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);
1394 // Custom lower several nodes.
1395 for (int i = MVT::FIRST_VECTOR_VALUETYPE;
1396 i <= MVT::LAST_VECTOR_VALUETYPE; ++i) {
1397 MVT VT = (MVT::SimpleValueType)i;
1399 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1400 // Extract subvector is special because the value type
1401 // (result) is 256/128-bit but the source is 512-bit wide.
1402 if (VT.is128BitVector() || VT.is256BitVector())
1403 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1405 if (VT.getVectorElementType() == MVT::i1)
1406 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
1408 // Do not attempt to custom lower other non-512-bit vectors
1409 if (!VT.is512BitVector())
1412 if (VT != MVT::v8i64) {
1413 setOperationAction(ISD::XOR, VT, Promote);
1414 AddPromotedToType (ISD::XOR, VT, MVT::v8i64);
1415 setOperationAction(ISD::OR, VT, Promote);
1416 AddPromotedToType (ISD::OR, VT, MVT::v8i64);
1417 setOperationAction(ISD::AND, VT, Promote);
1418 AddPromotedToType (ISD::AND, VT, MVT::v8i64);
1420 if ( EltSize >= 32) {
1421 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1422 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1423 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1424 setOperationAction(ISD::VSELECT, VT, Legal);
1425 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1426 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1427 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1430 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1431 MVT VT = (MVT::SimpleValueType)i;
1433 // Do not attempt to promote non-256-bit vectors
1434 if (!VT.is512BitVector())
1437 setOperationAction(ISD::LOAD, VT, Promote);
1438 AddPromotedToType (ISD::LOAD, VT, MVT::v8i64);
1439 setOperationAction(ISD::SELECT, VT, Promote);
1440 AddPromotedToType (ISD::SELECT, VT, MVT::v8i64);
1444 // SIGN_EXTEND_INREGs are evaluated by the extend type. Handle the expansion
1445 // of this type with custom code.
1446 for (int VT = MVT::FIRST_VECTOR_VALUETYPE;
1447 VT != MVT::LAST_VECTOR_VALUETYPE; VT++) {
1448 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,
1452 // We want to custom lower some of our intrinsics.
1453 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1454 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
1456 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1457 // handle type legalization for these operations here.
1459 // FIXME: We really should do custom legalization for addition and
1460 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1461 // than generic legalization for 64-bit multiplication-with-overflow, though.
1462 for (unsigned i = 0, e = 3+Subtarget->is64Bit(); i != e; ++i) {
1463 // Add/Sub/Mul with overflow operations are custom lowered.
1465 setOperationAction(ISD::SADDO, VT, Custom);
1466 setOperationAction(ISD::UADDO, VT, Custom);
1467 setOperationAction(ISD::SSUBO, VT, Custom);
1468 setOperationAction(ISD::USUBO, VT, Custom);
1469 setOperationAction(ISD::SMULO, VT, Custom);
1470 setOperationAction(ISD::UMULO, VT, Custom);
1473 // There are no 8-bit 3-address imul/mul instructions
1474 setOperationAction(ISD::SMULO, MVT::i8, Expand);
1475 setOperationAction(ISD::UMULO, MVT::i8, Expand);
1477 if (!Subtarget->is64Bit()) {
1478 // These libcalls are not available in 32-bit.
1479 setLibcallName(RTLIB::SHL_I128, 0);
1480 setLibcallName(RTLIB::SRL_I128, 0);
1481 setLibcallName(RTLIB::SRA_I128, 0);
1484 // Combine sin / cos into one node or libcall if possible.
1485 if (Subtarget->hasSinCos()) {
1486 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
1487 setLibcallName(RTLIB::SINCOS_F64, "sincos");
1488 if (Subtarget->isTargetDarwin()) {
1489 // For MacOSX, we don't want to the normal expansion of a libcall to
1490 // sincos. We want to issue a libcall to __sincos_stret to avoid memory
1492 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
1493 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
1497 // We have target-specific dag combine patterns for the following nodes:
1498 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1499 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1500 setTargetDAGCombine(ISD::VSELECT);
1501 setTargetDAGCombine(ISD::SELECT);
1502 setTargetDAGCombine(ISD::SHL);
1503 setTargetDAGCombine(ISD::SRA);
1504 setTargetDAGCombine(ISD::SRL);
1505 setTargetDAGCombine(ISD::OR);
1506 setTargetDAGCombine(ISD::AND);
1507 setTargetDAGCombine(ISD::ADD);
1508 setTargetDAGCombine(ISD::FADD);
1509 setTargetDAGCombine(ISD::FSUB);
1510 setTargetDAGCombine(ISD::FMA);
1511 setTargetDAGCombine(ISD::SUB);
1512 setTargetDAGCombine(ISD::LOAD);
1513 setTargetDAGCombine(ISD::STORE);
1514 setTargetDAGCombine(ISD::ZERO_EXTEND);
1515 setTargetDAGCombine(ISD::ANY_EXTEND);
1516 setTargetDAGCombine(ISD::SIGN_EXTEND);
1517 setTargetDAGCombine(ISD::SIGN_EXTEND_INREG);
1518 setTargetDAGCombine(ISD::TRUNCATE);
1519 setTargetDAGCombine(ISD::SINT_TO_FP);
1520 setTargetDAGCombine(ISD::SETCC);
1521 if (Subtarget->is64Bit())
1522 setTargetDAGCombine(ISD::MUL);
1523 setTargetDAGCombine(ISD::XOR);
1525 computeRegisterProperties();
1527 // On Darwin, -Os means optimize for size without hurting performance,
1528 // do not reduce the limit.
1529 MaxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1530 MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 16 : 8;
1531 MaxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1532 MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1533 MaxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1534 MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1535 setPrefLoopAlignment(4); // 2^4 bytes.
