * 2, -1);
9068 // Prepare truncation shuffle mask
9069 for (unsigned i = 0; i != NumElems; ++i)
9071 SDValue V = DAG.getVectorShuffle(NVT, DL,
9072 DAG.getNode(ISD::BITCAST, DL, NVT, In),
9073 DAG.getUNDEF(NVT), &MaskVec[0]);
9074 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V,
9075 DAG.getIntPtrConstant(0));
9078 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
9079 SelectionDAG &DAG) const {
9080 MVT VT = Op.getSimpleValueType();
9081 if (VT.isVector()) {
9082 if (VT == MVT::v8i16)
9083 return DAG.getNode(ISD::TRUNCATE, SDLoc(Op), VT,
9084 DAG.getNode(ISD::FP_TO_SINT, SDLoc(Op),
9085 MVT::v8i32, Op.getOperand(0)));
9089 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
9090 /*IsSigned=*/ true, /*IsReplace=*/ false);
9091 SDValue FIST = Vals.first, StackSlot = Vals.second;
9092 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
9093 if (FIST.getNode() == 0) return Op;
9095 if (StackSlot.getNode())
9097 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
9098 FIST, StackSlot, MachinePointerInfo(),
9099 false, false, false, 0);
9101 // The node is the result.
9105 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
9106 SelectionDAG &DAG) const {
9107 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
9108 /*IsSigned=*/ false, /*IsReplace=*/ false);
9109 SDValue FIST = Vals.first, StackSlot = Vals.second;
9110 assert(FIST.getNode() && "Unexpected failure");
9112 if (StackSlot.getNode())
9114 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
9115 FIST, StackSlot, MachinePointerInfo(),
9116 false, false, false, 0);
9118 // The node is the result.
9122 static SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) {
9124 MVT VT = Op.getSimpleValueType();
9125 SDValue In = Op.getOperand(0);
9126 MVT SVT = In.getSimpleValueType();
9128 assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!");
9130 return DAG.getNode(X86ISD::VFPEXT, DL, VT,
9131 DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32,
9132 In, DAG.getUNDEF(SVT)));
9135 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) const {
9136 LLVMContext *Context = DAG.getContext();
9138 MVT VT = Op.getSimpleValueType();
9140 unsigned NumElts = VT == MVT::f64 ? 2 : 4;
9141 if (VT.isVector()) {
9142 EltVT = VT.getVectorElementType();
9143 NumElts = VT.getVectorNumElements();
9146 if (EltVT == MVT::f64)
9147 C = ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
9148 APInt(64, ~(1ULL << 63))));
9150 C = ConstantFP::get(*Context, APFloat(APFloat::IEEEsingle,
9151 APInt(32, ~(1U << 31))));
9152 C = ConstantVector::getSplat(NumElts, C);
9153 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy());
9154 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
9155 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
9156 MachinePointerInfo::getConstantPool(),
9157 false, false, false, Alignment);
9158 if (VT.isVector()) {
9159 MVT ANDVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
9160 return DAG.getNode(ISD::BITCAST, dl, VT,
9161 DAG.getNode(ISD::AND, dl, ANDVT,
9162 DAG.getNode(ISD::BITCAST, dl, ANDVT,
9164 DAG.getNode(ISD::BITCAST, dl, ANDVT, Mask)));
9166 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
9169 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
9170 LLVMContext *Context = DAG.getContext();
9172 MVT VT = Op.getSimpleValueType();
9174 unsigned NumElts = VT == MVT::f64 ? 2 : 4;
9175 if (VT.isVector()) {
9176 EltVT = VT.getVectorElementType();
9177 NumElts = VT.getVectorNumElements();
9180 if (EltVT == MVT::f64)
9181 C = ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
9182 APInt(64, 1ULL << 63)));
9184 C = ConstantFP::get(*Context, APFloat(APFloat::IEEEsingle,
9185 APInt(32, 1U << 31)));
9186 C = ConstantVector::getSplat(NumElts, C);
9187 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy());
9188 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
9189 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
9190 MachinePointerInfo::getConstantPool(),
9191 false, false, false, Alignment);
9192 if (VT.isVector()) {
9193 MVT XORVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits()/64);
9194 return DAG.getNode(ISD::BITCAST, dl, VT,
9195 DAG.getNode(ISD::XOR, dl, XORVT,
9196 DAG.getNode(ISD::BITCAST, dl, XORVT,
9198 DAG.getNode(ISD::BITCAST, dl, XORVT, Mask)));
9201 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
9204 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
9205 LLVMContext *Context = DAG.getContext();
9206 SDValue Op0 = Op.getOperand(0);
9207 SDValue Op1 = Op.getOperand(1);
9209 MVT VT = Op.getSimpleValueType();
9210 MVT SrcVT = Op1.getSimpleValueType();
9212 // If second operand is smaller, extend it first.
9213 if (SrcVT.bitsLT(VT)) {
9214 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
9217 // And if it is bigger, shrink it first.
9218 if (SrcVT.bitsGT(VT)) {
9219 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
9223 // At this point the operands and the result should have the same
9224 // type, and that won't be f80 since that is not custom lowered.
9226 // First get the sign bit of second operand.
9227 SmallVector<Constant*,4> CV;
9228 if (SrcVT == MVT::f64) {
9229 const fltSemantics &Sem = APFloat::IEEEdouble;
9230 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(64, 1ULL << 63))));
9231 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(64, 0))));
9233 const fltSemantics &Sem = APFloat::IEEEsingle;
9234 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 1U << 31))));
9235 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 0))));
9236 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 0))));
9237 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 0))));
9239 Constant *C = ConstantVector::get(CV);
9240 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
9241 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
9242 MachinePointerInfo::getConstantPool(),
9243 false, false, false, 16);
9244 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
9246 // Shift sign bit right or left if the two operands have different types.
9247 if (SrcVT.bitsGT(VT)) {
9248 // Op0 is MVT::f32, Op1 is MVT::f64.
9249 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
9250 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
9251 DAG.getConstant(32, MVT::i32));
9252 SignBit = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, SignBit);
9253 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
9254 DAG.getIntPtrConstant(0));
9257 // Clear first operand sign bit.
9259 if (VT == MVT::f64) {
9260 const fltSemantics &Sem = APFloat::IEEEdouble;
9261 CV.push_back(ConstantFP::get(*Context, APFloat(Sem,
9262 APInt(64, ~(1ULL << 63)))));
9263 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(64, 0))));
9265 const fltSemantics &Sem = APFloat::IEEEsingle;
9266 CV.push_back(ConstantFP::get(*Context, APFloat(Sem,
9267 APInt(32, ~(1U << 31)))));
9268 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 0))));
9269 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 0))));
9270 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 0))));
9272 C = ConstantVector::get(CV);
9273 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
9274 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
9275 MachinePointerInfo::getConstantPool(),
9276 false, false, false, 16);
9277 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
9279 // Or the value with the sign bit.
9280 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
9283 static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) {
9284 SDValue N0 = Op.getOperand(0);
9286 MVT VT = Op.getSimpleValueType();
9288 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
9289 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
9290 DAG.getConstant(1, VT));
9291 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, VT));
9294 // LowerVectorAllZeroTest - Check whether an OR'd tree is PTEST-able.
9296 static SDValue LowerVectorAllZeroTest(SDValue Op, const X86Subtarget *Subtarget,
9297 SelectionDAG &DAG) {
9298 assert(Op.getOpcode() == ISD::OR && "Only check OR'd tree.");
9300 if (!Subtarget->hasSSE41())
9303 if (!Op->hasOneUse())
9306 SDNode *N = Op.getNode();
9309 SmallVector<SDValue, 8> Opnds;
9310 DenseMap<SDValue, unsigned> VecInMap;
9311 EVT VT = MVT::Other;
9313 // Recognize a special case where a vector is casted into wide integer to
9315 Opnds.push_back(N->getOperand(0));
9316 Opnds.push_back(N->getOperand(1));
9318 for (unsigned Slot = 0, e = Opnds.size(); Slot < e; ++Slot) {
9319 SmallVectorImpl<SDValue>::const_iterator I = Opnds.begin() + Slot;
9320 // BFS traverse all OR'd operands.
9321 if (I->getOpcode() == ISD::OR) {
9322 Opnds.push_back(I->getOperand(0));
9323 Opnds.push_back(I->getOperand(1));
9324 // Re-evaluate the number of nodes to be traversed.
9325 e += 2; // 2 more nodes (LHS and RHS) are pushed.
9329 // Quit if a non-EXTRACT_VECTOR_ELT
9330 if (I->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
9333 // Quit if without a constant index.
9334 SDValue Idx = I->getOperand(1);
9335 if (!isa<ConstantSDNode>(Idx))
9338 SDValue ExtractedFromVec = I->getOperand(0);
9339 DenseMap<SDValue, unsigned>::iterator M = VecInMap.find(ExtractedFromVec);
9340 if (M == VecInMap.end()) {
9341 VT = ExtractedFromVec.getValueType();
9342 // Quit if not 128/256-bit vector.
9343 if (!VT.is128BitVector() && !VT.is256BitVector())
9345 // Quit if not the same type.
9346 if (VecInMap.begin() != VecInMap.end() &&
9347 VT != VecInMap.begin()->first.getValueType())
9349 M = VecInMap.insert(std::make_pair(ExtractedFromVec, 0)).first;
9351 M->second |= 1U << cast<ConstantSDNode>(Idx)->getZExtValue();
9354 assert((VT.is128BitVector() || VT.is256BitVector()) &&
9355 "Not extracted from 128-/256-bit vector.");
9357 unsigned FullMask = (1U << VT.getVectorNumElements()) - 1U;
9358 SmallVector<SDValue, 8> VecIns;
9360 for (DenseMap<SDValue, unsigned>::const_iterator
9361 I = VecInMap.begin(), E = VecInMap.end(); I != E; ++I) {
9362 // Quit if not all elements are used.
9363 if (I->second != FullMask)
9365 VecIns.push_back(I->first);
9368 EVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
9370 // Cast all vectors into TestVT for PTEST.
9371 for (unsigned i = 0, e = VecIns.size(); i < e; ++i)
9372 VecIns[i] = DAG.getNode(ISD::BITCAST, DL, TestVT, VecIns[i]);
9374 // If more than one full vectors are evaluated, OR them first before PTEST.
9375 for (unsigned Slot = 0, e = VecIns.size(); e - Slot > 1; Slot += 2, e += 1) {
9376 // Each iteration will OR 2 nodes and append the result until there is only
9377 // 1 node left, i.e. the final OR'd value of all vectors.
9378 SDValue LHS = VecIns[Slot];
9379 SDValue RHS = VecIns[Slot + 1];
9380 VecIns.push_back(DAG.getNode(ISD::OR, DL, TestVT, LHS, RHS));
9383 return DAG.getNode(X86ISD::PTEST, DL, MVT::i32,
9384 VecIns.back(), VecIns.back());
9387 /// Emit nodes that will be selected as "test Op0,Op0", or something
9389 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
9390 SelectionDAG &DAG) const {
9393 // CF and OF aren't always set the way we want. Determine which
9394 // of these we need.
9395 bool NeedCF = false;
9396 bool NeedOF = false;
9399 case X86::COND_A: case X86::COND_AE:
9400 case X86::COND_B: case X86::COND_BE:
9403 case X86::COND_G: case X86::COND_GE:
9404 case X86::COND_L: case X86::COND_LE:
9405 case X86::COND_O: case X86::COND_NO:
9410 // See if we can use the EFLAGS value from the operand instead of
9411 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
9412 // we prove that the arithmetic won't overflow, we can't use OF or CF.
9413 if (Op.getResNo() != 0 || NeedOF || NeedCF)
9414 // Emit a CMP with 0, which is the TEST pattern.
9415 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
9416 DAG.getConstant(0, Op.getValueType()));
9418 unsigned Opcode = 0;
9419 unsigned NumOperands = 0;
9421 // Truncate operations may prevent the merge of the SETCC instruction
9422 // and the arithmetic intruction before it. Attempt to truncate the operands
9423 // of the arithmetic instruction and use a reduced bit-width instruction.
9424 bool NeedTruncation = false;
9425 SDValue ArithOp = Op;
9426 if (Op->getOpcode() == ISD::TRUNCATE && Op->hasOneUse()) {
9427 SDValue Arith = Op->getOperand(0);
9428 // Both the trunc and the arithmetic op need to have one user each.
9429 if (Arith->hasOneUse())
9430 switch (Arith.getOpcode()) {
9437 NeedTruncation = true;
9443 // NOTICE: In the code below we use ArithOp to hold the arithmetic operation
9444 // which may be the result of a CAST. We use the variable 'Op', which is the
9445 // non-casted variable when we check for possible users.
9446 switch (ArithOp.getOpcode()) {
9448 // Due to an isel shortcoming, be conservative if this add is likely to be
9449 // selected as part of a load-modify-store instruction. When the root node
9450 // in a match is a store, isel doesn't know how to remap non-chain non-flag
9451 // uses of other nodes in the match, such as the ADD in this case. This
9452 // leads to the ADD being left around and reselected, with the result being
9453 // two adds in the output. Alas, even if none our users are stores, that
9454 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
9455 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
9456 // climbing the DAG back to the root, and it doesn't seem to be worth the
9458 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
9459 UE = Op.getNode()->use_end(); UI != UE; ++UI)
9460 if (UI->getOpcode() != ISD::CopyToReg &&
9461 UI->getOpcode() != ISD::SETCC &&
9462 UI->getOpcode() != ISD::STORE)
9465 if (ConstantSDNode *C =
9466 dyn_cast<ConstantSDNode>(ArithOp.getNode()->getOperand(1))) {
9467 // An add of one will be selected as an INC.
9468 if (C->getAPIntValue() == 1) {
9469 Opcode = X86ISD::INC;
9474 // An add of negative one (subtract of one) will be selected as a DEC.
9475 if (C->getAPIntValue().isAllOnesValue()) {
9476 Opcode = X86ISD::DEC;
9482 // Otherwise use a regular EFLAGS-setting add.
9483 Opcode = X86ISD::ADD;
9487 // If the primary and result isn't used, don't bother using X86ISD::AND,
9488 // because a TEST instruction will be better.
9489 bool NonFlagUse = false;
9490 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
9491 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
9493 unsigned UOpNo = UI.getOperandNo();
9494 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
9495 // Look pass truncate.
9496 UOpNo = User->use_begin().getOperandNo();
9497 User = *User->use_begin();
9500 if (User->getOpcode() != ISD::BRCOND &&
9501 User->getOpcode() != ISD::SETCC &&
9502 !(User->getOpcode() == ISD::SELECT && UOpNo == 0)) {
9515 // Due to the ISEL shortcoming noted above, be conservative if this op is
9516 // likely to be selected as part of a load-modify-store instruction.
9517 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
9518 UE = Op.getNode()->use_end(); UI != UE; ++UI)
9519 if (UI->getOpcode() == ISD::STORE)
9522 // Otherwise use a regular EFLAGS-setting instruction.
9523 switch (ArithOp.getOpcode()) {
9524 default: llvm_unreachable("unexpected operator!");
9525 case ISD::SUB: Opcode = X86ISD::SUB; break;
9526 case ISD::XOR: Opcode = X86ISD::XOR; break;
9527 case ISD::AND: Opcode = X86ISD::AND; break;
9529 if (!NeedTruncation && (X86CC == X86::COND_E || X86CC == X86::COND_NE)) {
9530 SDValue EFLAGS = LowerVectorAllZeroTest(Op, Subtarget, DAG);
9531 if (EFLAGS.getNode())
9534 Opcode = X86ISD::OR;
9548 return SDValue(Op.getNode(), 1);
9554 // If we found that truncation is beneficial, perform the truncation and
9556 if (NeedTruncation) {
9557 EVT VT = Op.getValueType();
9558 SDValue WideVal = Op->getOperand(0);
9559 EVT WideVT = WideVal.getValueType();
9560 unsigned ConvertedOp = 0;
9561 // Use a target machine opcode to prevent further DAGCombine
9562 // optimizations that may separate the arithmetic operations
9563 // from the setcc node.
9564 switch (WideVal.getOpcode()) {
9566 case ISD::ADD: ConvertedOp = X86ISD::ADD; break;
9567 case ISD::SUB: ConvertedOp = X86ISD::SUB; break;
9568 case ISD::AND: ConvertedOp = X86ISD::AND; break;
9569 case ISD::OR: ConvertedOp = X86ISD::OR; break;
9570 case ISD::XOR: ConvertedOp = X86ISD::XOR; break;
9574 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9575 if (TLI.isOperationLegal(WideVal.getOpcode(), WideVT)) {
9576 SDValue V0 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(0));
9577 SDValue V1 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(1));
9578 Op = DAG.getNode(ConvertedOp, dl, VT, V0, V1);
9584 // Emit a CMP with 0, which is the TEST pattern.
9585 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
9586 DAG.getConstant(0, Op.getValueType()));
9588 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
9589 SmallVector<SDValue, 4> Ops;
9590 for (unsigned i = 0; i != NumOperands; ++i)
9591 Ops.push_back(Op.getOperand(i));
9593 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
9594 DAG.ReplaceAllUsesWith(Op, New);
9595 return SDValue(New.getNode(), 1);
9598 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
9600 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
9601 SelectionDAG &DAG) const {
9602 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
9603 if (C->getAPIntValue() == 0)
9604 return EmitTest(Op0, X86CC, DAG);
9607 if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 ||
9608 Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) {
9609 // Use SUB instead of CMP to enable CSE between SUB and CMP.
9610 SDVTList VTs = DAG.getVTList(Op0.getValueType(), MVT::i32);
9611 SDValue Sub = DAG.getNode(X86ISD::SUB, dl, VTs,
9613 return SDValue(Sub.getNode(), 1);
9615 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
9618 /// Convert a comparison if required by the subtarget.
9619 SDValue X86TargetLowering::ConvertCmpIfNecessary(SDValue Cmp,
9620 SelectionDAG &DAG) const {
9621 // If the subtarget does not support the FUCOMI instruction, floating-point
9622 // comparisons have to be converted.
9623 if (Subtarget->hasCMov() ||
9624 Cmp.getOpcode() != X86ISD::CMP ||
9625 !Cmp.getOperand(0).getValueType().isFloatingPoint() ||
9626 !Cmp.getOperand(1).getValueType().isFloatingPoint())
9629 // The instruction selector will select an FUCOM instruction instead of
9630 // FUCOMI, which writes the comparison result to FPSW instead of EFLAGS. Hence
9631 // build an SDNode sequence that transfers the result from FPSW into EFLAGS:
9632 // (X86sahf (trunc (srl (X86fp_stsw (trunc (X86cmp ...)), 8))))
9634 SDValue TruncFPSW = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, Cmp);
9635 SDValue FNStSW = DAG.getNode(X86ISD::FNSTSW16r, dl, MVT::i16, TruncFPSW);
9636 SDValue Srl = DAG.getNode(ISD::SRL, dl, MVT::i16, FNStSW,
9637 DAG.getConstant(8, MVT::i8));
9638 SDValue TruncSrl = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Srl);
9639 return DAG.getNode(X86ISD::SAHF, dl, MVT::i32, TruncSrl);
9642 static bool isAllOnes(SDValue V) {
9643 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
9644 return C && C->isAllOnesValue();
9647 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
9648 /// if it's possible.
9649 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
9650 SDLoc dl, SelectionDAG &DAG) const {
9651 SDValue Op0 = And.getOperand(0);
9652 SDValue Op1 = And.getOperand(1);
9653 if (Op0.getOpcode() == ISD::TRUNCATE)
9654 Op0 = Op0.getOperand(0);
9655 if (Op1.getOpcode() == ISD::TRUNCATE)
9656 Op1 = Op1.getOperand(0);
9659 if (Op1.getOpcode() == ISD::SHL)
9660 std::swap(Op0, Op1);
9661 if (Op0.getOpcode() == ISD::SHL) {
9662 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
9663 if (And00C->getZExtValue() == 1) {
9664 // If we looked past a truncate, check that it's only truncating away
9666 unsigned BitWidth = Op0.getValueSizeInBits();
9667 unsigned AndBitWidth = And.getValueSizeInBits();
9668 if (BitWidth > AndBitWidth) {
9670 DAG.ComputeMaskedBits(Op0, Zeros, Ones);
9671 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
9675 RHS = Op0.getOperand(1);
9677 } else if (Op1.getOpcode() == ISD::Constant) {
9678 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
9679 uint64_t AndRHSVal = AndRHS->getZExtValue();
9680 SDValue AndLHS = Op0;
9682 if (AndRHSVal == 1 && AndLHS.getOpcode() == ISD::SRL) {
9683 LHS = AndLHS.getOperand(0);
9684 RHS = AndLHS.getOperand(1);
9687 // Use BT if the immediate can't be encoded in a TEST instruction.
9688 if (!isUInt<32>(AndRHSVal) && isPowerOf2_64(AndRHSVal)) {
9690 RHS = DAG.getConstant(Log2_64_Ceil(AndRHSVal), LHS.getValueType());
9694 if (LHS.getNode()) {
9695 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
9696 // instruction. Since the shift amount is in-range-or-undefined, we know
9697 // that doing a bittest on the i32 value is ok. We extend to i32 because
9698 // the encoding for the i16 version is larger than the i32 version.
9699 // Also promote i16 to i32 for performance / code size reason.
9700 if (LHS.getValueType() == MVT::i8 ||
9701 LHS.getValueType() == MVT::i16)
9702 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
9704 // If the operand types disagree, extend the shift amount to match. Since
9705 // BT ignores high bits (like shifts) we can use anyextend.
9706 if (LHS.getValueType() != RHS.getValueType())
9707 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
9709 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
9710 X86::CondCode Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
9711 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
9712 DAG.getConstant(Cond, MVT::i8), BT);
9718 /// \brief - Turns an ISD::CondCode into a value suitable for SSE floating point
9720 static int translateX86FSETCC(ISD::CondCode SetCCOpcode, SDValue &Op0,
9725 // SSE Condition code mapping:
9734 switch (SetCCOpcode) {
9735 default: llvm_unreachable("Unexpected SETCC condition");
9737 case ISD::SETEQ: SSECC = 0; break;
9739 case ISD::SETGT: Swap = true; // Fallthrough
9741 case ISD::SETOLT: SSECC = 1; break;
9743 case ISD::SETGE: Swap = true; // Fallthrough
9745 case ISD::SETOLE: SSECC = 2; break;
9746 case ISD::SETUO: SSECC = 3; break;
9748 case ISD::SETNE: SSECC = 4; break;
9749 case ISD::SETULE: Swap = true; // Fallthrough
9750 case ISD::SETUGE: SSECC = 5; break;
9751 case ISD::SETULT: Swap = true; // Fallthrough
9752 case ISD::SETUGT: SSECC = 6; break;
9753 case ISD::SETO: SSECC = 7; break;
9755 case ISD::SETONE: SSECC = 8; break;
9758 std::swap(Op0, Op1);
9763 // Lower256IntVSETCC - Break a VSETCC 256-bit integer VSETCC into two new 128
9764 // ones, and then concatenate the result back.
9765 static SDValue Lower256IntVSETCC(SDValue Op, SelectionDAG &DAG) {
9766 MVT VT = Op.getSimpleValueType();
9768 assert(VT.is256BitVector() && Op.getOpcode() == ISD::SETCC &&
9769 "Unsupported value type for operation");
9771 unsigned NumElems = VT.getVectorNumElements();
9773 SDValue CC = Op.getOperand(2);
9775 // Extract the LHS vectors
9776 SDValue LHS = Op.getOperand(0);
9777 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
9778 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
9780 // Extract the RHS vectors
9781 SDValue RHS = Op.getOperand(1);
9782 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
9783 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
9785 // Issue the operation on the smaller types and concatenate the result back
9786 MVT EltVT = VT.getVectorElementType();
9787 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
9788 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
9789 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC),
9790 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
9793 static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG) {
9795 SDValue Op0 = Op.getOperand(0);
9796 SDValue Op1 = Op.getOperand(1);
9797 SDValue CC = Op.getOperand(2);
9798 MVT VT = Op.getSimpleValueType();
9800 assert(Op0.getValueType().getVectorElementType().getSizeInBits() >= 32 &&
9801 Op.getValueType().getScalarType() == MVT::i1 &&
9802 "Cannot set masked compare for this operation");
9804 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
9807 bool Unsigned = false;
9809 switch (SetCCOpcode) {
9810 default: llvm_unreachable("Unexpected SETCC condition");
9811 case ISD::SETNE: SSECC = 4; break;
9812 case ISD::SETEQ: SSECC = 0; break;
9813 case ISD::SETUGT: Unsigned = true;
9814 case ISD::SETGT: SSECC = 6; break; // NLE
9815 case ISD::SETULT: Unsigned = true;
9816 case ISD::SETLT: SSECC = 1; break;
9817 case ISD::SETUGE: Unsigned = true;
9818 case ISD::SETGE: SSECC = 5; break; // NLT
9819 case ISD::SETULE: Unsigned = true;
9820 case ISD::SETLE: SSECC = 2; break;
9822 unsigned Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM;
9823 return DAG.getNode(Opc, dl, VT, Op0, Op1,
9824 DAG.getConstant(SSECC, MVT::i8));
9828 static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
9829 SelectionDAG &DAG) {
9831 SDValue Op0 = Op.getOperand(0);
9832 SDValue Op1 = Op.getOperand(1);
9833 SDValue CC = Op.getOperand(2);
9834 MVT VT = Op.getSimpleValueType();
9835 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
9836 bool isFP = Op.getOperand(1).getSimpleValueType().isFloatingPoint();
9841 MVT EltVT = Op0.getSimpleValueType().getVectorElementType();
9842 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
9845 unsigned SSECC = translateX86FSETCC(SetCCOpcode, Op0, Op1);
9846 unsigned Opc = X86ISD::CMPP;
9847 if (Subtarget->hasAVX512() && VT.getVectorElementType() == MVT::i1) {
9848 assert(VT.getVectorNumElements() <= 16);
9851 // In the two special cases we can't handle, emit two comparisons.
9854 unsigned CombineOpc;
9855 if (SetCCOpcode == ISD::SETUEQ) {
9856 CC0 = 3; CC1 = 0; CombineOpc = ISD::OR;
9858 assert(SetCCOpcode == ISD::SETONE);
9859 CC0 = 7; CC1 = 4; CombineOpc = ISD::AND;
9862 SDValue Cmp0 = DAG.getNode(Opc, dl, VT, Op0, Op1,
9863 DAG.getConstant(CC0, MVT::i8));
9864 SDValue Cmp1 = DAG.getNode(Opc, dl, VT, Op0, Op1,
9865 DAG.getConstant(CC1, MVT::i8));
9866 return DAG.getNode(CombineOpc, dl, VT, Cmp0, Cmp1);
9868 // Handle all other FP comparisons here.
9869 return DAG.getNode(Opc, dl, VT, Op0, Op1,
9870 DAG.getConstant(SSECC, MVT::i8));
9873 // Break 256-bit integer vector compare into smaller ones.
9874 if (VT.is256BitVector() && !Subtarget->hasInt256())
9875 return Lower256IntVSETCC(Op, DAG);
9877 bool MaskResult = (VT.getVectorElementType() == MVT::i1);
9878 EVT OpVT = Op1.getValueType();
9879 if (Subtarget->hasAVX512()) {
9880 if (Op1.getValueType().is512BitVector() ||
9881 (MaskResult && OpVT.getVectorElementType().getSizeInBits() >= 32))
9882 return LowerIntVSETCC_AVX512(Op, DAG);
9884 // In AVX-512 architecture setcc returns mask with i1 elements,
9885 // But there is no compare instruction for i8 and i16 elements.
9886 // We are not talking about 512-bit operands in this case, these
9887 // types are illegal.
9889 (OpVT.getVectorElementType().getSizeInBits() < 32 &&
9890 OpVT.getVectorElementType().getSizeInBits() >= 8))
9891 return DAG.getNode(ISD::TRUNCATE, dl, VT,
9892 DAG.getNode(ISD::SETCC, dl, OpVT, Op0, Op1, CC));
9895 // We are handling one of the integer comparisons here. Since SSE only has
9896 // GT and EQ comparisons for integer, swapping operands and multiple
9897 // operations may be required for some comparisons.
9899 bool Swap = false, Invert = false, FlipSigns = false, MinMax = false;
9901 switch (SetCCOpcode) {
9902 default: llvm_unreachable("Unexpected SETCC condition");
9903 case ISD::SETNE: Invert = true;
9904 case ISD::SETEQ: Opc = MaskResult? X86ISD::PCMPEQM: X86ISD::PCMPEQ; break;
9905 case ISD::SETLT: Swap = true;
9906 case ISD::SETGT: Opc = MaskResult? X86ISD::PCMPGTM: X86ISD::PCMPGT; break;
9907 case ISD::SETGE: Swap = true;
9908 case ISD::SETLE: Opc = MaskResult? X86ISD::PCMPGTM: X86ISD::PCMPGT;
9909 Invert = true; break;
9910 case ISD::SETULT: Swap = true;
9911 case ISD::SETUGT: Opc = MaskResult? X86ISD::PCMPGTM: X86ISD::PCMPGT;
9912 FlipSigns = true; break;
9913 case ISD::SETUGE: Swap = true;
9914 case ISD::SETULE: Opc = MaskResult? X86ISD::PCMPGTM: X86ISD::PCMPGT;
9915 FlipSigns = true; Invert = true; break;
9918 // Special case: Use min/max operations for SETULE/SETUGE
9919 MVT VET = VT.getVectorElementType();
9921 (Subtarget->hasSSE41() && (VET >= MVT::i8 && VET <= MVT::i32))
9922 || (Subtarget->hasSSE2() && (VET == MVT::i8));
9925 switch (SetCCOpcode) {
9927 case ISD::SETULE: Opc = X86ISD::UMIN; MinMax = true; break;
9928 case ISD::SETUGE: Opc = X86ISD::UMAX; MinMax = true; break;
9931 if (MinMax) { Swap = false; Invert = false; FlipSigns = false; }
9935 std::swap(Op0, Op1);
9937 // Check that the operation in question is available (most are plain SSE2,
9938 // but PCMPGTQ and PCMPEQQ have different requirements).
9939 if (VT == MVT::v2i64) {
9940 if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42()) {
9941 assert(Subtarget->hasSSE2() && "Don't know how to lower!");
9943 // First cast everything to the right type.
9944 Op0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op0);
9945 Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op1);
9947 // Since SSE has no unsigned integer comparisons, we need to flip the sign
9948 // bits of the inputs before performing those operations. The lower
9949 // compare is always unsigned.
9952 SB = DAG.getConstant(0x80000000U, MVT::v4i32);
9954 SDValue Sign = DAG.getConstant(0x80000000U, MVT::i32);
9955 SDValue Zero = DAG.getConstant(0x00000000U, MVT::i32);
9956 SB = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
9957 Sign, Zero, Sign, Zero);
9959 Op0 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op0, SB);
9960 Op1 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op1, SB);
9962 // Emulate PCMPGTQ with (hi1 > hi2) | ((hi1 == hi2) & (lo1 > lo2))
9963 SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1);
9964 SDValue EQ = DAG.getNode(X86ISD::PCMPEQ, dl, MVT::v4i32, Op0, Op1);
9966 // Create masks for only the low parts/high parts of the 64 bit integers.
9967 static const int MaskHi[] = { 1, 1, 3, 3 };
9968 static const int MaskLo[] = { 0, 0, 2, 2 };
9969 SDValue EQHi = DAG.getVectorShuffle(MVT::v4i32, dl, EQ, EQ, MaskHi);
9970 SDValue GTLo = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskLo);
9971 SDValue GTHi = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi);
9973 SDValue Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, EQHi, GTLo);
9974 Result = DAG.getNode(ISD::OR, dl, MVT::v4i32, Result, GTHi);
9977 Result = DAG.getNOT(dl, Result, MVT::v4i32);
9979 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
9982 if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) {
9983 // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with
9984 // pcmpeqd + pshufd + pand.
9985 assert(Subtarget->hasSSE2() && !FlipSigns && "Don't know how to lower!");
9987 // First cast everything to the right type.
9988 Op0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op0);
9989 Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op1);
9992 SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1);
9994 // Make sure the lower and upper halves are both all-ones.
9995 static const int Mask[] = { 1, 0, 3, 2 };
9996 SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask);
9997 Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf);
10000 Result = DAG.getNOT(dl, Result, MVT::v4i32);
10002 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
10006 // Since SSE has no unsigned integer comparisons, we need to flip the sign
10007 // bits of the inputs before performing those operations.
10009 EVT EltVT = VT.getVectorElementType();
10010 SDValue SB = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()), VT);
10011 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SB);
10012 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SB);
10015 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
10017 // If the logical-not of the result is required, perform that now.
10019 Result = DAG.getNOT(dl, Result, VT);
10022 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Op0, Result);
10027 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
10029 MVT VT = Op.getSimpleValueType();
10031 if (VT.isVector()) return LowerVSETCC(Op, Subtarget, DAG);
10033 assert(VT == MVT::i8 && "SetCC type must be 8-bit integer");
10034 SDValue Op0 = Op.getOperand(0);
10035 SDValue Op1 = Op.getOperand(1);
10037 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
10039 // Optimize to BT if possible.
10040 // Lower (X & (1 << N)) == 0 to BT(X, N).
10041 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
10042 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
10043 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
10044 Op1.getOpcode() == ISD::Constant &&
10045 cast<ConstantSDNode>(Op1)->isNullValue() &&
10046 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
10047 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
10048 if (NewSetCC.getNode())
10052 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
10054 if (Op1.getOpcode() == ISD::Constant &&
10055 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
10056 cast<ConstantSDNode>(Op1)->isNullValue()) &&
10057 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
10059 // If the input is a setcc, then reuse the input setcc or use a new one with
10060 // the inverted condition.
10061 if (Op0.getOpcode() == X86ISD::SETCC) {
10062 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
10063 bool Invert = (CC == ISD::SETNE) ^
10064 cast<ConstantSDNode>(Op1)->isNullValue();
10065 if (!Invert) return Op0;
10067 CCode = X86::GetOppositeBranchCondition(CCode);
10068 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
10069 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
10073 bool isFP = Op1.getSimpleValueType().isFloatingPoint();
10074 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
10075 if (X86CC == X86::COND_INVALID)
10078 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, DAG);
10079 EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
10080 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
10081 DAG.getConstant(X86CC, MVT::i8), EFLAGS);
10084 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
10085 static bool isX86LogicalCmp(SDValue Op) {
10086 unsigned Opc = Op.getNode()->getOpcode();
10087 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI ||
10088 Opc == X86ISD::SAHF)
10090 if (Op.getResNo() == 1 &&
10091 (Opc == X86ISD::ADD ||
10092 Opc == X86ISD::SUB ||
10093 Opc == X86ISD::ADC ||
10094 Opc == X86ISD::SBB ||
10095 Opc == X86ISD::SMUL ||
10096 Opc == X86ISD::UMUL ||
10097 Opc == X86ISD::INC ||
10098 Opc == X86ISD::DEC ||
10099 Opc == X86ISD::OR ||
10100 Opc == X86ISD::XOR ||
10101 Opc == X86ISD::AND))
10104 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
10110 static bool isZero(SDValue V) {
10111 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
10112 return C && C->isNullValue();
10115 static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) {
10116 if (V.getOpcode() != ISD::TRUNCATE)
10119 SDValue VOp0 = V.getOperand(0);
10120 unsigned InBits = VOp0.getValueSizeInBits();
10121 unsigned Bits = V.getValueSizeInBits();
10122 return DAG.MaskedValueIsZero(VOp0, APInt::getHighBitsSet(InBits,InBits-Bits));
10125 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
10126 bool addTest = true;
10127 SDValue Cond = Op.getOperand(0);
10128 SDValue Op1 = Op.getOperand(1);
10129 SDValue Op2 = Op.getOperand(2);
10131 EVT VT = Op1.getValueType();
10134 // Lower fp selects into a CMP/AND/ANDN/OR sequence when the necessary SSE ops
10135 // are available. Otherwise fp cmovs get lowered into a less efficient branch
10136 // sequence later on.
10137 if (Cond.getOpcode() == ISD::SETCC &&
10138 ((Subtarget->hasSSE2() && (VT == MVT::f32 || VT == MVT::f64)) ||
10139 (Subtarget->hasSSE1() && VT == MVT::f32)) &&
10140 VT == Cond.getOperand(0).getValueType() && Cond->hasOneUse()) {
10141 SDValue CondOp0 = Cond.getOperand(0), CondOp1 = Cond.getOperand(1);
10142 int SSECC = translateX86FSETCC(
10143 cast<CondCodeSDNode>(Cond.getOperand(2))->get(), CondOp0, CondOp1);
10146 unsigned Opcode = VT == MVT::f32 ? X86ISD::FSETCCss : X86ISD::FSETCCsd;
10147 SDValue Cmp = DAG.getNode(Opcode, DL, VT, CondOp0, CondOp1,
10148 DAG.getConstant(SSECC, MVT::i8));
10149 SDValue AndN = DAG.getNode(X86ISD::FANDN, DL, VT, Cmp, Op2);
10150 SDValue And = DAG.getNode(X86ISD::FAND, DL, VT, Cmp, Op1);
10151 return DAG.getNode(X86ISD::FOR, DL, VT, AndN, And);
10155 if (Cond.getOpcode() == ISD::SETCC) {
10156 SDValue NewCond = LowerSETCC(Cond, DAG);
10157 if (NewCond.getNode())
10161 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
10162 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
10163 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
10164 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
10165 if (Cond.getOpcode() == X86ISD::SETCC &&
10166 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
10167 isZero(Cond.getOperand(1).getOperand(1))) {
10168 SDValue Cmp = Cond.getOperand(1);
10170 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
10172 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
10173 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
10174 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
10176 SDValue CmpOp0 = Cmp.getOperand(0);
10177 // Apply further optimizations for special cases
10178 // (select (x != 0), -1, 0) -> neg & sbb
10179 // (select (x == 0), 0, -1) -> neg & sbb
10180 if (ConstantSDNode *YC = dyn_cast<ConstantSDNode>(Y))
10181 if (YC->isNullValue() &&
10182 (isAllOnes(Op1) == (CondCode == X86::COND_NE))) {
10183 SDVTList VTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32);
10184 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, VTs,
10185 DAG.getConstant(0, CmpOp0.getValueType()),
10187 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
10188 DAG.getConstant(X86::COND_B, MVT::i8),
10189 SDValue(Neg.getNode(), 1));
10193 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
10194 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
10195 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
10197 SDValue Res = // Res = 0 or -1.
10198 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
10199 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
10201 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
10202 Res = DAG.getNOT(DL, Res, Res.getValueType());
10204 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
10205 if (N2C == 0 || !N2C->isNullValue())
10206 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
10211 // Look past (and (setcc_carry (cmp ...)), 1).
10212 if (Cond.getOpcode() == ISD::AND &&
10213 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
10214 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
10215 if (C && C->getAPIntValue() == 1)
10216 Cond = Cond.getOperand(0);
10219 // If condition flag is set by a X86ISD::CMP, then use it as the condition
10220 // setting operand in place of the X86ISD::SETCC.
10221 unsigned CondOpcode = Cond.getOpcode();
10222 if (CondOpcode == X86ISD::SETCC ||
10223 CondOpcode == X86ISD::SETCC_CARRY) {
10224 CC = Cond.getOperand(0);
10226 SDValue Cmp = Cond.getOperand(1);
10227 unsigned Opc = Cmp.getOpcode();
10228 MVT VT = Op.getSimpleValueType();
10230 bool IllegalFPCMov = false;
10231 if (VT.isFloatingPoint() && !VT.isVector() &&
10232 !isScalarFPTypeInSSEReg(VT)) // FPStack?
10233 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
10235 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
10236 Opc == X86ISD::BT) { // FIXME
10240 } else if (CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
10241 CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
10242 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
10243 Cond.getOperand(0).getValueType() != MVT::i8)) {
10244 SDValue LHS = Cond.getOperand(0);
10245 SDValue RHS = Cond.getOperand(1);
10246 unsigned X86Opcode;
10249 switch (CondOpcode) {
10250 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
10251 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
10252 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
10253 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
10254 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
10255 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
10256 default: llvm_unreachable("unexpected overflowing operator");
10258 if (CondOpcode == ISD::UMULO)
10259 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
10262 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
10264 SDValue X86Op = DAG.getNode(X86Opcode, DL, VTs, LHS, RHS);
10266 if (CondOpcode == ISD::UMULO)
10267 Cond = X86Op.getValue(2);
10269 Cond = X86Op.getValue(1);
10271 CC = DAG.getConstant(X86Cond, MVT::i8);
10276 // Look pass the truncate if the high bits are known zero.
10277 if (isTruncWithZeroHighBitsInput(Cond, DAG))
10278 Cond = Cond.getOperand(0);
10280 // We know the result of AND is compared against zero. Try to match
10282 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
10283 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
10284 if (NewSetCC.getNode()) {
10285 CC = NewSetCC.getOperand(0);
10286 Cond = NewSetCC.getOperand(1);
10293 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
10294 Cond = EmitTest(Cond, X86::COND_NE, DAG);
10297 // a < b ? -1 : 0 -> RES = ~setcc_carry
10298 // a < b ? 0 : -1 -> RES = setcc_carry
10299 // a >= b ? -1 : 0 -> RES = setcc_carry
10300 // a >= b ? 0 : -1 -> RES = ~setcc_carry
10301 if (Cond.getOpcode() == X86ISD::SUB) {
10302 Cond = ConvertCmpIfNecessary(Cond, DAG);
10303 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
10305 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
10306 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
10307 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
10308 DAG.getConstant(X86::COND_B, MVT::i8), Cond);
10309 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
10310 return DAG.getNOT(DL, Res, Res.getValueType());
10315 // X86 doesn't have an i8 cmov. If both operands are the result of a truncate
10316 // widen the cmov and push the truncate through. This avoids introducing a new
10317 // branch during isel and doesn't add any extensions.
10318 if (Op.getValueType() == MVT::i8 &&
10319 Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) {
10320 SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0);
10321 if (T1.getValueType() == T2.getValueType() &&
10322 // Blacklist CopyFromReg to avoid partial register stalls.
10323 T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){
10324 SDVTList VTs = DAG.getVTList(T1.getValueType(), MVT::Glue);
10325 SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, VTs, T2, T1, CC, Cond);
10326 return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov);
10330 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
10331 // condition is true.
10332 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
10333 SDValue Ops[] = { Op2, Op1, CC, Cond };
10334 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops, array_lengthof(Ops));
10337 static SDValue LowerSIGN_EXTEND_AVX512(SDValue Op, SelectionDAG &DAG) {
10338 MVT VT = Op->getSimpleValueType(0);
10339 SDValue In = Op->getOperand(0);
10340 MVT InVT = In.getSimpleValueType();
10343 unsigned int NumElts = VT.getVectorNumElements();
10344 if (NumElts != 8 && NumElts != 16)
10347 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1)
10348 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
10350 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
10351 assert (InVT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
10353 MVT ExtVT = (NumElts == 8) ? MVT::v8i64 : MVT::v16i32;
10354 Constant *C = ConstantInt::get(*DAG.getContext(),
10355 APInt::getAllOnesValue(ExtVT.getScalarType().getSizeInBits()));
10357 SDValue CP = DAG.getConstantPool(C, TLI.getPointerTy());
10358 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
10359 SDValue Ld = DAG.getLoad(ExtVT.getScalarType(), dl, DAG.getEntryNode(), CP,
10360 MachinePointerInfo::getConstantPool(),
10361 false, false, false, Alignment);
10362 SDValue Brcst = DAG.getNode(X86ISD::VBROADCASTM, dl, ExtVT, In, Ld);
10363 if (VT.is512BitVector())
10365 return DAG.getNode(X86ISD::VTRUNC, dl, VT, Brcst);
10368 static SDValue LowerSIGN_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
10369 SelectionDAG &DAG) {
10370 MVT VT = Op->getSimpleValueType(0);
10371 SDValue In = Op->getOperand(0);
10372 MVT InVT = In.getSimpleValueType();
10375 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
10376 return LowerSIGN_EXTEND_AVX512(Op, DAG);
10378 if ((VT != MVT::v4i64 || InVT != MVT::v4i32) &&
10379 (VT != MVT::v8i32 || InVT != MVT::v8i16))
10382 if (Subtarget->hasInt256())
10383 return DAG.getNode(X86ISD::VSEXT_MOVL, dl, VT, In);
10385 // Optimize vectors in AVX mode
10386 // Sign extend v8i16 to v8i32 and
10389 // Divide input vector into two parts
10390 // for v4i32 the shuffle mask will be { 0, 1, -1, -1} {2, 3, -1, -1}
10391 // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32
10392 // concat the vectors to original VT
10394 unsigned NumElems = InVT.getVectorNumElements();
10395 SDValue Undef = DAG.getUNDEF(InVT);
10397 SmallVector<int,8> ShufMask1(NumElems, -1);
10398 for (unsigned i = 0; i != NumElems/2; ++i)
10401 SDValue OpLo = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask1[0]);
10403 SmallVector<int,8> ShufMask2(NumElems, -1);
10404 for (unsigned i = 0; i != NumElems/2; ++i)
10405 ShufMask2[i] = i + NumElems/2;
10407 SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask2[0]);
10409 MVT HalfVT = MVT::getVectorVT(VT.getScalarType(),
10410 VT.getVectorNumElements()/2);
10412 OpLo = DAG.getNode(X86ISD::VSEXT_MOVL, dl, HalfVT, OpLo);
10413 OpHi = DAG.getNode(X86ISD::VSEXT_MOVL, dl, HalfVT, OpHi);
10415 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
10418 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
10419 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
10420 // from the AND / OR.