1537 // Predictable cmov don't hurt on atom because it's in-order.
1538 PredictableSelectIsExpensive = !Subtarget->isAtom();
1540 setPrefFunctionAlignment(4); // 2^4 bytes.
1543 EVT X86TargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
1544 if (!VT.isVector()) return MVT::i8;
1545 return VT.changeVectorElementTypeToInteger();
1548 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1549 /// the desired ByVal argument alignment.
1550 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1553 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1554 if (VTy->getBitWidth() == 128)
1556 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1557 unsigned EltAlign = 0;
1558 getMaxByValAlign(ATy->getElementType(), EltAlign);
1559 if (EltAlign > MaxAlign)
1560 MaxAlign = EltAlign;
1561 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1562 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1563 unsigned EltAlign = 0;
1564 getMaxByValAlign(STy->getElementType(i), EltAlign);
1565 if (EltAlign > MaxAlign)
1566 MaxAlign = EltAlign;
1573 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1574 /// function arguments in the caller parameter area. For X86, aggregates
1575 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1576 /// are at 4-byte boundaries.
1577 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty) const {
1578 if (Subtarget->is64Bit()) {
1579 // Max of 8 and alignment of type.
1580 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1587 if (Subtarget->hasSSE1())
1588 getMaxByValAlign(Ty, Align);
1592 /// getOptimalMemOpType - Returns the target specific optimal type for load
1593 /// and store operations as a result of memset, memcpy, and memmove
1594 /// lowering. If DstAlign is zero that means it's safe to destination
1595 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1596 /// means there isn't a need to check it against alignment requirement,
1597 /// probably because the source does not need to be loaded. If 'IsMemset' is
1598 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
1599 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
1600 /// source is constant so it does not need to be loaded.
1601 /// It returns EVT::Other if the type should be determined using generic
1602 /// target-independent logic.
1604 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1605 unsigned DstAlign, unsigned SrcAlign,
1606 bool IsMemset, bool ZeroMemset,
1608 MachineFunction &MF) const {
1609 const Function *F = MF.getFunction();
1610 if ((!IsMemset || ZeroMemset) &&
1611 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
1612 Attribute::NoImplicitFloat)) {
1614 (Subtarget->isUnalignedMemAccessFast() ||
1615 ((DstAlign == 0 || DstAlign >= 16) &&
1616 (SrcAlign == 0 || SrcAlign >= 16)))) {
1618 if (Subtarget->hasInt256())
1620 if (Subtarget->hasFp256())
1623 if (Subtarget->hasSSE2())
1625 if (Subtarget->hasSSE1())
1627 } else if (!MemcpyStrSrc && Size >= 8 &&
1628 !Subtarget->is64Bit() &&
1629 Subtarget->hasSSE2()) {
1630 // Do not use f64 to lower memcpy if source is string constant. It's
1631 // better to use i32 to avoid the loads.
1635 if (Subtarget->is64Bit() && Size >= 8)
1640 bool X86TargetLowering::isSafeMemOpType(MVT VT) const {
1642 return X86ScalarSSEf32;
1643 else if (VT == MVT::f64)
1644 return X86ScalarSSEf64;
1649 X86TargetLowering::allowsUnalignedMemoryAccesses(EVT VT, bool *Fast) const {
1651 *Fast = Subtarget->isUnalignedMemAccessFast();
1655 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1656 /// current function. The returned value is a member of the
1657 /// MachineJumpTableInfo::JTEntryKind enum.
1658 unsigned X86TargetLowering::getJumpTableEncoding() const {
1659 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1661 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1662 Subtarget->isPICStyleGOT())
1663 return MachineJumpTableInfo::EK_Custom32;
1665 // Otherwise, use the normal jump table encoding heuristics.
1666 return TargetLowering::getJumpTableEncoding();
1670 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1671 const MachineBasicBlock *MBB,
1672 unsigned uid,MCContext &Ctx) const{
1673 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1674 Subtarget->isPICStyleGOT());
1675 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1677 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1678 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1681 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1683 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1684 SelectionDAG &DAG) const {
1685 if (!Subtarget->is64Bit())
1686 // This doesn't have SDLoc associated with it, but is not really the
1687 // same as a Register.
1688 return DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), getPointerTy());
1692 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1693 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1695 const MCExpr *X86TargetLowering::
1696 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1697 MCContext &Ctx) const {
1698 // X86-64 uses RIP relative addressing based on the jump table label.
1699 if (Subtarget->isPICStyleRIPRel())
1700 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1702 // Otherwise, the reference is relative to the PIC base.
1703 return MCSymbolRefExpr::Create(MF->getPICBaseSymbol(), Ctx);
1706 // FIXME: Why this routine is here? Move to RegInfo!
1707 std::pair<const TargetRegisterClass*, uint8_t>
1708 X86TargetLowering::findRepresentativeClass(MVT VT) const{
1709 const TargetRegisterClass *RRC = 0;
1711 switch (VT.SimpleTy) {
1713 return TargetLowering::findRepresentativeClass(VT);
1714 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
1715 RRC = Subtarget->is64Bit() ?
1716 (const TargetRegisterClass*)&X86::GR64RegClass :
1717 (const TargetRegisterClass*)&X86::GR32RegClass;
1720 RRC = &X86::VR64RegClass;
1722 case MVT::f32: case MVT::f64:
1723 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1724 case MVT::v4f32: case MVT::v2f64:
1725 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
1727 RRC = &X86::VR128RegClass;
1730 return std::make_pair(RRC, Cost);
1733 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1734 unsigned &Offset) const {
1735 if (!Subtarget->isTargetLinux())
1738 if (Subtarget->is64Bit()) {
1739 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1741 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1753 //===----------------------------------------------------------------------===//
1754 // Return Value Calling Convention Implementation
1755 //===----------------------------------------------------------------------===//
1757 #include "X86GenCallingConv.inc"
1760 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv,
1761 MachineFunction &MF, bool isVarArg,
1762 const SmallVectorImpl<ISD::OutputArg> &Outs,
1763 LLVMContext &Context) const {
1764 SmallVector<CCValAssign, 16> RVLocs;
1765 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
1767 return CCInfo.CheckReturn(Outs, RetCC_X86);
1771 X86TargetLowering::LowerReturn(SDValue Chain,
1772 CallingConv::ID CallConv, bool isVarArg,
1773 const SmallVectorImpl<ISD::OutputArg> &Outs,
1774 const SmallVectorImpl<SDValue> &OutVals,
1775 SDLoc dl, SelectionDAG &DAG) const {
1776 MachineFunction &MF = DAG.getMachineFunction();
1777 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1779 SmallVector<CCValAssign, 16> RVLocs;
1780 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
1781 RVLocs, *DAG.getContext());
1782 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1785 SmallVector<SDValue, 6> RetOps;
1786 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1787 // Operand #1 = Bytes To Pop
1788 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
1791 // Copy the result values into the output registers.
1792 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1793 CCValAssign &VA = RVLocs[i];
1794 assert(VA.isRegLoc() && "Can only return in registers!");
1795 SDValue ValToCopy = OutVals[i];
1796 EVT ValVT = ValToCopy.getValueType();
1798 // Promote values to the appropriate types
1799 if (VA.getLocInfo() == CCValAssign::SExt)
1800 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
1801 else if (VA.getLocInfo() == CCValAssign::ZExt)
1802 ValToCopy = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), ValToCopy);
1803 else if (VA.getLocInfo() == CCValAssign::AExt)
1804 ValToCopy = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), ValToCopy);
1805 else if (VA.getLocInfo() == CCValAssign::BCvt)
1806 ValToCopy = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), ValToCopy);
1808 // If this is x86-64, and we disabled SSE, we can't return FP values,
1809 // or SSE or MMX vectors.
1810 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
1811 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
1812 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
1813 report_fatal_error("SSE register return with SSE disabled");
1815 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
1816 // llvm-gcc has never done it right and no one has noticed, so this
1817 // should be OK for now.
1818 if (ValVT == MVT::f64 &&
1819 (Subtarget->is64Bit() && !Subtarget->hasSSE2()))
1820 report_fatal_error("SSE2 register return with SSE2 disabled");
1822 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1823 // the RET instruction and handled by the FP Stackifier.
1824 if (VA.getLocReg() == X86::ST0 ||
1825 VA.getLocReg() == X86::ST1) {
1826 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1827 // change the value to the FP stack register class.
1828 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1829 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1830 RetOps.push_back(ValToCopy);
1831 // Don't emit a copytoreg.
1835 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1836 // which is returned in RAX / RDX.
1837 if (Subtarget->is64Bit()) {
1838 if (ValVT == MVT::x86mmx) {
1839 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1840 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ValToCopy);
1841 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
1843 // If we don't have SSE2 available, convert to v4f32 so the generated
1844 // register is legal.
1845 if (!Subtarget->hasSSE2())
1846 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32,ValToCopy);
1851 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1852 Flag = Chain.getValue(1);
1853 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
1856 // The x86-64 ABIs require that for returning structs by value we copy
1857 // the sret argument into %rax/%eax (depending on ABI) for the return.
1858 // Win32 requires us to put the sret argument to %eax as well.
1859 // We saved the argument into a virtual register in the entry block,
1860 // so now we copy the value out and into %rax/%eax.
1861 if (DAG.getMachineFunction().getFunction()->hasStructRetAttr() &&
1862 (Subtarget->is64Bit() || Subtarget->isTargetWindows())) {
1863 MachineFunction &MF = DAG.getMachineFunction();
1864 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1865 unsigned Reg = FuncInfo->getSRetReturnReg();
1867 "SRetReturnReg should have been set in LowerFormalArguments().");
1868 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1871 = (Subtarget->is64Bit() && !Subtarget->isTarget64BitILP32()) ?
1872 X86::RAX : X86::EAX;
1873 Chain = DAG.getCopyToReg(Chain, dl, RetValReg, Val, Flag);
1874 Flag = Chain.getValue(1);
1876 // RAX/EAX now acts like a return value.
1877 RetOps.push_back(DAG.getRegister(RetValReg, getPointerTy()));
1880 RetOps[0] = Chain; // Update chain.
1882 // Add the flag if we have it.
1884 RetOps.push_back(Flag);
1886 return DAG.getNode(X86ISD::RET_FLAG, dl,
1887 MVT::Other, &RetOps[0], RetOps.size());
1890 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
1891 if (N->getNumValues() != 1)
1893 if (!N->hasNUsesOfValue(1, 0))
1896 SDValue TCChain = Chain;
1897 SDNode *Copy = *N->use_begin();
1898 if (Copy->getOpcode() == ISD::CopyToReg) {
1899 // If the copy has a glue operand, we conservatively assume it isn't safe to
1900 // perform a tail call.
1901 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
1903 TCChain = Copy->getOperand(0);
1904 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
1907 bool HasRet = false;
1908 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
1910 if (UI->getOpcode() != X86ISD::RET_FLAG)
1923 X86TargetLowering::getTypeForExtArgOrReturn(MVT VT,
1924 ISD::NodeType ExtendKind) const {
1926 // TODO: Is this also valid on 32-bit?
1927 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
1928 ReturnMVT = MVT::i8;
1930 ReturnMVT = MVT::i32;
1932 MVT MinVT = getRegisterType(ReturnMVT);
1933 return VT.bitsLT(MinVT) ? MinVT : VT;
1936 /// LowerCallResult - Lower the result values of a call into the
1937 /// appropriate copies out of appropriate physical registers.