10421 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
10422 Opc = Op.getOpcode();
10423 if (Opc != ISD::OR && Opc != ISD::AND)
10425 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
10426 Op.getOperand(0).hasOneUse() &&
10427 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
10428 Op.getOperand(1).hasOneUse());
10431 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
10432 // 1 and that the SETCC node has a single use.
10433 static bool isXor1OfSetCC(SDValue Op) {
10434 if (Op.getOpcode() != ISD::XOR)
10436 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
10437 if (N1C && N1C->getAPIntValue() == 1) {
10438 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
10439 Op.getOperand(0).hasOneUse();
10444 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
10445 bool addTest = true;
10446 SDValue Chain = Op.getOperand(0);
10447 SDValue Cond = Op.getOperand(1);
10448 SDValue Dest = Op.getOperand(2);
10451 bool Inverted = false;
10453 if (Cond.getOpcode() == ISD::SETCC) {
10454 // Check for setcc([su]{add,sub,mul}o == 0).
10455 if (cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETEQ &&
10456 isa<ConstantSDNode>(Cond.getOperand(1)) &&
10457 cast<ConstantSDNode>(Cond.getOperand(1))->isNullValue() &&
10458 Cond.getOperand(0).getResNo() == 1 &&
10459 (Cond.getOperand(0).getOpcode() == ISD::SADDO ||
10460 Cond.getOperand(0).getOpcode() == ISD::UADDO ||
10461 Cond.getOperand(0).getOpcode() == ISD::SSUBO ||
10462 Cond.getOperand(0).getOpcode() == ISD::USUBO ||
10463 Cond.getOperand(0).getOpcode() == ISD::SMULO ||
10464 Cond.getOperand(0).getOpcode() == ISD::UMULO)) {
10466 Cond = Cond.getOperand(0);
10468 SDValue NewCond = LowerSETCC(Cond, DAG);
10469 if (NewCond.getNode())
10474 // FIXME: LowerXALUO doesn't handle these!!
10475 else if (Cond.getOpcode() == X86ISD::ADD ||
10476 Cond.getOpcode() == X86ISD::SUB ||
10477 Cond.getOpcode() == X86ISD::SMUL ||
10478 Cond.getOpcode() == X86ISD::UMUL)
10479 Cond = LowerXALUO(Cond, DAG);
10482 // Look pass (and (setcc_carry (cmp ...)), 1).
10483 if (Cond.getOpcode() == ISD::AND &&
10484 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
10485 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
10486 if (C && C->getAPIntValue() == 1)
10487 Cond = Cond.getOperand(0);
10490 // If condition flag is set by a X86ISD::CMP, then use it as the condition
10491 // setting operand in place of the X86ISD::SETCC.
10492 unsigned CondOpcode = Cond.getOpcode();
10493 if (CondOpcode == X86ISD::SETCC ||
10494 CondOpcode == X86ISD::SETCC_CARRY) {
10495 CC = Cond.getOperand(0);
10497 SDValue Cmp = Cond.getOperand(1);
10498 unsigned Opc = Cmp.getOpcode();
10499 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
10500 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
10504 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
10508 // These can only come from an arithmetic instruction with overflow,
10509 // e.g. SADDO, UADDO.
10510 Cond = Cond.getNode()->getOperand(1);
10516 CondOpcode = Cond.getOpcode();
10517 if (CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
10518 CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
10519 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
10520 Cond.getOperand(0).getValueType() != MVT::i8)) {
10521 SDValue LHS = Cond.getOperand(0);
10522 SDValue RHS = Cond.getOperand(1);
10523 unsigned X86Opcode;
10526 switch (CondOpcode) {
10527 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
10528 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
10529 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
10530 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
10531 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
10532 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
10533 default: llvm_unreachable("unexpected overflowing operator");
10536 X86Cond = X86::GetOppositeBranchCondition((X86::CondCode)X86Cond);
10537 if (CondOpcode == ISD::UMULO)
10538 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
10541 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
10543 SDValue X86Op = DAG.getNode(X86Opcode, dl, VTs, LHS, RHS);
10545 if (CondOpcode == ISD::UMULO)
10546 Cond = X86Op.getValue(2);
10548 Cond = X86Op.getValue(1);
10550 CC = DAG.getConstant(X86Cond, MVT::i8);
10554 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
10555 SDValue Cmp = Cond.getOperand(0).getOperand(1);
10556 if (CondOpc == ISD::OR) {
10557 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
10558 // two branches instead of an explicit OR instruction with a
10560 if (Cmp == Cond.getOperand(1).getOperand(1) &&
10561 isX86LogicalCmp(Cmp)) {
10562 CC = Cond.getOperand(0).getOperand(0);
10563 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
10564 Chain, Dest, CC, Cmp);
10565 CC = Cond.getOperand(1).getOperand(0);
10569 } else { // ISD::AND
10570 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
10571 // two branches instead of an explicit AND instruction with a
10572 // separate test. However, we only do this if this block doesn't
10573 // have a fall-through edge, because this requires an explicit
10574 // jmp when the condition is false.
10575 if (Cmp == Cond.getOperand(1).getOperand(1) &&
10576 isX86LogicalCmp(Cmp) &&
10577 Op.getNode()->hasOneUse()) {
10578 X86::CondCode CCode =
10579 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
10580 CCode = X86::GetOppositeBranchCondition(CCode);
10581 CC = DAG.getConstant(CCode, MVT::i8);
10582 SDNode *User = *Op.getNode()->use_begin();
10583 // Look for an unconditional branch following this conditional branch.
10584 // We need this because we need to reverse the successors in order
10585 // to implement FCMP_OEQ.
10586 if (User->getOpcode() == ISD::BR) {
10587 SDValue FalseBB = User->getOperand(1);
10589 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
10590 assert(NewBR == User);
10594 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
10595 Chain, Dest, CC, Cmp);
10596 X86::CondCode CCode =
10597 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
10598 CCode = X86::GetOppositeBranchCondition(CCode);
10599 CC = DAG.getConstant(CCode, MVT::i8);
10605 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
10606 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
10607 // It should be transformed during dag combiner except when the condition
10608 // is set by a arithmetics with overflow node.
10609 X86::CondCode CCode =
10610 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
10611 CCode = X86::GetOppositeBranchCondition(CCode);
10612 CC = DAG.getConstant(CCode, MVT::i8);
10613 Cond = Cond.getOperand(0).getOperand(1);
10615 } else if (Cond.getOpcode() == ISD::SETCC &&
10616 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETOEQ) {
10617 // For FCMP_OEQ, we can emit
10618 // two branches instead of an explicit AND instruction with a
10619 // separate test. However, we only do this if this block doesn't
10620 // have a fall-through edge, because this requires an explicit
10621 // jmp when the condition is false.
10622 if (Op.getNode()->hasOneUse()) {
10623 SDNode *User = *Op.getNode()->use_begin();
10624 // Look for an unconditional branch following this conditional branch.
10625 // We need this because we need to reverse the successors in order
10626 // to implement FCMP_OEQ.
10627 if (User->getOpcode() == ISD::BR) {
10628 SDValue FalseBB = User->getOperand(1);
10630 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
10631 assert(NewBR == User);
10635 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
10636 Cond.getOperand(0), Cond.getOperand(1));
10637 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
10638 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
10639 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
10640 Chain, Dest, CC, Cmp);
10641 CC = DAG.getConstant(X86::COND_P, MVT::i8);
10646 } else if (Cond.getOpcode() == ISD::SETCC &&
10647 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETUNE) {
10648 // For FCMP_UNE, we can emit
10649 // two branches instead of an explicit AND instruction with a
10650 // separate test. However, we only do this if this block doesn't
10651 // have a fall-through edge, because this requires an explicit
10652 // jmp when the condition is false.
10653 if (Op.getNode()->hasOneUse()) {
10654 SDNode *User = *Op.getNode()->use_begin();
10655 // Look for an unconditional branch following this conditional branch.
10656 // We need this because we need to reverse the successors in order
10657 // to implement FCMP_UNE.
10658 if (User->getOpcode() == ISD::BR) {
10659 SDValue FalseBB = User->getOperand(1);
10661 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
10662 assert(NewBR == User);
10665 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
10666 Cond.getOperand(0), Cond.getOperand(1));
10667 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
10668 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
10669 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
10670 Chain, Dest, CC, Cmp);
10671 CC = DAG.getConstant(X86::COND_NP, MVT::i8);
10681 // Look pass the truncate if the high bits are known zero.
10682 if (isTruncWithZeroHighBitsInput(Cond, DAG))
10683 Cond = Cond.getOperand(0);
10685 // We know the result of AND is compared against zero. Try to match
10687 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
10688 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
10689 if (NewSetCC.getNode()) {
10690 CC = NewSetCC.getOperand(0);
10691 Cond = NewSetCC.getOperand(1);
10698 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
10699 Cond = EmitTest(Cond, X86::COND_NE, DAG);
10701 Cond = ConvertCmpIfNecessary(Cond, DAG);
10702 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
10703 Chain, Dest, CC, Cond);
10706 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
10707 // Calls to _alloca is needed to probe the stack when allocating more than 4k
10708 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
10709 // that the guard pages used by the OS virtual memory manager are allocated in
10710 // correct sequence.
10712 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
10713 SelectionDAG &DAG) const {
10714 assert((Subtarget->isTargetCygMing() || Subtarget->isTargetWindows() ||
10715 getTargetMachine().Options.EnableSegmentedStacks) &&
10716 "This should be used only on Windows targets or when segmented stacks "
10718 assert(!Subtarget->isTargetEnvMacho() && "Not implemented");
10722 SDValue Chain = Op.getOperand(0);
10723 SDValue Size = Op.getOperand(1);
10724 // FIXME: Ensure alignment here
10726 bool Is64Bit = Subtarget->is64Bit();
10727 EVT SPTy = Is64Bit ? MVT::i64 : MVT::i32;
10729 if (getTargetMachine().Options.EnableSegmentedStacks) {
10730 MachineFunction &MF = DAG.getMachineFunction();
10731 MachineRegisterInfo &MRI = MF.getRegInfo();
10734 // The 64 bit implementation of segmented stacks needs to clobber both r10
10735 // r11. This makes it impossible to use it along with nested parameters.
10736 const Function *F = MF.getFunction();
10738 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
10740 if (I->hasNestAttr())
10741 report_fatal_error("Cannot use segmented stacks with functions that "
10742 "have nested arguments.");
10745 const TargetRegisterClass *AddrRegClass =
10746 getRegClassFor(Subtarget->is64Bit() ? MVT::i64:MVT::i32);
10747 unsigned Vreg = MRI.createVirtualRegister(AddrRegClass);
10748 Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size);
10749 SDValue Value = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain,
10750 DAG.getRegister(Vreg, SPTy));
10751 SDValue Ops1[2] = { Value, Chain };
10752 return DAG.getMergeValues(Ops1, 2, dl);
10755 unsigned Reg = (Subtarget->is64Bit() ? X86::RAX : X86::EAX);
10757 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
10758 Flag = Chain.getValue(1);
10759 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
10761 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
10762 Flag = Chain.getValue(1);
10764 const X86RegisterInfo *RegInfo =
10765 static_cast<const X86RegisterInfo*>(getTargetMachine().getRegisterInfo());
10766 Chain = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
10769 SDValue Ops1[2] = { Chain.getValue(0), Chain };
10770 return DAG.getMergeValues(Ops1, 2, dl);
10774 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
10775 MachineFunction &MF = DAG.getMachineFunction();
10776 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
10778 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
10781 if (!Subtarget->is64Bit() || Subtarget->isTargetWin64()) {
10782 // vastart just stores the address of the VarArgsFrameIndex slot into the
10783 // memory location argument.
10784 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
10786 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
10787 MachinePointerInfo(SV), false, false, 0);
10791 // gp_offset (0 - 6 * 8)
10792 // fp_offset (48 - 48 + 8 * 16)
10793 // overflow_arg_area (point to parameters coming in memory).
10795 SmallVector<SDValue, 8> MemOps;
10796 SDValue FIN = Op.getOperand(1);
10798 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
10799 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
10801 FIN, MachinePointerInfo(SV), false, false, 0);
10802 MemOps.push_back(Store);
10805 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
10806 FIN, DAG.getIntPtrConstant(4));
10807 Store = DAG.getStore(Op.getOperand(0), DL,
10808 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
10810 FIN, MachinePointerInfo(SV, 4), false, false, 0);
10811 MemOps.push_back(Store);
10813 // Store ptr to overflow_arg_area
10814 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
10815 FIN, DAG.getIntPtrConstant(4));
10816 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
10818 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
10819 MachinePointerInfo(SV, 8),
10821 MemOps.push_back(Store);
10823 // Store ptr to reg_save_area.
10824 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
10825 FIN, DAG.getIntPtrConstant(8));
10826 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
10828 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN,
10829 MachinePointerInfo(SV, 16), false, false, 0);
10830 MemOps.push_back(Store);
10831 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
10832 &MemOps[0], MemOps.size());
10835 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
10836 assert(Subtarget->is64Bit() &&
10837 "LowerVAARG only handles 64-bit va_arg!");
10838 assert((Subtarget->isTargetLinux() ||
10839 Subtarget->isTargetDarwin()) &&
10840 "Unhandled target in LowerVAARG");
10841 assert(Op.getNode()->getNumOperands() == 4);
10842 SDValue Chain = Op.getOperand(0);
10843 SDValue SrcPtr = Op.getOperand(1);
10844 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
10845 unsigned Align = Op.getConstantOperandVal(3);
10848 EVT ArgVT = Op.getNode()->getValueType(0);
10849 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
10850 uint32_t ArgSize = getDataLayout()->getTypeAllocSize(ArgTy);
10853 // Decide which area this value should be read from.
10854 // TODO: Implement the AMD64 ABI in its entirety. This simple
10855 // selection mechanism works only for the basic types.
10856 if (ArgVT == MVT::f80) {
10857 llvm_unreachable("va_arg for f80 not yet implemented");
10858 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
10859 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
10860 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
10861 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
10863 llvm_unreachable("Unhandled argument type in LowerVAARG");
10866 if (ArgMode == 2) {
10867 // Sanity Check: Make sure using fp_offset makes sense.
10868 assert(!getTargetMachine().Options.UseSoftFloat &&
10869 !(DAG.getMachineFunction()
10870 .getFunction()->getAttributes()
10871 .hasAttribute(AttributeSet::FunctionIndex,
10872 Attribute::NoImplicitFloat)) &&
10873 Subtarget->hasSSE1());
10876 // Insert VAARG_64 node into the DAG
10877 // VAARG_64 returns two values: Variable Argument Address, Chain
10878 SmallVector<SDValue, 11> InstOps;
10879 InstOps.push_back(Chain);
10880 InstOps.push_back(SrcPtr);
10881 InstOps.push_back(DAG.getConstant(ArgSize, MVT::i32));
10882 InstOps.push_back(DAG.getConstant(ArgMode, MVT::i8));
10883 InstOps.push_back(DAG.getConstant(Align, MVT::i32));
10884 SDVTList VTs = DAG.getVTList(getPointerTy(), MVT::Other);
10885 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
10886 VTs, &InstOps[0], InstOps.size(),
10888 MachinePointerInfo(SV),
10890 /*Volatile=*/false,
10892 /*WriteMem=*/true);
10893 Chain = VAARG.getValue(1);
10895 // Load the next argument and return it
10896 return DAG.getLoad(ArgVT, dl,
10899 MachinePointerInfo(),
10900 false, false, false, 0);
10903 static SDValue LowerVACOPY(SDValue Op, const X86Subtarget *Subtarget,
10904 SelectionDAG &DAG) {
10905 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
10906 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
10907 SDValue Chain = Op.getOperand(0);
10908 SDValue DstPtr = Op.getOperand(1);
10909 SDValue SrcPtr = Op.getOperand(2);
10910 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
10911 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
10914 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
10915 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
10917 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
10920 // getTargetVShiftNode - Handle vector element shifts where the shift amount
10921 // may or may not be a constant. Takes immediate version of shift as input.
10922 static SDValue getTargetVShiftNode(unsigned Opc, SDLoc dl, EVT VT,
10923 SDValue SrcOp, SDValue ShAmt,
10924 SelectionDAG &DAG) {
10925 assert(ShAmt.getValueType() == MVT::i32 && "ShAmt is not i32");
10927 if (isa<ConstantSDNode>(ShAmt)) {
10928 // Constant may be a TargetConstant. Use a regular constant.
10929 uint32_t ShiftAmt = cast<ConstantSDNode>(ShAmt)->getZExtValue();
10931 default: llvm_unreachable("Unknown target vector shift node");
10932 case X86ISD::VSHLI:
10933 case X86ISD::VSRLI:
10934 case X86ISD::VSRAI:
10935 return DAG.getNode(Opc, dl, VT, SrcOp,
10936 DAG.getConstant(ShiftAmt, MVT::i32));
10940 // Change opcode to non-immediate version
10942 default: llvm_unreachable("Unknown target vector shift node");
10943 case X86ISD::VSHLI: Opc = X86ISD::VSHL; break;
10944 case X86ISD::VSRLI: Opc = X86ISD::VSRL; break;
10945 case X86ISD::VSRAI: Opc = X86ISD::VSRA; break;
10948 // Need to build a vector containing shift amount
10949 // Shift amount is 32-bits, but SSE instructions read 64-bit, so fill with 0
10952 ShOps[1] = DAG.getConstant(0, MVT::i32);
10953 ShOps[2] = ShOps[3] = DAG.getUNDEF(MVT::i32);
10954 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, &ShOps[0], 4);
10956 // The return type has to be a 128-bit type with the same element
10957 // type as the input type.
10958 MVT EltVT = VT.getVectorElementType().getSimpleVT();
10959 EVT ShVT = MVT::getVectorVT(EltVT, 128/EltVT.getSizeInBits());
10961 ShAmt = DAG.getNode(ISD::BITCAST, dl, ShVT, ShAmt);
10962 return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt);
10965 static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
10967 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
10969 default: return SDValue(); // Don't custom lower most intrinsics.
10970 // Comparison intrinsics.
10971 case Intrinsic::x86_sse_comieq_ss:
10972 case Intrinsic::x86_sse_comilt_ss:
10973 case Intrinsic::x86_sse_comile_ss:
10974 case Intrinsic::x86_sse_comigt_ss:
10975 case Intrinsic::x86_sse_comige_ss:
10976 case Intrinsic::x86_sse_comineq_ss:
10977 case Intrinsic::x86_sse_ucomieq_ss:
10978 case Intrinsic::x86_sse_ucomilt_ss:
10979 case Intrinsic::x86_sse_ucomile_ss:
10980 case Intrinsic::x86_sse_ucomigt_ss:
10981 case Intrinsic::x86_sse_ucomige_ss:
10982 case Intrinsic::x86_sse_ucomineq_ss:
10983 case Intrinsic::x86_sse2_comieq_sd:
10984 case Intrinsic::x86_sse2_comilt_sd:
10985 case Intrinsic::x86_sse2_comile_sd:
10986 case Intrinsic::x86_sse2_comigt_sd:
10987 case Intrinsic::x86_sse2_comige_sd:
10988 case Intrinsic::x86_sse2_comineq_sd:
10989 case Intrinsic::x86_sse2_ucomieq_sd:
10990 case Intrinsic::x86_sse2_ucomilt_sd:
10991 case Intrinsic::x86_sse2_ucomile_sd:
10992 case Intrinsic::x86_sse2_ucomigt_sd:
10993 case Intrinsic::x86_sse2_ucomige_sd:
10994 case Intrinsic::x86_sse2_ucomineq_sd: {
10998 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
10999 case Intrinsic::x86_sse_comieq_ss:
11000 case Intrinsic::x86_sse2_comieq_sd:
11001 Opc = X86ISD::COMI;
11004 case Intrinsic::x86_sse_comilt_ss:
11005 case Intrinsic::x86_sse2_comilt_sd:
11006 Opc = X86ISD::COMI;
11009 case Intrinsic::x86_sse_comile_ss:
11010 case Intrinsic::x86_sse2_comile_sd:
11011 Opc = X86ISD::COMI;
11014 case Intrinsic::x86_sse_comigt_ss:
11015 case Intrinsic::x86_sse2_comigt_sd:
11016 Opc = X86ISD::COMI;
11019 case Intrinsic::x86_sse_comige_ss:
11020 case Intrinsic::x86_sse2_comige_sd:
11021 Opc = X86ISD::COMI;
11024 case Intrinsic::x86_sse_comineq_ss:
11025 case Intrinsic::x86_sse2_comineq_sd:
11026 Opc = X86ISD::COMI;
11029 case Intrinsic::x86_sse_ucomieq_ss:
11030 case Intrinsic::x86_sse2_ucomieq_sd:
11031 Opc = X86ISD::UCOMI;
11034 case Intrinsic::x86_sse_ucomilt_ss:
11035 case Intrinsic::x86_sse2_ucomilt_sd:
11036 Opc = X86ISD::UCOMI;
11039 case Intrinsic::x86_sse_ucomile_ss:
11040 case Intrinsic::x86_sse2_ucomile_sd:
11041 Opc = X86ISD::UCOMI;
11044 case Intrinsic::x86_sse_ucomigt_ss:
11045 case Intrinsic::x86_sse2_ucomigt_sd:
11046 Opc = X86ISD::UCOMI;
11049 case Intrinsic::x86_sse_ucomige_ss:
11050 case Intrinsic::x86_sse2_ucomige_sd:
11051 Opc = X86ISD::UCOMI;
11054 case Intrinsic::x86_sse_ucomineq_ss:
11055 case Intrinsic::x86_sse2_ucomineq_sd:
11056 Opc = X86ISD::UCOMI;
11061 SDValue LHS = Op.getOperand(1);
11062 SDValue RHS = Op.getOperand(2);
11063 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
11064 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
11065 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
11066 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
11067 DAG.getConstant(X86CC, MVT::i8), Cond);
11068 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
11071 // Arithmetic intrinsics.
11072 case Intrinsic::x86_sse2_pmulu_dq:
11073 case Intrinsic::x86_avx2_pmulu_dq:
11074 return DAG.getNode(X86ISD::PMULUDQ, dl, Op.getValueType(),
11075 Op.getOperand(1), Op.getOperand(2));
11077 // SSE2/AVX2 sub with unsigned saturation intrinsics
11078 case Intrinsic::x86_sse2_psubus_b:
11079 case Intrinsic::x86_sse2_psubus_w:
11080 case Intrinsic::x86_avx2_psubus_b:
11081 case Intrinsic::x86_avx2_psubus_w:
11082 return DAG.getNode(X86ISD::SUBUS, dl, Op.getValueType(),
11083 Op.getOperand(1), Op.getOperand(2));
11085 // SSE3/AVX horizontal add/sub intrinsics
11086 case Intrinsic::x86_sse3_hadd_ps:
11087 case Intrinsic::x86_sse3_hadd_pd:
11088 case Intrinsic::x86_avx_hadd_ps_256:
11089 case Intrinsic::x86_avx_hadd_pd_256:
11090 case Intrinsic::x86_sse3_hsub_ps:
11091 case Intrinsic::x86_sse3_hsub_pd:
11092 case Intrinsic::x86_avx_hsub_ps_256:
11093 case Intrinsic::x86_avx_hsub_pd_256:
11094 case Intrinsic::x86_ssse3_phadd_w_128:
11095 case Intrinsic::x86_ssse3_phadd_d_128:
11096 case Intrinsic::x86_avx2_phadd_w:
11097 case Intrinsic::x86_avx2_phadd_d:
11098 case Intrinsic::x86_ssse3_phsub_w_128:
11099 case Intrinsic::x86_ssse3_phsub_d_128:
11100 case Intrinsic::x86_avx2_phsub_w:
11101 case Intrinsic::x86_avx2_phsub_d: {
11104 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
11105 case Intrinsic::x86_sse3_hadd_ps:
11106 case Intrinsic::x86_sse3_hadd_pd:
11107 case Intrinsic::x86_avx_hadd_ps_256:
11108 case Intrinsic::x86_avx_hadd_pd_256:
11109 Opcode = X86ISD::FHADD;
11111 case Intrinsic::x86_sse3_hsub_ps:
11112 case Intrinsic::x86_sse3_hsub_pd:
11113 case Intrinsic::x86_avx_hsub_ps_256:
11114 case Intrinsic::x86_avx_hsub_pd_256:
11115 Opcode = X86ISD::FHSUB;
11117 case Intrinsic::x86_ssse3_phadd_w_128:
11118 case Intrinsic::x86_ssse3_phadd_d_128:
11119 case Intrinsic::x86_avx2_phadd_w:
11120 case Intrinsic::x86_avx2_phadd_d:
11121 Opcode = X86ISD::HADD;
11123 case Intrinsic::x86_ssse3_phsub_w_128:
11124 case Intrinsic::x86_ssse3_phsub_d_128:
11125 case Intrinsic::x86_avx2_phsub_w:
11126 case Intrinsic::x86_avx2_phsub_d:
11127 Opcode = X86ISD::HSUB;
11130 return DAG.getNode(Opcode, dl, Op.getValueType(),
11131 Op.getOperand(1), Op.getOperand(2));
11134 // SSE2/SSE41/AVX2 integer max/min intrinsics.
11135 case Intrinsic::x86_sse2_pmaxu_b:
11136 case Intrinsic::x86_sse41_pmaxuw:
11137 case Intrinsic::x86_sse41_pmaxud:
11138 case Intrinsic::x86_avx2_pmaxu_b:
11139 case Intrinsic::x86_avx2_pmaxu_w:
11140 case Intrinsic::x86_avx2_pmaxu_d:
11141 case Intrinsic::x86_sse2_pminu_b:
11142 case Intrinsic::x86_sse41_pminuw:
11143 case Intrinsic::x86_sse41_pminud:
11144 case Intrinsic::x86_avx2_pminu_b:
11145 case Intrinsic::x86_avx2_pminu_w:
11146 case Intrinsic::x86_avx2_pminu_d:
11147 case Intrinsic::x86_sse41_pmaxsb:
11148 case Intrinsic::x86_sse2_pmaxs_w:
11149 case Intrinsic::x86_sse41_pmaxsd:
11150 case Intrinsic::x86_avx2_pmaxs_b:
11151 case Intrinsic::x86_avx2_pmaxs_w:
11152 case Intrinsic::x86_avx2_pmaxs_d:
11153 case Intrinsic::x86_sse41_pminsb:
11154 case Intrinsic::x86_sse2_pmins_w:
11155 case Intrinsic::x86_sse41_pminsd:
11156 case Intrinsic::x86_avx2_pmins_b:
11157 case Intrinsic::x86_avx2_pmins_w:
11158 case Intrinsic::x86_avx2_pmins_d: {
11161 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
11162 case Intrinsic::x86_sse2_pmaxu_b:
11163 case Intrinsic::x86_sse41_pmaxuw:
11164 case Intrinsic::x86_sse41_pmaxud:
11165 case Intrinsic::x86_avx2_pmaxu_b:
11166 case Intrinsic::x86_avx2_pmaxu_w:
11167 case Intrinsic::x86_avx2_pmaxu_d:
11168 Opcode = X86ISD::UMAX;
11170 case Intrinsic::x86_sse2_pminu_b:
11171 case Intrinsic::x86_sse41_pminuw:
11172 case Intrinsic::x86_sse41_pminud:
11173 case Intrinsic::x86_avx2_pminu_b:
11174 case Intrinsic::x86_avx2_pminu_w:
11175 case Intrinsic::x86_avx2_pminu_d:
11176 Opcode = X86ISD::UMIN;
11178 case Intrinsic::x86_sse41_pmaxsb:
11179 case Intrinsic::x86_sse2_pmaxs_w:
11180 case Intrinsic::x86_sse41_pmaxsd:
11181 case Intrinsic::x86_avx2_pmaxs_b:
11182 case Intrinsic::x86_avx2_pmaxs_w:
11183 case Intrinsic::x86_avx2_pmaxs_d:
11184 Opcode = X86ISD::SMAX;
11186 case Intrinsic::x86_sse41_pminsb:
11187 case Intrinsic::x86_sse2_pmins_w:
11188 case Intrinsic::x86_sse41_pminsd:
11189 case Intrinsic::x86_avx2_pmins_b:
11190 case Intrinsic::x86_avx2_pmins_w:
11191 case Intrinsic::x86_avx2_pmins_d:
11192 Opcode = X86ISD::SMIN;
11195 return DAG.getNode(Opcode, dl, Op.getValueType(),
11196 Op.getOperand(1), Op.getOperand(2));
11199 // SSE/SSE2/AVX floating point max/min intrinsics.
11200 case Intrinsic::x86_sse_max_ps:
11201 case Intrinsic::x86_sse2_max_pd:
11202 case Intrinsic::x86_avx_max_ps_256:
11203 case Intrinsic::x86_avx_max_pd_256:
11204 case Intrinsic::x86_avx512_max_ps_512:
11205 case Intrinsic::x86_avx512_max_pd_512:
11206 case Intrinsic::x86_sse_min_ps:
11207 case Intrinsic::x86_sse2_min_pd:
11208 case Intrinsic::x86_avx_min_ps_256:
11209 case Intrinsic::x86_avx_min_pd_256:
11210 case Intrinsic::x86_avx512_min_ps_512:
11211 case Intrinsic::x86_avx512_min_pd_512: {
11214 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
11215 case Intrinsic::x86_sse_max_ps:
11216 case Intrinsic::x86_sse2_max_pd:
11217 case Intrinsic::x86_avx_max_ps_256:
11218 case Intrinsic::x86_avx_max_pd_256:
11219 case Intrinsic::x86_avx512_max_ps_512:
11220 case Intrinsic::x86_avx512_max_pd_512:
11221 Opcode = X86ISD::FMAX;
11223 case Intrinsic::x86_sse_min_ps:
11224 case Intrinsic::x86_sse2_min_pd:
11225 case Intrinsic::x86_avx_min_ps_256:
11226 case Intrinsic::x86_avx_min_pd_256:
11227 case Intrinsic::x86_avx512_min_ps_512:
11228 case Intrinsic::x86_avx512_min_pd_512:
11229 Opcode = X86ISD::FMIN;
11232 return DAG.getNode(Opcode, dl, Op.getValueType(),
11233 Op.getOperand(1), Op.getOperand(2));
11236 // AVX2 variable shift intrinsics
11237 case Intrinsic::x86_avx2_psllv_d:
11238 case Intrinsic::x86_avx2_psllv_q:
11239 case Intrinsic::x86_avx2_psllv_d_256:
11240 case Intrinsic::x86_avx2_psllv_q_256:
11241 case Intrinsic::x86_avx2_psrlv_d:
11242 case Intrinsic::x86_avx2_psrlv_q:
11243 case Intrinsic::x86_avx2_psrlv_d_256:
11244 case Intrinsic::x86_avx2_psrlv_q_256:
11245 case Intrinsic::x86_avx2_psrav_d:
11246 case Intrinsic::x86_avx2_psrav_d_256: {
11249 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
11250 case Intrinsic::x86_avx2_psllv_d:
11251 case Intrinsic::x86_avx2_psllv_q:
11252 case Intrinsic::x86_avx2_psllv_d_256:
11253 case Intrinsic::x86_avx2_psllv_q_256:
11256 case Intrinsic::x86_avx2_psrlv_d:
11257 case Intrinsic::x86_avx2_psrlv_q:
11258 case Intrinsic::x86_avx2_psrlv_d_256:
11259 case Intrinsic::x86_avx2_psrlv_q_256:
11262 case Intrinsic::x86_avx2_psrav_d:
11263 case Intrinsic::x86_avx2_psrav_d_256:
11267 return DAG.getNode(Opcode, dl, Op.getValueType(),
11268 Op.getOperand(1), Op.getOperand(2));
11271 case Intrinsic::x86_ssse3_pshuf_b_128:
11272 case Intrinsic::x86_avx2_pshuf_b:
11273 return DAG.getNode(X86ISD::PSHUFB, dl, Op.getValueType(),
11274 Op.getOperand(1), Op.getOperand(2));
11276 case Intrinsic::x86_ssse3_psign_b_128:
11277 case Intrinsic::x86_ssse3_psign_w_128:
11278 case Intrinsic::x86_ssse3_psign_d_128:
11279 case Intrinsic::x86_avx2_psign_b:
11280 case Intrinsic::x86_avx2_psign_w:
11281 case Intrinsic::x86_avx2_psign_d:
11282 return DAG.getNode(X86ISD::PSIGN, dl, Op.getValueType(),
11283 Op.getOperand(1), Op.getOperand(2));
11285 case Intrinsic::x86_sse41_insertps:
11286 return DAG.getNode(X86ISD::INSERTPS, dl, Op.getValueType(),
11287 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
11289 case Intrinsic::x86_avx_vperm2f128_ps_256:
11290 case Intrinsic::x86_avx_vperm2f128_pd_256:
11291 case Intrinsic::x86_avx_vperm2f128_si_256:
11292 case Intrinsic::x86_avx2_vperm2i128:
11293 return DAG.getNode(X86ISD::VPERM2X128, dl, Op.getValueType(),
11294 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
11296 case Intrinsic::x86_avx2_permd:
11297 case Intrinsic::x86_avx2_permps:
11298 // Operands intentionally swapped. Mask is last operand to intrinsic,
11299 // but second operand for node/intruction.
11300 return DAG.getNode(X86ISD::VPERMV, dl, Op.getValueType(),
11301 Op.getOperand(2), Op.getOperand(1));
11303 case Intrinsic::x86_sse_sqrt_ps:
11304 case Intrinsic::x86_sse2_sqrt_pd:
11305 case Intrinsic::x86_avx_sqrt_ps_256:
11306 case Intrinsic::x86_avx_sqrt_pd_256:
11307 return DAG.getNode(ISD::FSQRT, dl, Op.getValueType(), Op.getOperand(1));
11309 // ptest and testp intrinsics. The intrinsic these come from are designed to
11310 // return an integer value, not just an instruction so lower it to the ptest
11311 // or testp pattern and a setcc for the result.
11312 case Intrinsic::x86_sse41_ptestz:
11313 case Intrinsic::x86_sse41_ptestc:
11314 case Intrinsic::x86_sse41_ptestnzc:
11315 case Intrinsic::x86_avx_ptestz_256:
11316 case Intrinsic::x86_avx_ptestc_256:
11317 case Intrinsic::x86_avx_ptestnzc_256:
11318 case Intrinsic::x86_avx_vtestz_ps:
11319 case Intrinsic::x86_avx_vtestc_ps:
11320 case Intrinsic::x86_avx_vtestnzc_ps:
11321 case Intrinsic::x86_avx_vtestz_pd:
11322 case Intrinsic::x86_avx_vtestc_pd:
11323 case Intrinsic::x86_avx_vtestnzc_pd:
11324 case Intrinsic::x86_avx_vtestz_ps_256:
11325 case Intrinsic::x86_avx_vtestc_ps_256:
11326 case Intrinsic::x86_avx_vtestnzc_ps_256:
11327 case Intrinsic::x86_avx_vtestz_pd_256:
11328 case Intrinsic::x86_avx_vtestc_pd_256:
11329 case Intrinsic::x86_avx_vtestnzc_pd_256: {
11330 bool IsTestPacked = false;
11333 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
11334 case Intrinsic::x86_avx_vtestz_ps:
11335 case Intrinsic::x86_avx_vtestz_pd:
11336 case Intrinsic::x86_avx_vtestz_ps_256:
11337 case Intrinsic::x86_avx_vtestz_pd_256:
11338 IsTestPacked = true; // Fallthrough
11339 case Intrinsic::x86_sse41_ptestz:
11340 case Intrinsic::x86_avx_ptestz_256:
11342 X86CC = X86::COND_E;
11344 case Intrinsic::x86_avx_vtestc_ps:
11345 case Intrinsic::x86_avx_vtestc_pd:
11346 case Intrinsic::x86_avx_vtestc_ps_256:
11347 case Intrinsic::x86_avx_vtestc_pd_256:
11348 IsTestPacked = true; // Fallthrough
11349 case Intrinsic::x86_sse41_ptestc:
11350 case Intrinsic::x86_avx_ptestc_256:
11352 X86CC = X86::COND_B;
11354 case Intrinsic::x86_avx_vtestnzc_ps:
11355 case Intrinsic::x86_avx_vtestnzc_pd:
11356 case Intrinsic::x86_avx_vtestnzc_ps_256:
11357 case Intrinsic::x86_avx_vtestnzc_pd_256:
11358 IsTestPacked = true; // Fallthrough
11359 case Intrinsic::x86_sse41_ptestnzc:
11360 case Intrinsic::x86_avx_ptestnzc_256:
11362 X86CC = X86::COND_A;
11366 SDValue LHS = Op.getOperand(1);
11367 SDValue RHS = Op.getOperand(2);
11368 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
11369 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
11370 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
11371 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
11372 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
11374 case Intrinsic::x86_avx512_kortestz:
11375 case Intrinsic::x86_avx512_kortestc: {
11376 unsigned X86CC = (IntNo == Intrinsic::x86_avx512_kortestz)? X86::COND_E: X86::COND_B;
11377 SDValue LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1, Op.getOperand(1));
11378 SDValue RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1, Op.getOperand(2));
11379 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
11380 SDValue Test = DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, LHS, RHS);
11381 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
11382 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
11385 // SSE/AVX shift intrinsics
11386 case Intrinsic::x86_sse2_psll_w:
11387 case Intrinsic::x86_sse2_psll_d:
11388 case Intrinsic::x86_sse2_psll_q:
11389 case Intrinsic::x86_avx2_psll_w:
11390 case Intrinsic::x86_avx2_psll_d:
11391 case Intrinsic::x86_avx2_psll_q:
11392 case Intrinsic::x86_sse2_psrl_w:
11393 case Intrinsic::x86_sse2_psrl_d:
11394 case Intrinsic::x86_sse2_psrl_q:
11395 case Intrinsic::x86_avx2_psrl_w:
11396 case Intrinsic::x86_avx2_psrl_d:
11397 case Intrinsic::x86_avx2_psrl_q:
11398 case Intrinsic::x86_sse2_psra_w:
11399 case Intrinsic::x86_sse2_psra_d:
11400 case Intrinsic::x86_avx2_psra_w:
11401 case Intrinsic::x86_avx2_psra_d: {
11404 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
11405 case Intrinsic::x86_sse2_psll_w:
11406 case Intrinsic::x86_sse2_psll_d:
11407 case Intrinsic::x86_sse2_psll_q:
11408 case Intrinsic::x86_avx2_psll_w:
11409 case Intrinsic::x86_avx2_psll_d:
11410 case Intrinsic::x86_avx2_psll_q:
11411 Opcode = X86ISD::VSHL;
11413 case Intrinsic::x86_sse2_psrl_w:
11414 case Intrinsic::x86_sse2_psrl_d:
11415 case Intrinsic::x86_sse2_psrl_q:
11416 case Intrinsic::x86_avx2_psrl_w:
11417 case Intrinsic::x86_avx2_psrl_d:
11418 case Intrinsic::x86_avx2_psrl_q:
11419 Opcode = X86ISD::VSRL;
11421 case Intrinsic::x86_sse2_psra_w:
11422 case Intrinsic::x86_sse2_psra_d:
11423 case Intrinsic::x86_avx2_psra_w:
11424 case Intrinsic::x86_avx2_psra_d:
11425 Opcode = X86ISD::VSRA;
11428 return DAG.getNode(Opcode, dl, Op.getValueType(),
11429 Op.getOperand(1), Op.getOperand(2));
11432 // SSE/AVX immediate shift intrinsics
11433 case Intrinsic::x86_sse2_pslli_w:
11434 case Intrinsic::x86_sse2_pslli_d:
11435 case Intrinsic::x86_sse2_pslli_q:
11436 case Intrinsic::x86_avx2_pslli_w:
11437 case Intrinsic::x86_avx2_pslli_d:
11438 case Intrinsic::x86_avx2_pslli_q:
11439 case Intrinsic::x86_sse2_psrli_w:
11440 case Intrinsic::x86_sse2_psrli_d:
11441 case Intrinsic::x86_sse2_psrli_q:
11442 case Intrinsic::x86_avx2_psrli_w:
11443 case Intrinsic::x86_avx2_psrli_d:
11444 case Intrinsic::x86_avx2_psrli_q:
11445 case Intrinsic::x86_sse2_psrai_w:
11446 case Intrinsic::x86_sse2_psrai_d:
11447 case Intrinsic::x86_avx2_psrai_w:
11448 case Intrinsic::x86_avx2_psrai_d: {
11451 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
11452 case Intrinsic::x86_sse2_pslli_w:
11453 case Intrinsic::x86_sse2_pslli_d:
11454 case Intrinsic::x86_sse2_pslli_q:
11455 case Intrinsic::x86_avx2_pslli_w:
11456 case Intrinsic::x86_avx2_pslli_d:
11457 case Intrinsic::x86_avx2_pslli_q:
11458 Opcode = X86ISD::VSHLI;
11460 case Intrinsic::x86_sse2_psrli_w:
11461 case Intrinsic::x86_sse2_psrli_d:
11462 case Intrinsic::x86_sse2_psrli_q:
11463 case Intrinsic::x86_avx2_psrli_w:
11464 case Intrinsic::x86_avx2_psrli_d:
11465 case Intrinsic::x86_avx2_psrli_q:
11466 Opcode = X86ISD::VSRLI;
11468 case Intrinsic::x86_sse2_psrai_w:
11469 case Intrinsic::x86_sse2_psrai_d:
11470 case Intrinsic::x86_avx2_psrai_w:
11471 case Intrinsic::x86_avx2_psrai_d:
11472 Opcode = X86ISD::VSRAI;
11475 return getTargetVShiftNode(Opcode, dl, Op.getValueType(),
11476 Op.getOperand(1), Op.getOperand(2), DAG);
11479 case Intrinsic::x86_sse42_pcmpistria128:
11480 case Intrinsic::x86_sse42_pcmpestria128:
11481 case Intrinsic::x86_sse42_pcmpistric128:
11482 case Intrinsic::x86_sse42_pcmpestric128:
11483 case Intrinsic::x86_sse42_pcmpistrio128:
11484 case Intrinsic::x86_sse42_pcmpestrio128:
11485 case Intrinsic::x86_sse42_pcmpistris128:
11486 case Intrinsic::x86_sse42_pcmpestris128:
11487 case Intrinsic::x86_sse42_pcmpistriz128:
11488 case Intrinsic::x86_sse42_pcmpestriz128: {
11492 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
11493 case Intrinsic::x86_sse42_pcmpistria128:
11494 Opcode = X86ISD::PCMPISTRI;
11495 X86CC = X86::COND_A;
11497 case Intrinsic::x86_sse42_pcmpestria128:
11498 Opcode = X86ISD::PCMPESTRI;
11499 X86CC = X86::COND_A;
11501 case Intrinsic::x86_sse42_pcmpistric128:
11502 Opcode = X86ISD::PCMPISTRI;
11503 X86CC = X86::COND_B;
11505 case Intrinsic::x86_sse42_pcmpestric128:
11506 Opcode = X86ISD::PCMPESTRI;
11507 X86CC = X86::COND_B;
11509 case Intrinsic::x86_sse42_pcmpistrio128:
11510 Opcode = X86ISD::PCMPISTRI;
11511 X86CC = X86::COND_O;
11513 case Intrinsic::x86_sse42_pcmpestrio128:
11514 Opcode = X86ISD::PCMPESTRI;
11515 X86CC = X86::COND_O;
11517 case Intrinsic::x86_sse42_pcmpistris128:
11518 Opcode = X86ISD::PCMPISTRI;
11519 X86CC = X86::COND_S;
11521 case Intrinsic::x86_sse42_pcmpestris128:
11522 Opcode = X86ISD::PCMPESTRI;
11523 X86CC = X86::COND_S;
11525 case Intrinsic::x86_sse42_pcmpistriz128:
11526 Opcode = X86ISD::PCMPISTRI;
11527 X86CC = X86::COND_E;
11529 case Intrinsic::x86_sse42_pcmpestriz128:
11530 Opcode = X86ISD::PCMPESTRI;
11531 X86CC = X86::COND_E;
11534 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
11535 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
11536 SDValue PCMP = DAG.getNode(Opcode, dl, VTs, NewOps.data(), NewOps.size());
11537 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
11538 DAG.getConstant(X86CC, MVT::i8),
11539 SDValue(PCMP.getNode(), 1));
11540 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
11543 case Intrinsic::x86_sse42_pcmpistri128:
11544 case Intrinsic::x86_sse42_pcmpestri128: {
11546 if (IntNo == Intrinsic::x86_sse42_pcmpistri128)
11547 Opcode = X86ISD::PCMPISTRI;
11549 Opcode = X86ISD::PCMPESTRI;
11551 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
11552 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
11553 return DAG.getNode(Opcode, dl, VTs, NewOps.data(), NewOps.size());
11555 case Intrinsic::x86_fma_vfmadd_ps:
11556 case Intrinsic::x86_fma_vfmadd_pd:
11557 case Intrinsic::x86_fma_vfmsub_ps:
11558 case Intrinsic::x86_fma_vfmsub_pd:
11559 case Intrinsic::x86_fma_vfnmadd_ps:
11560 case Intrinsic::x86_fma_vfnmadd_pd:
11561 case Intrinsic::x86_fma_vfnmsub_ps:
11562 case Intrinsic::x86_fma_vfnmsub_pd:
11563 case Intrinsic::x86_fma_vfmaddsub_ps:
11564 case Intrinsic::x86_fma_vfmaddsub_pd:
11565 case Intrinsic::x86_fma_vfmsubadd_ps:
11566 case Intrinsic::x86_fma_vfmsubadd_pd:
11567 case Intrinsic::x86_fma_vfmadd_ps_256:
11568 case Intrinsic::x86_fma_vfmadd_pd_256:
11569 case Intrinsic::x86_fma_vfmsub_ps_256:
11570 case Intrinsic::x86_fma_vfmsub_pd_256:
11571 case Intrinsic::x86_fma_vfnmadd_ps_256:
11572 case Intrinsic::x86_fma_vfnmadd_pd_256:
11573 case Intrinsic::x86_fma_vfnmsub_ps_256:
11574 case Intrinsic::x86_fma_vfnmsub_pd_256:
11575 case Intrinsic::x86_fma_vfmaddsub_ps_256:
11576 case Intrinsic::x86_fma_vfmaddsub_pd_256:
11577 case Intrinsic::x86_fma_vfmsubadd_ps_256:
11578 case Intrinsic::x86_fma_vfmsubadd_pd_256: {
11581 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
11582 case Intrinsic::x86_fma_vfmadd_ps:
11583 case Intrinsic::x86_fma_vfmadd_pd:
11584 case Intrinsic::x86_fma_vfmadd_ps_256:
11585 case Intrinsic::x86_fma_vfmadd_pd_256:
11586 Opc = X86ISD::FMADD;
11588 case Intrinsic::x86_fma_vfmsub_ps:
11589 case Intrinsic::x86_fma_vfmsub_pd:
11590 case Intrinsic::x86_fma_vfmsub_ps_256:
11591 case Intrinsic::x86_fma_vfmsub_pd_256:
11592 Opc = X86ISD::FMSUB;
11594 case Intrinsic::x86_fma_vfnmadd_ps:
11595 case Intrinsic::x86_fma_vfnmadd_pd:
11596 case Intrinsic::x86_fma_vfnmadd_ps_256:
11597 case Intrinsic::x86_fma_vfnmadd_pd_256:
11598 Opc = X86ISD::FNMADD;
11600 case Intrinsic::x86_fma_vfnmsub_ps:
11601 case Intrinsic::x86_fma_vfnmsub_pd:
11602 case Intrinsic::x86_fma_vfnmsub_ps_256:
11603 case Intrinsic::x86_fma_vfnmsub_pd_256:
11604 Opc = X86ISD::FNMSUB;
11606 case Intrinsic::x86_fma_vfmaddsub_ps:
11607 case Intrinsic::x86_fma_vfmaddsub_pd:
11608 case Intrinsic::x86_fma_vfmaddsub_ps_256:
11609 case Intrinsic::x86_fma_vfmaddsub_pd_256:
11610 Opc = X86ISD::FMADDSUB;
11612 case Intrinsic::x86_fma_vfmsubadd_ps:
11613 case Intrinsic::x86_fma_vfmsubadd_pd:
11614 case Intrinsic::x86_fma_vfmsubadd_ps_256:
11615 case Intrinsic::x86_fma_vfmsubadd_pd_256:
11616 Opc = X86ISD::FMSUBADD;
11620 return DAG.getNode(Opc, dl, Op.getValueType(), Op.getOperand(1),
11621 Op.getOperand(2), Op.getOperand(3));
11626 static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, SelectionDAG &DAG) {
11628 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11630 default: return SDValue(); // Don't custom lower most intrinsics.