1940 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1941 CallingConv::ID CallConv, bool isVarArg,
1942 const SmallVectorImpl<ISD::InputArg> &Ins,
1943 SDLoc dl, SelectionDAG &DAG,
1944 SmallVectorImpl<SDValue> &InVals) const {
1946 // Assign locations to each value returned by this call.
1947 SmallVector<CCValAssign, 16> RVLocs;
1948 bool Is64Bit = Subtarget->is64Bit();
1949 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1950 getTargetMachine(), RVLocs, *DAG.getContext());
1951 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1953 // Copy all of the result registers out of their specified physreg.
1954 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
1955 CCValAssign &VA = RVLocs[i];
1956 EVT CopyVT = VA.getValVT();
1958 // If this is x86-64, and we disabled SSE, we can't return FP values
1959 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1960 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1961 report_fatal_error("SSE register return with SSE disabled");
1966 // If this is a call to a function that returns an fp value on the floating
1967 // point stack, we must guarantee the value is popped from the stack, so
1968 // a CopyFromReg is not good enough - the copy instruction may be eliminated
1969 // if the return value is not used. We use the FpPOP_RETVAL instruction
1971 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1) {
1972 // If we prefer to use the value in xmm registers, copy it out as f80 and
1973 // use a truncate to move it from fp stack reg to xmm reg.
1974 if (isScalarFPTypeInSSEReg(VA.getValVT())) CopyVT = MVT::f80;
1975 SDValue Ops[] = { Chain, InFlag };
1976 Chain = SDValue(DAG.getMachineNode(X86::FpPOP_RETVAL, dl, CopyVT,
1977 MVT::Other, MVT::Glue, Ops), 1);
1978 Val = Chain.getValue(0);
1980 // Round the f80 to the right size, which also moves it to the appropriate
1982 if (CopyVT != VA.getValVT())
1983 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1984 // This truncation won't change the value.
1985 DAG.getIntPtrConstant(1));
1987 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1988 CopyVT, InFlag).getValue(1);
1989 Val = Chain.getValue(0);
1991 InFlag = Chain.getValue(2);
1992 InVals.push_back(Val);
1998 //===----------------------------------------------------------------------===//
1999 // C & StdCall & Fast Calling Convention implementation
2000 //===----------------------------------------------------------------------===//
2001 // StdCall calling convention seems to be standard for many Windows' API
2002 // routines and around. It differs from C calling convention just a little:
2003 // callee should clean up the stack, not caller. Symbols should be also
2004 // decorated in some fancy way :) It doesn't support any vector arguments.
2005 // For info on fast calling convention see Fast Calling Convention (tail call)
2006 // implementation LowerX86_32FastCCCallTo.
2008 /// CallIsStructReturn - Determines whether a call uses struct return
2010 enum StructReturnType {
2015 static StructReturnType
2016 callIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
2018 return NotStructReturn;
2020 const ISD::ArgFlagsTy &Flags = Outs[0].Flags;
2021 if (!Flags.isSRet())
2022 return NotStructReturn;
2023 if (Flags.isInReg())
2024 return RegStructReturn;
2025 return StackStructReturn;
2028 /// ArgsAreStructReturn - Determines whether a function uses struct
2029 /// return semantics.
2030 static StructReturnType
2031 argsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
2033 return NotStructReturn;
2035 const ISD::ArgFlagsTy &Flags = Ins[0].Flags;
2036 if (!Flags.isSRet())
2037 return NotStructReturn;
2038 if (Flags.isInReg())
2039 return RegStructReturn;
2040 return StackStructReturn;
2043 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
2044 /// by "Src" to address "Dst" with size and alignment information specified by
2045 /// the specific parameter attribute. The copy will be passed as a byval
2046 /// function parameter.
2048 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2049 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2051 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
2053 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2054 /*isVolatile*/false, /*AlwaysInline=*/true,
2055 MachinePointerInfo(), MachinePointerInfo());
2058 /// IsTailCallConvention - Return true if the calling convention is one that
2059 /// supports tail call optimization.
2060 static bool IsTailCallConvention(CallingConv::ID CC) {
2061 return (CC == CallingConv::Fast || CC == CallingConv::GHC ||
2062 CC == CallingConv::HiPE);
2065 /// \brief Return true if the calling convention is a C calling convention.
2066 static bool IsCCallConvention(CallingConv::ID CC) {
2067 return (CC == CallingConv::C || CC == CallingConv::X86_64_Win64 ||
2068 CC == CallingConv::X86_64_SysV);
2071 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2072 if (!CI->isTailCall() || getTargetMachine().Options.DisableTailCalls)
2076 CallingConv::ID CalleeCC = CS.getCallingConv();
2077 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
2083 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
2084 /// a tailcall target by changing its ABI.
2085 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC,
2086 bool GuaranteedTailCallOpt) {
2087 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
2091 X86TargetLowering::LowerMemArgument(SDValue Chain,
2092 CallingConv::ID CallConv,
2093 const SmallVectorImpl<ISD::InputArg> &Ins,
2094 SDLoc dl, SelectionDAG &DAG,
2095 const CCValAssign &VA,
2096 MachineFrameInfo *MFI,
2098 // Create the nodes corresponding to a load from this parameter slot.
2099 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2100 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv,
2101 getTargetMachine().Options.GuaranteedTailCallOpt);
2102 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
2105 // If value is passed by pointer we have address passed instead of the value
2107 if (VA.getLocInfo() == CCValAssign::Indirect)
2108 ValVT = VA.getLocVT();
2110 ValVT = VA.getValVT();
2112 // FIXME: For now, all byval parameter objects are marked mutable. This can be
2113 // changed with more analysis.