11632 // RDRAND/RDSEED intrinsics.
11633 case Intrinsic::x86_rdrand_16:
11634 case Intrinsic::x86_rdrand_32:
11635 case Intrinsic::x86_rdrand_64:
11636 case Intrinsic::x86_rdseed_16:
11637 case Intrinsic::x86_rdseed_32:
11638 case Intrinsic::x86_rdseed_64: {
11639 unsigned Opcode = (IntNo == Intrinsic::x86_rdseed_16 ||
11640 IntNo == Intrinsic::x86_rdseed_32 ||
11641 IntNo == Intrinsic::x86_rdseed_64) ? X86ISD::RDSEED :
11643 // Emit the node with the right value type.
11644 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other);
11645 SDValue Result = DAG.getNode(Opcode, dl, VTs, Op.getOperand(0));
11647 // If the value returned by RDRAND/RDSEED was valid (CF=1), return 1.
11648 // Otherwise return the value from Rand, which is always 0, casted to i32.
11649 SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)),
11650 DAG.getConstant(1, Op->getValueType(1)),
11651 DAG.getConstant(X86::COND_B, MVT::i32),
11652 SDValue(Result.getNode(), 1) };
11653 SDValue isValid = DAG.getNode(X86ISD::CMOV, dl,
11654 DAG.getVTList(Op->getValueType(1), MVT::Glue),
11655 Ops, array_lengthof(Ops));
11657 // Return { result, isValid, chain }.
11658 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid,
11659 SDValue(Result.getNode(), 2));
11662 // XTEST intrinsics.
11663 case Intrinsic::x86_xtest: {
11664 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
11665 SDValue InTrans = DAG.getNode(X86ISD::XTEST, dl, VTs, Op.getOperand(0));
11666 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
11667 DAG.getConstant(X86::COND_NE, MVT::i8),
11669 SDValue Ret = DAG.getNode(ISD::ZERO_EXTEND, dl, Op->getValueType(0), SetCC);
11670 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(),
11671 Ret, SDValue(InTrans.getNode(), 1));
11676 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
11677 SelectionDAG &DAG) const {
11678 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
11679 MFI->setReturnAddressIsTaken(true);
11681 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
11683 EVT PtrVT = getPointerTy();
11686 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
11687 const X86RegisterInfo *RegInfo =
11688 static_cast<const X86RegisterInfo*>(getTargetMachine().getRegisterInfo());
11689 SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), PtrVT);
11690 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
11691 DAG.getNode(ISD::ADD, dl, PtrVT,
11692 FrameAddr, Offset),
11693 MachinePointerInfo(), false, false, false, 0);
11696 // Just load the return address.
11697 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
11698 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
11699 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
11702 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
11703 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
11704 MFI->setFrameAddressIsTaken(true);
11706 EVT VT = Op.getValueType();
11707 SDLoc dl(Op); // FIXME probably not meaningful
11708 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
11709 const X86RegisterInfo *RegInfo =
11710 static_cast<const X86RegisterInfo*>(getTargetMachine().getRegisterInfo());
11711 unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
11712 assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
11713 (FrameReg == X86::EBP && VT == MVT::i32)) &&
11714 "Invalid Frame Register!");
11715 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
11717 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
11718 MachinePointerInfo(),
11719 false, false, false, 0);
11723 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
11724 SelectionDAG &DAG) const {
11725 const X86RegisterInfo *RegInfo =
11726 static_cast<const X86RegisterInfo*>(getTargetMachine().getRegisterInfo());
11727 return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize());
11730 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
11731 SDValue Chain = Op.getOperand(0);
11732 SDValue Offset = Op.getOperand(1);
11733 SDValue Handler = Op.getOperand(2);
11736 EVT PtrVT = getPointerTy();
11737 const X86RegisterInfo *RegInfo =
11738 static_cast<const X86RegisterInfo*>(getTargetMachine().getRegisterInfo());
11739 unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
11740 assert(((FrameReg == X86::RBP && PtrVT == MVT::i64) ||
11741 (FrameReg == X86::EBP && PtrVT == MVT::i32)) &&
11742 "Invalid Frame Register!");
11743 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, PtrVT);
11744 unsigned StoreAddrReg = (PtrVT == MVT::i64) ? X86::RCX : X86::ECX;
11746 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Frame,
11747 DAG.getIntPtrConstant(RegInfo->getSlotSize()));
11748 StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, StoreAddr, Offset);
11749 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
11751 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
11753 return DAG.getNode(X86ISD::EH_RETURN, dl, MVT::Other, Chain,
11754 DAG.getRegister(StoreAddrReg, PtrVT));
11757 SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
11758 SelectionDAG &DAG) const {
11760 return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL,
11761 DAG.getVTList(MVT::i32, MVT::Other),
11762 Op.getOperand(0), Op.getOperand(1));
11765 SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
11766 SelectionDAG &DAG) const {
11768 return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
11769 Op.getOperand(0), Op.getOperand(1));
11772 static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
11773 return Op.getOperand(0);
11776 SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
11777 SelectionDAG &DAG) const {
11778 SDValue Root = Op.getOperand(0);
11779 SDValue Trmp = Op.getOperand(1); // trampoline
11780 SDValue FPtr = Op.getOperand(2); // nested function
11781 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
11784 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
11785 const TargetRegisterInfo* TRI = getTargetMachine().getRegisterInfo();
11787 if (Subtarget->is64Bit()) {
11788 SDValue OutChains[6];
11790 // Large code-model.
11791 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
11792 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
11794 const unsigned char N86R10 = TRI->getEncodingValue(X86::R10) & 0x7;
11795 const unsigned char N86R11 = TRI->getEncodingValue(X86::R11) & 0x7;
11797 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
11799 // Load the pointer to the nested function into R11.
11800 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
11801 SDValue Addr = Trmp;
11802 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
11803 Addr, MachinePointerInfo(TrmpAddr),
11806 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
11807 DAG.getConstant(2, MVT::i64));
11808 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
11809 MachinePointerInfo(TrmpAddr, 2),
11812 // Load the 'nest' parameter value into R10.
11813 // R10 is specified in X86CallingConv.td
11814 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
11815 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
11816 DAG.getConstant(10, MVT::i64));
11817 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
11818 Addr, MachinePointerInfo(TrmpAddr, 10),
11821 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
11822 DAG.getConstant(12, MVT::i64));
11823 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
11824 MachinePointerInfo(TrmpAddr, 12),
11827 // Jump to the nested function.
11828 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
11829 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
11830 DAG.getConstant(20, MVT::i64));
11831 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
11832 Addr, MachinePointerInfo(TrmpAddr, 20),
11835 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
11836 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
11837 DAG.getConstant(22, MVT::i64));
11838 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
11839 MachinePointerInfo(TrmpAddr, 22),
11842 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6);
11844 const Function *Func =
11845 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
11846 CallingConv::ID CC = Func->getCallingConv();
11851 llvm_unreachable("Unsupported calling convention");
11852 case CallingConv::C:
11853 case CallingConv::X86_StdCall: {
11854 // Pass 'nest' parameter in ECX.
11855 // Must be kept in sync with X86CallingConv.td
11856 NestReg = X86::ECX;
11858 // Check that ECX wasn't needed by an 'inreg' parameter.
11859 FunctionType *FTy = Func->getFunctionType();
11860 const AttributeSet &Attrs = Func->getAttributes();
11862 if (!Attrs.isEmpty() && !Func->isVarArg()) {
11863 unsigned InRegCount = 0;
11866 for (FunctionType::param_iterator I = FTy->param_begin(),
11867 E = FTy->param_end(); I != E; ++I, ++Idx)
11868 if (Attrs.hasAttribute(Idx, Attribute::InReg))
11869 // FIXME: should only count parameters that are lowered to integers.
11870 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
11872 if (InRegCount > 2) {
11873 report_fatal_error("Nest register in use - reduce number of inreg"
11879 case CallingConv::X86_FastCall:
11880 case CallingConv::X86_ThisCall:
11881 case CallingConv::Fast:
11882 // Pass 'nest' parameter in EAX.
11883 // Must be kept in sync with X86CallingConv.td
11884 NestReg = X86::EAX;
11888 SDValue OutChains[4];
11889 SDValue Addr, Disp;
11891 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
11892 DAG.getConstant(10, MVT::i32));
11893 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
11895 // This is storing the opcode for MOV32ri.
11896 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
11897 const unsigned char N86Reg = TRI->getEncodingValue(NestReg) & 0x7;
11898 OutChains[0] = DAG.getStore(Root, dl,
11899 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
11900 Trmp, MachinePointerInfo(TrmpAddr),
11903 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
11904 DAG.getConstant(1, MVT::i32));
11905 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
11906 MachinePointerInfo(TrmpAddr, 1),
11909 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
11910 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
11911 DAG.getConstant(5, MVT::i32));
11912 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
11913 MachinePointerInfo(TrmpAddr, 5),
11916 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
11917 DAG.getConstant(6, MVT::i32));
11918 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
11919 MachinePointerInfo(TrmpAddr, 6),
11922 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4);
11926 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
11927 SelectionDAG &DAG) const {
11929 The rounding mode is in bits 11:10 of FPSR, and has the following
11931 00 Round to nearest
11936 FLT_ROUNDS, on the other hand, expects the following:
11943 To perform the conversion, we do:
11944 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
11947 MachineFunction &MF = DAG.getMachineFunction();
11948 const TargetMachine &TM = MF.getTarget();
11949 const TargetFrameLowering &TFI = *TM.getFrameLowering();
11950 unsigned StackAlignment = TFI.getStackAlignment();
11951 EVT VT = Op.getValueType();
11954 // Save FP Control Word to stack slot
11955 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
11956 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
11958 MachineMemOperand *MMO =
11959 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11960 MachineMemOperand::MOStore, 2, 2);
11962 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
11963 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
11964 DAG.getVTList(MVT::Other),
11965 Ops, array_lengthof(Ops), MVT::i16,
11968 // Load FP Control Word from stack slot
11969 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
11970 MachinePointerInfo(), false, false, false, 0);
11972 // Transform as necessary
11974 DAG.getNode(ISD::SRL, DL, MVT::i16,
11975 DAG.getNode(ISD::AND, DL, MVT::i16,
11976 CWD, DAG.getConstant(0x800, MVT::i16)),
11977 DAG.getConstant(11, MVT::i8));
11979 DAG.getNode(ISD::SRL, DL, MVT::i16,
11980 DAG.getNode(ISD::AND, DL, MVT::i16,
11981 CWD, DAG.getConstant(0x400, MVT::i16)),
11982 DAG.getConstant(9, MVT::i8));
11985 DAG.getNode(ISD::AND, DL, MVT::i16,
11986 DAG.getNode(ISD::ADD, DL, MVT::i16,
11987 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
11988 DAG.getConstant(1, MVT::i16)),
11989 DAG.getConstant(3, MVT::i16));
11991 return DAG.getNode((VT.getSizeInBits() < 16 ?
11992 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
11995 static SDValue LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
11996 EVT VT = Op.getValueType();
11998 unsigned NumBits = VT.getSizeInBits();
12001 Op = Op.getOperand(0);
12002 if (VT == MVT::i8) {
12003 // Zero extend to i32 since there is not an i8 bsr.
12005 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
12008 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
12009 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
12010 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
12012 // If src is zero (i.e. bsr sets ZF), returns NumBits.
12015 DAG.getConstant(NumBits+NumBits-1, OpVT),
12016 DAG.getConstant(X86::COND_E, MVT::i8),
12019 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
12021 // Finally xor with NumBits-1.
12022 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
12025 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
12029 static SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, SelectionDAG &DAG) {
12030 EVT VT = Op.getValueType();
12032 unsigned NumBits = VT.getSizeInBits();
12035 Op = Op.getOperand(0);
12036 if (VT == MVT::i8) {
12037 // Zero extend to i32 since there is not an i8 bsr.
12039 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
12042 // Issue a bsr (scan bits in reverse).
12043 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
12044 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
12046 // And xor with NumBits-1.
12047 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
12050 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
12054 static SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
12055 EVT VT = Op.getValueType();
12056 unsigned NumBits = VT.getSizeInBits();
12058 Op = Op.getOperand(0);
12060 // Issue a bsf (scan bits forward) which also sets EFLAGS.
12061 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
12062 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
12064 // If src is zero (i.e. bsf sets ZF), returns NumBits.
12067 DAG.getConstant(NumBits, VT),
12068 DAG.getConstant(X86::COND_E, MVT::i8),
12071 return DAG.getNode(X86ISD::CMOV, dl, VT, Ops, array_lengthof(Ops));
12074 // Lower256IntArith - Break a 256-bit integer operation into two new 128-bit
12075 // ones, and then concatenate the result back.
12076 static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) {
12077 EVT VT = Op.getValueType();
12079 assert(VT.is256BitVector() && VT.isInteger() &&
12080 "Unsupported value type for operation");
12082 unsigned NumElems = VT.getVectorNumElements();
12085 // Extract the LHS vectors
12086 SDValue LHS = Op.getOperand(0);
12087 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
12088 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
12090 // Extract the RHS vectors
12091 SDValue RHS = Op.getOperand(1);
12092 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
12093 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
12095 MVT EltVT = VT.getVectorElementType().getSimpleVT();
12096 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
12098 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
12099 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1),
12100 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
12103 static SDValue LowerADD(SDValue Op, SelectionDAG &DAG) {
12104 assert(Op.getValueType().is256BitVector() &&
12105 Op.getValueType().isInteger() &&
12106 "Only handle AVX 256-bit vector integer operation");
12107 return Lower256IntArith(Op, DAG);
12110 static SDValue LowerSUB(SDValue Op, SelectionDAG &DAG) {
12111 assert(Op.getValueType().is256BitVector() &&
12112 Op.getValueType().isInteger() &&
12113 "Only handle AVX 256-bit vector integer operation");
12114 return Lower256IntArith(Op, DAG);
12117 static SDValue LowerMUL(SDValue Op, const X86Subtarget *Subtarget,
12118 SelectionDAG &DAG) {
12120 EVT VT = Op.getValueType();
12122 // Decompose 256-bit ops into smaller 128-bit ops.
12123 if (VT.is256BitVector() && !Subtarget->hasInt256())
12124 return Lower256IntArith(Op, DAG);
12126 SDValue A = Op.getOperand(0);
12127 SDValue B = Op.getOperand(1);
12129 // Lower v4i32 mul as 2x shuffle, 2x pmuludq, 2x shuffle.
12130 if (VT == MVT::v4i32) {
12131 assert(Subtarget->hasSSE2() && !Subtarget->hasSSE41() &&
12132 "Should not custom lower when pmuldq is available!");
12134 // Extract the odd parts.
12135 static const int UnpackMask[] = { 1, -1, 3, -1 };
12136 SDValue Aodds = DAG.getVectorShuffle(VT, dl, A, A, UnpackMask);
12137 SDValue Bodds = DAG.getVectorShuffle(VT, dl, B, B, UnpackMask);
12139 // Multiply the even parts.
12140 SDValue Evens = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, A, B);
12141 // Now multiply odd parts.
12142 SDValue Odds = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, Aodds, Bodds);
12144 Evens = DAG.getNode(ISD::BITCAST, dl, VT, Evens);
12145 Odds = DAG.getNode(ISD::BITCAST, dl, VT, Odds);
12147 // Merge the two vectors back together with a shuffle. This expands into 2
12149 static const int ShufMask[] = { 0, 4, 2, 6 };
12150 return DAG.getVectorShuffle(VT, dl, Evens, Odds, ShufMask);
12153 assert((VT == MVT::v2i64 || VT == MVT::v4i64) &&
12154 "Only know how to lower V2I64/V4I64 multiply");
12156 // Ahi = psrlqi(a, 32);
12157 // Bhi = psrlqi(b, 32);
12159 // AloBlo = pmuludq(a, b);
12160 // AloBhi = pmuludq(a, Bhi);
12161 // AhiBlo = pmuludq(Ahi, b);
12163 // AloBhi = psllqi(AloBhi, 32);
12164 // AhiBlo = psllqi(AhiBlo, 32);
12165 // return AloBlo + AloBhi + AhiBlo;
12167 SDValue ShAmt = DAG.getConstant(32, MVT::i32);
12169 SDValue Ahi = DAG.getNode(X86ISD::VSRLI, dl, VT, A, ShAmt);
12170 SDValue Bhi = DAG.getNode(X86ISD::VSRLI, dl, VT, B, ShAmt);
12172 // Bit cast to 32-bit vectors for MULUDQ
12173 EVT MulVT = (VT == MVT::v2i64) ? MVT::v4i32 : MVT::v8i32;
12174 A = DAG.getNode(ISD::BITCAST, dl, MulVT, A);
12175 B = DAG.getNode(ISD::BITCAST, dl, MulVT, B);
12176 Ahi = DAG.getNode(ISD::BITCAST, dl, MulVT, Ahi);
12177 Bhi = DAG.getNode(ISD::BITCAST, dl, MulVT, Bhi);
12179 SDValue AloBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, B);
12180 SDValue AloBhi = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, Bhi);
12181 SDValue AhiBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Ahi, B);
12183 AloBhi = DAG.getNode(X86ISD::VSHLI, dl, VT, AloBhi, ShAmt);
12184 AhiBlo = DAG.getNode(X86ISD::VSHLI, dl, VT, AhiBlo, ShAmt);
12186 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
12187 return DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
12190 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
12191 EVT VT = Op.getValueType();
12192 EVT EltTy = VT.getVectorElementType();
12193 unsigned NumElts = VT.getVectorNumElements();
12194 SDValue N0 = Op.getOperand(0);
12197 // Lower sdiv X, pow2-const.
12198 BuildVectorSDNode *C = dyn_cast<BuildVectorSDNode>(Op.getOperand(1));
12202 APInt SplatValue, SplatUndef;
12203 unsigned SplatBitSize;
12205 if (!C->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
12207 EltTy.getSizeInBits() < SplatBitSize)
12210 if ((SplatValue != 0) &&
12211 (SplatValue.isPowerOf2() || (-SplatValue).isPowerOf2())) {
12212 unsigned lg2 = SplatValue.countTrailingZeros();
12213 // Splat the sign bit.
12214 SDValue Sz = DAG.getConstant(EltTy.getSizeInBits()-1, MVT::i32);
12215 SDValue SGN = getTargetVShiftNode(X86ISD::VSRAI, dl, VT, N0, Sz, DAG);
12216 // Add (N0 < 0) ? abs2 - 1 : 0;
12217 SDValue Amt = DAG.getConstant(EltTy.getSizeInBits() - lg2, MVT::i32);
12218 SDValue SRL = getTargetVShiftNode(X86ISD::VSRLI, dl, VT, SGN, Amt, DAG);
12219 SDValue ADD = DAG.getNode(ISD::ADD, dl, VT, N0, SRL);
12220 SDValue Lg2Amt = DAG.getConstant(lg2, MVT::i32);
12221 SDValue SRA = getTargetVShiftNode(X86ISD::VSRAI, dl, VT, ADD, Lg2Amt, DAG);
12223 // If we're dividing by a positive value, we're done. Otherwise, we must
12224 // negate the result.
12225 if (SplatValue.isNonNegative())
12228 SmallVector<SDValue, 16> V(NumElts, DAG.getConstant(0, EltTy));
12229 SDValue Zero = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &V[0], NumElts);
12230 return DAG.getNode(ISD::SUB, dl, VT, Zero, SRA);
12235 static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
12236 const X86Subtarget *Subtarget) {
12237 EVT VT = Op.getValueType();
12239 SDValue R = Op.getOperand(0);
12240 SDValue Amt = Op.getOperand(1);
12242 // Optimize shl/srl/sra with constant shift amount.
12243 if (isSplatVector(Amt.getNode())) {
12244 SDValue SclrAmt = Amt->getOperand(0);
12245 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(SclrAmt)) {
12246 uint64_t ShiftAmt = C->getZExtValue();
12248 if (VT == MVT::v2i64 || VT == MVT::v4i32 || VT == MVT::v8i16 ||
12249 (Subtarget->hasInt256() &&
12250 (VT == MVT::v4i64 || VT == MVT::v8i32 || VT == MVT::v16i16)) ||
12251 (Subtarget->hasAVX512() &&
12252 (VT == MVT::v8i64 || VT == MVT::v16i32))) {
12253 if (Op.getOpcode() == ISD::SHL)
12254 return DAG.getNode(X86ISD::VSHLI, dl, VT, R,
12255 DAG.getConstant(ShiftAmt, MVT::i32));
12256 if (Op.getOpcode() == ISD::SRL)
12257 return DAG.getNode(X86ISD::VSRLI, dl, VT, R,
12258 DAG.getConstant(ShiftAmt, MVT::i32));
12259 if (Op.getOpcode() == ISD::SRA && VT != MVT::v2i64 && VT != MVT::v4i64)
12260 return DAG.getNode(X86ISD::VSRAI, dl, VT, R,
12261 DAG.getConstant(ShiftAmt, MVT::i32));
12264 if (VT == MVT::v16i8) {
12265 if (Op.getOpcode() == ISD::SHL) {
12266 // Make a large shift.
12267 SDValue SHL = DAG.getNode(X86ISD::VSHLI, dl, MVT::v8i16, R,
12268 DAG.getConstant(ShiftAmt, MVT::i32));
12269 SHL = DAG.getNode(ISD::BITCAST, dl, VT, SHL);
12270 // Zero out the rightmost bits.
12271 SmallVector<SDValue, 16> V(16,
12272 DAG.getConstant(uint8_t(-1U << ShiftAmt),
12274 return DAG.getNode(ISD::AND, dl, VT, SHL,
12275 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &V[0], 16));
12277 if (Op.getOpcode() == ISD::SRL) {
12278 // Make a large shift.
12279 SDValue SRL = DAG.getNode(X86ISD::VSRLI, dl, MVT::v8i16, R,
12280 DAG.getConstant(ShiftAmt, MVT::i32));
12281 SRL = DAG.getNode(ISD::BITCAST, dl, VT, SRL);
12282 // Zero out the leftmost bits.
12283 SmallVector<SDValue, 16> V(16,
12284 DAG.getConstant(uint8_t(-1U) >> ShiftAmt,
12286 return DAG.getNode(ISD::AND, dl, VT, SRL,
12287 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &V[0], 16));
12289 if (Op.getOpcode() == ISD::SRA) {
12290 if (ShiftAmt == 7) {
12291 // R s>> 7 === R s< 0
12292 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
12293 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
12296 // R s>> a === ((R u>> a) ^ m) - m
12297 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
12298 SmallVector<SDValue, 16> V(16, DAG.getConstant(128 >> ShiftAmt,
12300 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &V[0], 16);
12301 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
12302 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
12305 llvm_unreachable("Unknown shift opcode.");
12308 if (Subtarget->hasInt256() && VT == MVT::v32i8) {
12309 if (Op.getOpcode() == ISD::SHL) {
12310 // Make a large shift.
12311 SDValue SHL = DAG.getNode(X86ISD::VSHLI, dl, MVT::v16i16, R,
12312 DAG.getConstant(ShiftAmt, MVT::i32));
12313 SHL = DAG.getNode(ISD::BITCAST, dl, VT, SHL);
12314 // Zero out the rightmost bits.
12315 SmallVector<SDValue, 32> V(32,
12316 DAG.getConstant(uint8_t(-1U << ShiftAmt),
12318 return DAG.getNode(ISD::AND, dl, VT, SHL,
12319 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &V[0], 32));
12321 if (Op.getOpcode() == ISD::SRL) {
12322 // Make a large shift.
12323 SDValue SRL = DAG.getNode(X86ISD::VSRLI, dl, MVT::v16i16, R,
12324 DAG.getConstant(ShiftAmt, MVT::i32));
12325 SRL = DAG.getNode(ISD::BITCAST, dl, VT, SRL);
12326 // Zero out the leftmost bits.
12327 SmallVector<SDValue, 32> V(32,
12328 DAG.getConstant(uint8_t(-1U) >> ShiftAmt,
12330 return DAG.getNode(ISD::AND, dl, VT, SRL,
12331 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &V[0], 32));
12333 if (Op.getOpcode() == ISD::SRA) {
12334 if (ShiftAmt == 7) {
12335 // R s>> 7 === R s< 0
12336 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
12337 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
12340 // R s>> a === ((R u>> a) ^ m) - m
12341 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
12342 SmallVector<SDValue, 32> V(32, DAG.getConstant(128 >> ShiftAmt,
12344 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &V[0], 32);
12345 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
12346 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
12349 llvm_unreachable("Unknown shift opcode.");
12354 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
12355 if (!Subtarget->is64Bit() &&
12356 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
12357 Amt.getOpcode() == ISD::BITCAST &&
12358 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
12359 Amt = Amt.getOperand(0);
12360 unsigned Ratio = Amt.getValueType().getVectorNumElements() /
12361 VT.getVectorNumElements();
12362 unsigned RatioInLog2 = Log2_32_Ceil(Ratio);
12363 uint64_t ShiftAmt = 0;
12364 for (unsigned i = 0; i != Ratio; ++i) {
12365 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i));
12369 ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2)));
12371 // Check remaining shift amounts.
12372 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
12373 uint64_t ShAmt = 0;
12374 for (unsigned j = 0; j != Ratio; ++j) {
12375 ConstantSDNode *C =
12376 dyn_cast<ConstantSDNode>(Amt.getOperand(i + j));
12380 ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2)));
12382 if (ShAmt != ShiftAmt)
12385 switch (Op.getOpcode()) {
12387 llvm_unreachable("Unknown shift opcode!");
12389 return DAG.getNode(X86ISD::VSHLI, dl, VT, R,
12390 DAG.getConstant(ShiftAmt, MVT::i32));
12392 return DAG.getNode(X86ISD::VSRLI, dl, VT, R,
12393 DAG.getConstant(ShiftAmt, MVT::i32));
12395 return DAG.getNode(X86ISD::VSRAI, dl, VT, R,
12396 DAG.getConstant(ShiftAmt, MVT::i32));
12403 static SDValue LowerScalarVariableShift(SDValue Op, SelectionDAG &DAG,
12404 const X86Subtarget* Subtarget) {
12405 EVT VT = Op.getValueType();
12407 SDValue R = Op.getOperand(0);
12408 SDValue Amt = Op.getOperand(1);
12410 if ((VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) ||
12411 VT == MVT::v4i32 || VT == MVT::v8i16 ||
12412 (Subtarget->hasInt256() &&
12413 ((VT == MVT::v4i64 && Op.getOpcode() != ISD::SRA) ||
12414 VT == MVT::v8i32 || VT == MVT::v16i16)) ||
12415 (Subtarget->hasAVX512() && (VT == MVT::v8i64 || VT == MVT::v16i32))) {
12417 EVT EltVT = VT.getVectorElementType();
12419 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
12420 unsigned NumElts = VT.getVectorNumElements();
12422 for (i = 0; i != NumElts; ++i) {
12423 if (Amt.getOperand(i).getOpcode() == ISD::UNDEF)
12427 for (j = i; j != NumElts; ++j) {
12428 SDValue Arg = Amt.getOperand(j);
12429 if (Arg.getOpcode() == ISD::UNDEF) continue;
12430 if (Arg != Amt.getOperand(i))
12433 if (i != NumElts && j == NumElts)
12434 BaseShAmt = Amt.getOperand(i);
12436 if (Amt.getOpcode() == ISD::EXTRACT_SUBVECTOR)
12437 Amt = Amt.getOperand(0);
12438 if (Amt.getOpcode() == ISD::VECTOR_SHUFFLE &&
12439 cast<ShuffleVectorSDNode>(Amt)->isSplat()) {
12440 SDValue InVec = Amt.getOperand(0);
12441 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
12442 unsigned NumElts = InVec.getValueType().getVectorNumElements();
12444 for (; i != NumElts; ++i) {
12445 SDValue Arg = InVec.getOperand(i);
12446 if (Arg.getOpcode() == ISD::UNDEF) continue;
12450 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
12451 if (ConstantSDNode *C =
12452 dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
12453 unsigned SplatIdx =
12454 cast<ShuffleVectorSDNode>(Amt)->getSplatIndex();
12455 if (C->getZExtValue() == SplatIdx)
12456 BaseShAmt = InVec.getOperand(1);
12459 if (BaseShAmt.getNode() == 0)
12460 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, Amt,
12461 DAG.getIntPtrConstant(0));
12465 if (BaseShAmt.getNode()) {
12466 if (EltVT.bitsGT(MVT::i32))
12467 BaseShAmt = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BaseShAmt);
12468 else if (EltVT.bitsLT(MVT::i32))
12469 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt);
12471 switch (Op.getOpcode()) {
12473 llvm_unreachable("Unknown shift opcode!");
12475 switch (VT.getSimpleVT().SimpleTy) {
12476 default: return SDValue();
12485 return getTargetVShiftNode(X86ISD::VSHLI, dl, VT, R, BaseShAmt, DAG);
12488 switch (VT.getSimpleVT().SimpleTy) {
12489 default: return SDValue();
12496 return getTargetVShiftNode(X86ISD::VSRAI, dl, VT, R, BaseShAmt, DAG);
12499 switch (VT.getSimpleVT().SimpleTy) {
12500 default: return SDValue();
12509 return getTargetVShiftNode(X86ISD::VSRLI, dl, VT, R, BaseShAmt, DAG);
12515 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
12516 if (!Subtarget->is64Bit() &&
12517 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64) ||
12518 (Subtarget->hasAVX512() && VT == MVT::v8i64)) &&
12519 Amt.getOpcode() == ISD::BITCAST &&
12520 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
12521 Amt = Amt.getOperand(0);
12522 unsigned Ratio = Amt.getValueType().getVectorNumElements() /
12523 VT.getVectorNumElements();
12524 std::vector<SDValue> Vals(Ratio);
12525 for (unsigned i = 0; i != Ratio; ++i)
12526 Vals[i] = Amt.getOperand(i);
12527 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
12528 for (unsigned j = 0; j != Ratio; ++j)
12529 if (Vals[j] != Amt.getOperand(i + j))
12532 switch (Op.getOpcode()) {
12534 llvm_unreachable("Unknown shift opcode!");
12536 return DAG.getNode(X86ISD::VSHL, dl, VT, R, Op.getOperand(1));
12538 return DAG.getNode(X86ISD::VSRL, dl, VT, R, Op.getOperand(1));
12540 return DAG.getNode(X86ISD::VSRA, dl, VT, R, Op.getOperand(1));
12547 static SDValue LowerShift(SDValue Op, const X86Subtarget* Subtarget,
12548 SelectionDAG &DAG) {
12550 EVT VT = Op.getValueType();
12552 SDValue R = Op.getOperand(0);
12553 SDValue Amt = Op.getOperand(1);
12556 if (!Subtarget->hasSSE2())
12559 V = LowerScalarImmediateShift(Op, DAG, Subtarget);
12563 V = LowerScalarVariableShift(Op, DAG, Subtarget);
12567 if (Subtarget->hasAVX512() && (VT == MVT::v16i32 || VT == MVT::v8i64))
12569 // AVX2 has VPSLLV/VPSRAV/VPSRLV.
12570 if (Subtarget->hasInt256()) {
12571 if (Op.getOpcode() == ISD::SRL &&
12572 (VT == MVT::v2i64 || VT == MVT::v4i32 ||
12573 VT == MVT::v4i64 || VT == MVT::v8i32))
12575 if (Op.getOpcode() == ISD::SHL &&
12576 (VT == MVT::v2i64 || VT == MVT::v4i32 ||
12577 VT == MVT::v4i64 || VT == MVT::v8i32))
12579 if (Op.getOpcode() == ISD::SRA && (VT == MVT::v4i32 || VT == MVT::v8i32))
12583 // Lower SHL with variable shift amount.
12584 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
12585 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, VT));
12587 Op = DAG.getNode(ISD::ADD, dl, VT, Op, DAG.getConstant(0x3f800000U, VT));
12588 Op = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, Op);
12589 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
12590 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
12592 if (VT == MVT::v16i8 && Op->getOpcode() == ISD::SHL) {
12593 assert(Subtarget->hasSSE2() && "Need SSE2 for pslli/pcmpeq.");
12596 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(5, VT));
12597 Op = DAG.getNode(ISD::BITCAST, dl, VT, Op);
12599 // Turn 'a' into a mask suitable for VSELECT
12600 SDValue VSelM = DAG.getConstant(0x80, VT);
12601 SDValue OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
12602 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
12604 SDValue CM1 = DAG.getConstant(0x0f, VT);
12605 SDValue CM2 = DAG.getConstant(0x3f, VT);
12607 // r = VSELECT(r, psllw(r & (char16)15, 4), a);
12608 SDValue M = DAG.getNode(ISD::AND, dl, VT, R, CM1);
12609 M = getTargetVShiftNode(X86ISD::VSHLI, dl, MVT::v8i16, M,
12610 DAG.getConstant(4, MVT::i32), DAG);
12611 M = DAG.getNode(ISD::BITCAST, dl, VT, M);
12612 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, M, R);
12615 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
12616 OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
12617 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
12619 // r = VSELECT(r, psllw(r & (char16)63, 2), a);
12620 M = DAG.getNode(ISD::AND, dl, VT, R, CM2);
12621 M = getTargetVShiftNode(X86ISD::VSHLI, dl, MVT::v8i16, M,
12622 DAG.getConstant(2, MVT::i32), DAG);
12623 M = DAG.getNode(ISD::BITCAST, dl, VT, M);
12624 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, M, R);
12627 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
12628 OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
12629 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
12631 // return VSELECT(r, r+r, a);
12632 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel,
12633 DAG.getNode(ISD::ADD, dl, VT, R, R), R);
12637 // Decompose 256-bit shifts into smaller 128-bit shifts.
12638 if (VT.is256BitVector()) {
12639 unsigned NumElems = VT.getVectorNumElements();
12640 MVT EltVT = VT.getVectorElementType().getSimpleVT();
12641 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
12643 // Extract the two vectors
12644 SDValue V1 = Extract128BitVector(R, 0, DAG, dl);
12645 SDValue V2 = Extract128BitVector(R, NumElems/2, DAG, dl);
12647 // Recreate the shift amount vectors
12648 SDValue Amt1, Amt2;
12649 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
12650 // Constant shift amount
12651 SmallVector<SDValue, 4> Amt1Csts;
12652 SmallVector<SDValue, 4> Amt2Csts;
12653 for (unsigned i = 0; i != NumElems/2; ++i)
12654 Amt1Csts.push_back(Amt->getOperand(i));
12655 for (unsigned i = NumElems/2; i != NumElems; ++i)
12656 Amt2Csts.push_back(Amt->getOperand(i));
12658 Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT,
12659 &Amt1Csts[0], NumElems/2);
12660 Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT,
12661 &Amt2Csts[0], NumElems/2);
12663 // Variable shift amount
12664 Amt1 = Extract128BitVector(Amt, 0, DAG, dl);
12665 Amt2 = Extract128BitVector(Amt, NumElems/2, DAG, dl);
12668 // Issue new vector shifts for the smaller types
12669 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
12670 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
12672 // Concatenate the result back
12673 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
12679 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
12680 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
12681 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
12682 // looks for this combo and may remove the "setcc" instruction if the "setcc"
12683 // has only one use.
12684 SDNode *N = Op.getNode();
12685 SDValue LHS = N->getOperand(0);
12686 SDValue RHS = N->getOperand(1);
12687 unsigned BaseOp = 0;
12690 switch (Op.getOpcode()) {
12691 default: llvm_unreachable("Unknown ovf instruction!");
12693 // A subtract of one will be selected as a INC. Note that INC doesn't
12694 // set CF, so we can't do this for UADDO.
12695 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
12697 BaseOp = X86ISD::INC;
12698 Cond = X86::COND_O;
12701 BaseOp = X86ISD::ADD;
12702 Cond = X86::COND_O;
12705 BaseOp = X86ISD::ADD;
12706 Cond = X86::COND_B;
12709 // A subtract of one will be selected as a DEC. Note that DEC doesn't
12710 // set CF, so we can't do this for USUBO.
12711 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
12713 BaseOp = X86ISD::DEC;
12714 Cond = X86::COND_O;
12717 BaseOp = X86ISD::SUB;
12718 Cond = X86::COND_O;
12721 BaseOp = X86ISD::SUB;
12722 Cond = X86::COND_B;
12725 BaseOp = X86ISD::SMUL;
12726 Cond = X86::COND_O;
12728 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
12729 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
12731 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
12734 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
12735 DAG.getConstant(X86::COND_O, MVT::i32),
12736 SDValue(Sum.getNode(), 2));
12738 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
12742 // Also sets EFLAGS.
12743 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
12744 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
12747 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
12748 DAG.getConstant(Cond, MVT::i32),
12749 SDValue(Sum.getNode(), 1));
12751 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
12754 SDValue X86TargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
12755 SelectionDAG &DAG) const {
12757 EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
12758 EVT VT = Op.getValueType();
12760 if (!Subtarget->hasSSE2() || !VT.isVector())
12763 unsigned BitsDiff = VT.getScalarType().getSizeInBits() -
12764 ExtraVT.getScalarType().getSizeInBits();
12765 SDValue ShAmt = DAG.getConstant(BitsDiff, MVT::i32);
12767 switch (VT.getSimpleVT().SimpleTy) {
12768 default: return SDValue();
12771 if (!Subtarget->hasFp256())
12773 if (!Subtarget->hasInt256()) {
12774 // needs to be split
12775 unsigned NumElems = VT.getVectorNumElements();
12777 // Extract the LHS vectors
12778 SDValue LHS = Op.getOperand(0);
12779 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
12780 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
12782 MVT EltVT = VT.getVectorElementType().getSimpleVT();
12783 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
12785 EVT ExtraEltVT = ExtraVT.getVectorElementType();
12786 unsigned ExtraNumElems = ExtraVT.getVectorNumElements();
12787 ExtraVT = EVT::getVectorVT(*DAG.getContext(), ExtraEltVT,
12789 SDValue Extra = DAG.getValueType(ExtraVT);
12791 LHS1 = DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, Extra);
12792 LHS2 = DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, Extra);
12794 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, LHS1, LHS2);
12799 // (sext (vzext x)) -> (vsext x)
12800 SDValue Op0 = Op.getOperand(0);
12801 SDValue Op00 = Op0.getOperand(0);
12803 // Hopefully, this VECTOR_SHUFFLE is just a VZEXT.
12804 if (Op0.getOpcode() == ISD::BITCAST &&
12805 Op00.getOpcode() == ISD::VECTOR_SHUFFLE)
12806 Tmp1 = LowerVectorIntExtend(Op00, Subtarget, DAG);
12807 if (Tmp1.getNode()) {
12808 SDValue Tmp1Op0 = Tmp1.getOperand(0);
12809 assert(Tmp1Op0.getOpcode() == X86ISD::VZEXT &&
12810 "This optimization is invalid without a VZEXT.");
12811 return DAG.getNode(X86ISD::VSEXT, dl, VT, Tmp1Op0.getOperand(0));
12814 // If the above didn't work, then just use Shift-Left + Shift-Right.