2114 // In case of tail call optimization mark all arguments mutable. Since they
2115 // could be overwritten by lowering of arguments in case of a tail call.
2116 if (Flags.isByVal()) {
2117 unsigned Bytes = Flags.getByValSize();
2118 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2119 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
2120 return DAG.getFrameIndex(FI, getPointerTy());
2122 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2123 VA.getLocMemOffset(), isImmutable);
2124 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2125 return DAG.getLoad(ValVT, dl, Chain, FIN,
2126 MachinePointerInfo::getFixedStack(FI),
2127 false, false, false, 0);
2132 X86TargetLowering::LowerFormalArguments(SDValue Chain,
2133 CallingConv::ID CallConv,
2135 const SmallVectorImpl<ISD::InputArg> &Ins,
2138 SmallVectorImpl<SDValue> &InVals)
2140 MachineFunction &MF = DAG.getMachineFunction();
2141 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2143 const Function* Fn = MF.getFunction();
2144 if (Fn->hasExternalLinkage() &&
2145 Subtarget->isTargetCygMing() &&
2146 Fn->getName() == "main")
2147 FuncInfo->setForceFramePointer(true);
2149 MachineFrameInfo *MFI = MF.getFrameInfo();
2150 bool Is64Bit = Subtarget->is64Bit();
2151 bool IsWindows = Subtarget->isTargetWindows();
2152 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2154 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2155 "Var args not supported with calling convention fastcc, ghc or hipe");
2157 // Assign locations to all of the incoming arguments.
2158 SmallVector<CCValAssign, 16> ArgLocs;
2159 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
2160 ArgLocs, *DAG.getContext());
2162 // Allocate shadow area for Win64
2164 CCInfo.AllocateStack(32, 8);
2166 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
2168 unsigned LastVal = ~0U;
2170 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2171 CCValAssign &VA = ArgLocs[i];
2172 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
2174 assert(VA.getValNo() != LastVal &&
2175 "Don't support value assigned to multiple locs yet");
2177 LastVal = VA.getValNo();
2179 if (VA.isRegLoc()) {
2180 EVT RegVT = VA.getLocVT();
2181 const TargetRegisterClass *RC;
2182 if (RegVT == MVT::i32)
2183 RC = &X86::GR32RegClass;
2184 else if (Is64Bit && RegVT == MVT::i64)
2185 RC = &X86::GR64RegClass;
2186 else if (RegVT == MVT::f32)
2187 RC = &X86::FR32RegClass;
2188 else if (RegVT == MVT::f64)
2189 RC = &X86::FR64RegClass;
2190 else if (RegVT.is512BitVector())
2191 RC = &X86::VR512RegClass;
2192 else if (RegVT.is256BitVector())
2193 RC = &X86::VR256RegClass;
2194 else if (RegVT.is128BitVector())
2195 RC = &X86::VR128RegClass;
2196 else if (RegVT == MVT::x86mmx)
2197 RC = &X86::VR64RegClass;
2198 else if (RegVT == MVT::v8i1)
2199 RC = &X86::VK8RegClass;
2200 else if (RegVT == MVT::v16i1)
2201 RC = &X86::VK16RegClass;
2203 llvm_unreachable("Unknown argument type!");
2205 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2206 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2208 // If this is an 8 or 16-bit value, it is really passed promoted to 32
2209 // bits. Insert an assert[sz]ext to capture this, then truncate to the
2211 if (VA.getLocInfo() == CCValAssign::SExt)
2212 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2213 DAG.getValueType(VA.getValVT()));
2214 else if (VA.getLocInfo() == CCValAssign::ZExt)
2215 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2216 DAG.getValueType(VA.getValVT()));
2217 else if (VA.getLocInfo() == CCValAssign::BCvt)
2218 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2220 if (VA.isExtInLoc()) {
2221 // Handle MMX values passed in XMM regs.
2222 if (RegVT.isVector())
2223 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(), ArgValue);
2225 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2228 assert(VA.isMemLoc());
2229 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
2232 // If value is passed via pointer - do a load.
2233 if (VA.getLocInfo() == CCValAssign::Indirect)
2234 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
2235 MachinePointerInfo(), false, false, false, 0);
2237 InVals.push_back(ArgValue);
2240 // The x86-64 ABIs require that for returning structs by value we copy
2241 // the sret argument into %rax/%eax (depending on ABI) for the return.
2242 // Win32 requires us to put the sret argument to %eax as well.
2243 // Save the argument into a virtual register so that we can access it
2244 // from the return points.
2245 if (MF.getFunction()->hasStructRetAttr() &&
2246 (Subtarget->is64Bit() || Subtarget->isTargetWindows())) {
2247 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2248 unsigned Reg = FuncInfo->getSRetReturnReg();
2250 MVT PtrTy = getPointerTy();
2251 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrTy));
2252 FuncInfo->setSRetReturnReg(Reg);
2254 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
2255 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
2258 unsigned StackSize = CCInfo.getNextStackOffset();
2259 // Align stack specially for tail calls.
2260 if (FuncIsMadeTailCallSafe(CallConv,
2261 MF.getTarget().Options.GuaranteedTailCallOpt))
2262 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
2264 // If the function takes variable number of arguments, make a frame index for
2265 // the start of the first vararg value... for expansion of llvm.va_start.
2267 if (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
2268 CallConv != CallingConv::X86_ThisCall)) {
2269 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,true));
2272 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
2274 // FIXME: We should really autogenerate these arrays
2275 static const uint16_t GPR64ArgRegsWin64[] = {
2276 X86::RCX, X86::RDX, X86::R8, X86::R9
2278 static const uint16_t GPR64ArgRegs64Bit[] = {
2279 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
2281 static const uint16_t XMMArgRegs64Bit[] = {
2282 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2283 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2285 const uint16_t *GPR64ArgRegs;
2286 unsigned NumXMMRegs = 0;
2289 // The XMM registers which might contain var arg parameters are shadowed
2290 // in their paired GPR. So we only need to save the GPR to their home
2292 TotalNumIntRegs = 4;
2293 GPR64ArgRegs = GPR64ArgRegsWin64;
2295 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
2296 GPR64ArgRegs = GPR64ArgRegs64Bit;
2298 NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs64Bit,
2301 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
2304 bool NoImplicitFloatOps = Fn->getAttributes().
2305 hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat);
2306 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
2307 "SSE register cannot be used when SSE is disabled!");
2308 assert(!(NumXMMRegs && MF.getTarget().Options.UseSoftFloat &&
2309 NoImplicitFloatOps) &&
2310 "SSE register cannot be used when SSE is disabled!");
2311 if (MF.getTarget().Options.UseSoftFloat || NoImplicitFloatOps ||
2312 !Subtarget->hasSSE1())
2313 // Kernel mode asks for SSE to be disabled, so don't push them
2315 TotalNumXMMRegs = 0;
2318 const TargetFrameLowering &TFI = *getTargetMachine().getFrameLowering();
2319 // Get to the caller-allocated home save location. Add 8 to account
2320 // for the return address.
2321 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2322 FuncInfo->setRegSaveFrameIndex(
2323 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
2324 // Fixup to set vararg frame on shadow area (4 x i64).
2326 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
2328 // For X86-64, if there are vararg parameters that are passed via
2329 // registers, then we must store them to their spots on the stack so
2330 // they may be loaded by deferencing the result of va_next.
2331 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
2332 FuncInfo->setVarArgsFPOffset(TotalNumIntRegs * 8 + NumXMMRegs * 16);
2333 FuncInfo->setRegSaveFrameIndex(
2334 MFI->CreateStackObject(TotalNumIntRegs * 8 + TotalNumXMMRegs * 16, 16,
2338 // Store the integer parameter registers.
2339 SmallVector<SDValue, 8> MemOps;
2340 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
2342 unsigned Offset = FuncInfo->getVarArgsGPOffset();
2343 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
2344 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
2345 DAG.getIntPtrConstant(Offset));
2346 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
2347 &X86::GR64RegClass);
2348 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
2350 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2351 MachinePointerInfo::getFixedStack(
2352 FuncInfo->getRegSaveFrameIndex(), Offset),
2354 MemOps.push_back(Store);
2358 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
2359 // Now store the XMM (fp + vector) parameter registers.
2360 SmallVector<SDValue, 11> SaveXMMOps;
2361 SaveXMMOps.push_back(Chain);
2363 unsigned AL = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2364 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
2365 SaveXMMOps.push_back(ALVal);
2367 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2368 FuncInfo->getRegSaveFrameIndex()));
2369 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2370 FuncInfo->getVarArgsFPOffset()));
2372 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
2373 unsigned VReg = MF.addLiveIn(XMMArgRegs64Bit[NumXMMRegs],
2374 &X86::VR128RegClass);
2375 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
2376 SaveXMMOps.push_back(Val);
2378 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
2380 &SaveXMMOps[0], SaveXMMOps.size()));
2383 if (!MemOps.empty())
2384 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2385 &MemOps[0], MemOps.size());
2389 // Some CCs need callee pop.
2390 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
2391 MF.getTarget().Options.GuaranteedTailCallOpt)) {
2392 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
2394 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
2395 // If this is an sret function, the return should pop the hidden pointer.
2396 if (!Is64Bit && !IsTailCallConvention(CallConv) && !IsWindows &&
2397 argsAreStructReturn(Ins) == StackStructReturn)
2398 FuncInfo->setBytesToPopOnReturn(4);
2402 // RegSaveFrameIndex is X86-64 only.
2403 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
2404 if (CallConv == CallingConv::X86_FastCall ||
2405 CallConv == CallingConv::X86_ThisCall)
2406 // fastcc functions can't have varargs.
2407 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
2410 FuncInfo->setArgumentStackSize(StackSize);
2416 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
2417 SDValue StackPtr, SDValue Arg,
2418 SDLoc dl, SelectionDAG &DAG,
2419 const CCValAssign &VA,
2420 ISD::ArgFlagsTy Flags) const {
2421 unsigned LocMemOffset = VA.getLocMemOffset();
2422 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
2423 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
2424 if (Flags.isByVal())
2425 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
2427 return DAG.getStore(Chain, dl, Arg, PtrOff,
2428 MachinePointerInfo::getStack(LocMemOffset),
2432 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
2433 /// optimization is performed and it is required.
2435 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
2436 SDValue &OutRetAddr, SDValue Chain,
2437 bool IsTailCall, bool Is64Bit,
2438 int FPDiff, SDLoc dl) const {
2439 // Adjust the Return address stack slot.
2440 EVT VT = getPointerTy();
2441 OutRetAddr = getReturnAddressFrameIndex(DAG);
2443 // Load the "old" Return address.
2444 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
2445 false, false, false, 0);
2446 return SDValue(OutRetAddr.getNode(), 1);
2449 /// EmitTailCallStoreRetAddr - Emit a store of the return address if tail call
2450 /// optimization is performed and it is required (FPDiff!=0).
2452 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
2453 SDValue Chain, SDValue RetAddrFrIdx, EVT PtrVT,
2454 unsigned SlotSize, int FPDiff, SDLoc dl) {
2455 // Store the return address to the appropriate stack slot.
2456 if (!FPDiff) return Chain;
2457 // Calculate the new stack slot for the return address.