12815 Tmp1 = getTargetVShiftNode(X86ISD::VSHLI, dl, VT, Op0, ShAmt, DAG);
12816 return getTargetVShiftNode(X86ISD::VSRAI, dl, VT, Tmp1, ShAmt, DAG);
12821 static SDValue LowerATOMIC_FENCE(SDValue Op, const X86Subtarget *Subtarget,
12822 SelectionDAG &DAG) {
12824 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
12825 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
12826 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
12827 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
12829 // The only fence that needs an instruction is a sequentially-consistent
12830 // cross-thread fence.
12831 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
12832 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
12833 // no-sse2). There isn't any reason to disable it if the target processor
12835 if (Subtarget->hasSSE2() || Subtarget->is64Bit())
12836 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
12838 SDValue Chain = Op.getOperand(0);
12839 SDValue Zero = DAG.getConstant(0, MVT::i32);
12841 DAG.getRegister(X86::ESP, MVT::i32), // Base
12842 DAG.getTargetConstant(1, MVT::i8), // Scale
12843 DAG.getRegister(0, MVT::i32), // Index
12844 DAG.getTargetConstant(0, MVT::i32), // Disp
12845 DAG.getRegister(0, MVT::i32), // Segment.
12849 SDNode *Res = DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops);
12850 return SDValue(Res, 0);
12853 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
12854 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
12857 static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget *Subtarget,
12858 SelectionDAG &DAG) {
12859 EVT T = Op.getValueType();
12863 switch(T.getSimpleVT().SimpleTy) {
12864 default: llvm_unreachable("Invalid value type!");
12865 case MVT::i8: Reg = X86::AL; size = 1; break;
12866 case MVT::i16: Reg = X86::AX; size = 2; break;
12867 case MVT::i32: Reg = X86::EAX; size = 4; break;
12869 assert(Subtarget->is64Bit() && "Node not type legal!");
12870 Reg = X86::RAX; size = 8;
12873 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
12874 Op.getOperand(2), SDValue());
12875 SDValue Ops[] = { cpIn.getValue(0),
12878 DAG.getTargetConstant(size, MVT::i8),
12879 cpIn.getValue(1) };
12880 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
12881 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
12882 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
12883 Ops, array_lengthof(Ops), T, MMO);
12885 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
12889 static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget *Subtarget,
12890 SelectionDAG &DAG) {
12891 assert(Subtarget->is64Bit() && "Result not type legalized?");
12892 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
12893 SDValue TheChain = Op.getOperand(0);
12895 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
12896 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
12897 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
12899 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
12900 DAG.getConstant(32, MVT::i8));
12902 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
12905 return DAG.getMergeValues(Ops, array_lengthof(Ops), dl);
12908 static SDValue LowerBITCAST(SDValue Op, const X86Subtarget *Subtarget,
12909 SelectionDAG &DAG) {
12910 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
12911 MVT DstVT = Op.getSimpleValueType();
12912 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
12913 Subtarget->hasMMX() && "Unexpected custom BITCAST");
12914 assert((DstVT == MVT::i64 ||
12915 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
12916 "Unexpected custom BITCAST");
12917 // i64 <=> MMX conversions are Legal.
12918 if (SrcVT==MVT::i64 && DstVT.isVector())
12920 if (DstVT==MVT::i64 && SrcVT.isVector())
12922 // MMX <=> MMX conversions are Legal.
12923 if (SrcVT.isVector() && DstVT.isVector())
12925 // All other conversions need to be expanded.
12929 static SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
12930 SDNode *Node = Op.getNode();
12932 EVT T = Node->getValueType(0);
12933 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
12934 DAG.getConstant(0, T), Node->getOperand(2));
12935 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
12936 cast<AtomicSDNode>(Node)->getMemoryVT(),
12937 Node->getOperand(0),
12938 Node->getOperand(1), negOp,
12939 cast<AtomicSDNode>(Node)->getSrcValue(),
12940 cast<AtomicSDNode>(Node)->getAlignment(),
12941 cast<AtomicSDNode>(Node)->getOrdering(),
12942 cast<AtomicSDNode>(Node)->getSynchScope());
12945 static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) {
12946 SDNode *Node = Op.getNode();
12948 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
12950 // Convert seq_cst store -> xchg
12951 // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b)
12952 // FIXME: On 32-bit, store -> fist or movq would be more efficient
12953 // (The only way to get a 16-byte store is cmpxchg16b)
12954 // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment.
12955 if (cast<AtomicSDNode>(Node)->getOrdering() == SequentiallyConsistent ||
12956 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
12957 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
12958 cast<AtomicSDNode>(Node)->getMemoryVT(),
12959 Node->getOperand(0),
12960 Node->getOperand(1), Node->getOperand(2),
12961 cast<AtomicSDNode>(Node)->getMemOperand(),
12962 cast<AtomicSDNode>(Node)->getOrdering(),
12963 cast<AtomicSDNode>(Node)->getSynchScope());
12964 return Swap.getValue(1);
12966 // Other atomic stores have a simple pattern.
12970 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
12971 EVT VT = Op.getNode()->getValueType(0);
12973 // Let legalize expand this if it isn't a legal type yet.
12974 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
12977 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
12980 bool ExtraOp = false;
12981 switch (Op.getOpcode()) {
12982 default: llvm_unreachable("Invalid code");
12983 case ISD::ADDC: Opc = X86ISD::ADD; break;
12984 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
12985 case ISD::SUBC: Opc = X86ISD::SUB; break;
12986 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
12990 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
12992 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
12993 Op.getOperand(1), Op.getOperand(2));
12996 static SDValue LowerFSINCOS(SDValue Op, const X86Subtarget *Subtarget,
12997 SelectionDAG &DAG) {
12998 assert(Subtarget->isTargetDarwin() && Subtarget->is64Bit());
13000 // For MacOSX, we want to call an alternative entry point: __sincos_stret,
13001 // which returns the values as { float, float } (in XMM0) or
13002 // { double, double } (which is returned in XMM0, XMM1).
13004 SDValue Arg = Op.getOperand(0);
13005 EVT ArgVT = Arg.getValueType();
13006 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
13008 TargetLowering::ArgListTy Args;
13009 TargetLowering::ArgListEntry Entry;
13013 Entry.isSExt = false;
13014 Entry.isZExt = false;
13015 Args.push_back(Entry);
13017 bool isF64 = ArgVT == MVT::f64;
13018 // Only optimize x86_64 for now. i386 is a bit messy. For f32,
13019 // the small struct {f32, f32} is returned in (eax, edx). For f64,
13020 // the results are returned via SRet in memory.
13021 const char *LibcallName = isF64 ? "__sincos_stret" : "__sincosf_stret";
13022 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13023 SDValue Callee = DAG.getExternalSymbol(LibcallName, TLI.getPointerTy());
13025 Type *RetTy = isF64
13026 ? (Type*)StructType::get(ArgTy, ArgTy, NULL)
13027 : (Type*)VectorType::get(ArgTy, 4);
13029 CallLoweringInfo CLI(DAG.getEntryNode(), RetTy,
13030 false, false, false, false, 0,
13031 CallingConv::C, /*isTaillCall=*/false,
13032 /*doesNotRet=*/false, /*isReturnValueUsed*/true,
13033 Callee, Args, DAG, dl);
13034 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
13037 // Returned in xmm0 and xmm1.
13038 return CallResult.first;
13040 // Returned in bits 0:31 and 32:64 xmm0.
13041 SDValue SinVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
13042 CallResult.first, DAG.getIntPtrConstant(0));
13043 SDValue CosVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
13044 CallResult.first, DAG.getIntPtrConstant(1));
13045 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
13046 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys, SinVal, CosVal);
13049 /// LowerOperation - Provide custom lowering hooks for some operations.
13051 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
13052 switch (Op.getOpcode()) {
13053 default: llvm_unreachable("Should not custom lower this!");
13054 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op,DAG);
13055 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, Subtarget, DAG);
13056 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op, Subtarget, DAG);
13057 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
13058 case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG);
13059 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
13060 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
13061 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
13062 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
13063 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
13064 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op,Subtarget,DAG);
13065 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, Subtarget,DAG);
13066 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
13067 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
13068 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
13069 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
13070 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
13071 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
13072 case ISD::SHL_PARTS:
13073 case ISD::SRA_PARTS:
13074 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
13075 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
13076 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
13077 case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG);
13078 case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, Subtarget, DAG);
13079 case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, Subtarget, DAG);
13080 case ISD::ANY_EXTEND: return LowerANY_EXTEND(Op, Subtarget, DAG);
13081 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
13082 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
13083 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
13084 case ISD::FABS: return LowerFABS(Op, DAG);
13085 case ISD::FNEG: return LowerFNEG(Op, DAG);
13086 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
13087 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
13088 case ISD::SETCC: return LowerSETCC(Op, DAG);
13089 case ISD::SELECT: return LowerSELECT(Op, DAG);
13090 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
13091 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
13092 case ISD::VASTART: return LowerVASTART(Op, DAG);
13093 case ISD::VAARG: return LowerVAARG(Op, DAG);
13094 case ISD::VACOPY: return LowerVACOPY(Op, Subtarget, DAG);
13095 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
13096 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, DAG);
13097 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
13098 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
13099 case ISD::FRAME_TO_ARGS_OFFSET:
13100 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
13101 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
13102 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
13103 case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
13104 case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
13105 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
13106 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
13107 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
13108 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
13109 case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, DAG);
13110 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
13111 case ISD::MUL: return LowerMUL(Op, Subtarget, DAG);
13114 case ISD::SHL: return LowerShift(Op, Subtarget, DAG);
13120 case ISD::UMULO: return LowerXALUO(Op, DAG);
13121 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, Subtarget,DAG);
13122 case ISD::BITCAST: return LowerBITCAST(Op, Subtarget, DAG);
13126 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
13127 case ISD::ADD: return LowerADD(Op, DAG);
13128 case ISD::SUB: return LowerSUB(Op, DAG);
13129 case ISD::SDIV: return LowerSDIV(Op, DAG);
13130 case ISD::FSINCOS: return LowerFSINCOS(Op, Subtarget, DAG);
13134 static void ReplaceATOMIC_LOAD(SDNode *Node,
13135 SmallVectorImpl<SDValue> &Results,
13136 SelectionDAG &DAG) {
13138 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
13140 // Convert wide load -> cmpxchg8b/cmpxchg16b
13141 // FIXME: On 32-bit, load -> fild or movq would be more efficient
13142 // (The only way to get a 16-byte load is cmpxchg16b)
13143 // FIXME: 16-byte ATOMIC_CMP_SWAP isn't actually hooked up at the moment.
13144 SDValue Zero = DAG.getConstant(0, VT);
13145 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_CMP_SWAP, dl, VT,
13146 Node->getOperand(0),
13147 Node->getOperand(1), Zero, Zero,
13148 cast<AtomicSDNode>(Node)->getMemOperand(),
13149 cast<AtomicSDNode>(Node)->getOrdering(),
13150 cast<AtomicSDNode>(Node)->getSynchScope());
13151 Results.push_back(Swap.getValue(0));
13152 Results.push_back(Swap.getValue(1));
13156 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
13157 SelectionDAG &DAG, unsigned NewOp) {
13159 assert (Node->getValueType(0) == MVT::i64 &&
13160 "Only know how to expand i64 atomics");
13162 SDValue Chain = Node->getOperand(0);
13163 SDValue In1 = Node->getOperand(1);
13164 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
13165 Node->getOperand(2), DAG.getIntPtrConstant(0));
13166 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
13167 Node->getOperand(2), DAG.getIntPtrConstant(1));
13168 SDValue Ops[] = { Chain, In1, In2L, In2H };
13169 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
13171 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, array_lengthof(Ops), MVT::i64,
13172 cast<MemSDNode>(Node)->getMemOperand());
13173 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
13174 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
13175 Results.push_back(Result.getValue(2));
13178 /// ReplaceNodeResults - Replace a node with an illegal result type
13179 /// with a new node built out of custom code.
13180 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
13181 SmallVectorImpl<SDValue>&Results,
13182 SelectionDAG &DAG) const {
13184 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13185 switch (N->getOpcode()) {
13187 llvm_unreachable("Do not know how to custom type legalize this operation!");
13188 case ISD::SIGN_EXTEND_INREG:
13193 // We don't want to expand or promote these.
13195 case ISD::FP_TO_SINT:
13196 case ISD::FP_TO_UINT: {
13197 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
13199 if (!IsSigned && !isIntegerTypeFTOL(SDValue(N, 0).getValueType()))
13202 std::pair<SDValue,SDValue> Vals =
13203 FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true);
13204 SDValue FIST = Vals.first, StackSlot = Vals.second;
13205 if (FIST.getNode() != 0) {
13206 EVT VT = N->getValueType(0);
13207 // Return a load from the stack slot.
13208 if (StackSlot.getNode() != 0)
13209 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
13210 MachinePointerInfo(),
13211 false, false, false, 0));
13213 Results.push_back(FIST);
13217 case ISD::UINT_TO_FP: {
13218 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
13219 if (N->getOperand(0).getValueType() != MVT::v2i32 ||
13220 N->getValueType(0) != MVT::v2f32)
13222 SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64,
13224 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
13226 SDValue VBias = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2f64, Bias, Bias);
13227 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn,
13228 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, VBias));
13229 Or = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or);
13230 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias);
13231 Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub));
13234 case ISD::FP_ROUND: {
13235 if (!TLI.isTypeLegal(N->getOperand(0).getValueType()))
13237 SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0));
13238 Results.push_back(V);
13241 case ISD::READCYCLECOUNTER: {
13242 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
13243 SDValue TheChain = N->getOperand(0);
13244 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
13245 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
13247 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
13249 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
13250 SDValue Ops[] = { eax, edx };
13251 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops,
13252 array_lengthof(Ops)));
13253 Results.push_back(edx.getValue(1));
13256 case ISD::ATOMIC_CMP_SWAP: {
13257 EVT T = N->getValueType(0);
13258 assert((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair");
13259 bool Regs64bit = T == MVT::i128;
13260 EVT HalfT = Regs64bit ? MVT::i64 : MVT::i32;
13261 SDValue cpInL, cpInH;
13262 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
13263 DAG.getConstant(0, HalfT));
13264 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
13265 DAG.getConstant(1, HalfT));
13266 cpInL = DAG.getCopyToReg(N->getOperand(0), dl,
13267 Regs64bit ? X86::RAX : X86::EAX,
13269 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl,
13270 Regs64bit ? X86::RDX : X86::EDX,
13271 cpInH, cpInL.getValue(1));
13272 SDValue swapInL, swapInH;
13273 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
13274 DAG.getConstant(0, HalfT));
13275 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
13276 DAG.getConstant(1, HalfT));
13277 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl,
13278 Regs64bit ? X86::RBX : X86::EBX,
13279 swapInL, cpInH.getValue(1));
13280 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl,
13281 Regs64bit ? X86::RCX : X86::ECX,
13282 swapInH, swapInL.getValue(1));
13283 SDValue Ops[] = { swapInH.getValue(0),
13285 swapInH.getValue(1) };
13286 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
13287 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
13288 unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_DAG :
13289 X86ISD::LCMPXCHG8_DAG;
13290 SDValue Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys,
13291 Ops, array_lengthof(Ops), T, MMO);
13292 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl,
13293 Regs64bit ? X86::RAX : X86::EAX,
13294 HalfT, Result.getValue(1));
13295 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl,
13296 Regs64bit ? X86::RDX : X86::EDX,
13297 HalfT, cpOutL.getValue(2));
13298 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
13299 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF, 2));
13300 Results.push_back(cpOutH.getValue(1));
13303 case ISD::ATOMIC_LOAD_ADD:
13304 case ISD::ATOMIC_LOAD_AND:
13305 case ISD::ATOMIC_LOAD_NAND:
13306 case ISD::ATOMIC_LOAD_OR:
13307 case ISD::ATOMIC_LOAD_SUB:
13308 case ISD::ATOMIC_LOAD_XOR:
13309 case ISD::ATOMIC_LOAD_MAX:
13310 case ISD::ATOMIC_LOAD_MIN:
13311 case ISD::ATOMIC_LOAD_UMAX:
13312 case ISD::ATOMIC_LOAD_UMIN:
13313 case ISD::ATOMIC_SWAP: {
13315 switch (N->getOpcode()) {
13316 default: llvm_unreachable("Unexpected opcode");
13317 case ISD::ATOMIC_LOAD_ADD:
13318 Opc = X86ISD::ATOMADD64_DAG;
13320 case ISD::ATOMIC_LOAD_AND:
13321 Opc = X86ISD::ATOMAND64_DAG;
13323 case ISD::ATOMIC_LOAD_NAND:
13324 Opc = X86ISD::ATOMNAND64_DAG;
13326 case ISD::ATOMIC_LOAD_OR:
13327 Opc = X86ISD::ATOMOR64_DAG;
13329 case ISD::ATOMIC_LOAD_SUB:
13330 Opc = X86ISD::ATOMSUB64_DAG;
13332 case ISD::ATOMIC_LOAD_XOR:
13333 Opc = X86ISD::ATOMXOR64_DAG;
13335 case ISD::ATOMIC_LOAD_MAX:
13336 Opc = X86ISD::ATOMMAX64_DAG;
13338 case ISD::ATOMIC_LOAD_MIN:
13339 Opc = X86ISD::ATOMMIN64_DAG;
13341 case ISD::ATOMIC_LOAD_UMAX:
13342 Opc = X86ISD::ATOMUMAX64_DAG;
13344 case ISD::ATOMIC_LOAD_UMIN:
13345 Opc = X86ISD::ATOMUMIN64_DAG;
13347 case ISD::ATOMIC_SWAP:
13348 Opc = X86ISD::ATOMSWAP64_DAG;
13351 ReplaceATOMIC_BINARY_64(N, Results, DAG, Opc);
13354 case ISD::ATOMIC_LOAD:
13355 ReplaceATOMIC_LOAD(N, Results, DAG);
13359 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
13361 default: return NULL;
13362 case X86ISD::BSF: return "X86ISD::BSF";
13363 case X86ISD::BSR: return "X86ISD::BSR";
13364 case X86ISD::SHLD: return "X86ISD::SHLD";
13365 case X86ISD::SHRD: return "X86ISD::SHRD";
13366 case X86ISD::FAND: return "X86ISD::FAND";
13367 case X86ISD::FANDN: return "X86ISD::FANDN";
13368 case X86ISD::FOR: return "X86ISD::FOR";
13369 case X86ISD::FXOR: return "X86ISD::FXOR";
13370 case X86ISD::FSRL: return "X86ISD::FSRL";
13371 case X86ISD::FILD: return "X86ISD::FILD";
13372 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
13373 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
13374 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
13375 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
13376 case X86ISD::FLD: return "X86ISD::FLD";
13377 case X86ISD::FST: return "X86ISD::FST";
13378 case X86ISD::CALL: return "X86ISD::CALL";
13379 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
13380 case X86ISD::BT: return "X86ISD::BT";
13381 case X86ISD::CMP: return "X86ISD::CMP";
13382 case X86ISD::COMI: return "X86ISD::COMI";
13383 case X86ISD::UCOMI: return "X86ISD::UCOMI";
13384 case X86ISD::CMPM: return "X86ISD::CMPM";
13385 case X86ISD::CMPMU: return "X86ISD::CMPMU";
13386 case X86ISD::SETCC: return "X86ISD::SETCC";
13387 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
13388 case X86ISD::FSETCCsd: return "X86ISD::FSETCCsd";
13389 case X86ISD::FSETCCss: return "X86ISD::FSETCCss";
13390 case X86ISD::CMOV: return "X86ISD::CMOV";
13391 case X86ISD::BRCOND: return "X86ISD::BRCOND";
13392 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
13393 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
13394 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
13395 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
13396 case X86ISD::Wrapper: return "X86ISD::Wrapper";
13397 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
13398 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
13399 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
13400 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
13401 case X86ISD::PINSRB: return "X86ISD::PINSRB";
13402 case X86ISD::PINSRW: return "X86ISD::PINSRW";
13403 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
13404 case X86ISD::ANDNP: return "X86ISD::ANDNP";
13405 case X86ISD::PSIGN: return "X86ISD::PSIGN";
13406 case X86ISD::BLENDV: return "X86ISD::BLENDV";
13407 case X86ISD::BLENDI: return "X86ISD::BLENDI";
13408 case X86ISD::SUBUS: return "X86ISD::SUBUS";
13409 case X86ISD::HADD: return "X86ISD::HADD";
13410 case X86ISD::HSUB: return "X86ISD::HSUB";
13411 case X86ISD::FHADD: return "X86ISD::FHADD";
13412 case X86ISD::FHSUB: return "X86ISD::FHSUB";
13413 case X86ISD::UMAX: return "X86ISD::UMAX";
13414 case X86ISD::UMIN: return "X86ISD::UMIN";
13415 case X86ISD::SMAX: return "X86ISD::SMAX";
13416 case X86ISD::SMIN: return "X86ISD::SMIN";
13417 case X86ISD::FMAX: return "X86ISD::FMAX";
13418 case X86ISD::FMIN: return "X86ISD::FMIN";
13419 case X86ISD::FMAXC: return "X86ISD::FMAXC";
13420 case X86ISD::FMINC: return "X86ISD::FMINC";
13421 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
13422 case X86ISD::FRCP: return "X86ISD::FRCP";
13423 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
13424 case X86ISD::TLSBASEADDR: return "X86ISD::TLSBASEADDR";
13425 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
13426 case X86ISD::EH_SJLJ_SETJMP: return "X86ISD::EH_SJLJ_SETJMP";
13427 case X86ISD::EH_SJLJ_LONGJMP: return "X86ISD::EH_SJLJ_LONGJMP";
13428 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
13429 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
13430 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
13431 case X86ISD::FNSTSW16r: return "X86ISD::FNSTSW16r";
13432 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
13433 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
13434 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
13435 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
13436 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
13437 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
13438 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
13439 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
13440 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
13441 case X86ISD::VSEXT_MOVL: return "X86ISD::VSEXT_MOVL";
13442 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
13443 case X86ISD::VZEXT: return "X86ISD::VZEXT";
13444 case X86ISD::VSEXT: return "X86ISD::VSEXT";
13445 case X86ISD::VTRUNC: return "X86ISD::VTRUNC";
13446 case X86ISD::VTRUNCM: return "X86ISD::VTRUNCM";
13447 case X86ISD::VFPEXT: return "X86ISD::VFPEXT";
13448 case X86ISD::VFPROUND: return "X86ISD::VFPROUND";
13449 case X86ISD::VSHLDQ: return "X86ISD::VSHLDQ";
13450 case X86ISD::VSRLDQ: return "X86ISD::VSRLDQ";
13451 case X86ISD::VSHL: return "X86ISD::VSHL";
13452 case X86ISD::VSRL: return "X86ISD::VSRL";
13453 case X86ISD::VSRA: return "X86ISD::VSRA";
13454 case X86ISD::VSHLI: return "X86ISD::VSHLI";
13455 case X86ISD::VSRLI: return "X86ISD::VSRLI";
13456 case X86ISD::VSRAI: return "X86ISD::VSRAI";
13457 case X86ISD::CMPP: return "X86ISD::CMPP";
13458 case X86ISD::PCMPEQ: return "X86ISD::PCMPEQ";
13459 case X86ISD::PCMPGT: return "X86ISD::PCMPGT";
13460 case X86ISD::PCMPEQM: return "X86ISD::PCMPEQM";
13461 case X86ISD::PCMPGTM: return "X86ISD::PCMPGTM";
13462 case X86ISD::ADD: return "X86ISD::ADD";
13463 case X86ISD::SUB: return "X86ISD::SUB";
13464 case X86ISD::ADC: return "X86ISD::ADC";
13465 case X86ISD::SBB: return "X86ISD::SBB";
13466 case X86ISD::SMUL: return "X86ISD::SMUL";
13467 case X86ISD::UMUL: return "X86ISD::UMUL";
13468 case X86ISD::INC: return "X86ISD::INC";
13469 case X86ISD::DEC: return "X86ISD::DEC";
13470 case X86ISD::OR: return "X86ISD::OR";
13471 case X86ISD::XOR: return "X86ISD::XOR";
13472 case X86ISD::AND: return "X86ISD::AND";
13473 case X86ISD::BLSI: return "X86ISD::BLSI";
13474 case X86ISD::BLSMSK: return "X86ISD::BLSMSK";
13475 case X86ISD::BLSR: return "X86ISD::BLSR";
13476 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
13477 case X86ISD::PTEST: return "X86ISD::PTEST";
13478 case X86ISD::TESTP: return "X86ISD::TESTP";
13479 case X86ISD::TESTM: return "X86ISD::TESTM";
13480 case X86ISD::KORTEST: return "X86ISD::KORTEST";
13481 case X86ISD::KTEST: return "X86ISD::KTEST";
13482 case X86ISD::PALIGNR: return "X86ISD::PALIGNR";
13483 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
13484 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
13485 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
13486 case X86ISD::SHUFP: return "X86ISD::SHUFP";
13487 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
13488 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
13489 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
13490 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
13491 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
13492 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
13493 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
13494 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
13495 case X86ISD::MOVSD: return "X86ISD::MOVSD";
13496 case X86ISD::MOVSS: return "X86ISD::MOVSS";
13497 case X86ISD::UNPCKL: return "X86ISD::UNPCKL";
13498 case X86ISD::UNPCKH: return "X86ISD::UNPCKH";
13499 case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
13500 case X86ISD::VBROADCASTM: return "X86ISD::VBROADCASTM";
13501 case X86ISD::VPERMILP: return "X86ISD::VPERMILP";
13502 case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128";
13503 case X86ISD::VPERMV: return "X86ISD::VPERMV";
13504 case X86ISD::VPERMV3: return "X86ISD::VPERMV3";
13505 case X86ISD::VPERMI: return "X86ISD::VPERMI";
13506 case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
13507 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
13508 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
13509 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
13510 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
13511 case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA";
13512 case X86ISD::WIN_FTOL: return "X86ISD::WIN_FTOL";
13513 case X86ISD::SAHF: return "X86ISD::SAHF";
13514 case X86ISD::RDRAND: return "X86ISD::RDRAND";
13515 case X86ISD::RDSEED: return "X86ISD::RDSEED";
13516 case X86ISD::FMADD: return "X86ISD::FMADD";
13517 case X86ISD::FMSUB: return "X86ISD::FMSUB";
13518 case X86ISD::FNMADD: return "X86ISD::FNMADD";
13519 case X86ISD::FNMSUB: return "X86ISD::FNMSUB";
13520 case X86ISD::FMADDSUB: return "X86ISD::FMADDSUB";
13521 case X86ISD::FMSUBADD: return "X86ISD::FMSUBADD";
13522 case X86ISD::PCMPESTRI: return "X86ISD::PCMPESTRI";
13523 case X86ISD::PCMPISTRI: return "X86ISD::PCMPISTRI";
13524 case X86ISD::XTEST: return "X86ISD::XTEST";
13528 // isLegalAddressingMode - Return true if the addressing mode represented
13529 // by AM is legal for this target, for a load/store of the specified type.
13530 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
13532 // X86 supports extremely general addressing modes.
13533 CodeModel::Model M = getTargetMachine().getCodeModel();
13534 Reloc::Model R = getTargetMachine().getRelocationModel();
13536 // X86 allows a sign-extended 32-bit immediate field as a displacement.
13537 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
13542 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
13544 // If a reference to this global requires an extra load, we can't fold it.
13545 if (isGlobalStubReference(GVFlags))
13548 // If BaseGV requires a register for the PIC base, we cannot also have a
13549 // BaseReg specified.
13550 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
13553 // If lower 4G is not available, then we must use rip-relative addressing.
13554 if ((M != CodeModel::Small || R != Reloc::Static) &&
13555 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
13559 switch (AM.Scale) {
13565 // These scales always work.
13570 // These scales are formed with basereg+scalereg. Only accept if there is
13575 default: // Other stuff never works.
13582 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
13583 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
13585 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
13586 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
13587 return NumBits1 > NumBits2;
13590 bool X86TargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
13591 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
13594 if (!isTypeLegal(EVT::getEVT(Ty1)))
13597 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
13599 // Assuming the caller doesn't have a zeroext or signext return parameter,
13600 // truncation all the way down to i1 is valid.
13604 bool X86TargetLowering::isLegalICmpImmediate(int64_t Imm) const {
13605 return isInt<32>(Imm);
13608 bool X86TargetLowering::isLegalAddImmediate(int64_t Imm) const {
13609 // Can also use sub to handle negated immediates.
13610 return isInt<32>(Imm);
13613 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
13614 if (!VT1.isInteger() || !VT2.isInteger())
13616 unsigned NumBits1 = VT1.getSizeInBits();
13617 unsigned NumBits2 = VT2.getSizeInBits();
13618 return NumBits1 > NumBits2;
13621 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
13622 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
13623 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
13626 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
13627 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
13628 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
13631 bool X86TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
13632 EVT VT1 = Val.getValueType();
13633 if (isZExtFree(VT1, VT2))
13636 if (Val.getOpcode() != ISD::LOAD)
13639 if (!VT1.isSimple() || !VT1.isInteger() ||
13640 !VT2.isSimple() || !VT2.isInteger())
13643 switch (VT1.getSimpleVT().SimpleTy) {
13648 // X86 has 8, 16, and 32-bit zero-extending loads.
13656 X86TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
13657 if (!(Subtarget->hasFMA() || Subtarget->hasFMA4()))
13660 VT = VT.getScalarType();
13662 if (!VT.isSimple())
13665 switch (VT.getSimpleVT().SimpleTy) {
13676 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
13677 // i16 instructions are longer (0x66 prefix) and potentially slower.
13678 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
13681 /// isShuffleMaskLegal - Targets can use this to indicate that they only
13682 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
13683 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
13684 /// are assumed to be legal.
13686 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
13688 if (!VT.isSimple())
13691 MVT SVT = VT.getSimpleVT();
13693 // Very little shuffling can be done for 64-bit vectors right now.
13694 if (VT.getSizeInBits() == 64)
13697 // FIXME: pshufb, blends, shifts.
13698 return (SVT.getVectorNumElements() == 2 ||
13699 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
13700 isMOVLMask(M, SVT) ||
13701 isSHUFPMask(M, SVT) ||
13702 isPSHUFDMask(M, SVT) ||
13703 isPSHUFHWMask(M, SVT, Subtarget->hasInt256()) ||
13704 isPSHUFLWMask(M, SVT, Subtarget->hasInt256()) ||
13705 isPALIGNRMask(M, SVT, Subtarget) ||
13706 isUNPCKLMask(M, SVT, Subtarget->hasInt256()) ||
13707 isUNPCKHMask(M, SVT, Subtarget->hasInt256()) ||
13708 isUNPCKL_v_undef_Mask(M, SVT, Subtarget->hasInt256()) ||
13709 isUNPCKH_v_undef_Mask(M, SVT, Subtarget->hasInt256()));
13713 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
13715 if (!VT.isSimple())
13718 MVT SVT = VT.getSimpleVT();
13719 unsigned NumElts = SVT.getVectorNumElements();
13720 // FIXME: This collection of masks seems suspect.
13723 if (NumElts == 4 && SVT.is128BitVector()) {
13724 return (isMOVLMask(Mask, SVT) ||
13725 isCommutedMOVLMask(Mask, SVT, true) ||
13726 isSHUFPMask(Mask, SVT) ||
13727 isSHUFPMask(Mask, SVT, /* Commuted */ true));
13732 //===----------------------------------------------------------------------===//
13733 // X86 Scheduler Hooks
13734 //===----------------------------------------------------------------------===//
13736 /// Utility function to emit xbegin specifying the start of an RTM region.
13737 static MachineBasicBlock *EmitXBegin(MachineInstr *MI, MachineBasicBlock *MBB,
13738 const TargetInstrInfo *TII) {
13739 DebugLoc DL = MI->getDebugLoc();
13741 const BasicBlock *BB = MBB->getBasicBlock();
13742 MachineFunction::iterator I = MBB;
13745 // For the v = xbegin(), we generate
13756 MachineBasicBlock *thisMBB = MBB;
13757 MachineFunction *MF = MBB->getParent();
13758 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
13759 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
13760 MF->insert(I, mainMBB);
13761 MF->insert(I, sinkMBB);
13763 // Transfer the remainder of BB and its successor edges to sinkMBB.
13764 sinkMBB->splice(sinkMBB->begin(), MBB,
13765 llvm::next(MachineBasicBlock::iterator(MI)), MBB->end());
13766 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
13770 // # fallthrough to mainMBB
13771 // # abortion to sinkMBB
13772 BuildMI(thisMBB, DL, TII->get(X86::XBEGIN_4)).addMBB(sinkMBB);
13773 thisMBB->addSuccessor(mainMBB);
13774 thisMBB->addSuccessor(sinkMBB);
13778 BuildMI(mainMBB, DL, TII->get(X86::MOV32ri), X86::EAX).addImm(-1);
13779 mainMBB->addSuccessor(sinkMBB);
13782 // EAX is live into the sinkMBB
13783 sinkMBB->addLiveIn(X86::EAX);
13784 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
13785 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
13788 MI->eraseFromParent();
13792 // Get CMPXCHG opcode for the specified data type.
13793 static unsigned getCmpXChgOpcode(EVT VT) {
13794 switch (VT.getSimpleVT().SimpleTy) {
13795 case MVT::i8: return X86::LCMPXCHG8;
13796 case MVT::i16: return X86::LCMPXCHG16;
13797 case MVT::i32: return X86::LCMPXCHG32;
13798 case MVT::i64: return X86::LCMPXCHG64;
13802 llvm_unreachable("Invalid operand size!");
13805 // Get LOAD opcode for the specified data type.
13806 static unsigned getLoadOpcode(EVT VT) {
13807 switch (VT.getSimpleVT().SimpleTy) {
13808 case MVT::i8: return X86::MOV8rm;
13809 case MVT::i16: return X86::MOV16rm;
13810 case MVT::i32: return X86::MOV32rm;
13811 case MVT::i64: return X86::MOV64rm;
13815 llvm_unreachable("Invalid operand size!");
13818 // Get opcode of the non-atomic one from the specified atomic instruction.
13819 static unsigned getNonAtomicOpcode(unsigned Opc) {
13821 case X86::ATOMAND8: return X86::AND8rr;
13822 case X86::ATOMAND16: return X86::AND16rr;
13823 case X86::ATOMAND32: return X86::AND32rr;
13824 case X86::ATOMAND64: return X86::AND64rr;
13825 case X86::ATOMOR8: return X86::OR8rr;
13826 case X86::ATOMOR16: return X86::OR16rr;
13827 case X86::ATOMOR32: return X86::OR32rr;
13828 case X86::ATOMOR64: return X86::OR64rr;
13829 case X86::ATOMXOR8: return X86::XOR8rr;
13830 case X86::ATOMXOR16: return X86::XOR16rr;
13831 case X86::ATOMXOR32: return X86::XOR32rr;
13832 case X86::ATOMXOR64: return X86::XOR64rr;
13834 llvm_unreachable("Unhandled atomic-load-op opcode!");
13837 // Get opcode of the non-atomic one from the specified atomic instruction with
13839 static unsigned getNonAtomicOpcodeWithExtraOpc(unsigned Opc,
13840 unsigned &ExtraOpc) {
13842 case X86::ATOMNAND8: ExtraOpc = X86::NOT8r; return X86::AND8rr;
13843 case X86::ATOMNAND16: ExtraOpc = X86::NOT16r; return X86::AND16rr;
13844 case X86::ATOMNAND32: ExtraOpc = X86::NOT32r; return X86::AND32rr;
13845 case X86::ATOMNAND64: ExtraOpc = X86::NOT64r; return X86::AND64rr;
13846 case X86::ATOMMAX8: ExtraOpc = X86::CMP8rr; return X86::CMOVL32rr;
13847 case X86::ATOMMAX16: ExtraOpc = X86::CMP16rr; return X86::CMOVL16rr;
13848 case X86::ATOMMAX32: ExtraOpc = X86::CMP32rr; return X86::CMOVL32rr;
13849 case X86::ATOMMAX64: ExtraOpc = X86::CMP64rr; return X86::CMOVL64rr;
13850 case X86::ATOMMIN8: ExtraOpc = X86::CMP8rr; return X86::CMOVG32rr;
13851 case X86::ATOMMIN16: ExtraOpc = X86::CMP16rr; return X86::CMOVG16rr;
13852 case X86::ATOMMIN32: ExtraOpc = X86::CMP32rr; return X86::CMOVG32rr;
13853 case X86::ATOMMIN64: ExtraOpc = X86::CMP64rr; return X86::CMOVG64rr;
13854 case X86::ATOMUMAX8: ExtraOpc = X86::CMP8rr; return X86::CMOVB32rr;
13855 case X86::ATOMUMAX16: ExtraOpc = X86::CMP16rr; return X86::CMOVB16rr;
13856 case X86::ATOMUMAX32: ExtraOpc = X86::CMP32rr; return X86::CMOVB32rr;
13857 case X86::ATOMUMAX64: ExtraOpc = X86::CMP64rr; return X86::CMOVB64rr;
13858 case X86::ATOMUMIN8: ExtraOpc = X86::CMP8rr; return X86::CMOVA32rr;
13859 case X86::ATOMUMIN16: ExtraOpc = X86::CMP16rr; return X86::CMOVA16rr;
13860 case X86::ATOMUMIN32: ExtraOpc = X86::CMP32rr; return X86::CMOVA32rr;
13861 case X86::ATOMUMIN64: ExtraOpc = X86::CMP64rr; return X86::CMOVA64rr;
13863 llvm_unreachable("Unhandled atomic-load-op opcode!");
13866 // Get opcode of the non-atomic one from the specified atomic instruction for
13867 // 64-bit data type on 32-bit target.
13868 static unsigned getNonAtomic6432Opcode(unsigned Opc, unsigned &HiOpc) {
13870 case X86::ATOMAND6432: HiOpc = X86::AND32rr; return X86::AND32rr;
13871 case X86::ATOMOR6432: HiOpc = X86::OR32rr; return X86::OR32rr;
13872 case X86::ATOMXOR6432: HiOpc = X86::XOR32rr; return X86::XOR32rr;
13873 case X86::ATOMADD6432: HiOpc = X86::ADC32rr; return X86::ADD32rr;
13874 case X86::ATOMSUB6432: HiOpc = X86::SBB32rr; return X86::SUB32rr;
13875 case X86::ATOMSWAP6432: HiOpc = X86::MOV32rr; return X86::MOV32rr;
13876 case X86::ATOMMAX6432: HiOpc = X86::SETLr; return X86::SETLr;
13877 case X86::ATOMMIN6432: HiOpc = X86::SETGr; return X86::SETGr;
13878 case X86::ATOMUMAX6432: HiOpc = X86::SETBr; return X86::SETBr;
13879 case X86::ATOMUMIN6432: HiOpc = X86::SETAr; return X86::SETAr;
13881 llvm_unreachable("Unhandled atomic-load-op opcode!");
13884 // Get opcode of the non-atomic one from the specified atomic instruction for
13885 // 64-bit data type on 32-bit target with extra opcode.
13886 static unsigned getNonAtomic6432OpcodeWithExtraOpc(unsigned Opc,
13888 unsigned &ExtraOpc) {
13890 case X86::ATOMNAND6432:
13891 ExtraOpc = X86::NOT32r;
13892 HiOpc = X86::AND32rr;
13893 return X86::AND32rr;
13895 llvm_unreachable("Unhandled atomic-load-op opcode!");
13898 // Get pseudo CMOV opcode from the specified data type.
13899 static unsigned getPseudoCMOVOpc(EVT VT) {
13900 switch (VT.getSimpleVT().SimpleTy) {
13901 case MVT::i8: return X86::CMOV_GR8;
13902 case MVT::i16: return X86::CMOV_GR16;
13903 case MVT::i32: return X86::CMOV_GR32;
13907 llvm_unreachable("Unknown CMOV opcode!");
13910 // EmitAtomicLoadArith - emit the code sequence for pseudo atomic instructions.
13911 // They will be translated into a spin-loop or compare-exchange loop from
13914 // dst = atomic-fetch-op MI.addr, MI.val
13920 // t1 = LOAD MI.addr
13922 // t4 = phi(t1, t3 / loop)
13923 // t2 = OP MI.val, t4
13925 // LCMPXCHG [MI.addr], t2, [EAX is implicitly used & defined]
13931 MachineBasicBlock *
13932 X86TargetLowering::EmitAtomicLoadArith(MachineInstr *MI,
13933 MachineBasicBlock *MBB) const {
13934 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
13935 DebugLoc DL = MI->getDebugLoc();
13937 MachineFunction *MF = MBB->getParent();
13938 MachineRegisterInfo &MRI = MF->getRegInfo();
13940 const BasicBlock *BB = MBB->getBasicBlock();
13941 MachineFunction::iterator I = MBB;
13944 assert(MI->getNumOperands() <= X86::AddrNumOperands + 4 &&
13945 "Unexpected number of operands");
13947 assert(MI->hasOneMemOperand() &&
13948 "Expected atomic-load-op to have one memoperand");
13950 // Memory Reference
13951 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
13952 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
13954 unsigned DstReg, SrcReg;
13955 unsigned MemOpndSlot;
13957 unsigned CurOp = 0;
13959 DstReg = MI->getOperand(CurOp++).getReg();
13960 MemOpndSlot = CurOp;
13961 CurOp += X86::AddrNumOperands;
13962 SrcReg = MI->getOperand(CurOp++).getReg();
13964 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
13965 MVT::SimpleValueType VT = *RC->vt_begin();
13966 unsigned t1 = MRI.createVirtualRegister(RC);
13967 unsigned t2 = MRI.createVirtualRegister(RC);
13968 unsigned t3 = MRI.createVirtualRegister(RC);
13969 unsigned t4 = MRI.createVirtualRegister(RC);
13970 unsigned PhyReg = getX86SubSuperRegister(X86::EAX, VT);
13972 unsigned LCMPXCHGOpc = getCmpXChgOpcode(VT);
13973 unsigned LOADOpc = getLoadOpcode(VT);
13975 // For the atomic load-arith operator, we generate
13978 // t1 = LOAD [MI.addr]
13980 // t4 = phi(t1 / thisMBB, t3 / mainMBB)
13981 // t1 = OP MI.val, EAX
13983 // LCMPXCHG [MI.addr], t1, [EAX is implicitly used & defined]
13989 MachineBasicBlock *thisMBB = MBB;
13990 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
13991 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
13992 MF->insert(I, mainMBB);
13993 MF->insert(I, sinkMBB);
13995 MachineInstrBuilder MIB;
13997 // Transfer the remainder of BB and its successor edges to sinkMBB.