2458 int NewReturnAddrFI =
2459 MF.getFrameInfo()->CreateFixedObject(SlotSize, (int64_t)FPDiff - SlotSize,
2461 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, PtrVT);
2462 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
2463 MachinePointerInfo::getFixedStack(NewReturnAddrFI),
2469 X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2470 SmallVectorImpl<SDValue> &InVals) const {
2471 SelectionDAG &DAG = CLI.DAG;
2473 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
2474 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
2475 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
2476 SDValue Chain = CLI.Chain;
2477 SDValue Callee = CLI.Callee;
2478 CallingConv::ID CallConv = CLI.CallConv;
2479 bool &isTailCall = CLI.IsTailCall;
2480 bool isVarArg = CLI.IsVarArg;
2482 MachineFunction &MF = DAG.getMachineFunction();
2483 bool Is64Bit = Subtarget->is64Bit();
2484 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2485 bool IsWindows = Subtarget->isTargetWindows();
2486 StructReturnType SR = callIsStructReturn(Outs);
2487 bool IsSibcall = false;
2489 if (MF.getTarget().Options.DisableTailCalls)
2493 // Check if it's really possible to do a tail call.
2494 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
2495 isVarArg, SR != NotStructReturn,
2496 MF.getFunction()->hasStructRetAttr(), CLI.RetTy,
2497 Outs, OutVals, Ins, DAG);
2499 // Sibcalls are automatically detected tailcalls which do not require
2501 if (!MF.getTarget().Options.GuaranteedTailCallOpt && isTailCall)
2508 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2509 "Var args not supported with calling convention fastcc, ghc or hipe");
2511 // Analyze operands of the call, assigning locations to each operand.
2512 SmallVector<CCValAssign, 16> ArgLocs;
2513 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
2514 ArgLocs, *DAG.getContext());
2516 // Allocate shadow area for Win64
2518 CCInfo.AllocateStack(32, 8);
2520 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2522 // Get a count of how many bytes are to be pushed on the stack.
2523 unsigned NumBytes = CCInfo.getNextStackOffset();
2525 // This is a sibcall. The memory operands are available in caller's
2526 // own caller's stack.
2528 else if (getTargetMachine().Options.GuaranteedTailCallOpt &&
2529 IsTailCallConvention(CallConv))
2530 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
2533 if (isTailCall && !IsSibcall) {
2534 // Lower arguments at fp - stackoffset + fpdiff.
2535 X86MachineFunctionInfo *X86Info = MF.getInfo<X86MachineFunctionInfo>();
2536 unsigned NumBytesCallerPushed = X86Info->getBytesToPopOnReturn();
2538 FPDiff = NumBytesCallerPushed - NumBytes;
2540 // Set the delta of movement of the returnaddr stackslot.
2541 // But only set if delta is greater than previous delta.
2542 if (FPDiff < X86Info->getTCReturnAddrDelta())
2543 X86Info->setTCReturnAddrDelta(FPDiff);
2547 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true),
2550 SDValue RetAddrFrIdx;
2551 // Load return address for tail calls.
2552 if (isTailCall && FPDiff)
2553 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
2554 Is64Bit, FPDiff, dl);
2556 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2557 SmallVector<SDValue, 8> MemOpChains;
2560 // Walk the register/memloc assignments, inserting copies/loads. In the case
2561 // of tail call optimization arguments are handle later.
2562 const X86RegisterInfo *RegInfo =
2563 static_cast<const X86RegisterInfo*>(getTargetMachine().getRegisterInfo());
2564 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2565 CCValAssign &VA = ArgLocs[i];
2566 EVT RegVT = VA.getLocVT();
2567 SDValue Arg = OutVals[i];
2568 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2569 bool isByVal = Flags.isByVal();
2571 // Promote the value if needed.
2572 switch (VA.getLocInfo()) {
2573 default: llvm_unreachable("Unknown loc info!");
2574 case CCValAssign::Full: break;
2575 case CCValAssign::SExt:
2576 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
2578 case CCValAssign::ZExt:
2579 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
2581 case CCValAssign::AExt:
2582 if (RegVT.is128BitVector()) {
2583 // Special case: passing MMX values in XMM registers.
2584 Arg = DAG.getNode(ISD::BITCAST, dl, MVT::i64, Arg);
2585 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
2586 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
2588 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
2590 case CCValAssign::BCvt:
2591 Arg = DAG.getNode(ISD::BITCAST, dl, RegVT, Arg);
2593 case CCValAssign::Indirect: {
2594 // Store the argument.
2595 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
2596 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
2597 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
2598 MachinePointerInfo::getFixedStack(FI),
2605 if (VA.isRegLoc()) {
2606 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2607 if (isVarArg && IsWin64) {
2608 // Win64 ABI requires argument XMM reg to be copied to the corresponding
2609 // shadow reg if callee is a varargs function.
2610 unsigned ShadowReg = 0;
2611 switch (VA.getLocReg()) {
2612 case X86::XMM0: ShadowReg = X86::RCX; break;
2613 case X86::XMM1: ShadowReg = X86::RDX; break;
2614 case X86::XMM2: ShadowReg = X86::R8; break;
2615 case X86::XMM3: ShadowReg = X86::R9; break;
2618 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
2620 } else if (!IsSibcall && (!isTailCall || isByVal)) {
2621 assert(VA.isMemLoc());
2622 if (StackPtr.getNode() == 0)
2623 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
2625 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
2626 dl, DAG, VA, Flags));
2630 if (!MemOpChains.empty())
2631 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2632 &MemOpChains[0], MemOpChains.size());
2634 if (Subtarget->isPICStyleGOT()) {
2635 // ELF / PIC requires GOT in the EBX register before function calls via PLT
2638 RegsToPass.push_back(std::make_pair(unsigned(X86::EBX),
2639 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), getPointerTy())));
2641 // If we are tail calling and generating PIC/GOT style code load the
2642 // address of the callee into ECX. The value in ecx is used as target of
2643 // the tail jump. This is done to circumvent the ebx/callee-saved problem
2644 // for tail calls on PIC/GOT architectures. Normally we would just put the
2645 // address of GOT into ebx and then call target@PLT. But for tail calls
2646 // ebx would be restored (since ebx is callee saved) before jumping to the
2649 // Note: The actual moving to ECX is done further down.