13998 sinkMBB->splice(sinkMBB->begin(), MBB,
13999 llvm::next(MachineBasicBlock::iterator(MI)), MBB->end());
14000 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
14003 MIB = BuildMI(thisMBB, DL, TII->get(LOADOpc), t1);
14004 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
14005 MachineOperand NewMO = MI->getOperand(MemOpndSlot + i);
14007 NewMO.setIsKill(false);
14008 MIB.addOperand(NewMO);
14010 for (MachineInstr::mmo_iterator MMOI = MMOBegin; MMOI != MMOEnd; ++MMOI) {
14011 unsigned flags = (*MMOI)->getFlags();
14012 flags = (flags & ~MachineMemOperand::MOStore) | MachineMemOperand::MOLoad;
14013 MachineMemOperand *MMO =
14014 MF->getMachineMemOperand((*MMOI)->getPointerInfo(), flags,
14015 (*MMOI)->getSize(),
14016 (*MMOI)->getBaseAlignment(),
14017 (*MMOI)->getTBAAInfo(),
14018 (*MMOI)->getRanges());
14019 MIB.addMemOperand(MMO);
14022 thisMBB->addSuccessor(mainMBB);
14025 MachineBasicBlock *origMainMBB = mainMBB;
14028 MachineInstr *Phi = BuildMI(mainMBB, DL, TII->get(X86::PHI), t4)
14029 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(mainMBB);
14031 unsigned Opc = MI->getOpcode();
14034 llvm_unreachable("Unhandled atomic-load-op opcode!");
14035 case X86::ATOMAND8:
14036 case X86::ATOMAND16:
14037 case X86::ATOMAND32:
14038 case X86::ATOMAND64:
14040 case X86::ATOMOR16:
14041 case X86::ATOMOR32:
14042 case X86::ATOMOR64:
14043 case X86::ATOMXOR8:
14044 case X86::ATOMXOR16:
14045 case X86::ATOMXOR32:
14046 case X86::ATOMXOR64: {
14047 unsigned ARITHOpc = getNonAtomicOpcode(Opc);
14048 BuildMI(mainMBB, DL, TII->get(ARITHOpc), t2).addReg(SrcReg)
14052 case X86::ATOMNAND8:
14053 case X86::ATOMNAND16:
14054 case X86::ATOMNAND32:
14055 case X86::ATOMNAND64: {
14056 unsigned Tmp = MRI.createVirtualRegister(RC);
14058 unsigned ANDOpc = getNonAtomicOpcodeWithExtraOpc(Opc, NOTOpc);
14059 BuildMI(mainMBB, DL, TII->get(ANDOpc), Tmp).addReg(SrcReg)
14061 BuildMI(mainMBB, DL, TII->get(NOTOpc), t2).addReg(Tmp);
14064 case X86::ATOMMAX8:
14065 case X86::ATOMMAX16:
14066 case X86::ATOMMAX32:
14067 case X86::ATOMMAX64:
14068 case X86::ATOMMIN8:
14069 case X86::ATOMMIN16:
14070 case X86::ATOMMIN32:
14071 case X86::ATOMMIN64:
14072 case X86::ATOMUMAX8:
14073 case X86::ATOMUMAX16:
14074 case X86::ATOMUMAX32:
14075 case X86::ATOMUMAX64:
14076 case X86::ATOMUMIN8:
14077 case X86::ATOMUMIN16:
14078 case X86::ATOMUMIN32:
14079 case X86::ATOMUMIN64: {
14081 unsigned CMOVOpc = getNonAtomicOpcodeWithExtraOpc(Opc, CMPOpc);
14083 BuildMI(mainMBB, DL, TII->get(CMPOpc))
14087 if (Subtarget->hasCMov()) {
14088 if (VT != MVT::i8) {
14090 BuildMI(mainMBB, DL, TII->get(CMOVOpc), t2)
14094 // Promote i8 to i32 to use CMOV32
14095 const TargetRegisterInfo* TRI = getTargetMachine().getRegisterInfo();
14096 const TargetRegisterClass *RC32 =
14097 TRI->getSubClassWithSubReg(getRegClassFor(MVT::i32), X86::sub_8bit);
14098 unsigned SrcReg32 = MRI.createVirtualRegister(RC32);
14099 unsigned AccReg32 = MRI.createVirtualRegister(RC32);
14100 unsigned Tmp = MRI.createVirtualRegister(RC32);
14102 unsigned Undef = MRI.createVirtualRegister(RC32);
14103 BuildMI(mainMBB, DL, TII->get(TargetOpcode::IMPLICIT_DEF), Undef);
14105 BuildMI(mainMBB, DL, TII->get(TargetOpcode::INSERT_SUBREG), SrcReg32)
14108 .addImm(X86::sub_8bit);
14109 BuildMI(mainMBB, DL, TII->get(TargetOpcode::INSERT_SUBREG), AccReg32)
14112 .addImm(X86::sub_8bit);
14114 BuildMI(mainMBB, DL, TII->get(CMOVOpc), Tmp)
14118 BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), t2)
14119 .addReg(Tmp, 0, X86::sub_8bit);
14122 // Use pseudo select and lower them.
14123 assert((VT == MVT::i8 || VT == MVT::i16 || VT == MVT::i32) &&
14124 "Invalid atomic-load-op transformation!");
14125 unsigned SelOpc = getPseudoCMOVOpc(VT);
14126 X86::CondCode CC = X86::getCondFromCMovOpc(CMOVOpc);
14127 assert(CC != X86::COND_INVALID && "Invalid atomic-load-op transformation!");
14128 MIB = BuildMI(mainMBB, DL, TII->get(SelOpc), t2)
14129 .addReg(SrcReg).addReg(t4)
14131 mainMBB = EmitLoweredSelect(MIB, mainMBB);
14132 // Replace the original PHI node as mainMBB is changed after CMOV
14134 BuildMI(*origMainMBB, Phi, DL, TII->get(X86::PHI), t4)
14135 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(mainMBB);
14136 Phi->eraseFromParent();
14142 // Copy PhyReg back from virtual register.
14143 BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), PhyReg)
14146 MIB = BuildMI(mainMBB, DL, TII->get(LCMPXCHGOpc));
14147 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
14148 MachineOperand NewMO = MI->getOperand(MemOpndSlot + i);
14150 NewMO.setIsKill(false);
14151 MIB.addOperand(NewMO);
14154 MIB.setMemRefs(MMOBegin, MMOEnd);
14156 // Copy PhyReg back to virtual register.
14157 BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), t3)
14160 BuildMI(mainMBB, DL, TII->get(X86::JNE_4)).addMBB(origMainMBB);
14162 mainMBB->addSuccessor(origMainMBB);
14163 mainMBB->addSuccessor(sinkMBB);
14166 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
14167 TII->get(TargetOpcode::COPY), DstReg)
14170 MI->eraseFromParent();
14174 // EmitAtomicLoadArith6432 - emit the code sequence for pseudo atomic
14175 // instructions. They will be translated into a spin-loop or compare-exchange
14179 // dst = atomic-fetch-op MI.addr, MI.val
14185 // t1L = LOAD [MI.addr + 0]
14186 // t1H = LOAD [MI.addr + 4]
14188 // t4L = phi(t1L, t3L / loop)
14189 // t4H = phi(t1H, t3H / loop)
14190 // t2L = OP MI.val.lo, t4L
14191 // t2H = OP MI.val.hi, t4H
14196 // LCMPXCHG8B [MI.addr], [ECX:EBX & EDX:EAX are implicitly used and EDX:EAX is implicitly defined]
14204 MachineBasicBlock *
14205 X86TargetLowering::EmitAtomicLoadArith6432(MachineInstr *MI,
14206 MachineBasicBlock *MBB) const {
14207 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
14208 DebugLoc DL = MI->getDebugLoc();
14210 MachineFunction *MF = MBB->getParent();
14211 MachineRegisterInfo &MRI = MF->getRegInfo();
14213 const BasicBlock *BB = MBB->getBasicBlock();
14214 MachineFunction::iterator I = MBB;
14217 assert(MI->getNumOperands() <= X86::AddrNumOperands + 7 &&
14218 "Unexpected number of operands");
14220 assert(MI->hasOneMemOperand() &&
14221 "Expected atomic-load-op32 to have one memoperand");
14223 // Memory Reference
14224 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
14225 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
14227 unsigned DstLoReg, DstHiReg;
14228 unsigned SrcLoReg, SrcHiReg;
14229 unsigned MemOpndSlot;
14231 unsigned CurOp = 0;
14233 DstLoReg = MI->getOperand(CurOp++).getReg();
14234 DstHiReg = MI->getOperand(CurOp++).getReg();
14235 MemOpndSlot = CurOp;
14236 CurOp += X86::AddrNumOperands;
14237 SrcLoReg = MI->getOperand(CurOp++).getReg();
14238 SrcHiReg = MI->getOperand(CurOp++).getReg();
14240 const TargetRegisterClass *RC = &X86::GR32RegClass;
14241 const TargetRegisterClass *RC8 = &X86::GR8RegClass;
14243 unsigned t1L = MRI.createVirtualRegister(RC);
14244 unsigned t1H = MRI.createVirtualRegister(RC);
14245 unsigned t2L = MRI.createVirtualRegister(RC);
14246 unsigned t2H = MRI.createVirtualRegister(RC);
14247 unsigned t3L = MRI.createVirtualRegister(RC);
14248 unsigned t3H = MRI.createVirtualRegister(RC);
14249 unsigned t4L = MRI.createVirtualRegister(RC);
14250 unsigned t4H = MRI.createVirtualRegister(RC);
14252 unsigned LCMPXCHGOpc = X86::LCMPXCHG8B;
14253 unsigned LOADOpc = X86::MOV32rm;
14255 // For the atomic load-arith operator, we generate
14258 // t1L = LOAD [MI.addr + 0]
14259 // t1H = LOAD [MI.addr + 4]
14261 // t4L = phi(t1L / thisMBB, t3L / mainMBB)
14262 // t4H = phi(t1H / thisMBB, t3H / mainMBB)
14263 // t2L = OP MI.val.lo, t4L
14264 // t2H = OP MI.val.hi, t4H
14267 // LCMPXCHG8B [MI.addr], [ECX:EBX & EDX:EAX are implicitly used and EDX:EAX is implicitly defined]
14275 MachineBasicBlock *thisMBB = MBB;
14276 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
14277 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
14278 MF->insert(I, mainMBB);
14279 MF->insert(I, sinkMBB);
14281 MachineInstrBuilder MIB;
14283 // Transfer the remainder of BB and its successor edges to sinkMBB.
14284 sinkMBB->splice(sinkMBB->begin(), MBB,
14285 llvm::next(MachineBasicBlock::iterator(MI)), MBB->end());
14286 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
14290 MIB = BuildMI(thisMBB, DL, TII->get(LOADOpc), t1L);
14291 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
14292 MachineOperand NewMO = MI->getOperand(MemOpndSlot + i);
14294 NewMO.setIsKill(false);
14295 MIB.addOperand(NewMO);
14297 for (MachineInstr::mmo_iterator MMOI = MMOBegin; MMOI != MMOEnd; ++MMOI) {
14298 unsigned flags = (*MMOI)->getFlags();
14299 flags = (flags & ~MachineMemOperand::MOStore) | MachineMemOperand::MOLoad;
14300 MachineMemOperand *MMO =
14301 MF->getMachineMemOperand((*MMOI)->getPointerInfo(), flags,
14302 (*MMOI)->getSize(),
14303 (*MMOI)->getBaseAlignment(),
14304 (*MMOI)->getTBAAInfo(),
14305 (*MMOI)->getRanges());
14306 MIB.addMemOperand(MMO);
14308 MachineInstr *LowMI = MIB;
14311 MIB = BuildMI(thisMBB, DL, TII->get(LOADOpc), t1H);
14312 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
14313 if (i == X86::AddrDisp) {
14314 MIB.addDisp(MI->getOperand(MemOpndSlot + i), 4); // 4 == sizeof(i32)
14316 MachineOperand NewMO = MI->getOperand(MemOpndSlot + i);
14318 NewMO.setIsKill(false);
14319 MIB.addOperand(NewMO);
14322 MIB.setMemRefs(LowMI->memoperands_begin(), LowMI->memoperands_end());
14324 thisMBB->addSuccessor(mainMBB);
14327 MachineBasicBlock *origMainMBB = mainMBB;
14330 MachineInstr *PhiL = BuildMI(mainMBB, DL, TII->get(X86::PHI), t4L)
14331 .addReg(t1L).addMBB(thisMBB).addReg(t3L).addMBB(mainMBB);
14332 MachineInstr *PhiH = BuildMI(mainMBB, DL, TII->get(X86::PHI), t4H)
14333 .addReg(t1H).addMBB(thisMBB).addReg(t3H).addMBB(mainMBB);
14335 unsigned Opc = MI->getOpcode();
14338 llvm_unreachable("Unhandled atomic-load-op6432 opcode!");
14339 case X86::ATOMAND6432:
14340 case X86::ATOMOR6432:
14341 case X86::ATOMXOR6432:
14342 case X86::ATOMADD6432:
14343 case X86::ATOMSUB6432: {
14345 unsigned LoOpc = getNonAtomic6432Opcode(Opc, HiOpc);
14346 BuildMI(mainMBB, DL, TII->get(LoOpc), t2L).addReg(t4L)
14348 BuildMI(mainMBB, DL, TII->get(HiOpc), t2H).addReg(t4H)
14352 case X86::ATOMNAND6432: {
14353 unsigned HiOpc, NOTOpc;
14354 unsigned LoOpc = getNonAtomic6432OpcodeWithExtraOpc(Opc, HiOpc, NOTOpc);
14355 unsigned TmpL = MRI.createVirtualRegister(RC);
14356 unsigned TmpH = MRI.createVirtualRegister(RC);
14357 BuildMI(mainMBB, DL, TII->get(LoOpc), TmpL).addReg(SrcLoReg)
14359 BuildMI(mainMBB, DL, TII->get(HiOpc), TmpH).addReg(SrcHiReg)
14361 BuildMI(mainMBB, DL, TII->get(NOTOpc), t2L).addReg(TmpL);
14362 BuildMI(mainMBB, DL, TII->get(NOTOpc), t2H).addReg(TmpH);
14365 case X86::ATOMMAX6432:
14366 case X86::ATOMMIN6432:
14367 case X86::ATOMUMAX6432:
14368 case X86::ATOMUMIN6432: {
14370 unsigned LoOpc = getNonAtomic6432Opcode(Opc, HiOpc);
14371 unsigned cL = MRI.createVirtualRegister(RC8);
14372 unsigned cH = MRI.createVirtualRegister(RC8);
14373 unsigned cL32 = MRI.createVirtualRegister(RC);
14374 unsigned cH32 = MRI.createVirtualRegister(RC);
14375 unsigned cc = MRI.createVirtualRegister(RC);
14376 // cl := cmp src_lo, lo
14377 BuildMI(mainMBB, DL, TII->get(X86::CMP32rr))
14378 .addReg(SrcLoReg).addReg(t4L);
14379 BuildMI(mainMBB, DL, TII->get(LoOpc), cL);
14380 BuildMI(mainMBB, DL, TII->get(X86::MOVZX32rr8), cL32).addReg(cL);
14381 // ch := cmp src_hi, hi
14382 BuildMI(mainMBB, DL, TII->get(X86::CMP32rr))
14383 .addReg(SrcHiReg).addReg(t4H);
14384 BuildMI(mainMBB, DL, TII->get(HiOpc), cH);
14385 BuildMI(mainMBB, DL, TII->get(X86::MOVZX32rr8), cH32).addReg(cH);
14386 // cc := if (src_hi == hi) ? cl : ch;
14387 if (Subtarget->hasCMov()) {
14388 BuildMI(mainMBB, DL, TII->get(X86::CMOVE32rr), cc)
14389 .addReg(cH32).addReg(cL32);
14391 MIB = BuildMI(mainMBB, DL, TII->get(X86::CMOV_GR32), cc)
14392 .addReg(cH32).addReg(cL32)
14393 .addImm(X86::COND_E);
14394 mainMBB = EmitLoweredSelect(MIB, mainMBB);
14396 BuildMI(mainMBB, DL, TII->get(X86::TEST32rr)).addReg(cc).addReg(cc);
14397 if (Subtarget->hasCMov()) {
14398 BuildMI(mainMBB, DL, TII->get(X86::CMOVNE32rr), t2L)
14399 .addReg(SrcLoReg).addReg(t4L);
14400 BuildMI(mainMBB, DL, TII->get(X86::CMOVNE32rr), t2H)
14401 .addReg(SrcHiReg).addReg(t4H);
14403 MIB = BuildMI(mainMBB, DL, TII->get(X86::CMOV_GR32), t2L)
14404 .addReg(SrcLoReg).addReg(t4L)
14405 .addImm(X86::COND_NE);
14406 mainMBB = EmitLoweredSelect(MIB, mainMBB);
14407 // As the lowered CMOV won't clobber EFLAGS, we could reuse it for the
14408 // 2nd CMOV lowering.
14409 mainMBB->addLiveIn(X86::EFLAGS);
14410 MIB = BuildMI(mainMBB, DL, TII->get(X86::CMOV_GR32), t2H)
14411 .addReg(SrcHiReg).addReg(t4H)
14412 .addImm(X86::COND_NE);
14413 mainMBB = EmitLoweredSelect(MIB, mainMBB);
14414 // Replace the original PHI node as mainMBB is changed after CMOV
14416 BuildMI(*origMainMBB, PhiL, DL, TII->get(X86::PHI), t4L)
14417 .addReg(t1L).addMBB(thisMBB).addReg(t3L).addMBB(mainMBB);
14418 BuildMI(*origMainMBB, PhiH, DL, TII->get(X86::PHI), t4H)
14419 .addReg(t1H).addMBB(thisMBB).addReg(t3H).addMBB(mainMBB);
14420 PhiL->eraseFromParent();
14421 PhiH->eraseFromParent();
14425 case X86::ATOMSWAP6432: {
14427 unsigned LoOpc = getNonAtomic6432Opcode(Opc, HiOpc);
14428 BuildMI(mainMBB, DL, TII->get(LoOpc), t2L).addReg(SrcLoReg);
14429 BuildMI(mainMBB, DL, TII->get(HiOpc), t2H).addReg(SrcHiReg);
14434 // Copy EDX:EAX back from HiReg:LoReg
14435 BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), X86::EAX).addReg(t4L);
14436 BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), X86::EDX).addReg(t4H);
14437 // Copy ECX:EBX from t1H:t1L
14438 BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), X86::EBX).addReg(t2L);
14439 BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), X86::ECX).addReg(t2H);
14441 MIB = BuildMI(mainMBB, DL, TII->get(LCMPXCHGOpc));
14442 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
14443 MachineOperand NewMO = MI->getOperand(MemOpndSlot + i);
14445 NewMO.setIsKill(false);
14446 MIB.addOperand(NewMO);
14448 MIB.setMemRefs(MMOBegin, MMOEnd);
14450 // Copy EDX:EAX back to t3H:t3L
14451 BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), t3L).addReg(X86::EAX);
14452 BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), t3H).addReg(X86::EDX);
14454 BuildMI(mainMBB, DL, TII->get(X86::JNE_4)).addMBB(origMainMBB);
14456 mainMBB->addSuccessor(origMainMBB);
14457 mainMBB->addSuccessor(sinkMBB);
14460 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
14461 TII->get(TargetOpcode::COPY), DstLoReg)
14463 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
14464 TII->get(TargetOpcode::COPY), DstHiReg)
14467 MI->eraseFromParent();
14471 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
14472 // or XMM0_V32I8 in AVX all of this code can be replaced with that
14473 // in the .td file.
14474 static MachineBasicBlock *EmitPCMPSTRM(MachineInstr *MI, MachineBasicBlock *BB,
14475 const TargetInstrInfo *TII) {
14477 switch (MI->getOpcode()) {
14478 default: llvm_unreachable("illegal opcode!");
14479 case X86::PCMPISTRM128REG: Opc = X86::PCMPISTRM128rr; break;
14480 case X86::VPCMPISTRM128REG: Opc = X86::VPCMPISTRM128rr; break;
14481 case X86::PCMPISTRM128MEM: Opc = X86::PCMPISTRM128rm; break;
14482 case X86::VPCMPISTRM128MEM: Opc = X86::VPCMPISTRM128rm; break;
14483 case X86::PCMPESTRM128REG: Opc = X86::PCMPESTRM128rr; break;
14484 case X86::VPCMPESTRM128REG: Opc = X86::VPCMPESTRM128rr; break;
14485 case X86::PCMPESTRM128MEM: Opc = X86::PCMPESTRM128rm; break;
14486 case X86::VPCMPESTRM128MEM: Opc = X86::VPCMPESTRM128rm; break;
14489 DebugLoc dl = MI->getDebugLoc();
14490 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
14492 unsigned NumArgs = MI->getNumOperands();
14493 for (unsigned i = 1; i < NumArgs; ++i) {
14494 MachineOperand &Op = MI->getOperand(i);
14495 if (!(Op.isReg() && Op.isImplicit()))
14496 MIB.addOperand(Op);
14498 if (MI->hasOneMemOperand())
14499 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
14501 BuildMI(*BB, MI, dl,
14502 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
14503 .addReg(X86::XMM0);
14505 MI->eraseFromParent();
14509 // FIXME: Custom handling because TableGen doesn't support multiple implicit
14510 // defs in an instruction pattern
14511 static MachineBasicBlock *EmitPCMPSTRI(MachineInstr *MI, MachineBasicBlock *BB,
14512 const TargetInstrInfo *TII) {
14514 switch (MI->getOpcode()) {
14515 default: llvm_unreachable("illegal opcode!");
14516 case X86::PCMPISTRIREG: Opc = X86::PCMPISTRIrr; break;
14517 case X86::VPCMPISTRIREG: Opc = X86::VPCMPISTRIrr; break;
14518 case X86::PCMPISTRIMEM: Opc = X86::PCMPISTRIrm; break;
14519 case X86::VPCMPISTRIMEM: Opc = X86::VPCMPISTRIrm; break;
14520 case X86::PCMPESTRIREG: Opc = X86::PCMPESTRIrr; break;
14521 case X86::VPCMPESTRIREG: Opc = X86::VPCMPESTRIrr; break;
14522 case X86::PCMPESTRIMEM: Opc = X86::PCMPESTRIrm; break;
14523 case X86::VPCMPESTRIMEM: Opc = X86::VPCMPESTRIrm; break;
14526 DebugLoc dl = MI->getDebugLoc();
14527 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
14529 unsigned NumArgs = MI->getNumOperands(); // remove the results
14530 for (unsigned i = 1; i < NumArgs; ++i) {
14531 MachineOperand &Op = MI->getOperand(i);
14532 if (!(Op.isReg() && Op.isImplicit()))
14533 MIB.addOperand(Op);
14535 if (MI->hasOneMemOperand())
14536 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
14538 BuildMI(*BB, MI, dl,
14539 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
14542 MI->eraseFromParent();
14546 static MachineBasicBlock * EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB,
14547 const TargetInstrInfo *TII,
14548 const X86Subtarget* Subtarget) {
14549 DebugLoc dl = MI->getDebugLoc();
14551 // Address into RAX/EAX, other two args into ECX, EDX.
14552 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
14553 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
14554 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
14555 for (int i = 0; i < X86::AddrNumOperands; ++i)
14556 MIB.addOperand(MI->getOperand(i));
14558 unsigned ValOps = X86::AddrNumOperands;
14559 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
14560 .addReg(MI->getOperand(ValOps).getReg());
14561 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
14562 .addReg(MI->getOperand(ValOps+1).getReg());
14564 // The instruction doesn't actually take any operands though.
14565 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
14567 MI->eraseFromParent(); // The pseudo is gone now.
14571 MachineBasicBlock *
14572 X86TargetLowering::EmitVAARG64WithCustomInserter(
14574 MachineBasicBlock *MBB) const {
14575 // Emit va_arg instruction on X86-64.
14577 // Operands to this pseudo-instruction:
14578 // 0 ) Output : destination address (reg)
14579 // 1-5) Input : va_list address (addr, i64mem)
14580 // 6 ) ArgSize : Size (in bytes) of vararg type
14581 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
14582 // 8 ) Align : Alignment of type
14583 // 9 ) EFLAGS (implicit-def)
14585 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
14586 assert(X86::AddrNumOperands == 5 && "VAARG_64 assumes 5 address operands");
14588 unsigned DestReg = MI->getOperand(0).getReg();
14589 MachineOperand &Base = MI->getOperand(1);
14590 MachineOperand &Scale = MI->getOperand(2);
14591 MachineOperand &Index = MI->getOperand(3);
14592 MachineOperand &Disp = MI->getOperand(4);
14593 MachineOperand &Segment = MI->getOperand(5);
14594 unsigned ArgSize = MI->getOperand(6).getImm();
14595 unsigned ArgMode = MI->getOperand(7).getImm();
14596 unsigned Align = MI->getOperand(8).getImm();
14598 // Memory Reference
14599 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
14600 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
14601 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
14603 // Machine Information
14604 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
14605 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
14606 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
14607 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
14608 DebugLoc DL = MI->getDebugLoc();
14610 // struct va_list {
14613 // i64 overflow_area (address)
14614 // i64 reg_save_area (address)
14616 // sizeof(va_list) = 24
14617 // alignment(va_list) = 8
14619 unsigned TotalNumIntRegs = 6;
14620 unsigned TotalNumXMMRegs = 8;
14621 bool UseGPOffset = (ArgMode == 1);
14622 bool UseFPOffset = (ArgMode == 2);
14623 unsigned MaxOffset = TotalNumIntRegs * 8 +
14624 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
14626 /* Align ArgSize to a multiple of 8 */
14627 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
14628 bool NeedsAlign = (Align > 8);
14630 MachineBasicBlock *thisMBB = MBB;
14631 MachineBasicBlock *overflowMBB;
14632 MachineBasicBlock *offsetMBB;
14633 MachineBasicBlock *endMBB;
14635 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
14636 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
14637 unsigned OffsetReg = 0;
14639 if (!UseGPOffset && !UseFPOffset) {
14640 // If we only pull from the overflow region, we don't create a branch.
14641 // We don't need to alter control flow.
14642 OffsetDestReg = 0; // unused
14643 OverflowDestReg = DestReg;
14646 overflowMBB = thisMBB;
14649 // First emit code to check if gp_offset (or fp_offset) is below the bound.
14650 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
14651 // If not, pull from overflow_area. (branch to overflowMBB)
14656 // offsetMBB overflowMBB
14661 // Registers for the PHI in endMBB
14662 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
14663 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
14665 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
14666 MachineFunction *MF = MBB->getParent();
14667 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
14668 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
14669 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
14671 MachineFunction::iterator MBBIter = MBB;
14674 // Insert the new basic blocks
14675 MF->insert(MBBIter, offsetMBB);
14676 MF->insert(MBBIter, overflowMBB);
14677 MF->insert(MBBIter, endMBB);
14679 // Transfer the remainder of MBB and its successor edges to endMBB.
14680 endMBB->splice(endMBB->begin(), thisMBB,
14681 llvm::next(MachineBasicBlock::iterator(MI)),
14683 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
14685 // Make offsetMBB and overflowMBB successors of thisMBB
14686 thisMBB->addSuccessor(offsetMBB);
14687 thisMBB->addSuccessor(overflowMBB);
14689 // endMBB is a successor of both offsetMBB and overflowMBB
14690 offsetMBB->addSuccessor(endMBB);
14691 overflowMBB->addSuccessor(endMBB);
14693 // Load the offset value into a register
14694 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
14695 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
14699 .addDisp(Disp, UseFPOffset ? 4 : 0)
14700 .addOperand(Segment)
14701 .setMemRefs(MMOBegin, MMOEnd);
14703 // Check if there is enough room left to pull this argument.
14704 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
14706 .addImm(MaxOffset + 8 - ArgSizeA8);
14708 // Branch to "overflowMBB" if offset >= max
14709 // Fall through to "offsetMBB" otherwise
14710 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
14711 .addMBB(overflowMBB);
14714 // In offsetMBB, emit code to use the reg_save_area.
14716 assert(OffsetReg != 0);
14718 // Read the reg_save_area address.
14719 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
14720 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
14725 .addOperand(Segment)
14726 .setMemRefs(MMOBegin, MMOEnd);
14728 // Zero-extend the offset
14729 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
14730 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
14733 .addImm(X86::sub_32bit);
14735 // Add the offset to the reg_save_area to get the final address.
14736 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
14737 .addReg(OffsetReg64)
14738 .addReg(RegSaveReg);
14740 // Compute the offset for the next argument
14741 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
14742 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
14744 .addImm(UseFPOffset ? 16 : 8);
14746 // Store it back into the va_list.
14747 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
14751 .addDisp(Disp, UseFPOffset ? 4 : 0)
14752 .addOperand(Segment)
14753 .addReg(NextOffsetReg)
14754 .setMemRefs(MMOBegin, MMOEnd);
14757 BuildMI(offsetMBB, DL, TII->get(X86::JMP_4))
14762 // Emit code to use overflow area
14765 // Load the overflow_area address into a register.
14766 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
14767 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
14772 .addOperand(Segment)
14773 .setMemRefs(MMOBegin, MMOEnd);
14775 // If we need to align it, do so. Otherwise, just copy the address
14776 // to OverflowDestReg.
14778 // Align the overflow address
14779 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
14780 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
14782 // aligned_addr = (addr + (align-1)) & ~(align-1)
14783 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
14784 .addReg(OverflowAddrReg)
14787 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
14789 .addImm(~(uint64_t)(Align-1));
14791 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
14792 .addReg(OverflowAddrReg);
14795 // Compute the next overflow address after this argument.
14796 // (the overflow address should be kept 8-byte aligned)
14797 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
14798 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
14799 .addReg(OverflowDestReg)
14800 .addImm(ArgSizeA8);
14802 // Store the new overflow address.
14803 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
14808 .addOperand(Segment)
14809 .addReg(NextAddrReg)
14810 .setMemRefs(MMOBegin, MMOEnd);
14812 // If we branched, emit the PHI to the front of endMBB.
14814 BuildMI(*endMBB, endMBB->begin(), DL,
14815 TII->get(X86::PHI), DestReg)
14816 .addReg(OffsetDestReg).addMBB(offsetMBB)
14817 .addReg(OverflowDestReg).addMBB(overflowMBB);
14820 // Erase the pseudo instruction
14821 MI->eraseFromParent();
14826 MachineBasicBlock *
14827 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
14829 MachineBasicBlock *MBB) const {
14830 // Emit code to save XMM registers to the stack. The ABI says that the
14831 // number of registers to save is given in %al, so it's theoretically
14832 // possible to do an indirect jump trick to avoid saving all of them,
14833 // however this code takes a simpler approach and just executes all
14834 // of the stores if %al is non-zero. It's less code, and it's probably
14835 // easier on the hardware branch predictor, and stores aren't all that
14836 // expensive anyway.
14838 // Create the new basic blocks. One block contains all the XMM stores,
14839 // and one block is the final destination regardless of whether any
14840 // stores were performed.
14841 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
14842 MachineFunction *F = MBB->getParent();
14843 MachineFunction::iterator MBBIter = MBB;
14845 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
14846 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
14847 F->insert(MBBIter, XMMSaveMBB);
14848 F->insert(MBBIter, EndMBB);
14850 // Transfer the remainder of MBB and its successor edges to EndMBB.
14851 EndMBB->splice(EndMBB->begin(), MBB,
14852 llvm::next(MachineBasicBlock::iterator(MI)),
14854 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
14856 // The original block will now fall through to the XMM save block.
14857 MBB->addSuccessor(XMMSaveMBB);
14858 // The XMMSaveMBB will fall through to the end block.
14859 XMMSaveMBB->addSuccessor(EndMBB);
14861 // Now add the instructions.
14862 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
14863 DebugLoc DL = MI->getDebugLoc();
14865 unsigned CountReg = MI->getOperand(0).getReg();
14866 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
14867 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
14869 if (!Subtarget->isTargetWin64()) {
14870 // If %al is 0, branch around the XMM save block.
14871 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
14872 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
14873 MBB->addSuccessor(EndMBB);
14876 unsigned MOVOpc = Subtarget->hasFp256() ? X86::VMOVAPSmr : X86::MOVAPSmr;
14877 // In the XMM save block, save all the XMM argument registers.
14878 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
14879 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
14880 MachineMemOperand *MMO =
14881 F->getMachineMemOperand(
14882 MachinePointerInfo::getFixedStack(RegSaveFrameIndex, Offset),
14883 MachineMemOperand::MOStore,
14884 /*Size=*/16, /*Align=*/16);
14885 BuildMI(XMMSaveMBB, DL, TII->get(MOVOpc))
14886 .addFrameIndex(RegSaveFrameIndex)
14887 .addImm(/*Scale=*/1)
14888 .addReg(/*IndexReg=*/0)
14889 .addImm(/*Disp=*/Offset)
14890 .addReg(/*Segment=*/0)
14891 .addReg(MI->getOperand(i).getReg())
14892 .addMemOperand(MMO);
14895 MI->eraseFromParent(); // The pseudo instruction is gone now.
14900 // The EFLAGS operand of SelectItr might be missing a kill marker
14901 // because there were multiple uses of EFLAGS, and ISel didn't know
14902 // which to mark. Figure out whether SelectItr should have had a
14903 // kill marker, and set it if it should. Returns the correct kill
14905 static bool checkAndUpdateEFLAGSKill(MachineBasicBlock::iterator SelectItr,
14906 MachineBasicBlock* BB,
14907 const TargetRegisterInfo* TRI) {
14908 // Scan forward through BB for a use/def of EFLAGS.
14909 MachineBasicBlock::iterator miI(llvm::next(SelectItr));
14910 for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) {
14911 const MachineInstr& mi = *miI;
14912 if (mi.readsRegister(X86::EFLAGS))
14914 if (mi.definesRegister(X86::EFLAGS))
14915 break; // Should have kill-flag - update below.
14918 // If we hit the end of the block, check whether EFLAGS is live into a
14920 if (miI == BB->end()) {
14921 for (MachineBasicBlock::succ_iterator sItr = BB->succ_begin(),
14922 sEnd = BB->succ_end();
14923 sItr != sEnd; ++sItr) {
14924 MachineBasicBlock* succ = *sItr;
14925 if (succ->isLiveIn(X86::EFLAGS))
14930 // We found a def, or hit the end of the basic block and EFLAGS wasn't live
14931 // out. SelectMI should have a kill flag on EFLAGS.
14932 SelectItr->addRegisterKilled(X86::EFLAGS, TRI);
14936 MachineBasicBlock *
14937 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
14938 MachineBasicBlock *BB) const {
14939 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
14940 DebugLoc DL = MI->getDebugLoc();
14942 // To "insert" a SELECT_CC instruction, we actually have to insert the
14943 // diamond control-flow pattern. The incoming instruction knows the
14944 // destination vreg to set, the condition code register to branch on, the
14945 // true/false values to select between, and a branch opcode to use.
14946 const BasicBlock *LLVM_BB = BB->getBasicBlock();
14947 MachineFunction::iterator It = BB;
14953 // cmpTY ccX, r1, r2
14955 // fallthrough --> copy0MBB
14956 MachineBasicBlock *thisMBB = BB;
14957 MachineFunction *F = BB->getParent();
14958 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
14959 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
14960 F->insert(It, copy0MBB);
14961 F->insert(It, sinkMBB);
14963 // If the EFLAGS register isn't dead in the terminator, then claim that it's
14964 // live into the sink and copy blocks.
14965 const TargetRegisterInfo* TRI = getTargetMachine().getRegisterInfo();
14966 if (!MI->killsRegister(X86::EFLAGS) &&
14967 !checkAndUpdateEFLAGSKill(MI, BB, TRI)) {
14968 copy0MBB->addLiveIn(X86::EFLAGS);
14969 sinkMBB->addLiveIn(X86::EFLAGS);
14972 // Transfer the remainder of BB and its successor edges to sinkMBB.
14973 sinkMBB->splice(sinkMBB->begin(), BB,
14974 llvm::next(MachineBasicBlock::iterator(MI)),
14976 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
14978 // Add the true and fallthrough blocks as its successors.
14979 BB->addSuccessor(copy0MBB);
14980 BB->addSuccessor(sinkMBB);
14982 // Create the conditional branch instruction.
14984 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
14985 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
14988 // %FalseValue = ...
14989 // # fallthrough to sinkMBB
14990 copy0MBB->addSuccessor(sinkMBB);
14993 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
14995 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
14996 TII->get(X86::PHI), MI->getOperand(0).getReg())
14997 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
14998 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
15000 MI->eraseFromParent(); // The pseudo instruction is gone now.
15004 MachineBasicBlock *
15005 X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI, MachineBasicBlock *BB,
15006 bool Is64Bit) const {
15007 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
15008 DebugLoc DL = MI->getDebugLoc();
15009 MachineFunction *MF = BB->getParent();
15010 const BasicBlock *LLVM_BB = BB->getBasicBlock();
15012 assert(getTargetMachine().Options.EnableSegmentedStacks);
15014 unsigned TlsReg = Is64Bit ? X86::FS : X86::GS;
15015 unsigned TlsOffset = Is64Bit ? 0x70 : 0x30;
15018 // ... [Till the alloca]
15019 // If stacklet is not large enough, jump to mallocMBB
15022 // Allocate by subtracting from RSP
15023 // Jump to continueMBB
15026 // Allocate by call to runtime
15030 // [rest of original BB]
15033 MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB);
15034 MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB);
15035 MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB);
15037 MachineRegisterInfo &MRI = MF->getRegInfo();
15038 const TargetRegisterClass *AddrRegClass =
15039 getRegClassFor(Is64Bit ? MVT::i64:MVT::i32);
15041 unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass),
15042 bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass),
15043 tmpSPVReg = MRI.createVirtualRegister(AddrRegClass),
15044 SPLimitVReg = MRI.createVirtualRegister(AddrRegClass),
15045 sizeVReg = MI->getOperand(1).getReg(),
15046 physSPReg = Is64Bit ? X86::RSP : X86::ESP;
15048 MachineFunction::iterator MBBIter = BB;
15051 MF->insert(MBBIter, bumpMBB);
15052 MF->insert(MBBIter, mallocMBB);
15053 MF->insert(MBBIter, continueMBB);
15055 continueMBB->splice(continueMBB->begin(), BB, llvm::next
15056 (MachineBasicBlock::iterator(MI)), BB->end());
15057 continueMBB->transferSuccessorsAndUpdatePHIs(BB);
15059 // Add code to the main basic block to check if the stack limit has been hit,
15060 // and if so, jump to mallocMBB otherwise to bumpMBB.
15061 BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg);
15062 BuildMI(BB, DL, TII->get(Is64Bit ? X86::SUB64rr:X86::SUB32rr), SPLimitVReg)
15063 .addReg(tmpSPVReg).addReg(sizeVReg);
15064 BuildMI(BB, DL, TII->get(Is64Bit ? X86::CMP64mr:X86::CMP32mr))
15065 .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg)
15066 .addReg(SPLimitVReg);
15067 BuildMI(BB, DL, TII->get(X86::JG_4)).addMBB(mallocMBB);
15069 // bumpMBB simply decreases the stack pointer, since we know the current
15070 // stacklet has enough space.
15071 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg)
15072 .addReg(SPLimitVReg);
15073 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg)
15074 .addReg(SPLimitVReg);
15075 BuildMI(bumpMBB, DL, TII->get(X86::JMP_4)).addMBB(continueMBB);
15077 // Calls into a routine in libgcc to allocate more space from the heap.
15078 const uint32_t *RegMask =
15079 getTargetMachine().getRegisterInfo()->getCallPreservedMask(CallingConv::C);
15081 BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
15083 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
15084 .addExternalSymbol("__morestack_allocate_stack_space")
15085 .addRegMask(RegMask)
15086 .addReg(X86::RDI, RegState::Implicit)
15087 .addReg(X86::RAX, RegState::ImplicitDefine);
15089 BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg)
15091 BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg);
15092 BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32))
15093 .addExternalSymbol("__morestack_allocate_stack_space")
15094 .addRegMask(RegMask)
15095 .addReg(X86::EAX, RegState::ImplicitDefine);
15099 BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg)
15102 BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg)
15103 .addReg(Is64Bit ? X86::RAX : X86::EAX);
15104 BuildMI(mallocMBB, DL, TII->get(X86::JMP_4)).addMBB(continueMBB);
15106 // Set up the CFG correctly.
15107 BB->addSuccessor(bumpMBB);
15108 BB->addSuccessor(mallocMBB);
15109 mallocMBB->addSuccessor(continueMBB);
15110 bumpMBB->addSuccessor(continueMBB);
15112 // Take care of the PHI nodes.
15113 BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI),
15114 MI->getOperand(0).getReg())
15115 .addReg(mallocPtrVReg).addMBB(mallocMBB)
15116 .addReg(bumpSPPtrVReg).addMBB(bumpMBB);
15118 // Delete the original pseudo instruction.
15119 MI->eraseFromParent();
15122 return continueMBB;
15125 MachineBasicBlock *
15126 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
15127 MachineBasicBlock *BB) const {
15128 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
15129 DebugLoc DL = MI->getDebugLoc();
15131 assert(!Subtarget->isTargetEnvMacho());
15133 // The lowering is pretty easy: we're just emitting the call to _alloca. The
15134 // non-trivial part is impdef of ESP.
15136 if (Subtarget->isTargetWin64()) {
15137 if (Subtarget->isTargetCygMing()) {
15138 // ___chkstk(Mingw64):
15139 // Clobbers R10, R11, RAX and EFLAGS.
15141 BuildMI(*BB, MI, DL, TII->get(X86::W64ALLOCA))
15142 .addExternalSymbol("___chkstk")
15143 .addReg(X86::RAX, RegState::Implicit)
15144 .addReg(X86::RSP, RegState::Implicit)
15145 .addReg(X86::RAX, RegState::Define | RegState::Implicit)
15146 .addReg(X86::RSP, RegState::Define | RegState::Implicit)
15147 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
15149 // __chkstk(MSVCRT): does not update stack pointer.
15150 // Clobbers R10, R11 and EFLAGS.
15151 BuildMI(*BB, MI, DL, TII->get(X86::W64ALLOCA))
15152 .addExternalSymbol("__chkstk")
15153 .addReg(X86::RAX, RegState::Implicit)
15154 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
15155 // RAX has the offset to be subtracted from RSP.
15156 BuildMI(*BB, MI, DL, TII->get(X86::SUB64rr), X86::RSP)
15161 const char *StackProbeSymbol =
15162 Subtarget->isTargetWindows() ? "_chkstk" : "_alloca";
15164 BuildMI(*BB, MI, DL, TII->get(X86::CALLpcrel32))
15165 .addExternalSymbol(StackProbeSymbol)
15166 .addReg(X86::EAX, RegState::Implicit)
15167 .addReg(X86::ESP, RegState::Implicit)
15168 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
15169 .addReg(X86::ESP, RegState::Define | RegState::Implicit)
15170 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
15173 MI->eraseFromParent(); // The pseudo instruction is gone now.
15177 MachineBasicBlock *
15178 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
15179 MachineBasicBlock *BB) const {
15180 // This is pretty easy. We're taking the value that we received from
15181 // our load from the relocation, sticking it in either RDI (x86-64)
15182 // or EAX and doing an indirect call. The return value will then
15183 // be in the normal return register.
15184 const X86InstrInfo *TII
15185 = static_cast<const X86InstrInfo*>(getTargetMachine().getInstrInfo());
15186 DebugLoc DL = MI->getDebugLoc();
15187 MachineFunction *F = BB->getParent();
15189 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
15190 assert(MI->getOperand(3).isGlobal() && "This should be a global");
15192 // Get a register mask for the lowered call.
15193 // FIXME: The 32-bit calls have non-standard calling conventions. Use a
15194 // proper register mask.
15195 const uint32_t *RegMask =
15196 getTargetMachine().getRegisterInfo()->getCallPreservedMask(CallingConv::C);
15197 if (Subtarget->is64Bit()) {
15198 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
15199 TII->get(X86::MOV64rm), X86::RDI)
15201 .addImm(0).addReg(0)
15202 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
15203 MI->getOperand(3).getTargetFlags())
15205 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
15206 addDirectMem(MIB, X86::RDI);
15207 MIB.addReg(X86::RAX, RegState::ImplicitDefine).addRegMask(RegMask);
15208 } else if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
15209 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
15210 TII->get(X86::MOV32rm), X86::EAX)
15212 .addImm(0).addReg(0)
15213 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
15214 MI->getOperand(3).getTargetFlags())
15216 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
15217 addDirectMem(MIB, X86::EAX);
15218 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
15220 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
15221 TII->get(X86::MOV32rm), X86::EAX)
15222 .addReg(TII->getGlobalBaseReg(F))
15223 .addImm(0).addReg(0)
15224 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
15225 MI->getOperand(3).getTargetFlags())
15227 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
15228 addDirectMem(MIB, X86::EAX);
15229 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
15232 MI->eraseFromParent(); // The pseudo instruction is gone now.
15236 MachineBasicBlock *
15237 X86TargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
15238 MachineBasicBlock *MBB) const {
15239 DebugLoc DL = MI->getDebugLoc();
15240 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
15242 MachineFunction *MF = MBB->getParent();
15243 MachineRegisterInfo &MRI = MF->getRegInfo();
15245 const BasicBlock *BB = MBB->getBasicBlock();
15246 MachineFunction::iterator I = MBB;
15249 // Memory Reference
15250 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
15251 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
15254 unsigned MemOpndSlot = 0;
15256 unsigned CurOp = 0;
15258 DstReg = MI->getOperand(CurOp++).getReg();
15259 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
15260 assert(RC->hasType(MVT::i32) && "Invalid destination!");
15261 unsigned mainDstReg = MRI.createVirtualRegister(RC);
15262 unsigned restoreDstReg = MRI.createVirtualRegister(RC);
15264 MemOpndSlot = CurOp;
15266 MVT PVT = getPointerTy();
15267 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
15268 "Invalid Pointer Size!");
15270 // For v = setjmp(buf), we generate
15273 // buf[LabelOffset] = restoreMBB
15274 // SjLjSetup restoreMBB
15280 // v = phi(main, restore)
15285 MachineBasicBlock *thisMBB = MBB;
15286 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
15287 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
15288 MachineBasicBlock *restoreMBB = MF->CreateMachineBasicBlock(BB);
15289 MF->insert(I, mainMBB);
15290 MF->insert(I, sinkMBB);
15291 MF->push_back(restoreMBB);
15293 MachineInstrBuilder MIB;
15295 // Transfer the remainder of BB and its successor edges to sinkMBB.
15296 sinkMBB->splice(sinkMBB->begin(), MBB,
15297 llvm::next(MachineBasicBlock::iterator(MI)), MBB->end());
15298 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
15301 unsigned PtrStoreOpc = 0;
15302 unsigned LabelReg = 0;
15303 const int64_t LabelOffset = 1 * PVT.getStoreSize();
15304 Reloc::Model RM = getTargetMachine().getRelocationModel();
15305 bool UseImmLabel = (getTargetMachine().getCodeModel() == CodeModel::Small) &&
15306 (RM == Reloc::Static || RM == Reloc::DynamicNoPIC);
15308 // Prepare IP either in reg or imm.
15309 if (!UseImmLabel) {
15310 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr;
15311 const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
15312 LabelReg = MRI.createVirtualRegister(PtrRC);
15313 if (Subtarget->is64Bit()) {
15314 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA64r), LabelReg)
15318 .addMBB(restoreMBB)
15321 const X86InstrInfo *XII = static_cast<const X86InstrInfo*>(TII);
15322 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA32r), LabelReg)
15323 .addReg(XII->getGlobalBaseReg(MF))
15326 .addMBB(restoreMBB, Subtarget->ClassifyBlockAddressReference())
15330 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi;
15332 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PtrStoreOpc));
15333 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
15334 if (i == X86::AddrDisp)
15335 MIB.addDisp(MI->getOperand(MemOpndSlot + i), LabelOffset);
15337 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
15340 MIB.addReg(LabelReg);
15342 MIB.addMBB(restoreMBB);
15343 MIB.setMemRefs(MMOBegin, MMOEnd);
15345 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
15346 .addMBB(restoreMBB);
15348 const X86RegisterInfo *RegInfo =
15349 static_cast<const X86RegisterInfo*>(getTargetMachine().getRegisterInfo());
15350 MIB.addRegMask(RegInfo->getNoPreservedMask());
15351 thisMBB->addSuccessor(mainMBB);
15352 thisMBB->addSuccessor(restoreMBB);
15356 BuildMI(mainMBB, DL, TII->get(X86::MOV32r0), mainDstReg);
15357 mainMBB->addSuccessor(sinkMBB);
15360 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
15361 TII->get(X86::PHI), DstReg)
15362 .addReg(mainDstReg).addMBB(mainMBB)
15363 .addReg(restoreDstReg).addMBB(restoreMBB);
15366 BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1);
15367 BuildMI(restoreMBB, DL, TII->get(X86::JMP_4)).addMBB(sinkMBB);
15368 restoreMBB->addSuccessor(sinkMBB);
15370 MI->eraseFromParent();
15374 MachineBasicBlock *
15375 X86TargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
15376 MachineBasicBlock *MBB) const {
15377 DebugLoc DL = MI->getDebugLoc();
15378 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
15380 MachineFunction *MF = MBB->getParent();
15381 MachineRegisterInfo &MRI = MF->getRegInfo();
15383 // Memory Reference
15384 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
15385 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
15387 MVT PVT = getPointerTy();
15388 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
15389 "Invalid Pointer Size!");
15391 const TargetRegisterClass *RC =
15392 (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass;
15393 unsigned Tmp = MRI.createVirtualRegister(RC);
15394 // Since FP is only updated here but NOT referenced, it's treated as GPR.
15395 const X86RegisterInfo *RegInfo =
15396 static_cast<const X86RegisterInfo*>(getTargetMachine().getRegisterInfo());
15397 unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
15398 unsigned SP = RegInfo->getStackRegister();
15400 MachineInstrBuilder MIB;
15402 const int64_t LabelOffset = 1 * PVT.getStoreSize();
15403 const int64_t SPOffset = 2 * PVT.getStoreSize();
15405 unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm;
15406 unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r;
15409 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP);
15410 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
15411 MIB.addOperand(MI->getOperand(i));
15412 MIB.setMemRefs(MMOBegin, MMOEnd);
15414 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp);
15415 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
15416 if (i == X86::AddrDisp)
15417 MIB.addDisp(MI->getOperand(i), LabelOffset);
15419 MIB.addOperand(MI->getOperand(i));
15421 MIB.setMemRefs(MMOBegin, MMOEnd);
15423 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP);
15424 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
15425 if (i == X86::AddrDisp)
15426 MIB.addDisp(MI->getOperand(i), SPOffset);
15428 MIB.addOperand(MI->getOperand(i));
15430 MIB.setMemRefs(MMOBegin, MMOEnd);
15432 BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp);
15434 MI->eraseFromParent();
15438 MachineBasicBlock *
15439 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
15440 MachineBasicBlock *BB) const {
15441 switch (MI->getOpcode()) {
15442 default: llvm_unreachable("Unexpected instr type to insert");
15443 case X86::TAILJMPd64:
15444 case X86::TAILJMPr64:
15445 case X86::TAILJMPm64:
15446 llvm_unreachable("TAILJMP64 would not be touched here.");
15447 case X86::TCRETURNdi64:
15448 case X86::TCRETURNri64:
15449 case X86::TCRETURNmi64:
15451 case X86::WIN_ALLOCA:
15452 return EmitLoweredWinAlloca(MI, BB);
15453 case X86::SEG_ALLOCA_32:
15454 return EmitLoweredSegAlloca(MI, BB, false);
15455 case X86::SEG_ALLOCA_64:
15456 return EmitLoweredSegAlloca(MI, BB, true);
15457 case X86::TLSCall_32:
15458 case X86::TLSCall_64:
15459 return EmitLoweredTLSCall(MI, BB);
15460 case X86::CMOV_GR8:
15461 case X86::CMOV_FR32:
15462 case X86::CMOV_FR64:
15463 case X86::CMOV_V4F32:
15464 case X86::CMOV_V2F64:
15465 case X86::CMOV_V2I64:
15466 case X86::CMOV_V8F32:
15467 case X86::CMOV_V4F64:
15468 case X86::CMOV_V4I64:
15469 case X86::CMOV_GR16:
15470 case X86::CMOV_GR32:
15471 case X86::CMOV_RFP32:
15472 case X86::CMOV_RFP64:
15473 case X86::CMOV_RFP80:
15474 return EmitLoweredSelect(MI, BB);
15476 case X86::FP32_TO_INT16_IN_MEM:
15477 case X86::FP32_TO_INT32_IN_MEM:
15478 case X86::FP32_TO_INT64_IN_MEM:
15479 case X86::FP64_TO_INT16_IN_MEM:
15480 case X86::FP64_TO_INT32_IN_MEM:
15481 case X86::FP64_TO_INT64_IN_MEM:
15482 case X86::FP80_TO_INT16_IN_MEM:
15483 case X86::FP80_TO_INT32_IN_MEM:
15484 case X86::FP80_TO_INT64_IN_MEM: {
15485 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
15486 DebugLoc DL = MI->getDebugLoc();
15488 // Change the floating point control register to use "round towards zero"
15489 // mode when truncating to an integer value.
15490 MachineFunction *F = BB->getParent();
15491 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
15492 addFrameReference(BuildMI(*BB, MI, DL,
15493 TII->get(X86::FNSTCW16m)), CWFrameIdx);
15495 // Load the old value of the high byte of the control word...
15497 F->getRegInfo().createVirtualRegister(&X86::GR16RegClass);
15498 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
15501 // Set the high part to be round to zero...
15502 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
15505 // Reload the modified control word now...
15506 addFrameReference(BuildMI(*BB, MI, DL,
15507 TII->get(X86::FLDCW16m)), CWFrameIdx);
15509 // Restore the memory image of control word to original value
15510 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
15513 // Get the X86 opcode to use.
15515 switch (MI->getOpcode()) {
15516 default: llvm_unreachable("illegal opcode!");
15517 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
15518 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
15519 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
15520 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
15521 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
15522 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
15523 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
15524 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
15525 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
15529 MachineOperand &Op = MI->getOperand(0);
15531 AM.BaseType = X86AddressMode::RegBase;
15532 AM.Base.Reg = Op.getReg();
15534 AM.BaseType = X86AddressMode::FrameIndexBase;
15535 AM.Base.FrameIndex = Op.getIndex();
15537 Op = MI->getOperand(1);
15539 AM.Scale = Op.getImm();
15540 Op = MI->getOperand(2);
15542 AM.IndexReg = Op.getImm();
15543 Op = MI->getOperand(3);
15544 if (Op.isGlobal()) {
15545 AM.GV = Op.getGlobal();
15547 AM.Disp = Op.getImm();
15549 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
15550 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
15552 // Reload the original control word now.
15553 addFrameReference(BuildMI(*BB, MI, DL,
15554 TII->get(X86::FLDCW16m)), CWFrameIdx);
15556 MI->eraseFromParent(); // The pseudo instruction is gone now.
15559 // String/text processing lowering.
15560 case X86::PCMPISTRM128REG:
15561 case X86::VPCMPISTRM128REG:
15562 case X86::PCMPISTRM128MEM:
15563 case X86::VPCMPISTRM128MEM:
15564 case X86::PCMPESTRM128REG:
15565 case X86::VPCMPESTRM128REG:
15566 case X86::PCMPESTRM128MEM:
15567 case X86::VPCMPESTRM128MEM:
15568 assert(Subtarget->hasSSE42() &&
15569 "Target must have SSE4.2 or AVX features enabled");
15570 return EmitPCMPSTRM(MI, BB, getTargetMachine().getInstrInfo());
15572 // String/text processing lowering.
15573 case X86::PCMPISTRIREG:
15574 case X86::VPCMPISTRIREG:
15575 case X86::PCMPISTRIMEM:
15576 case X86::VPCMPISTRIMEM:
15577 case X86::PCMPESTRIREG:
15578 case X86::VPCMPESTRIREG:
15579 case X86::PCMPESTRIMEM:
15580 case X86::VPCMPESTRIMEM:
15581 assert(Subtarget->hasSSE42() &&
15582 "Target must have SSE4.2 or AVX features enabled");
15583 return EmitPCMPSTRI(MI, BB, getTargetMachine().getInstrInfo());
15585 // Thread synchronization.
15587 return EmitMonitor(MI, BB, getTargetMachine().getInstrInfo(), Subtarget);
15591 return EmitXBegin(MI, BB, getTargetMachine().getInstrInfo());
15593 // Atomic Lowering.
15594 case X86::ATOMAND8:
15595 case X86::ATOMAND16:
15596 case X86::ATOMAND32:
15597 case X86::ATOMAND64:
15600 case X86::ATOMOR16:
15601 case X86::ATOMOR32:
15602 case X86::ATOMOR64:
15604 case X86::ATOMXOR16:
15605 case X86::ATOMXOR8:
15606 case X86::ATOMXOR32:
15607 case X86::ATOMXOR64:
15609 case X86::ATOMNAND8:
15610 case X86::ATOMNAND16:
15611 case X86::ATOMNAND32:
15612 case X86::ATOMNAND64:
15614 case X86::ATOMMAX8:
15615 case X86::ATOMMAX16:
15616 case X86::ATOMMAX32:
15617 case X86::ATOMMAX64:
15619 case X86::ATOMMIN8:
15620 case X86::ATOMMIN16:
15621 case X86::ATOMMIN32:
15622 case X86::ATOMMIN64:
15624 case X86::ATOMUMAX8:
15625 case X86::ATOMUMAX16:
15626 case X86::ATOMUMAX32:
15627 case X86::ATOMUMAX64:
15629 case X86::ATOMUMIN8:
15630 case X86::ATOMUMIN16:
15631 case X86::ATOMUMIN32:
15632 case X86::ATOMUMIN64:
15633 return EmitAtomicLoadArith(MI, BB);
15635 // This group does 64-bit operations on a 32-bit host.
15636 case X86::ATOMAND6432:
15637 case X86::ATOMOR6432:
15638 case X86::ATOMXOR6432:
15639 case X86::ATOMNAND6432:
15640 case X86::ATOMADD6432:
15641 case X86::ATOMSUB6432:
15642 case X86::ATOMMAX6432:
15643 case X86::ATOMMIN6432:
15644 case X86::ATOMUMAX6432:
15645 case X86::ATOMUMIN6432:
15646 case X86::ATOMSWAP6432:
15647 return EmitAtomicLoadArith6432(MI, BB);
15649 case X86::VASTART_SAVE_XMM_REGS:
15650 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
15652 case X86::VAARG_64:
15653 return EmitVAARG64WithCustomInserter(MI, BB);
15655 case X86::EH_SjLj_SetJmp32:
15656 case X86::EH_SjLj_SetJmp64:
15657 return emitEHSjLjSetJmp(MI, BB);
15659 case X86::EH_SjLj_LongJmp32:
15660 case X86::EH_SjLj_LongJmp64:
15661 return emitEHSjLjLongJmp(MI, BB);
15665 //===----------------------------------------------------------------------===//
15666 // X86 Optimization Hooks
15667 //===----------------------------------------------------------------------===//
15669 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
15672 const SelectionDAG &DAG,
15673 unsigned Depth) const {
15674 unsigned BitWidth = KnownZero.getBitWidth();
15675 unsigned Opc = Op.getOpcode();
15676 assert((Opc >= ISD::BUILTIN_OP_END ||
15677 Opc == ISD::INTRINSIC_WO_CHAIN ||
15678 Opc == ISD::INTRINSIC_W_CHAIN ||
15679 Opc == ISD::INTRINSIC_VOID) &&
15680 "Should use MaskedValueIsZero if you don't know whether Op"
15681 " is a target node!");
15683 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
15697 // These nodes' second result is a boolean.
15698 if (Op.getResNo() == 0)
15701 case X86ISD::SETCC:
15702 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
15704 case ISD::INTRINSIC_WO_CHAIN: {
15705 unsigned IntId = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
15706 unsigned NumLoBits = 0;
15709 case Intrinsic::x86_sse_movmsk_ps:
15710 case Intrinsic::x86_avx_movmsk_ps_256:
15711 case Intrinsic::x86_sse2_movmsk_pd:
15712 case Intrinsic::x86_avx_movmsk_pd_256:
15713 case Intrinsic::x86_mmx_pmovmskb:
15714 case Intrinsic::x86_sse2_pmovmskb_128:
15715 case Intrinsic::x86_avx2_pmovmskb: {
15716 // High bits of movmskp{s|d}, pmovmskb are known zero.
15718 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
15719 case Intrinsic::x86_sse_movmsk_ps: NumLoBits = 4; break;
15720 case Intrinsic::x86_avx_movmsk_ps_256: NumLoBits = 8; break;
15721 case Intrinsic::x86_sse2_movmsk_pd: NumLoBits = 2; break;
15722 case Intrinsic::x86_avx_movmsk_pd_256: NumLoBits = 4; break;
15723 case Intrinsic::x86_mmx_pmovmskb: NumLoBits = 8; break;
15724 case Intrinsic::x86_sse2_pmovmskb_128: NumLoBits = 16; break;
15725 case Intrinsic::x86_avx2_pmovmskb: NumLoBits = 32; break;
15727 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - NumLoBits);
15736 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
15737 unsigned Depth) const {
15738 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
15739 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
15740 return Op.getValueType().getScalarType().getSizeInBits();
15746 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
15747 /// node is a GlobalAddress + offset.
15748 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
15749 const GlobalValue* &GA,
15750 int64_t &Offset) const {
15751 if (N->getOpcode() == X86ISD::Wrapper) {
15752 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
15753 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
15754 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
15758 return TargetLowering::isGAPlusOffset(N, GA, Offset);
15761 /// isShuffleHigh128VectorInsertLow - Checks whether the shuffle node is the
15762 /// same as extracting the high 128-bit part of 256-bit vector and then
15763 /// inserting the result into the low part of a new 256-bit vector
15764 static bool isShuffleHigh128VectorInsertLow(ShuffleVectorSDNode *SVOp) {
15765 EVT VT = SVOp->getValueType(0);
15766 unsigned NumElems = VT.getVectorNumElements();
15768 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
15769 for (unsigned i = 0, j = NumElems/2; i != NumElems/2; ++i, ++j)
15770 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
15771 SVOp->getMaskElt(j) >= 0)
15777 /// isShuffleLow128VectorInsertHigh - Checks whether the shuffle node is the
15778 /// same as extracting the low 128-bit part of 256-bit vector and then
15779 /// inserting the result into the high part of a new 256-bit vector
15780 static bool isShuffleLow128VectorInsertHigh(ShuffleVectorSDNode *SVOp) {
15781 EVT VT = SVOp->getValueType(0);
15782 unsigned NumElems = VT.getVectorNumElements();
15784 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
15785 for (unsigned i = NumElems/2, j = 0; i != NumElems; ++i, ++j)
15786 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
15787 SVOp->getMaskElt(j) >= 0)
15793 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
15794 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
15795 TargetLowering::DAGCombinerInfo &DCI,
15796 const X86Subtarget* Subtarget) {
15798 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
15799 SDValue V1 = SVOp->getOperand(0);
15800 SDValue V2 = SVOp->getOperand(1);
15801 EVT VT = SVOp->getValueType(0);
15802 unsigned NumElems = VT.getVectorNumElements();
15804 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
15805 V2.getOpcode() == ISD::CONCAT_VECTORS) {
15809 // V UNDEF BUILD_VECTOR UNDEF
15811 // CONCAT_VECTOR CONCAT_VECTOR
15814 // RESULT: V + zero extended
15816 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
15817 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
15818 V1.getOperand(1).getOpcode() != ISD::UNDEF)
15821 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
15824 // To match the shuffle mask, the first half of the mask should
15825 // be exactly the first vector, and all the rest a splat with the
15826 // first element of the second one.
15827 for (unsigned i = 0; i != NumElems/2; ++i)
15828 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
15829 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
15832 // If V1 is coming from a vector load then just fold to a VZEXT_LOAD.
15833 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(V1.getOperand(0))) {
15834 if (Ld->hasNUsesOfValue(1, 0)) {
15835 SDVTList Tys = DAG.getVTList(MVT::v4i64, MVT::Other);
15836 SDValue Ops[] = { Ld->getChain(), Ld->getBasePtr() };
15838 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops,
15839 array_lengthof(Ops),
15841 Ld->getPointerInfo(),
15842 Ld->getAlignment(),
15843 false/*isVolatile*/, true/*ReadMem*/,
15844 false/*WriteMem*/);
15846 // Make sure the newly-created LOAD is in the same position as Ld in
15847 // terms of dependency. We create a TokenFactor for Ld and ResNode,
15848 // and update uses of Ld's output chain to use the TokenFactor.
15849 if (Ld->hasAnyUseOfValue(1)) {
15850 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
15851 SDValue(Ld, 1), SDValue(ResNode.getNode(), 1));
15852 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), NewChain);
15853 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(Ld, 1),
15854 SDValue(ResNode.getNode(), 1));
15857 return DAG.getNode(ISD::BITCAST, dl, VT, ResNode);
15861 // Emit a zeroed vector and insert the desired subvector on its
15863 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
15864 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0), 0, DAG, dl);
15865 return DCI.CombineTo(N, InsV);
15868 //===--------------------------------------------------------------------===//
15869 // Combine some shuffles into subvector extracts and inserts:
15872 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
15873 if (isShuffleHigh128VectorInsertLow(SVOp)) {
15874 SDValue V = Extract128BitVector(V1, NumElems/2, DAG, dl);
15875 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, 0, DAG, dl);
15876 return DCI.CombineTo(N, InsV);
15879 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
15880 if (isShuffleLow128VectorInsertHigh(SVOp)) {
15881 SDValue V = Extract128BitVector(V1, 0, DAG, dl);
15882 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, NumElems/2, DAG, dl);
15883 return DCI.CombineTo(N, InsV);
15889 /// PerformShuffleCombine - Performs several different shuffle combines.
15890 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
15891 TargetLowering::DAGCombinerInfo &DCI,
15892 const X86Subtarget *Subtarget) {
15894 EVT VT = N->getValueType(0);
15896 // Don't create instructions with illegal types after legalize types has run.
15897 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15898 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
15901 // Combine 256-bit vector shuffles. This is only profitable when in AVX mode
15902 if (Subtarget->hasFp256() && VT.is256BitVector() &&
15903 N->getOpcode() == ISD::VECTOR_SHUFFLE)
15904 return PerformShuffleCombine256(N, DAG, DCI, Subtarget);
15906 // Only handle 128 wide vector from here on.
15907 if (!VT.is128BitVector())
15910 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
15911 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
15912 // consecutive, non-overlapping, and in the right order.
15913 SmallVector<SDValue, 16> Elts;
15914 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
15915 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
15917 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
15920 /// PerformTruncateCombine - Converts truncate operation to
15921 /// a sequence of vector shuffle operations.
15922 /// It is possible when we truncate 256-bit vector to 128-bit vector
15923 static SDValue PerformTruncateCombine(SDNode *N, SelectionDAG &DAG,
15924 TargetLowering::DAGCombinerInfo &DCI,
15925 const X86Subtarget *Subtarget) {
15929 /// XFormVExtractWithShuffleIntoLoad - Check if a vector extract from a target
15930 /// specific shuffle of a load can be folded into a single element load.
15931 /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
15932 /// shuffles have been customed lowered so we need to handle those here.
15933 static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
15934 TargetLowering::DAGCombinerInfo &DCI) {
15935 if (DCI.isBeforeLegalizeOps())
15938 SDValue InVec = N->getOperand(0);
15939 SDValue EltNo = N->getOperand(1);
15941 if (!isa<ConstantSDNode>(EltNo))
15944 EVT VT = InVec.getValueType();
15946 bool HasShuffleIntoBitcast = false;
15947 if (InVec.getOpcode() == ISD::BITCAST) {
15948 // Don't duplicate a load with other uses.
15949 if (!InVec.hasOneUse())
15951 EVT BCVT = InVec.getOperand(0).getValueType();
15952 if (BCVT.getVectorNumElements() != VT.getVectorNumElements())
15954 InVec = InVec.getOperand(0);
15955 HasShuffleIntoBitcast = true;
15958 if (!isTargetShuffle(InVec.getOpcode()))
15961 // Don't duplicate a load with other uses.
15962 if (!InVec.hasOneUse())
15965 SmallVector<int, 16> ShuffleMask;
15967 if (!getTargetShuffleMask(InVec.getNode(), VT.getSimpleVT(), ShuffleMask,
15971 // Select the input vector, guarding against out of range extract vector.
15972 unsigned NumElems = VT.getVectorNumElements();
15973 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
15974 int Idx = (Elt > (int)NumElems) ? -1 : ShuffleMask[Elt];
15975 SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0)
15976 : InVec.getOperand(1);
15978 // If inputs to shuffle are the same for both ops, then allow 2 uses
15979 unsigned AllowedUses = InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
15981 if (LdNode.getOpcode() == ISD::BITCAST) {
15982 // Don't duplicate a load with other uses.
15983 if (!LdNode.getNode()->hasNUsesOfValue(AllowedUses, 0))
15986 AllowedUses = 1; // only allow 1 load use if we have a bitcast
15987 LdNode = LdNode.getOperand(0);
15990 if (!ISD::isNormalLoad(LdNode.getNode()))
15993 LoadSDNode *LN0 = cast<LoadSDNode>(LdNode);
15995 if (!LN0 ||!LN0->hasNUsesOfValue(AllowedUses, 0) || LN0->isVolatile())
15998 if (HasShuffleIntoBitcast) {
15999 // If there's a bitcast before the shuffle, check if the load type and
16000 // alignment is valid.
16001 unsigned Align = LN0->getAlignment();
16002 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16003 unsigned NewAlign = TLI.getDataLayout()->
16004 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
16006 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, VT))
16010 // All checks match so transform back to vector_shuffle so that DAG combiner
16011 // can finish the job
16014 // Create shuffle node taking into account the case that its a unary shuffle
16015 SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(VT) : InVec.getOperand(1);
16016 Shuffle = DAG.getVectorShuffle(InVec.getValueType(), dl,
16017 InVec.getOperand(0), Shuffle,
16019 Shuffle = DAG.getNode(ISD::BITCAST, dl, VT, Shuffle);
16020 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle,
16024 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
16025 /// generation and convert it from being a bunch of shuffles and extracts
16026 /// to a simple store and scalar loads to extract the elements.
16027 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
16028 TargetLowering::DAGCombinerInfo &DCI) {
16029 SDValue NewOp = XFormVExtractWithShuffleIntoLoad(N, DAG, DCI);
16030 if (NewOp.getNode())
16033 SDValue InputVector = N->getOperand(0);
16034 // Detect whether we are trying to convert from mmx to i32 and the bitcast
16035 // from mmx to v2i32 has a single usage.
16036 if (InputVector.getNode()->getOpcode() == llvm::ISD::BITCAST &&
16037 InputVector.getNode()->getOperand(0).getValueType() == MVT::x86mmx &&
16038 InputVector.hasOneUse() && N->getValueType(0) == MVT::i32)
16039 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
16040 N->getValueType(0),
16041 InputVector.getNode()->getOperand(0));
16043 // Only operate on vectors of 4 elements, where the alternative shuffling
16044 // gets to be more expensive.
16045 if (InputVector.getValueType() != MVT::v4i32)
16048 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
16049 // single use which is a sign-extend or zero-extend, and all elements are
16051 SmallVector<SDNode *, 4> Uses;
16052 unsigned ExtractedElements = 0;
16053 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
16054 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
16055 if (UI.getUse().getResNo() != InputVector.getResNo())
16058 SDNode *Extract = *UI;
16059 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
16062 if (Extract->getValueType(0) != MVT::i32)
16064 if (!Extract->hasOneUse())
16066 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
16067 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
16069 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
16072 // Record which element was extracted.
16073 ExtractedElements |=
16074 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
16076 Uses.push_back(Extract);
16079 // If not all the elements were used, this may not be worthwhile.
16080 if (ExtractedElements != 15)
16083 // Ok, we've now decided to do the transformation.
16084 SDLoc dl(InputVector);
16086 // Store the value to a temporary stack slot.
16087 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
16088 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
16089 MachinePointerInfo(), false, false, 0);
16091 // Replace each use (extract) with a load of the appropriate element.
16092 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
16093 UE = Uses.end(); UI != UE; ++UI) {
16094 SDNode *Extract = *UI;
16096 // cOMpute the element's address.
16097 SDValue Idx = Extract->getOperand(1);
16099 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
16100 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
16101 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16102 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
16104 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(),
16105 StackPtr, OffsetVal);
16107 // Load the scalar.
16108 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch,
16109 ScalarAddr, MachinePointerInfo(),
16110 false, false, false, 0);
16112 // Replace the exact with the load.
16113 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
16116 // The replacement was made in place; don't return anything.
16120 /// \brief Matches a VSELECT onto min/max or return 0 if the node doesn't match.
16121 static unsigned matchIntegerMINMAX(SDValue Cond, EVT VT, SDValue LHS,
16122 SDValue RHS, SelectionDAG &DAG,
16123 const X86Subtarget *Subtarget) {
16124 if (!VT.isVector())
16127 switch (VT.getSimpleVT().SimpleTy) {
16132 if (!Subtarget->hasAVX2())
16137 if (!Subtarget->hasSSE2())
16141 // SSE2 has only a small subset of the operations.
16142 bool hasUnsigned = Subtarget->hasSSE41() ||
16143 (Subtarget->hasSSE2() && VT == MVT::v16i8);
16144 bool hasSigned = Subtarget->hasSSE41() ||
16145 (Subtarget->hasSSE2() && VT == MVT::v8i16);
16147 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
16149 // Check for x CC y ? x : y.
16150 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
16151 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
16156 return hasUnsigned ? X86ISD::UMIN : 0;
16159 return hasUnsigned ? X86ISD::UMAX : 0;
16162 return hasSigned ? X86ISD::SMIN : 0;
16165 return hasSigned ? X86ISD::SMAX : 0;
16167 // Check for x CC y ? y : x -- a min/max with reversed arms.
16168 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
16169 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
16174 return hasUnsigned ? X86ISD::UMAX : 0;
16177 return hasUnsigned ? X86ISD::UMIN : 0;
16180 return hasSigned ? X86ISD::SMAX : 0;
16183 return hasSigned ? X86ISD::SMIN : 0;
16190 /// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT
16192 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
16193 TargetLowering::DAGCombinerInfo &DCI,
16194 const X86Subtarget *Subtarget) {
16196 SDValue Cond = N->getOperand(0);
16197 // Get the LHS/RHS of the select.
16198 SDValue LHS = N->getOperand(1);
16199 SDValue RHS = N->getOperand(2);
16200 EVT VT = LHS.getValueType();
16202 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
16203 // instructions match the semantics of the common C idiom x<y?x:y but not
16204 // x<=y?x:y, because of how they handle negative zero (which can be
16205 // ignored in unsafe-math mode).
16206 if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() &&
16207 VT != MVT::f80 && DAG.getTargetLoweringInfo().isTypeLegal(VT) &&
16208 (Subtarget->hasSSE2() ||
16209 (Subtarget->hasSSE1() && VT.getScalarType() == MVT::f32))) {
16210 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
16212 unsigned Opcode = 0;
16213 // Check for x CC y ? x : y.
16214 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
16215 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
16219 // Converting this to a min would handle NaNs incorrectly, and swapping
16220 // the operands would cause it to handle comparisons between positive
16221 // and negative zero incorrectly.
16222 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
16223 if (!DAG.getTarget().Options.UnsafeFPMath &&
16224 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
16226 std::swap(LHS, RHS);
16228 Opcode = X86ISD::FMIN;
16231 // Converting this to a min would handle comparisons between positive
16232 // and negative zero incorrectly.
16233 if (!DAG.getTarget().Options.UnsafeFPMath &&
16234 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
16236 Opcode = X86ISD::FMIN;
16239 // Converting this to a min would handle both negative zeros and NaNs
16240 // incorrectly, but we can swap the operands to fix both.
16241 std::swap(LHS, RHS);
16245 Opcode = X86ISD::FMIN;
16249 // Converting this to a max would handle comparisons between positive
16250 // and negative zero incorrectly.
16251 if (!DAG.getTarget().Options.UnsafeFPMath &&
16252 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
16254 Opcode = X86ISD::FMAX;
16257 // Converting this to a max would handle NaNs incorrectly, and swapping
16258 // the operands would cause it to handle comparisons between positive
16259 // and negative zero incorrectly.
16260 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
16261 if (!DAG.getTarget().Options.UnsafeFPMath &&
16262 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
16264 std::swap(LHS, RHS);
16266 Opcode = X86ISD::FMAX;
16269 // Converting this to a max would handle both negative zeros and NaNs
16270 // incorrectly, but we can swap the operands to fix both.
16271 std::swap(LHS, RHS);
16275 Opcode = X86ISD::FMAX;
16278 // Check for x CC y ? y : x -- a min/max with reversed arms.
16279 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
16280 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
16284 // Converting this to a min would handle comparisons between positive
16285 // and negative zero incorrectly, and swapping the operands would
16286 // cause it to handle NaNs incorrectly.
16287 if (!DAG.getTarget().Options.UnsafeFPMath &&
16288 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
16289 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
16291 std::swap(LHS, RHS);
16293 Opcode = X86ISD::FMIN;
16296 // Converting this to a min would handle NaNs incorrectly.
16297 if (!DAG.getTarget().Options.UnsafeFPMath &&
16298 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
16300 Opcode = X86ISD::FMIN;
16303 // Converting this to a min would handle both negative zeros and NaNs
16304 // incorrectly, but we can swap the operands to fix both.
16305 std::swap(LHS, RHS);
16309 Opcode = X86ISD::FMIN;
16313 // Converting this to a max would handle NaNs incorrectly.
16314 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
16316 Opcode = X86ISD::FMAX;
16319 // Converting this to a max would handle comparisons between positive
16320 // and negative zero incorrectly, and swapping the operands would
16321 // cause it to handle NaNs incorrectly.
16322 if (!DAG.getTarget().Options.UnsafeFPMath &&
16323 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
16324 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
16326 std::swap(LHS, RHS);
16328 Opcode = X86ISD::FMAX;
16331 // Converting this to a max would handle both negative zeros and NaNs
16332 // incorrectly, but we can swap the operands to fix both.
16333 std::swap(LHS, RHS);
16337 Opcode = X86ISD::FMAX;
16343 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
16346 if (Subtarget->hasAVX512() && VT.isVector() &&
16347 Cond.getValueType().getVectorElementType() == MVT::i1) {
16348 // v16i8 (select v16i1, v16i8, v16i8) does not have a proper
16349 // lowering on AVX-512. In this case we convert it to
16350 // v16i8 (select v16i8, v16i8, v16i8) and use AVX instruction.
16351 // The same situation for all 128 and 256-bit vectors of i8 and i16
16352 EVT OpVT = LHS.getValueType();
16353 if ((OpVT.is128BitVector() || OpVT.is256BitVector()) &&
16354 (OpVT.getVectorElementType() == MVT::i8 ||
16355 OpVT.getVectorElementType() == MVT::i16)) {
16356 Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, OpVT, Cond);
16357 DCI.AddToWorklist(Cond.getNode());
16358 return DAG.getNode(N->getOpcode(), DL, OpVT, Cond, LHS, RHS);
16363 // If this is a select between two integer constants, try to do some
16365 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
16366 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
16367 // Don't do this for crazy integer types.
16368 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
16369 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
16370 // so that TrueC (the true value) is larger than FalseC.
16371 bool NeedsCondInvert = false;
16373 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
16374 // Efficiently invertible.
16375 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
16376 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
16377 isa<ConstantSDNode>(Cond.getOperand(1))))) {
16378 NeedsCondInvert = true;
16379 std::swap(TrueC, FalseC);
16382 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
16383 if (FalseC->getAPIntValue() == 0 &&
16384 TrueC->getAPIntValue().isPowerOf2()) {
16385 if (NeedsCondInvert) // Invert the condition if needed.
16386 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
16387 DAG.getConstant(1, Cond.getValueType()));
16389 // Zero extend the condition if needed.
16390 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
16392 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
16393 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
16394 DAG.getConstant(ShAmt, MVT::i8));
16397 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
16398 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
16399 if (NeedsCondInvert) // Invert the condition if needed.
16400 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
16401 DAG.getConstant(1, Cond.getValueType()));
16403 // Zero extend the condition if needed.
16404 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
16405 FalseC->getValueType(0), Cond);
16406 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
16407 SDValue(FalseC, 0));
16410 // Optimize cases that will turn into an LEA instruction. This requires
16411 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
16412 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
16413 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
16414 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
16416 bool isFastMultiplier = false;
16418 switch ((unsigned char)Diff) {
16420 case 1: // result = add base, cond
16421 case 2: // result = lea base( , cond*2)
16422 case 3: // result = lea base(cond, cond*2)
16423 case 4: // result = lea base( , cond*4)
16424 case 5: // result = lea base(cond, cond*4)
16425 case 8: // result = lea base( , cond*8)
16426 case 9: // result = lea base(cond, cond*8)
16427 isFastMultiplier = true;
16432 if (isFastMultiplier) {
16433 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
16434 if (NeedsCondInvert) // Invert the condition if needed.
16435 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
16436 DAG.getConstant(1, Cond.getValueType()));
16438 // Zero extend the condition if needed.
16439 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
16441 // Scale the condition by the difference.
16443 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
16444 DAG.getConstant(Diff, Cond.getValueType()));
16446 // Add the base if non-zero.
16447 if (FalseC->getAPIntValue() != 0)
16448 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
16449 SDValue(FalseC, 0));
16456 // Canonicalize max and min:
16457 // (x > y) ? x : y -> (x >= y) ? x : y
16458 // (x < y) ? x : y -> (x <= y) ? x : y
16459 // This allows use of COND_S / COND_NS (see TranslateX86CC) which eliminates
16460 // the need for an extra compare
16461 // against zero. e.g.
16462 // (x - y) > 0 : (x - y) ? 0 -> (x - y) >= 0 : (x - y) ? 0
16464 // testl %edi, %edi
16466 // cmovgl %edi, %eax
16470 // cmovsl %eax, %edi
16471 if (N->getOpcode() == ISD::SELECT && Cond.getOpcode() == ISD::SETCC &&
16472 DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
16473 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
16474 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
16479 ISD::CondCode NewCC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGE;
16480 Cond = DAG.getSetCC(SDLoc(Cond), Cond.getValueType(),
16481 Cond.getOperand(0), Cond.getOperand(1), NewCC);
16482 return DAG.getNode(ISD::SELECT, DL, VT, Cond, LHS, RHS);
16487 // Match VSELECTs into subs with unsigned saturation.
16488 if (!DCI.isBeforeLegalize() &&
16489 N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
16490 // psubus is available in SSE2 and AVX2 for i8 and i16 vectors.
16491 ((Subtarget->hasSSE2() && (VT == MVT::v16i8 || VT == MVT::v8i16)) ||
16492 (Subtarget->hasAVX2() && (VT == MVT::v32i8 || VT == MVT::v16i16)))) {
16493 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
16495 // Check if one of the arms of the VSELECT is a zero vector. If it's on the
16496 // left side invert the predicate to simplify logic below.
16498 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
16500 CC = ISD::getSetCCInverse(CC, true);
16501 } else if (ISD::isBuildVectorAllZeros(RHS.getNode())) {
16505 if (Other.getNode() && Other->getNumOperands() == 2 &&
16506 DAG.isEqualTo(Other->getOperand(0), Cond.getOperand(0))) {
16507 SDValue OpLHS = Other->getOperand(0), OpRHS = Other->getOperand(1);
16508 SDValue CondRHS = Cond->getOperand(1);
16510 // Look for a general sub with unsigned saturation first.
16511 // x >= y ? x-y : 0 --> subus x, y
16512 // x > y ? x-y : 0 --> subus x, y
16513 if ((CC == ISD::SETUGE || CC == ISD::SETUGT) &&
16514 Other->getOpcode() == ISD::SUB && DAG.isEqualTo(OpRHS, CondRHS))
16515 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
16517 // If the RHS is a constant we have to reverse the const canonicalization.
16518 // x > C-1 ? x+-C : 0 --> subus x, C
16519 if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD &&
16520 isSplatVector(CondRHS.getNode()) && isSplatVector(OpRHS.getNode())) {
16521 APInt A = cast<ConstantSDNode>(OpRHS.getOperand(0))->getAPIntValue();
16522 if (CondRHS.getConstantOperandVal(0) == -A-1)
16523 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS,
16524 DAG.getConstant(-A, VT));
16527 // Another special case: If C was a sign bit, the sub has been
16528 // canonicalized into a xor.
16529 // FIXME: Would it be better to use ComputeMaskedBits to determine whether
16530 // it's safe to decanonicalize the xor?
16531 // x s< 0 ? x^C : 0 --> subus x, C
16532 if (CC == ISD::SETLT && Other->getOpcode() == ISD::XOR &&
16533 ISD::isBuildVectorAllZeros(CondRHS.getNode()) &&
16534 isSplatVector(OpRHS.getNode())) {
16535 APInt A = cast<ConstantSDNode>(OpRHS.getOperand(0))->getAPIntValue();
16537 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
16542 // Try to match a min/max vector operation.
16543 if (!DCI.isBeforeLegalize() &&
16544 N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC)
16545 if (unsigned Op = matchIntegerMINMAX(Cond, VT, LHS, RHS, DAG, Subtarget))
16546 return DAG.getNode(Op, DL, N->getValueType(0), LHS, RHS);
16548 // Simplify vector selection if the selector will be produced by CMPP*/PCMP*.
16549 if (!DCI.isBeforeLegalize() && N->getOpcode() == ISD::VSELECT &&
16550 Cond.getOpcode() == ISD::SETCC) {
16552 assert(Cond.getValueType().isVector() &&
16553 "vector select expects a vector selector!");
16555 EVT IntVT = Cond.getValueType();
16556 bool TValIsAllOnes = ISD::isBuildVectorAllOnes(LHS.getNode());
16557 bool FValIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
16559 if (!TValIsAllOnes && !FValIsAllZeros) {
16560 // Try invert the condition if true value is not all 1s and false value
16562 bool TValIsAllZeros = ISD::isBuildVectorAllZeros(LHS.getNode());
16563 bool FValIsAllOnes = ISD::isBuildVectorAllOnes(RHS.getNode());
16565 if (TValIsAllZeros || FValIsAllOnes) {
16566 SDValue CC = Cond.getOperand(2);
16567 ISD::CondCode NewCC =
16568 ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
16569 Cond.getOperand(0).getValueType().isInteger());
16570 Cond = DAG.getSetCC(DL, IntVT, Cond.getOperand(0), Cond.getOperand(1), NewCC);
16571 std::swap(LHS, RHS);
16572 TValIsAllOnes = FValIsAllOnes;
16573 FValIsAllZeros = TValIsAllZeros;
16577 if (TValIsAllOnes || FValIsAllZeros) {
16580 if (TValIsAllOnes && FValIsAllZeros)
16582 else if (TValIsAllOnes)
16583 Ret = DAG.getNode(ISD::OR, DL, IntVT, Cond,
16584 DAG.getNode(ISD::BITCAST, DL, IntVT, RHS));
16585 else if (FValIsAllZeros)
16586 Ret = DAG.getNode(ISD::AND, DL, IntVT, Cond,
16587 DAG.getNode(ISD::BITCAST, DL, IntVT, LHS));
16589 return DAG.getNode(ISD::BITCAST, DL, VT, Ret);
16593 // If we know that this node is legal then we know that it is going to be
16594 // matched by one of the SSE/AVX BLEND instructions. These instructions only
16595 // depend on the highest bit in each word. Try to use SimplifyDemandedBits
16596 // to simplify previous instructions.
16597 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16598 if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() &&
16599 !DCI.isBeforeLegalize() && TLI.isOperationLegal(ISD::VSELECT, VT)) {
16600 unsigned BitWidth = Cond.getValueType().getScalarType().getSizeInBits();
16602 // Don't optimize vector selects that map to mask-registers.
16606 assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size");
16607 APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);
16609 APInt KnownZero, KnownOne;
16610 TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(),
16611 DCI.isBeforeLegalizeOps());
16612 if (TLO.ShrinkDemandedConstant(Cond, DemandedMask) ||
16613 TLI.SimplifyDemandedBits(Cond, DemandedMask, KnownZero, KnownOne, TLO))
16614 DCI.CommitTargetLoweringOpt(TLO);
16620 // Check whether a boolean test is testing a boolean value generated by
16621 // X86ISD::SETCC. If so, return the operand of that SETCC and proper condition
16624 // Simplify the following patterns:
16625 // (Op (CMP (SETCC Cond EFLAGS) 1) EQ) or
16626 // (Op (CMP (SETCC Cond EFLAGS) 0) NEQ)
16627 // to (Op EFLAGS Cond)
16629 // (Op (CMP (SETCC Cond EFLAGS) 0) EQ) or
16630 // (Op (CMP (SETCC Cond EFLAGS) 1) NEQ)
16631 // to (Op EFLAGS !Cond)
16633 // where Op could be BRCOND or CMOV.
16635 static SDValue checkBoolTestSetCCCombine(SDValue Cmp, X86::CondCode &CC) {
16636 // Quit if not CMP and SUB with its value result used.
16637 if (Cmp.getOpcode() != X86ISD::CMP &&
16638 (Cmp.getOpcode() != X86ISD::SUB || Cmp.getNode()->hasAnyUseOfValue(0)))
16641 // Quit if not used as a boolean value.
16642 if (CC != X86::COND_E && CC != X86::COND_NE)
16645 // Check CMP operands. One of them should be 0 or 1 and the other should be
16646 // an SetCC or extended from it.
16647 SDValue Op1 = Cmp.getOperand(0);
16648 SDValue Op2 = Cmp.getOperand(1);
16651 const ConstantSDNode* C = 0;
16652 bool needOppositeCond = (CC == X86::COND_E);
16653 bool checkAgainstTrue = false; // Is it a comparison against 1?
16655 if ((C = dyn_cast<ConstantSDNode>(Op1)))
16657 else if ((C = dyn_cast<ConstantSDNode>(Op2)))
16659 else // Quit if all operands are not constants.
16662 if (C->getZExtValue() == 1) {
16663 needOppositeCond = !needOppositeCond;
16664 checkAgainstTrue = true;
16665 } else if (C->getZExtValue() != 0)
16666 // Quit if the constant is neither 0 or 1.
16669 bool truncatedToBoolWithAnd = false;
16670 // Skip (zext $x), (trunc $x), or (and $x, 1) node.
16671 while (SetCC.getOpcode() == ISD::ZERO_EXTEND ||
16672 SetCC.getOpcode() == ISD::TRUNCATE ||
16673 SetCC.getOpcode() == ISD::AND) {
16674 if (SetCC.getOpcode() == ISD::AND) {
16676 ConstantSDNode *CS;
16677 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(0))) &&
16678 CS->getZExtValue() == 1)
16680 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(1))) &&
16681 CS->getZExtValue() == 1)
16685 SetCC = SetCC.getOperand(OpIdx);
16686 truncatedToBoolWithAnd = true;
16688 SetCC = SetCC.getOperand(0);
16691 switch (SetCC.getOpcode()) {
16692 case X86ISD::SETCC_CARRY:
16693 // Since SETCC_CARRY gives output based on R = CF ? ~0 : 0, it's unsafe to
16694 // simplify it if the result of SETCC_CARRY is not canonicalized to 0 or 1,
16695 // i.e. it's a comparison against true but the result of SETCC_CARRY is not
16696 // truncated to i1 using 'and'.
16697 if (checkAgainstTrue && !truncatedToBoolWithAnd)
16699 assert(X86::CondCode(SetCC.getConstantOperandVal(0)) == X86::COND_B &&
16700 "Invalid use of SETCC_CARRY!");
16702 case X86ISD::SETCC:
16703 // Set the condition code or opposite one if necessary.
16704 CC = X86::CondCode(SetCC.getConstantOperandVal(0));
16705 if (needOppositeCond)
16706 CC = X86::GetOppositeBranchCondition(CC);
16707 return SetCC.getOperand(1);
16708 case X86ISD::CMOV: {
16709 // Check whether false/true value has canonical one, i.e. 0 or 1.
16710 ConstantSDNode *FVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(0));
16711 ConstantSDNode *TVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(1));
16712 // Quit if true value is not a constant.
16715 // Quit if false value is not a constant.
16717 SDValue Op = SetCC.getOperand(0);
16718 // Skip 'zext' or 'trunc' node.
16719 if (Op.getOpcode() == ISD::ZERO_EXTEND ||
16720 Op.getOpcode() == ISD::TRUNCATE)
16721 Op = Op.getOperand(0);
16722 // A special case for rdrand/rdseed, where 0 is set if false cond is
16724 if ((Op.getOpcode() != X86ISD::RDRAND &&
16725 Op.getOpcode() != X86ISD::RDSEED) || Op.getResNo() != 0)
16728 // Quit if false value is not the constant 0 or 1.
16729 bool FValIsFalse = true;
16730 if (FVal && FVal->getZExtValue() != 0) {
16731 if (FVal->getZExtValue() != 1)
16733 // If FVal is 1, opposite cond is needed.
16734 needOppositeCond = !needOppositeCond;
16735 FValIsFalse = false;
16737 // Quit if TVal is not the constant opposite of FVal.
16738 if (FValIsFalse && TVal->getZExtValue() != 1)
16740 if (!FValIsFalse && TVal->getZExtValue() != 0)
16742 CC = X86::CondCode(SetCC.getConstantOperandVal(2));
16743 if (needOppositeCond)
16744 CC = X86::GetOppositeBranchCondition(CC);
16745 return SetCC.getOperand(3);
16752 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
16753 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
16754 TargetLowering::DAGCombinerInfo &DCI,
16755 const X86Subtarget *Subtarget) {
16758 // If the flag operand isn't dead, don't touch this CMOV.
16759 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
16762 SDValue FalseOp = N->getOperand(0);
16763 SDValue TrueOp = N->getOperand(1);
16764 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
16765 SDValue Cond = N->getOperand(3);
16767 if (CC == X86::COND_E || CC == X86::COND_NE) {
16768 switch (Cond.getOpcode()) {
16772 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
16773 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
16774 return (CC == X86::COND_E) ? FalseOp : TrueOp;
16780 Flags = checkBoolTestSetCCCombine(Cond, CC);
16781 if (Flags.getNode() &&
16782 // Extra check as FCMOV only supports a subset of X86 cond.
16783 (FalseOp.getValueType() != MVT::f80 || hasFPCMov(CC))) {
16784 SDValue Ops[] = { FalseOp, TrueOp,
16785 DAG.getConstant(CC, MVT::i8), Flags };
16786 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList(),
16787 Ops, array_lengthof(Ops));
16790 // If this is a select between two integer constants, try to do some
16791 // optimizations. Note that the operands are ordered the opposite of SELECT
16793 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
16794 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
16795 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
16796 // larger than FalseC (the false value).
16797 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
16798 CC = X86::GetOppositeBranchCondition(CC);
16799 std::swap(TrueC, FalseC);
16800 std::swap(TrueOp, FalseOp);
16803 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
16804 // This is efficient for any integer data type (including i8/i16) and
16806 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
16807 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
16808 DAG.getConstant(CC, MVT::i8), Cond);
16810 // Zero extend the condition if needed.
16811 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
16813 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
16814 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
16815 DAG.getConstant(ShAmt, MVT::i8));
16816 if (N->getNumValues() == 2) // Dead flag value?
16817 return DCI.CombineTo(N, Cond, SDValue());
16821 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
16822 // for any integer data type, including i8/i16.
16823 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
16824 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
16825 DAG.getConstant(CC, MVT::i8), Cond);
16827 // Zero extend the condition if needed.
16828 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
16829 FalseC->getValueType(0), Cond);
16830 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
16831 SDValue(FalseC, 0));
16833 if (N->getNumValues() == 2) // Dead flag value?
16834 return DCI.CombineTo(N, Cond, SDValue());
16838 // Optimize cases that will turn into an LEA instruction. This requires
16839 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
16840 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
16841 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
16842 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
16844 bool isFastMultiplier = false;
16846 switch ((unsigned char)Diff) {
16848 case 1: // result = add base, cond
16849 case 2: // result = lea base( , cond*2)
16850 case 3: // result = lea base(cond, cond*2)
16851 case 4: // result = lea base( , cond*4)
16852 case 5: // result = lea base(cond, cond*4)
16853 case 8: // result = lea base( , cond*8)
16854 case 9: // result = lea base(cond, cond*8)
16855 isFastMultiplier = true;
16860 if (isFastMultiplier) {
16861 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
16862 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
16863 DAG.getConstant(CC, MVT::i8), Cond);
16864 // Zero extend the condition if needed.
16865 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
16867 // Scale the condition by the difference.
16869 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
16870 DAG.getConstant(Diff, Cond.getValueType()));
16872 // Add the base if non-zero.
16873 if (FalseC->getAPIntValue() != 0)
16874 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
16875 SDValue(FalseC, 0));
16876 if (N->getNumValues() == 2) // Dead flag value?
16877 return DCI.CombineTo(N, Cond, SDValue());
16884 // Handle these cases:
16885 // (select (x != c), e, c) -> select (x != c), e, x),
16886 // (select (x == c), c, e) -> select (x == c), x, e)
16887 // where the c is an integer constant, and the "select" is the combination
16888 // of CMOV and CMP.
16890 // The rationale for this change is that the conditional-move from a constant
16891 // needs two instructions, however, conditional-move from a register needs
16892 // only one instruction.
16894 // CAVEAT: By replacing a constant with a symbolic value, it may obscure
16895 // some instruction-combining opportunities. This opt needs to be
16896 // postponed as late as possible.
16898 if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) {
16899 // the DCI.xxxx conditions are provided to postpone the optimization as
16900 // late as possible.
16902 ConstantSDNode *CmpAgainst = 0;
16903 if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) &&
16904 (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) &&
16905 !isa<ConstantSDNode>(Cond.getOperand(0))) {
16907 if (CC == X86::COND_NE &&
16908 CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) {
16909 CC = X86::GetOppositeBranchCondition(CC);
16910 std::swap(TrueOp, FalseOp);
16913 if (CC == X86::COND_E &&
16914 CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) {
16915 SDValue Ops[] = { FalseOp, Cond.getOperand(0),
16916 DAG.getConstant(CC, MVT::i8), Cond };
16917 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops,
16918 array_lengthof(Ops));
16926 /// PerformMulCombine - Optimize a single multiply with constant into two
16927 /// in order to implement it with two cheaper instructions, e.g.
16928 /// LEA + SHL, LEA + LEA.
16929 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
16930 TargetLowering::DAGCombinerInfo &DCI) {
16931 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
16934 EVT VT = N->getValueType(0);
16935 if (VT != MVT::i64)
16938 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
16941 uint64_t MulAmt = C->getZExtValue();
16942 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
16945 uint64_t MulAmt1 = 0;
16946 uint64_t MulAmt2 = 0;
16947 if ((MulAmt % 9) == 0) {
16949 MulAmt2 = MulAmt / 9;
16950 } else if ((MulAmt % 5) == 0) {
16952 MulAmt2 = MulAmt / 5;
16953 } else if ((MulAmt % 3) == 0) {
16955 MulAmt2 = MulAmt / 3;
16958 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
16961 if (isPowerOf2_64(MulAmt2) &&
16962 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
16963 // If second multiplifer is pow2, issue it first. We want the multiply by
16964 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
16966 std::swap(MulAmt1, MulAmt2);
16969 if (isPowerOf2_64(MulAmt1))
16970 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
16971 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
16973 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
16974 DAG.getConstant(MulAmt1, VT));
16976 if (isPowerOf2_64(MulAmt2))
16977 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
16978 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
16980 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
16981 DAG.getConstant(MulAmt2, VT));
16983 // Do not add new nodes to DAG combiner worklist.
16984 DCI.CombineTo(N, NewMul, false);
16989 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
16990 SDValue N0 = N->getOperand(0);
16991 SDValue N1 = N->getOperand(1);
16992 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
16993 EVT VT = N0.getValueType();
16995 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
16996 // since the result of setcc_c is all zero's or all ones.
16997 if (VT.isInteger() && !VT.isVector() &&
16998 N1C && N0.getOpcode() == ISD::AND &&
16999 N0.getOperand(1).getOpcode() == ISD::Constant) {
17000 SDValue N00 = N0.getOperand(0);
17001 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
17002 ((N00.getOpcode() == ISD::ANY_EXTEND ||
17003 N00.getOpcode() == ISD::ZERO_EXTEND) &&
17004 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
17005 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
17006 APInt ShAmt = N1C->getAPIntValue();
17007 Mask = Mask.shl(ShAmt);
17009 return DAG.getNode(ISD::AND, SDLoc(N), VT,
17010 N00, DAG.getConstant(Mask, VT));
17014 // Hardware support for vector shifts is sparse which makes us scalarize the
17015 // vector operations in many cases. Also, on sandybridge ADD is faster than
17017 // (shl V, 1) -> add V,V
17018 if (isSplatVector(N1.getNode())) {
17019 assert(N0.getValueType().isVector() && "Invalid vector shift type");
17020 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1->getOperand(0));
17021 // We shift all of the values by one. In many cases we do not have
17022 // hardware support for this operation. This is better expressed as an ADD
17024 if (N1C && (1 == N1C->getZExtValue())) {
17025 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N0);
17032 /// \brief Returns a vector of 0s if the node in input is a vector logical
17033 /// shift by a constant amount which is known to be bigger than or equal
17034 /// to the vector element size in bits.
17035 static SDValue performShiftToAllZeros(SDNode *N, SelectionDAG &DAG,
17036 const X86Subtarget *Subtarget) {
17037 EVT VT = N->getValueType(0);
17039 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16 &&
17040 (!Subtarget->hasInt256() ||
17041 (VT != MVT::v4i64 && VT != MVT::v8i32 && VT != MVT::v16i16)))
17044 SDValue Amt = N->getOperand(1);
17046 if (isSplatVector(Amt.getNode())) {
17047 SDValue SclrAmt = Amt->getOperand(0);
17048 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(SclrAmt)) {
17049 APInt ShiftAmt = C->getAPIntValue();
17050 unsigned MaxAmount = VT.getVectorElementType().getSizeInBits();
17052 // SSE2/AVX2 logical shifts always return a vector of 0s
17053 // if the shift amount is bigger than or equal to
17054 // the element size. The constant shift amount will be
17055 // encoded as a 8-bit immediate.
17056 if (ShiftAmt.trunc(8).uge(MaxAmount))
17057 return getZeroVector(VT, Subtarget, DAG, DL);
17064 /// PerformShiftCombine - Combine shifts.
17065 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
17066 TargetLowering::DAGCombinerInfo &DCI,
17067 const X86Subtarget *Subtarget) {
17068 if (N->getOpcode() == ISD::SHL) {
17069 SDValue V = PerformSHLCombine(N, DAG);
17070 if (V.getNode()) return V;
17073 if (N->getOpcode() != ISD::SRA) {
17074 // Try to fold this logical shift into a zero vector.
17075 SDValue V = performShiftToAllZeros(N, DAG, Subtarget);
17076 if (V.getNode()) return V;
17082 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
17083 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
17084 // and friends. Likewise for OR -> CMPNEQSS.
17085 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
17086 TargetLowering::DAGCombinerInfo &DCI,
17087 const X86Subtarget *Subtarget) {
17090 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
17091 // we're requiring SSE2 for both.
17092 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
17093 SDValue N0 = N->getOperand(0);
17094 SDValue N1 = N->getOperand(1);
17095 SDValue CMP0 = N0->getOperand(1);
17096 SDValue CMP1 = N1->getOperand(1);
17099 // The SETCCs should both refer to the same CMP.
17100 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
17103 SDValue CMP00 = CMP0->getOperand(0);
17104 SDValue CMP01 = CMP0->getOperand(1);
17105 EVT VT = CMP00.getValueType();
17107 if (VT == MVT::f32 || VT == MVT::f64) {
17108 bool ExpectingFlags = false;
17109 // Check for any users that want flags:
17110 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
17111 !ExpectingFlags && UI != UE; ++UI)
17112 switch (UI->getOpcode()) {
17117 ExpectingFlags = true;
17119 case ISD::CopyToReg:
17120 case ISD::SIGN_EXTEND:
17121 case ISD::ZERO_EXTEND:
17122 case ISD::ANY_EXTEND:
17126 if (!ExpectingFlags) {
17127 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
17128 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
17130 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
17131 X86::CondCode tmp = cc0;
17136 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
17137 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
17138 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
17139 X86ISD::NodeType NTOperator = is64BitFP ?
17140 X86ISD::FSETCCsd : X86ISD::FSETCCss;
17141 // FIXME: need symbolic constants for these magic numbers.
17142 // See X86ATTInstPrinter.cpp:printSSECC().
17143 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
17144 SDValue OnesOrZeroesF = DAG.getNode(NTOperator, DL, MVT::f32, CMP00, CMP01,
17145 DAG.getConstant(x86cc, MVT::i8));
17146 SDValue OnesOrZeroesI = DAG.getNode(ISD::BITCAST, DL, MVT::i32,
17148 SDValue ANDed = DAG.getNode(ISD::AND, DL, MVT::i32, OnesOrZeroesI,
17149 DAG.getConstant(1, MVT::i32));
17150 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8, ANDed);
17151 return OneBitOfTruth;
17159 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
17160 /// so it can be folded inside ANDNP.
17161 static bool CanFoldXORWithAllOnes(const SDNode *N) {
17162 EVT VT = N->getValueType(0);
17164 // Match direct AllOnes for 128 and 256-bit vectors
17165 if (ISD::isBuildVectorAllOnes(N))
17168 // Look through a bit convert.
17169 if (N->getOpcode() == ISD::BITCAST)
17170 N = N->getOperand(0).getNode();
17172 // Sometimes the operand may come from a insert_subvector building a 256-bit
17174 if (VT.is256BitVector() &&
17175 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
17176 SDValue V1 = N->getOperand(0);
17177 SDValue V2 = N->getOperand(1);
17179 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
17180 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
17181 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
17182 ISD::isBuildVectorAllOnes(V2.getNode()))
17189 // On AVX/AVX2 the type v8i1 is legalized to v8i16, which is an XMM sized
17190 // register. In most cases we actually compare or select YMM-sized registers
17191 // and mixing the two types creates horrible code. This method optimizes
17192 // some of the transition sequences.
17193 static SDValue WidenMaskArithmetic(SDNode *N, SelectionDAG &DAG,
17194 TargetLowering::DAGCombinerInfo &DCI,
17195 const X86Subtarget *Subtarget) {
17196 EVT VT = N->getValueType(0);
17197 if (!VT.is256BitVector())
17200 assert((N->getOpcode() == ISD::ANY_EXTEND ||
17201 N->getOpcode() == ISD::ZERO_EXTEND ||
17202 N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node");
17204 SDValue Narrow = N->getOperand(0);
17205 EVT NarrowVT = Narrow->getValueType(0);
17206 if (!NarrowVT.is128BitVector())
17209 if (Narrow->getOpcode() != ISD::XOR &&
17210 Narrow->getOpcode() != ISD::AND &&
17211 Narrow->getOpcode() != ISD::OR)
17214 SDValue N0 = Narrow->getOperand(0);
17215 SDValue N1 = Narrow->getOperand(1);
17218 // The Left side has to be a trunc.
17219 if (N0.getOpcode() != ISD::TRUNCATE)
17222 // The type of the truncated inputs.
17223 EVT WideVT = N0->getOperand(0)->getValueType(0);
17227 // The right side has to be a 'trunc' or a constant vector.
17228 bool RHSTrunc = N1.getOpcode() == ISD::TRUNCATE;
17229 bool RHSConst = (isSplatVector(N1.getNode()) &&
17230 isa<ConstantSDNode>(N1->getOperand(0)));
17231 if (!RHSTrunc && !RHSConst)
17234 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
17236 if (!TLI.isOperationLegalOrPromote(Narrow->getOpcode(), WideVT))
17239 // Set N0 and N1 to hold the inputs to the new wide operation.
17240 N0 = N0->getOperand(0);
17242 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT.getScalarType(),
17243 N1->getOperand(0));
17244 SmallVector<SDValue, 8> C(WideVT.getVectorNumElements(), N1);
17245 N1 = DAG.getNode(ISD::BUILD_VECTOR, DL, WideVT, &C[0], C.size());
17246 } else if (RHSTrunc) {
17247 N1 = N1->getOperand(0);
17250 // Generate the wide operation.
17251 SDValue Op = DAG.getNode(Narrow->getOpcode(), DL, WideVT, N0, N1);
17252 unsigned Opcode = N->getOpcode();
17254 case ISD::ANY_EXTEND:
17256 case ISD::ZERO_EXTEND: {
17257 unsigned InBits = NarrowVT.getScalarType().getSizeInBits();
17258 APInt Mask = APInt::getAllOnesValue(InBits);
17259 Mask = Mask.zext(VT.getScalarType().getSizeInBits());
17260 return DAG.getNode(ISD::AND, DL, VT,
17261 Op, DAG.getConstant(Mask, VT));
17263 case ISD::SIGN_EXTEND:
17264 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT,
17265 Op, DAG.getValueType(NarrowVT));
17267 llvm_unreachable("Unexpected opcode");
17271 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
17272 TargetLowering::DAGCombinerInfo &DCI,
17273 const X86Subtarget *Subtarget) {
17274 EVT VT = N->getValueType(0);
17275 if (DCI.isBeforeLegalizeOps())
17278 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
17282 // Create BLSI, and BLSR instructions
17283 // BLSI is X & (-X)
17284 // BLSR is X & (X-1)
17285 if (Subtarget->hasBMI() && (VT == MVT::i32 || VT == MVT::i64)) {
17286 SDValue N0 = N->getOperand(0);
17287 SDValue N1 = N->getOperand(1);
17290 // Check LHS for neg
17291 if (N0.getOpcode() == ISD::SUB && N0.getOperand(1) == N1 &&
17292 isZero(N0.getOperand(0)))
17293 return DAG.getNode(X86ISD::BLSI, DL, VT, N1);
17295 // Check RHS for neg
17296 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1) == N0 &&
17297 isZero(N1.getOperand(0)))
17298 return DAG.getNode(X86ISD::BLSI, DL, VT, N0);
17300 // Check LHS for X-1
17301 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1 &&
17302 isAllOnes(N0.getOperand(1)))
17303 return DAG.getNode(X86ISD::BLSR, DL, VT, N1);
17305 // Check RHS for X-1
17306 if (N1.getOpcode() == ISD::ADD && N1.getOperand(0) == N0 &&
17307 isAllOnes(N1.getOperand(1)))
17308 return DAG.getNode(X86ISD::BLSR, DL, VT, N0);
17313 // Want to form ANDNP nodes:
17314 // 1) In the hopes of then easily combining them with OR and AND nodes
17315 // to form PBLEND/PSIGN.
17316 // 2) To match ANDN packed intrinsics
17317 if (VT != MVT::v2i64 && VT != MVT::v4i64)
17320 SDValue N0 = N->getOperand(0);
17321 SDValue N1 = N->getOperand(1);
17324 // Check LHS for vnot
17325 if (N0.getOpcode() == ISD::XOR &&
17326 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
17327 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
17328 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
17330 // Check RHS for vnot
17331 if (N1.getOpcode() == ISD::XOR &&
17332 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
17333 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
17334 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
17339 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
17340 TargetLowering::DAGCombinerInfo &DCI,
17341 const X86Subtarget *Subtarget) {
17342 EVT VT = N->getValueType(0);
17343 if (DCI.isBeforeLegalizeOps())
17346 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
17350 SDValue N0 = N->getOperand(0);
17351 SDValue N1 = N->getOperand(1);
17353 // look for psign/blend
17354 if (VT == MVT::v2i64 || VT == MVT::v4i64) {
17355 if (!Subtarget->hasSSSE3() ||
17356 (VT == MVT::v4i64 && !Subtarget->hasInt256()))
17359 // Canonicalize pandn to RHS
17360 if (N0.getOpcode() == X86ISD::ANDNP)
17362 // or (and (m, y), (pandn m, x))
17363 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
17364 SDValue Mask = N1.getOperand(0);
17365 SDValue X = N1.getOperand(1);
17367 if (N0.getOperand(0) == Mask)
17368 Y = N0.getOperand(1);
17369 if (N0.getOperand(1) == Mask)
17370 Y = N0.getOperand(0);
17372 // Check to see if the mask appeared in both the AND and ANDNP and
17376 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
17377 // Look through mask bitcast.
17378 if (Mask.getOpcode() == ISD::BITCAST)
17379 Mask = Mask.getOperand(0);
17380 if (X.getOpcode() == ISD::BITCAST)
17381 X = X.getOperand(0);
17382 if (Y.getOpcode() == ISD::BITCAST)
17383 Y = Y.getOperand(0);
17385 EVT MaskVT = Mask.getValueType();
17387 // Validate that the Mask operand is a vector sra node.
17388 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
17389 // there is no psrai.b
17390 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
17391 unsigned SraAmt = ~0;
17392 if (Mask.getOpcode() == ISD::SRA) {
17393 SDValue Amt = Mask.getOperand(1);
17394 if (isSplatVector(Amt.getNode())) {
17395 SDValue SclrAmt = Amt->getOperand(0);
17396 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(SclrAmt))
17397 SraAmt = C->getZExtValue();
17399 } else if (Mask.getOpcode() == X86ISD::VSRAI) {
17400 SDValue SraC = Mask.getOperand(1);
17401 SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
17403 if ((SraAmt + 1) != EltBits)
17408 // Now we know we at least have a plendvb with the mask val. See if
17409 // we can form a psignb/w/d.
17410 // psign = x.type == y.type == mask.type && y = sub(0, x);
17411 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
17412 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
17413 X.getValueType() == MaskVT && Y.getValueType() == MaskVT) {
17414 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
17415 "Unsupported VT for PSIGN");
17416 Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask.getOperand(0));
17417 return DAG.getNode(ISD::BITCAST, DL, VT, Mask);
17419 // PBLENDVB only available on SSE 4.1
17420 if (!Subtarget->hasSSE41())
17423 EVT BlendVT = (VT == MVT::v4i64) ? MVT::v32i8 : MVT::v16i8;
17425 X = DAG.getNode(ISD::BITCAST, DL, BlendVT, X);
17426 Y = DAG.getNode(ISD::BITCAST, DL, BlendVT, Y);
17427 Mask = DAG.getNode(ISD::BITCAST, DL, BlendVT, Mask);
17428 Mask = DAG.getNode(ISD::VSELECT, DL, BlendVT, Mask, Y, X);
17429 return DAG.getNode(ISD::BITCAST, DL, VT, Mask);
17433 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
17436 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
17437 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
17439 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
17441 if (!N0.hasOneUse() || !N1.hasOneUse())
17444 SDValue ShAmt0 = N0.getOperand(1);
17445 if (ShAmt0.getValueType() != MVT::i8)
17447 SDValue ShAmt1 = N1.getOperand(1);
17448 if (ShAmt1.getValueType() != MVT::i8)
17450 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
17451 ShAmt0 = ShAmt0.getOperand(0);
17452 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
17453 ShAmt1 = ShAmt1.getOperand(0);
17456 unsigned Opc = X86ISD::SHLD;
17457 SDValue Op0 = N0.getOperand(0);
17458 SDValue Op1 = N1.getOperand(0);
17459 if (ShAmt0.getOpcode() == ISD::SUB) {
17460 Opc = X86ISD::SHRD;
17461 std::swap(Op0, Op1);
17462 std::swap(ShAmt0, ShAmt1);
17465 unsigned Bits = VT.getSizeInBits();
17466 if (ShAmt1.getOpcode() == ISD::SUB) {
17467 SDValue Sum = ShAmt1.getOperand(0);
17468 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
17469 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
17470 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
17471 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
17472 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
17473 return DAG.getNode(Opc, DL, VT,
17475 DAG.getNode(ISD::TRUNCATE, DL,
17478 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
17479 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
17481 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
17482 return DAG.getNode(Opc, DL, VT,
17483 N0.getOperand(0), N1.getOperand(0),
17484 DAG.getNode(ISD::TRUNCATE, DL,
17491 // Generate NEG and CMOV for integer abs.
17492 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
17493 EVT VT = N->getValueType(0);
17495 // Since X86 does not have CMOV for 8-bit integer, we don't convert
17496 // 8-bit integer abs to NEG and CMOV.
17497 if (VT.isInteger() && VT.getSizeInBits() == 8)
17500 SDValue N0 = N->getOperand(0);
17501 SDValue N1 = N->getOperand(1);
17504 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
17505 // and change it to SUB and CMOV.
17506 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
17507 N0.getOpcode() == ISD::ADD &&
17508 N0.getOperand(1) == N1 &&
17509 N1.getOpcode() == ISD::SRA &&
17510 N1.getOperand(0) == N0.getOperand(0))
17511 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
17512 if (Y1C->getAPIntValue() == VT.getSizeInBits()-1) {
17513 // Generate SUB & CMOV.
17514 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32),
17515 DAG.getConstant(0, VT), N0.getOperand(0));
17517 SDValue Ops[] = { N0.getOperand(0), Neg,
17518 DAG.getConstant(X86::COND_GE, MVT::i8),
17519 SDValue(Neg.getNode(), 1) };
17520 return DAG.getNode(X86ISD::CMOV, DL, DAG.getVTList(VT, MVT::Glue),
17521 Ops, array_lengthof(Ops));
17526 // PerformXorCombine - Attempts to turn XOR nodes into BLSMSK nodes
17527 static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
17528 TargetLowering::DAGCombinerInfo &DCI,
17529 const X86Subtarget *Subtarget) {
17530 EVT VT = N->getValueType(0);
17531 if (DCI.isBeforeLegalizeOps())
17534 if (Subtarget->hasCMov()) {
17535 SDValue RV = performIntegerAbsCombine(N, DAG);
17540 // Try forming BMI if it is available.
17541 if (!Subtarget->hasBMI())
17544 if (VT != MVT::i32 && VT != MVT::i64)
17547 assert(Subtarget->hasBMI() && "Creating BLSMSK requires BMI instructions");
17549 // Create BLSMSK instructions by finding X ^ (X-1)
17550 SDValue N0 = N->getOperand(0);
17551 SDValue N1 = N->getOperand(1);
17554 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1 &&
17555 isAllOnes(N0.getOperand(1)))
17556 return DAG.getNode(X86ISD::BLSMSK, DL, VT, N1);
17558 if (N1.getOpcode() == ISD::ADD && N1.getOperand(0) == N0 &&
17559 isAllOnes(N1.getOperand(1)))
17560 return DAG.getNode(X86ISD::BLSMSK, DL, VT, N0);
17565 /// PerformLOADCombine - Do target-specific dag combines on LOAD nodes.
17566 static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
17567 TargetLowering::DAGCombinerInfo &DCI,
17568 const X86Subtarget *Subtarget) {
17569 LoadSDNode *Ld = cast<LoadSDNode>(N);
17570 EVT RegVT = Ld->getValueType(0);
17571 EVT MemVT = Ld->getMemoryVT();
17573 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
17574 unsigned RegSz = RegVT.getSizeInBits();
17576 // On Sandybridge unaligned 256bit loads are inefficient.
17577 ISD::LoadExtType Ext = Ld->getExtensionType();
17578 unsigned Alignment = Ld->getAlignment();
17579 bool IsAligned = Alignment == 0 || Alignment >= MemVT.getSizeInBits()/8;
17580 if (RegVT.is256BitVector() && !Subtarget->hasInt256() &&
17581 !DCI.isBeforeLegalizeOps() && !IsAligned && Ext == ISD::NON_EXTLOAD) {
17582 unsigned NumElems = RegVT.getVectorNumElements();
17586 SDValue Ptr = Ld->getBasePtr();
17587 SDValue Increment = DAG.getConstant(16, TLI.getPointerTy());
17589 EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
17591 SDValue Load1 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
17592 Ld->getPointerInfo(), Ld->isVolatile(),
17593 Ld->isNonTemporal(), Ld->isInvariant(),
17595 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
17596 SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
17597 Ld->getPointerInfo(), Ld->isVolatile(),
17598 Ld->isNonTemporal(), Ld->isInvariant(),
17599 std::min(16U, Alignment));
17600 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
17602 Load2.getValue(1));
17604 SDValue NewVec = DAG.getUNDEF(RegVT);
17605 NewVec = Insert128BitVector(NewVec, Load1, 0, DAG, dl);
17606 NewVec = Insert128BitVector(NewVec, Load2, NumElems/2, DAG, dl);
17607 return DCI.CombineTo(N, NewVec, TF, true);
17610 // If this is a vector EXT Load then attempt to optimize it using a
17611 // shuffle. If SSSE3 is not available we may emit an illegal shuffle but the
17612 // expansion is still better than scalar code.
17613 // We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise we'll
17614 // emit a shuffle and a arithmetic shift.
17615 // TODO: It is possible to support ZExt by zeroing the undef values
17616 // during the shuffle phase or after the shuffle.
17617 if (RegVT.isVector() && RegVT.isInteger() && Subtarget->hasSSE2() &&
17618 (Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)) {
17619 assert(MemVT != RegVT && "Cannot extend to the same type");
17620 assert(MemVT.isVector() && "Must load a vector from memory");
17622 unsigned NumElems = RegVT.getVectorNumElements();
17623 unsigned MemSz = MemVT.getSizeInBits();
17624 assert(RegSz > MemSz && "Register size must be greater than the mem size");
17626 if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256())
17629 // All sizes must be a power of two.
17630 if (!isPowerOf2_32(RegSz * MemSz * NumElems))
17633 // Attempt to load the original value using scalar loads.
17634 // Find the largest scalar type that divides the total loaded size.
17635 MVT SclrLoadTy = MVT::i8;
17636 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
17637 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
17638 MVT Tp = (MVT::SimpleValueType)tp;
17639 if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
17644 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
17645 if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
17647 SclrLoadTy = MVT::f64;
17649 // Calculate the number of scalar loads that we need to perform
17650 // in order to load our vector from memory.
17651 unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
17652 if (Ext == ISD::SEXTLOAD && NumLoads > 1)
17655 unsigned loadRegZize = RegSz;
17656 if (Ext == ISD::SEXTLOAD && RegSz == 256)
17659 // Represent our vector as a sequence of elements which are the
17660 // largest scalar that we can load.
17661 EVT LoadUnitVecVT = EVT::getVectorVT(*DAG.getContext(), SclrLoadTy,
17662 loadRegZize/SclrLoadTy.getSizeInBits());
17664 // Represent the data using the same element type that is stored in
17665 // memory. In practice, we ''widen'' MemVT.
17667 EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
17668 loadRegZize/MemVT.getScalarType().getSizeInBits());
17670 assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
17671 "Invalid vector type");
17673 // We can't shuffle using an illegal type.
17674 if (!TLI.isTypeLegal(WideVecVT))
17677 SmallVector<SDValue, 8> Chains;
17678 SDValue Ptr = Ld->getBasePtr();
17679 SDValue Increment = DAG.getConstant(SclrLoadTy.getSizeInBits()/8,
17680 TLI.getPointerTy());
17681 SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
17683 for (unsigned i = 0; i < NumLoads; ++i) {
17684 // Perform a single load.
17685 SDValue ScalarLoad = DAG.getLoad(SclrLoadTy, dl, Ld->getChain(),
17686 Ptr, Ld->getPointerInfo(),
17687 Ld->isVolatile(), Ld->isNonTemporal(),
17688 Ld->isInvariant(), Ld->getAlignment());
17689 Chains.push_back(ScalarLoad.getValue(1));
17690 // Create the first element type using SCALAR_TO_VECTOR in order to avoid
17691 // another round of DAGCombining.
17693 Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
17695 Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
17696 ScalarLoad, DAG.getIntPtrConstant(i));
17698 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
17701 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Chains[0],
17704 // Bitcast the loaded value to a vector of the original element type, in
17705 // the size of the target vector type.
17706 SDValue SlicedVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Res);
17707 unsigned SizeRatio = RegSz/MemSz;
17709 if (Ext == ISD::SEXTLOAD) {
17710 // If we have SSE4.1 we can directly emit a VSEXT node.
17711 if (Subtarget->hasSSE41()) {
17712 SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
17713 return DCI.CombineTo(N, Sext, TF, true);
17716 // Otherwise we'll shuffle the small elements in the high bits of the
17717 // larger type and perform an arithmetic shift. If the shift is not legal
17718 // it's better to scalarize.
17719 if (!TLI.isOperationLegalOrCustom(ISD::SRA, RegVT))
17722 // Redistribute the loaded elements into the different locations.
17723 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
17724 for (unsigned i = 0; i != NumElems; ++i)
17725 ShuffleVec[i*SizeRatio + SizeRatio-1] = i;
17727 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
17728 DAG.getUNDEF(WideVecVT),
17731 Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
17733 // Build the arithmetic shift.
17734 unsigned Amt = RegVT.getVectorElementType().getSizeInBits() -
17735 MemVT.getVectorElementType().getSizeInBits();
17736 Shuff = DAG.getNode(ISD::SRA, dl, RegVT, Shuff,
17737 DAG.getConstant(Amt, RegVT));
17739 return DCI.CombineTo(N, Shuff, TF, true);
17742 // Redistribute the loaded elements into the different locations.
17743 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
17744 for (unsigned i = 0; i != NumElems; ++i)
17745 ShuffleVec[i*SizeRatio] = i;
17747 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
17748 DAG.getUNDEF(WideVecVT),
17751 // Bitcast to the requested type.
17752 Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
17753 // Replace the original load with the new sequence
17754 // and return the new chain.
17755 return DCI.CombineTo(N, Shuff, TF, true);
17761 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
17762 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
17763 const X86Subtarget *Subtarget) {
17764 StoreSDNode *St = cast<StoreSDNode>(N);
17765 EVT VT = St->getValue().getValueType();
17766 EVT StVT = St->getMemoryVT();
17768 SDValue StoredVal = St->getOperand(1);
17769 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
17771 // If we are saving a concatenation of two XMM registers, perform two stores.
17772 // On Sandy Bridge, 256-bit memory operations are executed by two
17773 // 128-bit ports. However, on Haswell it is better to issue a single 256-bit
17774 // memory operation.
17775 unsigned Alignment = St->getAlignment();
17776 bool IsAligned = Alignment == 0 || Alignment >= VT.getSizeInBits()/8;
17777 if (VT.is256BitVector() && !Subtarget->hasInt256() &&
17778 StVT == VT && !IsAligned) {
17779 unsigned NumElems = VT.getVectorNumElements();
17783 SDValue Value0 = Extract128BitVector(StoredVal, 0, DAG, dl);
17784 SDValue Value1 = Extract128BitVector(StoredVal, NumElems/2, DAG, dl);
17786 SDValue Stride = DAG.getConstant(16, TLI.getPointerTy());
17787 SDValue Ptr0 = St->getBasePtr();
17788 SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
17790 SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
17791 St->getPointerInfo(), St->isVolatile(),
17792 St->isNonTemporal(), Alignment);
17793 SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
17794 St->getPointerInfo(), St->isVolatile(),
17795 St->isNonTemporal(),
17796 std::min(16U, Alignment));
17797 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
17800 // Optimize trunc store (of multiple scalars) to shuffle and store.
17801 // First, pack all of the elements in one place. Next, store to memory
17802 // in fewer chunks.
17803 if (St->isTruncatingStore() && VT.isVector()) {
17804 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
17805 unsigned NumElems = VT.getVectorNumElements();
17806 assert(StVT != VT && "Cannot truncate to the same type");
17807 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
17808 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
17810 // From, To sizes and ElemCount must be pow of two
17811 if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
17812 // We are going to use the original vector elt for storing.
17813 // Accumulated smaller vector elements must be a multiple of the store size.
17814 if (0 != (NumElems * FromSz) % ToSz) return SDValue();
17816 unsigned SizeRatio = FromSz / ToSz;
17818 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
17820 // Create a type on which we perform the shuffle
17821 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
17822 StVT.getScalarType(), NumElems*SizeRatio);
17824 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
17826 SDValue WideVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, St->getValue());
17827 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
17828 for (unsigned i = 0; i != NumElems; ++i)
17829 ShuffleVec[i] = i * SizeRatio;
17831 // Can't shuffle using an illegal type.
17832 if (!TLI.isTypeLegal(WideVecVT))
17835 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
17836 DAG.getUNDEF(WideVecVT),
17838 // At this point all of the data is stored at the bottom of the
17839 // register. We now need to save it to mem.
17841 // Find the largest store unit
17842 MVT StoreType = MVT::i8;
17843 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
17844 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
17845 MVT Tp = (MVT::SimpleValueType)tp;
17846 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToSz)
17850 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
17851 if (TLI.isTypeLegal(MVT::f64) && StoreType.getSizeInBits() < 64 &&
17852 (64 <= NumElems * ToSz))
17853 StoreType = MVT::f64;
17855 // Bitcast the original vector into a vector of store-size units
17856 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
17857 StoreType, VT.getSizeInBits()/StoreType.getSizeInBits());
17858 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
17859 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, dl, StoreVecVT, Shuff);
17860 SmallVector<SDValue, 8> Chains;
17861 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
17862 TLI.getPointerTy());
17863 SDValue Ptr = St->getBasePtr();
17865 // Perform one or more big stores into memory.
17866 for (unsigned i=0, e=(ToSz*NumElems)/StoreType.getSizeInBits(); i!=e; ++i) {
17867 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
17868 StoreType, ShuffWide,
17869 DAG.getIntPtrConstant(i));
17870 SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr,
17871 St->getPointerInfo(), St->isVolatile(),
17872 St->isNonTemporal(), St->getAlignment());
17873 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
17874 Chains.push_back(Ch);
17877 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Chains[0],
17881 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
17882 // the FP state in cases where an emms may be missing.
17883 // A preferable solution to the general problem is to figure out the right
17884 // places to insert EMMS. This qualifies as a quick hack.
17886 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
17887 if (VT.getSizeInBits() != 64)
17890 const Function *F = DAG.getMachineFunction().getFunction();
17891 bool NoImplicitFloatOps = F->getAttributes().
17892 hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat);
17893 bool F64IsLegal = !DAG.getTarget().Options.UseSoftFloat && !NoImplicitFloatOps
17894 && Subtarget->hasSSE2();
17895 if ((VT.isVector() ||
17896 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
17897 isa<LoadSDNode>(St->getValue()) &&
17898 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
17899 St->getChain().hasOneUse() && !St->isVolatile()) {
17900 SDNode* LdVal = St->getValue().getNode();
17901 LoadSDNode *Ld = 0;
17902 int TokenFactorIndex = -1;
17903 SmallVector<SDValue, 8> Ops;
17904 SDNode* ChainVal = St->getChain().getNode();
17905 // Must be a store of a load. We currently handle two cases: the load
17906 // is a direct child, and it's under an intervening TokenFactor. It is
17907 // possible to dig deeper under nested TokenFactors.
17908 if (ChainVal == LdVal)
17909 Ld = cast<LoadSDNode>(St->getChain());
17910 else if (St->getValue().hasOneUse() &&
17911 ChainVal->getOpcode() == ISD::TokenFactor) {
17912 for (unsigned i = 0, e = ChainVal->getNumOperands(); i != e; ++i) {
17913 if (ChainVal->getOperand(i).getNode() == LdVal) {
17914 TokenFactorIndex = i;
17915 Ld = cast<LoadSDNode>(St->getValue());
17917 Ops.push_back(ChainVal->getOperand(i));
17921 if (!Ld || !ISD::isNormalLoad(Ld))
17924 // If this is not the MMX case, i.e. we are just turning i64 load/store
17925 // into f64 load/store, avoid the transformation if there are multiple
17926 // uses of the loaded value.
17927 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
17932 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
17933 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
17935 if (Subtarget->is64Bit() || F64IsLegal) {
17936 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
17937 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
17938 Ld->getPointerInfo(), Ld->isVolatile(),
17939 Ld->isNonTemporal(), Ld->isInvariant(),
17940 Ld->getAlignment());
17941 SDValue NewChain = NewLd.getValue(1);
17942 if (TokenFactorIndex != -1) {
17943 Ops.push_back(NewChain);
17944 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
17947 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
17948 St->getPointerInfo(),
17949 St->isVolatile(), St->isNonTemporal(),
17950 St->getAlignment());
17953 // Otherwise, lower to two pairs of 32-bit loads / stores.
17954 SDValue LoAddr = Ld->getBasePtr();
17955 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
17956 DAG.getConstant(4, MVT::i32));
17958 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
17959 Ld->getPointerInfo(),
17960 Ld->isVolatile(), Ld->isNonTemporal(),
17961 Ld->isInvariant(), Ld->getAlignment());
17962 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
17963 Ld->getPointerInfo().getWithOffset(4),
17964 Ld->isVolatile(), Ld->isNonTemporal(),
17966 MinAlign(Ld->getAlignment(), 4));
17968 SDValue NewChain = LoLd.getValue(1);
17969 if (TokenFactorIndex != -1) {
17970 Ops.push_back(LoLd);
17971 Ops.push_back(HiLd);
17972 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
17976 LoAddr = St->getBasePtr();
17977 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
17978 DAG.getConstant(4, MVT::i32));
17980 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
17981 St->getPointerInfo(),
17982 St->isVolatile(), St->isNonTemporal(),
17983 St->getAlignment());
17984 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
17985 St->getPointerInfo().getWithOffset(4),
17987 St->isNonTemporal(),
17988 MinAlign(St->getAlignment(), 4));
17989 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
17994 /// isHorizontalBinOp - Return 'true' if this vector operation is "horizontal"
17995 /// and return the operands for the horizontal operation in LHS and RHS. A
17996 /// horizontal operation performs the binary operation on successive elements
17997 /// of its first operand, then on successive elements of its second operand,
17998 /// returning the resulting values in a vector. For example, if
17999 /// A = < float a0, float a1, float a2, float a3 >
18001 /// B = < float b0, float b1, float b2, float b3 >
18002 /// then the result of doing a horizontal operation on A and B is
18003 /// A horizontal-op B = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >.
18004 /// In short, LHS and RHS are inspected to see if LHS op RHS is of the form
18005 /// A horizontal-op B, for some already available A and B, and if so then LHS is
18006 /// set to A, RHS to B, and the routine returns 'true'.
18007 /// Note that the binary operation should have the property that if one of the
18008 /// operands is UNDEF then the result is UNDEF.
18009 static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool IsCommutative) {
18010 // Look for the following pattern: if
18011 // A = < float a0, float a1, float a2, float a3 >
18012 // B = < float b0, float b1, float b2, float b3 >
18014 // LHS = VECTOR_SHUFFLE A, B, <0, 2, 4, 6>
18015 // RHS = VECTOR_SHUFFLE A, B, <1, 3, 5, 7>
18016 // then LHS op RHS = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >
18017 // which is A horizontal-op B.
18019 // At least one of the operands should be a vector shuffle.
18020 if (LHS.getOpcode() != ISD::VECTOR_SHUFFLE &&
18021 RHS.getOpcode() != ISD::VECTOR_SHUFFLE)
18024 MVT VT = LHS.getSimpleValueType();
18026 assert((VT.is128BitVector() || VT.is256BitVector()) &&
18027 "Unsupported vector type for horizontal add/sub");
18029 // Handle 128 and 256-bit vector lengths. AVX defines horizontal add/sub to
18030 // operate independently on 128-bit lanes.
18031 unsigned NumElts = VT.getVectorNumElements();
18032 unsigned NumLanes = VT.getSizeInBits()/128;
18033 unsigned NumLaneElts = NumElts / NumLanes;
18034 assert((NumLaneElts % 2 == 0) &&
18035 "Vector type should have an even number of elements in each lane");
18036 unsigned HalfLaneElts = NumLaneElts/2;
18038 // View LHS in the form
18039 // LHS = VECTOR_SHUFFLE A, B, LMask
18040 // If LHS is not a shuffle then pretend it is the shuffle
18041 // LHS = VECTOR_SHUFFLE LHS, undef, <0, 1, ..., N-1>
18042 // NOTE: in what follows a default initialized SDValue represents an UNDEF of
18045 SmallVector<int, 16> LMask(NumElts);
18046 if (LHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
18047 if (LHS.getOperand(0).getOpcode() != ISD::UNDEF)
18048 A = LHS.getOperand(0);
18049 if (LHS.getOperand(1).getOpcode() != ISD::UNDEF)
18050 B = LHS.getOperand(1);
18051 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(LHS.getNode())->getMask();
18052 std::copy(Mask.begin(), Mask.end(), LMask.begin());
18054 if (LHS.getOpcode() != ISD::UNDEF)
18056 for (unsigned i = 0; i != NumElts; ++i)
18060 // Likewise, view RHS in the form
18061 // RHS = VECTOR_SHUFFLE C, D, RMask
18063 SmallVector<int, 16> RMask(NumElts);
18064 if (RHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
18065 if (RHS.getOperand(0).getOpcode() != ISD::UNDEF)
18066 C = RHS.getOperand(0);
18067 if (RHS.getOperand(1).getOpcode() != ISD::UNDEF)
18068 D = RHS.getOperand(1);
18069 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(RHS.getNode())->getMask();
18070 std::copy(Mask.begin(), Mask.end(), RMask.begin());
18072 if (RHS.getOpcode() != ISD::UNDEF)
18074 for (unsigned i = 0; i != NumElts; ++i)
18078 // Check that the shuffles are both shuffling the same vectors.
18079 if (!(A == C && B == D) && !(A == D && B == C))
18082 // If everything is UNDEF then bail out: it would be better to fold to UNDEF.
18083 if (!A.getNode() && !B.getNode())
18086 // If A and B occur in reverse order in RHS, then "swap" them (which means
18087 // rewriting the mask).
18089 CommuteVectorShuffleMask(RMask, NumElts);
18091 // At this point LHS and RHS are equivalent to
18092 // LHS = VECTOR_SHUFFLE A, B, LMask
18093 // RHS = VECTOR_SHUFFLE A, B, RMask
18094 // Check that the masks correspond to performing a horizontal operation.
18095 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
18096 for (unsigned i = 0; i != NumLaneElts; ++i) {
18097 int LIdx = LMask[i+l], RIdx = RMask[i+l];
18099 // Ignore any UNDEF components.
18100 if (LIdx < 0 || RIdx < 0 ||
18101 (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) ||
18102 (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts)))
18105 // Check that successive elements are being operated on. If not, this is
18106 // not a horizontal operation.
18107 unsigned Src = (i/HalfLaneElts); // each lane is split between srcs
18108 int Index = 2*(i%HalfLaneElts) + NumElts*Src + l;
18109 if (!(LIdx == Index && RIdx == Index + 1) &&
18110 !(IsCommutative && LIdx == Index + 1 && RIdx == Index))
18115 LHS = A.getNode() ? A : B; // If A is 'UNDEF', use B for it.
18116 RHS = B.getNode() ? B : A; // If B is 'UNDEF', use A for it.
18120 /// PerformFADDCombine - Do target-specific dag combines on floating point adds.
18121 static SDValue PerformFADDCombine(SDNode *N, SelectionDAG &DAG,
18122 const X86Subtarget *Subtarget) {
18123 EVT VT = N->getValueType(0);
18124 SDValue LHS = N->getOperand(0);
18125 SDValue RHS = N->getOperand(1);
18127 // Try to synthesize horizontal adds from adds of shuffles.
18128 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
18129 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
18130 isHorizontalBinOp(LHS, RHS, true))
18131 return DAG.getNode(X86ISD::FHADD, SDLoc(N), VT, LHS, RHS);
18135 /// PerformFSUBCombine - Do target-specific dag combines on floating point subs.
18136 static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG,
18137 const X86Subtarget *Subtarget) {
18138 EVT VT = N->getValueType(0);
18139 SDValue LHS = N->getOperand(0);
18140 SDValue RHS = N->getOperand(1);
18142 // Try to synthesize horizontal subs from subs of shuffles.
18143 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
18144 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
18145 isHorizontalBinOp(LHS, RHS, false))
18146 return DAG.getNode(X86ISD::FHSUB, SDLoc(N), VT, LHS, RHS);
18150 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
18151 /// X86ISD::FXOR nodes.
18152 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
18153 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
18154 // F[X]OR(0.0, x) -> x
18155 // F[X]OR(x, 0.0) -> x
18156 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
18157 if (C->getValueAPF().isPosZero())
18158 return N->getOperand(1);
18159 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
18160 if (C->getValueAPF().isPosZero())
18161 return N->getOperand(0);
18165 /// PerformFMinFMaxCombine - Do target-specific dag combines on X86ISD::FMIN and
18166 /// X86ISD::FMAX nodes.
18167 static SDValue PerformFMinFMaxCombine(SDNode *N, SelectionDAG &DAG) {
18168 assert(N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD::FMAX);
18170 // Only perform optimizations if UnsafeMath is used.
18171 if (!DAG.getTarget().Options.UnsafeFPMath)
18174 // If we run in unsafe-math mode, then convert the FMAX and FMIN nodes
18175 // into FMINC and FMAXC, which are Commutative operations.
18176 unsigned NewOp = 0;
18177 switch (N->getOpcode()) {
18178 default: llvm_unreachable("unknown opcode");
18179 case X86ISD::FMIN: NewOp = X86ISD::FMINC; break;
18180 case X86ISD::FMAX: NewOp = X86ISD::FMAXC; break;
18183 return DAG.getNode(NewOp, SDLoc(N), N->getValueType(0),
18184 N->getOperand(0), N->getOperand(1));
18187 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
18188 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
18189 // FAND(0.0, x) -> 0.0
18190 // FAND(x, 0.0) -> 0.0
18191 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
18192 if (C->getValueAPF().isPosZero())
18193 return N->getOperand(0);
18194 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
18195 if (C->getValueAPF().isPosZero())
18196 return N->getOperand(1);
18200 /// PerformFANDNCombine - Do target-specific dag combines on X86ISD::FANDN nodes
18201 static SDValue PerformFANDNCombine(SDNode *N, SelectionDAG &DAG) {
18202 // FANDN(x, 0.0) -> 0.0
18203 // FANDN(0.0, x) -> x
18204 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
18205 if (C->getValueAPF().isPosZero())
18206 return N->getOperand(1);
18207 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
18208 if (C->getValueAPF().isPosZero())
18209 return N->getOperand(1);
18213 static SDValue PerformBTCombine(SDNode *N,
18215 TargetLowering::DAGCombinerInfo &DCI) {
18216 // BT ignores high bits in the bit index operand.
18217 SDValue Op1 = N->getOperand(1);
18218 if (Op1.hasOneUse()) {
18219 unsigned BitWidth = Op1.getValueSizeInBits();
18220 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
18221 APInt KnownZero, KnownOne;
18222 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
18223 !DCI.isBeforeLegalizeOps());
18224 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
18225 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
18226 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
18227 DCI.CommitTargetLoweringOpt(TLO);
18232 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
18233 SDValue Op = N->getOperand(0);
18234 if (Op.getOpcode() == ISD::BITCAST)
18235 Op = Op.getOperand(0);
18236 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
18237 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
18238 VT.getVectorElementType().getSizeInBits() ==
18239 OpVT.getVectorElementType().getSizeInBits()) {
18240 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
18245 static SDValue PerformSIGN_EXTEND_INREGCombine(SDNode *N, SelectionDAG &DAG,
18246 const X86Subtarget *Subtarget) {
18247 EVT VT = N->getValueType(0);
18248 if (!VT.isVector())
18251 SDValue N0 = N->getOperand(0);
18252 SDValue N1 = N->getOperand(1);
18253 EVT ExtraVT = cast<VTSDNode>(N1)->getVT();
18256 // The SIGN_EXTEND_INREG to v4i64 is expensive operation on the
18257 // both SSE and AVX2 since there is no sign-extended shift right
18258 // operation on a vector with 64-bit elements.
18259 //(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) ->
18260 // (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT)))
18261 if (VT == MVT::v4i64 && (N0.getOpcode() == ISD::ANY_EXTEND ||
18262 N0.getOpcode() == ISD::SIGN_EXTEND)) {
18263 SDValue N00 = N0.getOperand(0);
18265 // EXTLOAD has a better solution on AVX2,
18266 // it may be replaced with X86ISD::VSEXT node.
18267 if (N00.getOpcode() == ISD::LOAD && Subtarget->hasInt256())
18268 if (!ISD::isNormalLoad(N00.getNode()))
18271 if (N00.getValueType() == MVT::v4i32 && ExtraVT.getSizeInBits() < 128) {
18272 SDValue Tmp = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32,
18274 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i64, Tmp);
18280 static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
18281 TargetLowering::DAGCombinerInfo &DCI,
18282 const X86Subtarget *Subtarget) {
18283 if (!DCI.isBeforeLegalizeOps())
18286 if (!Subtarget->hasFp256())
18289 EVT VT = N->getValueType(0);
18290 if (VT.isVector() && VT.getSizeInBits() == 256) {
18291 SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget);
18299 static SDValue PerformFMACombine(SDNode *N, SelectionDAG &DAG,
18300 const X86Subtarget* Subtarget) {
18302 EVT VT = N->getValueType(0);
18304 // Let legalize expand this if it isn't a legal type yet.
18305 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
18308 EVT ScalarVT = VT.getScalarType();
18309 if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) ||
18310 (!Subtarget->hasFMA() && !Subtarget->hasFMA4()))
18313 SDValue A = N->getOperand(0);
18314 SDValue B = N->getOperand(1);
18315 SDValue C = N->getOperand(2);
18317 bool NegA = (A.getOpcode() == ISD::FNEG);
18318 bool NegB = (B.getOpcode() == ISD::FNEG);
18319 bool NegC = (C.getOpcode() == ISD::FNEG);
18321 // Negative multiplication when NegA xor NegB
18322 bool NegMul = (NegA != NegB);
18324 A = A.getOperand(0);
18326 B = B.getOperand(0);
18328 C = C.getOperand(0);
18332 Opcode = (!NegC) ? X86ISD::FMADD : X86ISD::FMSUB;
18334 Opcode = (!NegC) ? X86ISD::FNMADD : X86ISD::FNMSUB;
18336 return DAG.getNode(Opcode, dl, VT, A, B, C);
18339 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG,
18340 TargetLowering::DAGCombinerInfo &DCI,
18341 const X86Subtarget *Subtarget) {
18342 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
18343 // (and (i32 x86isd::setcc_carry), 1)
18344 // This eliminates the zext. This transformation is necessary because
18345 // ISD::SETCC is always legalized to i8.
18347 SDValue N0 = N->getOperand(0);
18348 EVT VT = N->getValueType(0);
18350 if (N0.getOpcode() == ISD::AND &&
18352 N0.getOperand(0).hasOneUse()) {
18353 SDValue N00 = N0.getOperand(0);
18354 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
18355 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
18356 if (!C || C->getZExtValue() != 1)
18358 return DAG.getNode(ISD::AND, dl, VT,
18359 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
18360 N00.getOperand(0), N00.getOperand(1)),
18361 DAG.getConstant(1, VT));
18365 if (VT.is256BitVector()) {
18366 SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget);
18374 // Optimize x == -y --> x+y == 0
18375 // x != -y --> x+y != 0
18376 static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG) {
18377 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
18378 SDValue LHS = N->getOperand(0);
18379 SDValue RHS = N->getOperand(1);
18381 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB)
18382 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(LHS.getOperand(0)))
18383 if (C->getAPIntValue() == 0 && LHS.hasOneUse()) {
18384 SDValue addV = DAG.getNode(ISD::ADD, SDLoc(N),
18385 LHS.getValueType(), RHS, LHS.getOperand(1));
18386 return DAG.getSetCC(SDLoc(N), N->getValueType(0),
18387 addV, DAG.getConstant(0, addV.getValueType()), CC);
18389 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && RHS.getOpcode() == ISD::SUB)
18390 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS.getOperand(0)))
18391 if (C->getAPIntValue() == 0 && RHS.hasOneUse()) {
18392 SDValue addV = DAG.getNode(ISD::ADD, SDLoc(N),
18393 RHS.getValueType(), LHS, RHS.getOperand(1));
18394 return DAG.getSetCC(SDLoc(N), N->getValueType(0),
18395 addV, DAG.getConstant(0, addV.getValueType()), CC);
18400 // Helper function of PerformSETCCCombine. It is to materialize "setb reg"
18401 // as "sbb reg,reg", since it can be extended without zext and produces
18402 // an all-ones bit which is more useful than 0/1 in some cases.
18403 static SDValue MaterializeSETB(SDLoc DL, SDValue EFLAGS, SelectionDAG &DAG) {
18404 return DAG.getNode(ISD::AND, DL, MVT::i8,
18405 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
18406 DAG.getConstant(X86::COND_B, MVT::i8), EFLAGS),
18407 DAG.getConstant(1, MVT::i8));
18410 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
18411 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG,
18412 TargetLowering::DAGCombinerInfo &DCI,
18413 const X86Subtarget *Subtarget) {
18415 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0));
18416 SDValue EFLAGS = N->getOperand(1);
18418 if (CC == X86::COND_A) {
18419 // Try to convert COND_A into COND_B in an attempt to facilitate
18420 // materializing "setb reg".
18422 // Do not flip "e > c", where "c" is a constant, because Cmp instruction
18423 // cannot take an immediate as its first operand.
18425 if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() &&
18426 EFLAGS.getValueType().isInteger() &&
18427 !isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
18428 SDValue NewSub = DAG.getNode(X86ISD::SUB, SDLoc(EFLAGS),
18429 EFLAGS.getNode()->getVTList(),
18430 EFLAGS.getOperand(1), EFLAGS.getOperand(0));
18431 SDValue NewEFLAGS = SDValue(NewSub.getNode(), EFLAGS.getResNo());
18432 return MaterializeSETB(DL, NewEFLAGS, DAG);
18436 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
18437 // a zext and produces an all-ones bit which is more useful than 0/1 in some
18439 if (CC == X86::COND_B)
18440 return MaterializeSETB(DL, EFLAGS, DAG);
18444 Flags = checkBoolTestSetCCCombine(EFLAGS, CC);
18445 if (Flags.getNode()) {
18446 SDValue Cond = DAG.getConstant(CC, MVT::i8);
18447 return DAG.getNode(X86ISD::SETCC, DL, N->getVTList(), Cond, Flags);
18453 // Optimize branch condition evaluation.
18455 static SDValue PerformBrCondCombine(SDNode *N, SelectionDAG &DAG,
18456 TargetLowering::DAGCombinerInfo &DCI,
18457 const X86Subtarget *Subtarget) {
18459 SDValue Chain = N->getOperand(0);
18460 SDValue Dest = N->getOperand(1);
18461 SDValue EFLAGS = N->getOperand(3);
18462 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(2));
18466 Flags = checkBoolTestSetCCCombine(EFLAGS, CC);
18467 if (Flags.getNode()) {
18468 SDValue Cond = DAG.getConstant(CC, MVT::i8);
18469 return DAG.getNode(X86ISD::BRCOND, DL, N->getVTList(), Chain, Dest, Cond,
18476 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
18477 const X86TargetLowering *XTLI) {
18478 SDValue Op0 = N->getOperand(0);
18479 EVT InVT = Op0->getValueType(0);
18481 // SINT_TO_FP(v4i8) -> SINT_TO_FP(SEXT(v4i8 to v4i32))
18482 if (InVT == MVT::v8i8 || InVT == MVT::v4i8) {
18484 MVT DstVT = InVT == MVT::v4i8 ? MVT::v4i32 : MVT::v8i32;
18485 SDValue P = DAG.getNode(ISD::SIGN_EXTEND, dl, DstVT, Op0);
18486 return DAG.getNode(ISD::SINT_TO_FP, dl, N->getValueType(0), P);
18489 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
18490 // a 32-bit target where SSE doesn't support i64->FP operations.
18491 if (Op0.getOpcode() == ISD::LOAD) {
18492 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
18493 EVT VT = Ld->getValueType(0);
18494 if (!Ld->isVolatile() && !N->getValueType(0).isVector() &&
18495 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
18496 !XTLI->getSubtarget()->is64Bit() &&
18497 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
18498 SDValue FILDChain = XTLI->BuildFILD(SDValue(N, 0), Ld->getValueType(0),
18499 Ld->getChain(), Op0, DAG);
18500 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
18507 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
18508 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
18509 X86TargetLowering::DAGCombinerInfo &DCI) {
18510 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
18511 // the result is either zero or one (depending on the input carry bit).
18512 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
18513 if (X86::isZeroNode(N->getOperand(0)) &&
18514 X86::isZeroNode(N->getOperand(1)) &&
18515 // We don't have a good way to replace an EFLAGS use, so only do this when
18517 SDValue(N, 1).use_empty()) {
18519 EVT VT = N->getValueType(0);
18520 SDValue CarryOut = DAG.getConstant(0, N->getValueType(1));
18521 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
18522 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
18523 DAG.getConstant(X86::COND_B,MVT::i8),
18525 DAG.getConstant(1, VT));
18526 return DCI.CombineTo(N, Res1, CarryOut);
18532 // fold (add Y, (sete X, 0)) -> adc 0, Y
18533 // (add Y, (setne X, 0)) -> sbb -1, Y
18534 // (sub (sete X, 0), Y) -> sbb 0, Y
18535 // (sub (setne X, 0), Y) -> adc -1, Y
18536 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
18539 // Look through ZExts.
18540 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
18541 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
18544 SDValue SetCC = Ext.getOperand(0);
18545 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
18548 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
18549 if (CC != X86::COND_E && CC != X86::COND_NE)
18552 SDValue Cmp = SetCC.getOperand(1);
18553 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
18554 !X86::isZeroNode(Cmp.getOperand(1)) ||
18555 !Cmp.getOperand(0).getValueType().isInteger())
18558 SDValue CmpOp0 = Cmp.getOperand(0);
18559 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
18560 DAG.getConstant(1, CmpOp0.getValueType()));
18562 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
18563 if (CC == X86::COND_NE)
18564 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
18565 DL, OtherVal.getValueType(), OtherVal,
18566 DAG.getConstant(-1ULL, OtherVal.getValueType()), NewCmp);
18567 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
18568 DL, OtherVal.getValueType(), OtherVal,
18569 DAG.getConstant(0, OtherVal.getValueType()), NewCmp);
18572 /// PerformADDCombine - Do target-specific dag combines on integer adds.
18573 static SDValue PerformAddCombine(SDNode *N, SelectionDAG &DAG,
18574 const X86Subtarget *Subtarget) {
18575 EVT VT = N->getValueType(0);
18576 SDValue Op0 = N->getOperand(0);
18577 SDValue Op1 = N->getOperand(1);
18579 // Try to synthesize horizontal adds from adds of shuffles.
18580 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
18581 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
18582 isHorizontalBinOp(Op0, Op1, true))
18583 return DAG.getNode(X86ISD::HADD, SDLoc(N), VT, Op0, Op1);
18585 return OptimizeConditionalInDecrement(N, DAG);
18588 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG,
18589 const X86Subtarget *Subtarget) {
18590 SDValue Op0 = N->getOperand(0);
18591 SDValue Op1 = N->getOperand(1);
18593 // X86 can't encode an immediate LHS of a sub. See if we can push the
18594 // negation into a preceding instruction.
18595 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
18596 // If the RHS of the sub is a XOR with one use and a constant, invert the
18597 // immediate. Then add one to the LHS of the sub so we can turn
18598 // X-Y -> X+~Y+1, saving one register.
18599 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
18600 isa<ConstantSDNode>(Op1.getOperand(1))) {
18601 APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue();
18602 EVT VT = Op0.getValueType();
18603 SDValue NewXor = DAG.getNode(ISD::XOR, SDLoc(Op1), VT,
18605 DAG.getConstant(~XorC, VT));
18606 return DAG.getNode(ISD::ADD, SDLoc(N), VT, NewXor,
18607 DAG.getConstant(C->getAPIntValue()+1, VT));
18611 // Try to synthesize horizontal adds from adds of shuffles.
18612 EVT VT = N->getValueType(0);
18613 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
18614 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
18615 isHorizontalBinOp(Op0, Op1, true))
18616 return DAG.getNode(X86ISD::HSUB, SDLoc(N), VT, Op0, Op1);
18618 return OptimizeConditionalInDecrement(N, DAG);
18621 /// performVZEXTCombine - Performs build vector combines
18622 static SDValue performVZEXTCombine(SDNode *N, SelectionDAG &DAG,
18623 TargetLowering::DAGCombinerInfo &DCI,
18624 const X86Subtarget *Subtarget) {
18625 // (vzext (bitcast (vzext (x)) -> (vzext x)
18626 SDValue In = N->getOperand(0);
18627 while (In.getOpcode() == ISD::BITCAST)
18628 In = In.getOperand(0);
18630 if (In.getOpcode() != X86ISD::VZEXT)
18633 return DAG.getNode(X86ISD::VZEXT, SDLoc(N), N->getValueType(0),
18637 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
18638 DAGCombinerInfo &DCI) const {
18639 SelectionDAG &DAG = DCI.DAG;
18640 switch (N->getOpcode()) {
18642 case ISD::EXTRACT_VECTOR_ELT:
18643 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, DCI);
18645 case ISD::SELECT: return PerformSELECTCombine(N, DAG, DCI, Subtarget);
18646 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI, Subtarget);
18647 case ISD::ADD: return PerformAddCombine(N, DAG, Subtarget);
18648 case ISD::SUB: return PerformSubCombine(N, DAG, Subtarget);
18649 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
18650 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
18653 case ISD::SRL: return PerformShiftCombine(N, DAG, DCI, Subtarget);
18654 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
18655 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
18656 case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget);
18657 case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget);
18658 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
18659 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, this);
18660 case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
18661 case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
18663 case X86ISD::FOR: return PerformFORCombine(N, DAG);
18665 case X86ISD::FMAX: return PerformFMinFMaxCombine(N, DAG);
18666 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
18667 case X86ISD::FANDN: return PerformFANDNCombine(N, DAG);
18668 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
18669 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
18670 case ISD::ANY_EXTEND:
18671 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG, DCI, Subtarget);
18672 case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
18673 case ISD::SIGN_EXTEND_INREG: return PerformSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
18674 case ISD::TRUNCATE: return PerformTruncateCombine(N, DAG,DCI,Subtarget);
18675 case ISD::SETCC: return PerformISDSETCCCombine(N, DAG);
18676 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
18677 case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
18678 case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);
18679 case X86ISD::SHUFP: // Handle all target specific shuffles
18680 case X86ISD::PALIGNR:
18681 case X86ISD::UNPCKH:
18682 case X86ISD::UNPCKL:
18683 case X86ISD::MOVHLPS:
18684 case X86ISD::MOVLHPS:
18685 case X86ISD::PSHUFD:
18686 case X86ISD::PSHUFHW:
18687 case X86ISD::PSHUFLW:
18688 case X86ISD::MOVSS:
18689 case X86ISD::MOVSD:
18690 case X86ISD::VPERMILP:
18691 case X86ISD::VPERM2X128:
18692 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
18693 case ISD::FMA: return PerformFMACombine(N, DAG, Subtarget);
18699 /// isTypeDesirableForOp - Return true if the target has native support for
18700 /// the specified value type and it is 'desirable' to use the type for the
18701 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
18702 /// instruction encodings are longer and some i16 instructions are slow.
18703 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
18704 if (!isTypeLegal(VT))
18706 if (VT != MVT::i16)
18713 case ISD::SIGN_EXTEND:
18714 case ISD::ZERO_EXTEND:
18715 case ISD::ANY_EXTEND:
18728 /// IsDesirableToPromoteOp - This method query the target whether it is
18729 /// beneficial for dag combiner to promote the specified node. If true, it
18730 /// should return the desired promotion type by reference.
18731 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
18732 EVT VT = Op.getValueType();
18733 if (VT != MVT::i16)
18736 bool Promote = false;
18737 bool Commute = false;
18738 switch (Op.getOpcode()) {
18741 LoadSDNode *LD = cast<LoadSDNode>(Op);
18742 // If the non-extending load has a single use and it's not live out, then it
18743 // might be folded.
18744 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
18745 Op.hasOneUse()*/) {
18746 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
18747 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
18748 // The only case where we'd want to promote LOAD (rather then it being
18749 // promoted as an operand is when it's only use is liveout.
18750 if (UI->getOpcode() != ISD::CopyToReg)
18757 case ISD::SIGN_EXTEND:
18758 case ISD::ZERO_EXTEND:
18759 case ISD::ANY_EXTEND:
18764 SDValue N0 = Op.getOperand(0);
18765 // Look out for (store (shl (load), x)).
18766 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
18779 SDValue N0 = Op.getOperand(0);
18780 SDValue N1 = Op.getOperand(1);
18781 if (!Commute && MayFoldLoad(N1))
18783 // Avoid disabling potential load folding opportunities.
18784 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
18786 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
18796 //===----------------------------------------------------------------------===//
18797 // X86 Inline Assembly Support
18798 //===----------------------------------------------------------------------===//
18801 // Helper to match a string separated by whitespace.
18802 bool matchAsmImpl(StringRef s, ArrayRef<const StringRef *> args) {
18803 s = s.substr(s.find_first_not_of(" \t")); // Skip leading whitespace.
18805 for (unsigned i = 0, e = args.size(); i != e; ++i) {
18806 StringRef piece(*args[i]);
18807 if (!s.startswith(piece)) // Check if the piece matches.
18810 s = s.substr(piece.size());
18811 StringRef::size_type pos = s.find_first_not_of(" \t");
18812 if (pos == 0) // We matched a prefix.
18820 const VariadicFunction1<bool, StringRef, StringRef, matchAsmImpl> matchAsm={};
18823 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
18824 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
18826 std::string AsmStr = IA->getAsmString();
18828 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
18829 if (!Ty || Ty->getBitWidth() % 16 != 0)
18832 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
18833 SmallVector<StringRef, 4> AsmPieces;
18834 SplitString(AsmStr, AsmPieces, ";\n");
18836 switch (AsmPieces.size()) {
18837 default: return false;
18839 // FIXME: this should verify that we are targeting a 486 or better. If not,
18840 // we will turn this bswap into something that will be lowered to logical
18841 // ops instead of emitting the bswap asm. For now, we don't support 486 or
18842 // lower so don't worry about this.
18844 if (matchAsm(AsmPieces[0], "bswap", "$0") ||
18845 matchAsm(AsmPieces[0], "bswapl", "$0") ||
18846 matchAsm(AsmPieces[0], "bswapq", "$0") ||
18847 matchAsm(AsmPieces[0], "bswap", "${0:q}") ||
18848 matchAsm(AsmPieces[0], "bswapl", "${0:q}") ||
18849 matchAsm(AsmPieces[0], "bswapq", "${0:q}")) {
18850 // No need to check constraints, nothing other than the equivalent of
18851 // "=r,0" would be valid here.
18852 return IntrinsicLowering::LowerToByteSwap(CI);
18855 // rorw $$8, ${0:w} --> llvm.bswap.i16
18856 if (CI->getType()->isIntegerTy(16) &&
18857 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
18858 (matchAsm(AsmPieces[0], "rorw", "$$8,", "${0:w}") ||
18859 matchAsm(AsmPieces[0], "rolw", "$$8,", "${0:w}"))) {
18861 const std::string &ConstraintsStr = IA->getConstraintString();
18862 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
18863 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
18864 if (AsmPieces.size() == 4 &&
18865 AsmPieces[0] == "~{cc}" &&
18866 AsmPieces[1] == "~{dirflag}" &&
18867 AsmPieces[2] == "~{flags}" &&
18868 AsmPieces[3] == "~{fpsr}")
18869 return IntrinsicLowering::LowerToByteSwap(CI);
18873 if (CI->getType()->isIntegerTy(32) &&
18874 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
18875 matchAsm(AsmPieces[0], "rorw", "$$8,", "${0:w}") &&
18876 matchAsm(AsmPieces[1], "rorl", "$$16,", "$0") &&
18877 matchAsm(AsmPieces[2], "rorw", "$$8,", "${0:w}")) {
18879 const std::string &ConstraintsStr = IA->getConstraintString();
18880 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
18881 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
18882 if (AsmPieces.size() == 4 &&
18883 AsmPieces[0] == "~{cc}" &&
18884 AsmPieces[1] == "~{dirflag}" &&
18885 AsmPieces[2] == "~{flags}" &&
18886 AsmPieces[3] == "~{fpsr}")
18887 return IntrinsicLowering::LowerToByteSwap(CI);
18890 if (CI->getType()->isIntegerTy(64)) {
18891 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
18892 if (Constraints.size() >= 2 &&
18893 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
18894 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
18895 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
18896 if (matchAsm(AsmPieces[0], "bswap", "%eax") &&
18897 matchAsm(AsmPieces[1], "bswap", "%edx") &&
18898 matchAsm(AsmPieces[2], "xchgl", "%eax,", "%edx"))
18899 return IntrinsicLowering::LowerToByteSwap(CI);
18907 /// getConstraintType - Given a constraint letter, return the type of
18908 /// constraint it is for this target.
18909 X86TargetLowering::ConstraintType
18910 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
18911 if (Constraint.size() == 1) {
18912 switch (Constraint[0]) {
18923 return C_RegisterClass;
18947 return TargetLowering::getConstraintType(Constraint);
18950 /// Examine constraint type and operand type and determine a weight value.
18951 /// This object must already have been set up with the operand type
18952 /// and the current alternative constraint selected.
18953 TargetLowering::ConstraintWeight
18954 X86TargetLowering::getSingleConstraintMatchWeight(
18955 AsmOperandInfo &info, const char *constraint) const {
18956 ConstraintWeight weight = CW_Invalid;
18957 Value *CallOperandVal = info.CallOperandVal;
18958 // If we don't have a value, we can't do a match,
18959 // but allow it at the lowest weight.
18960 if (CallOperandVal == NULL)
18962 Type *type = CallOperandVal->getType();
18963 // Look at the constraint type.
18964 switch (*constraint) {
18966 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
18977 if (CallOperandVal->getType()->isIntegerTy())
18978 weight = CW_SpecificReg;
18983 if (type->isFloatingPointTy())
18984 weight = CW_SpecificReg;
18987 if (type->isX86_MMXTy() && Subtarget->hasMMX())
18988 weight = CW_SpecificReg;
18992 if (((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasSSE1()) ||
18993 ((type->getPrimitiveSizeInBits() == 256) && Subtarget->hasFp256()))
18994 weight = CW_Register;
18997 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
18998 if (C->getZExtValue() <= 31)
18999 weight = CW_Constant;
19003 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
19004 if (C->getZExtValue() <= 63)
19005 weight = CW_Constant;
19009 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
19010 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
19011 weight = CW_Constant;
19015 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
19016 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
19017 weight = CW_Constant;
19021 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
19022 if (C->getZExtValue() <= 3)
19023 weight = CW_Constant;
19027 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
19028 if (C->getZExtValue() <= 0xff)
19029 weight = CW_Constant;
19034 if (dyn_cast<ConstantFP>(CallOperandVal)) {
19035 weight = CW_Constant;
19039 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
19040 if ((C->getSExtValue() >= -0x80000000LL) &&
19041 (C->getSExtValue() <= 0x7fffffffLL))
19042 weight = CW_Constant;
19046 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
19047 if (C->getZExtValue() <= 0xffffffff)
19048 weight = CW_Constant;
19055 /// LowerXConstraint - try to replace an X constraint, which matches anything,
19056 /// with another that has more specific requirements based on the type of the
19057 /// corresponding operand.
19058 const char *X86TargetLowering::
19059 LowerXConstraint(EVT ConstraintVT) const {
19060 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
19061 // 'f' like normal targets.
19062 if (ConstraintVT.isFloatingPoint()) {
19063 if (Subtarget->hasSSE2())
19065 if (Subtarget->hasSSE1())
19069 return TargetLowering::LowerXConstraint(ConstraintVT);
19072 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
19073 /// vector. If it is invalid, don't add anything to Ops.
19074 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
19075 std::string &Constraint,
19076 std::vector<SDValue>&Ops,
19077 SelectionDAG &DAG) const {
19078 SDValue Result(0, 0);
19080 // Only support length 1 constraints for now.
19081 if (Constraint.length() > 1) return;
19083 char ConstraintLetter = Constraint[0];
19084 switch (ConstraintLetter) {
19087 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
19088 if (C->getZExtValue() <= 31) {
19089 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
19095 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
19096 if (C->getZExtValue() <= 63) {
19097 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
19103 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
19104 if (isInt<8>(C->getSExtValue())) {
19105 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
19111 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
19112 if (C->getZExtValue() <= 255) {
19113 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
19119 // 32-bit signed value
19120 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
19121 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
19122 C->getSExtValue())) {
19123 // Widen to 64 bits here to get it sign extended.
19124 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
19127 // FIXME gcc accepts some relocatable values here too, but only in certain
19128 // memory models; it's complicated.
19133 // 32-bit unsigned value
19134 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
19135 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
19136 C->getZExtValue())) {
19137 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
19141 // FIXME gcc accepts some relocatable values here too, but only in certain
19142 // memory models; it's complicated.
19146 // Literal immediates are always ok.
19147 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
19148 // Widen to 64 bits here to get it sign extended.
19149 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
19153 // In any sort of PIC mode addresses need to be computed at runtime by
19154 // adding in a register or some sort of table lookup. These can't
19155 // be used as immediates.
19156 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
19159 // If we are in non-pic codegen mode, we allow the address of a global (with
19160 // an optional displacement) to be used with 'i'.
19161 GlobalAddressSDNode *GA = 0;
19162 int64_t Offset = 0;
19164 // Match either (GA), (GA+C), (GA+C1+C2), etc.
19166 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
19167 Offset += GA->getOffset();
19169 } else if (Op.getOpcode() == ISD::ADD) {
19170 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
19171 Offset += C->getZExtValue();
19172 Op = Op.getOperand(0);
19175 } else if (Op.getOpcode() == ISD::SUB) {
19176 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
19177 Offset += -C->getZExtValue();
19178 Op = Op.getOperand(0);
19183 // Otherwise, this isn't something we can handle, reject it.
19187 const GlobalValue *GV = GA->getGlobal();
19188 // If we require an extra load to get this address, as in PIC mode, we
19189 // can't accept it.
19190 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
19191 getTargetMachine())))
19194 Result = DAG.getTargetGlobalAddress(GV, SDLoc(Op),
19195 GA->getValueType(0), Offset);
19200 if (Result.getNode()) {
19201 Ops.push_back(Result);
19204 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
19207 std::pair<unsigned, const TargetRegisterClass*>
19208 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
19210 // First, see if this is a constraint that directly corresponds to an LLVM
19212 if (Constraint.size() == 1) {
19213 // GCC Constraint Letters
19214 switch (Constraint[0]) {
19216 // TODO: Slight differences here in allocation order and leaving
19217 // RIP in the class. Do they matter any more here than they do
19218 // in the normal allocation?
19219 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
19220 if (Subtarget->is64Bit()) {
19221 if (VT == MVT::i32 || VT == MVT::f32)
19222 return std::make_pair(0U, &X86::GR32RegClass);
19223 if (VT == MVT::i16)
19224 return std::make_pair(0U, &X86::GR16RegClass);
19225 if (VT == MVT::i8 || VT == MVT::i1)
19226 return std::make_pair(0U, &X86::GR8RegClass);
19227 if (VT == MVT::i64 || VT == MVT::f64)
19228 return std::make_pair(0U, &X86::GR64RegClass);
19231 // 32-bit fallthrough
19232 case 'Q': // Q_REGS
19233 if (VT == MVT::i32 || VT == MVT::f32)
19234 return std::make_pair(0U, &X86::GR32_ABCDRegClass);
19235 if (VT == MVT::i16)
19236 return std::make_pair(0U, &X86::GR16_ABCDRegClass);
19237 if (VT == MVT::i8 || VT == MVT::i1)
19238 return std::make_pair(0U, &X86::GR8_ABCD_LRegClass);
19239 if (VT == MVT::i64)
19240 return std::make_pair(0U, &X86::GR64_ABCDRegClass);
19242 case 'r': // GENERAL_REGS
19243 case 'l': // INDEX_REGS
19244 if (VT == MVT::i8 || VT == MVT::i1)
19245 return std::make_pair(0U, &X86::GR8RegClass);
19246 if (VT == MVT::i16)
19247 return std::make_pair(0U, &X86::GR16RegClass);
19248 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
19249 return std::make_pair(0U, &X86::GR32RegClass);
19250 return std::make_pair(0U, &X86::GR64RegClass);
19251 case 'R': // LEGACY_REGS
19252 if (VT == MVT::i8 || VT == MVT::i1)
19253 return std::make_pair(0U, &X86::GR8_NOREXRegClass);
19254 if (VT == MVT::i16)
19255 return std::make_pair(0U, &X86::GR16_NOREXRegClass);
19256 if (VT == MVT::i32 || !Subtarget->is64Bit())
19257 return std::make_pair(0U, &X86::GR32_NOREXRegClass);
19258 return std::make_pair(0U, &X86::GR64_NOREXRegClass);
19259 case 'f': // FP Stack registers.
19260 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
19261 // value to the correct fpstack register class.
19262 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
19263 return std::make_pair(0U, &X86::RFP32RegClass);
19264 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
19265 return std::make_pair(0U, &X86::RFP64RegClass);
19266 return std::make_pair(0U, &X86::RFP80RegClass);
19267 case 'y': // MMX_REGS if MMX allowed.
19268 if (!Subtarget->hasMMX()) break;
19269 return std::make_pair(0U, &X86::VR64RegClass);
19270 case 'Y': // SSE_REGS if SSE2 allowed
19271 if (!Subtarget->hasSSE2()) break;
19273 case 'x': // SSE_REGS if SSE1 allowed or AVX_REGS if AVX allowed
19274 if (!Subtarget->hasSSE1()) break;
19276 switch (VT.SimpleTy) {
19278 // Scalar SSE types.
19281 return std::make_pair(0U, &X86::FR32RegClass);
19284 return std::make_pair(0U, &X86::FR64RegClass);
19292 return std::make_pair(0U, &X86::VR128RegClass);
19300 return std::make_pair(0U, &X86::VR256RegClass);
19305 return std::make_pair(0U, &X86::VR512RegClass);
19311 // Use the default implementation in TargetLowering to convert the register
19312 // constraint into a member of a register class.
19313 std::pair<unsigned, const TargetRegisterClass*> Res;
19314 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
19316 // Not found as a standard register?
19317 if (Res.second == 0) {
19318 // Map st(0) -> st(7) -> ST0
19319 if (Constraint.size() == 7 && Constraint[0] == '{' &&
19320 tolower(Constraint[1]) == 's' &&
19321 tolower(Constraint[2]) == 't' &&
19322 Constraint[3] == '(' &&
19323 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
19324 Constraint[5] == ')' &&
19325 Constraint[6] == '}') {
19327 Res.first = X86::ST0+Constraint[4]-'0';
19328 Res.second = &X86::RFP80RegClass;
19332 // GCC allows "st(0)" to be called just plain "st".
19333 if (StringRef("{st}").equals_lower(Constraint)) {
19334 Res.first = X86::ST0;
19335 Res.second = &X86::RFP80RegClass;
19340 if (StringRef("{flags}").equals_lower(Constraint)) {
19341 Res.first = X86::EFLAGS;
19342 Res.second = &X86::CCRRegClass;
19346 // 'A' means EAX + EDX.
19347 if (Constraint == "A") {
19348 Res.first = X86::EAX;
19349 Res.second = &X86::GR32_ADRegClass;
19355 // Otherwise, check to see if this is a register class of the wrong value
19356 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
19357 // turn into {ax},{dx}.
19358 if (Res.second->hasType(VT))
19359 return Res; // Correct type already, nothing to do.
19361 // All of the single-register GCC register classes map their values onto
19362 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
19363 // really want an 8-bit or 32-bit register, map to the appropriate register
19364 // class and return the appropriate register.
19365 if (Res.second == &X86::GR16RegClass) {
19366 if (VT == MVT::i8 || VT == MVT::i1) {
19367 unsigned DestReg = 0;
19368 switch (Res.first) {
19370 case X86::AX: DestReg = X86::AL; break;
19371 case X86::DX: DestReg = X86::DL; break;
19372 case X86::CX: DestReg = X86::CL; break;
19373 case X86::BX: DestReg = X86::BL; break;
19376 Res.first = DestReg;
19377 Res.second = &X86::GR8RegClass;
19379 } else if (VT == MVT::i32 || VT == MVT::f32) {
19380 unsigned DestReg = 0;
19381 switch (Res.first) {
19383 case X86::AX: DestReg = X86::EAX; break;
19384 case X86::DX: DestReg = X86::EDX; break;
19385 case X86::CX: DestReg = X86::ECX; break;
19386 case X86::BX: DestReg = X86::EBX; break;
19387 case X86::SI: DestReg = X86::ESI; break;
19388 case X86::DI: DestReg = X86::EDI; break;
19389 case X86::BP: DestReg = X86::EBP; break;
19390 case X86::SP: DestReg = X86::ESP; break;
19393 Res.first = DestReg;
19394 Res.second = &X86::GR32RegClass;
19396 } else if (VT == MVT::i64 || VT == MVT::f64) {
19397 unsigned DestReg = 0;
19398 switch (Res.first) {
19400 case X86::AX: DestReg = X86::RAX; break;
19401 case X86::DX: DestReg = X86::RDX; break;
19402 case X86::CX: DestReg = X86::RCX; break;
19403 case X86::BX: DestReg = X86::RBX; break;
19404 case X86::SI: DestReg = X86::RSI; break;
19405 case X86::DI: DestReg = X86::RDI; break;
19406 case X86::BP: DestReg = X86::RBP; break;
19407 case X86::SP: DestReg = X86::RSP; break;
19410 Res.first = DestReg;
19411 Res.second = &X86::GR64RegClass;
19414 } else if (Res.second == &X86::FR32RegClass ||
19415 Res.second == &X86::FR64RegClass ||
19416 Res.second == &X86::VR128RegClass ||
19417 Res.second == &X86::VR256RegClass ||
19418 Res.second == &X86::FR32XRegClass ||
19419 Res.second == &X86::FR64XRegClass ||
19420 Res.second == &X86::VR128XRegClass ||
19421 Res.second == &X86::VR256XRegClass ||
19422 Res.second == &X86::VR512RegClass) {
19423 // Handle references to XMM physical registers that got mapped into the
19424 // wrong class. This can happen with constraints like {xmm0} where the
19425 // target independent register mapper will just pick the first match it can
19426 // find, ignoring the required type.
19428 if (VT == MVT::f32 || VT == MVT::i32)
19429 Res.second = &X86::FR32RegClass;
19430 else if (VT == MVT::f64 || VT == MVT::i64)
19431 Res.second = &X86::FR64RegClass;
19432 else if (X86::VR128RegClass.hasType(VT))
19433 Res.second = &X86::VR128RegClass;
19434 else if (X86::VR256RegClass.hasType(VT))
19435 Res.second = &X86::VR256RegClass;
19436 else if (X86::VR512RegClass.hasType(VT))
19437 Res.second = &X86::VR512RegClass;