2650 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2651 if (G && !G->getGlobal()->hasHiddenVisibility() &&
2652 !G->getGlobal()->hasProtectedVisibility())
2653 Callee = LowerGlobalAddress(Callee, DAG);
2654 else if (isa<ExternalSymbolSDNode>(Callee))
2655 Callee = LowerExternalSymbol(Callee, DAG);
2659 if (Is64Bit && isVarArg && !IsWin64) {
2660 // From AMD64 ABI document:
2661 // For calls that may call functions that use varargs or stdargs
2662 // (prototype-less calls or calls to functions containing ellipsis (...) in
2663 // the declaration) %al is used as hidden argument to specify the number
2664 // of SSE registers used. The contents of %al do not need to match exactly
2665 // the number of registers, but must be an ubound on the number of SSE
2666 // registers used and is in the range 0 - 8 inclusive.
2668 // Count the number of XMM registers allocated.
2669 static const uint16_t XMMArgRegs[] = {
2670 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2671 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2673 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
2674 assert((Subtarget->hasSSE1() || !NumXMMRegs)
2675 && "SSE registers cannot be used when SSE is disabled");
2677 RegsToPass.push_back(std::make_pair(unsigned(X86::AL),
2678 DAG.getConstant(NumXMMRegs, MVT::i8)));
2681 // For tail calls lower the arguments to the 'real' stack slot.
2683 // Force all the incoming stack arguments to be loaded from the stack
2684 // before any new outgoing arguments are stored to the stack, because the
2685 // outgoing stack slots may alias the incoming argument stack slots, and
2686 // the alias isn't otherwise explicit. This is slightly more conservative
2687 // than necessary, because it means that each store effectively depends
2688 // on every argument instead of just those arguments it would clobber.
2689 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2691 SmallVector<SDValue, 8> MemOpChains2;
2694 if (getTargetMachine().Options.GuaranteedTailCallOpt) {
2695 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2696 CCValAssign &VA = ArgLocs[i];
2699 assert(VA.isMemLoc());
2700 SDValue Arg = OutVals[i];
2701 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2702 // Create frame index.
2703 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2704 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2705 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2706 FIN = DAG.getFrameIndex(FI, getPointerTy());
2708 if (Flags.isByVal()) {
2709 // Copy relative to framepointer.
2710 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2711 if (StackPtr.getNode() == 0)
2712 StackPtr = DAG.getCopyFromReg(Chain, dl,
2713 RegInfo->getStackRegister(),
2715 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2717 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2721 // Store relative to framepointer.
2722 MemOpChains2.push_back(
2723 DAG.getStore(ArgChain, dl, Arg, FIN,
2724 MachinePointerInfo::getFixedStack(FI),
2730 if (!MemOpChains2.empty())
2731 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2732 &MemOpChains2[0], MemOpChains2.size());
2734 // Store the return address to the appropriate stack slot.
2735 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx,
2736 getPointerTy(), RegInfo->getSlotSize(),
2740 // Build a sequence of copy-to-reg nodes chained together with token chain
2741 // and flag operands which copy the outgoing args into registers.
2743 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2744 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2745 RegsToPass[i].second, InFlag);
2746 InFlag = Chain.getValue(1);
2749 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2750 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2751 // In the 64-bit large code model, we have to make all calls
2752 // through a register, since the call instruction's 32-bit
2753 // pc-relative offset may not be large enough to hold the whole
2755 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2756 // If the callee is a GlobalAddress node (quite common, every direct call
2757 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2760 // We should use extra load for direct calls to dllimported functions in
2762 const GlobalValue *GV = G->getGlobal();
2763 if (!GV->hasDLLImportLinkage()) {
2764 unsigned char OpFlags = 0;
2765 bool ExtraLoad = false;
2766 unsigned WrapperKind = ISD::DELETED_NODE;
2768 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2769 // external symbols most go through the PLT in PIC mode. If the symbol
2770 // has hidden or protected visibility, or if it is static or local, then
2771 // we don't need to use the PLT - we can directly call it.
2772 if (Subtarget->isTargetELF() &&
2773 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2774 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2775 OpFlags = X86II::MO_PLT;
2776 } else if (Subtarget->isPICStyleStubAny() &&
2777 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2778 (!Subtarget->getTargetTriple().isMacOSX() ||
2779 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
2780 // PC-relative references to external symbols should go through $stub,
2781 // unless we're building with the leopard linker or later, which
2782 // automatically synthesizes these stubs.
2783 OpFlags = X86II::MO_DARWIN_STUB;
2784 } else if (Subtarget->isPICStyleRIPRel() &&
2785 isa<Function>(GV) &&
2786 cast<Function>(GV)->getAttributes().
2787 hasAttribute(AttributeSet::FunctionIndex,
2788 Attribute::NonLazyBind)) {
2789 // If the function is marked as non-lazy, generate an indirect call
2790 // which loads from the GOT directly. This avoids runtime overhead
2791 // at the cost of eager binding (and one extra byte of encoding).
2792 OpFlags = X86II::MO_GOTPCREL;
2793 WrapperKind = X86ISD::WrapperRIP;
2797 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
2798 G->getOffset(), OpFlags);
2800 // Add a wrapper if needed.
2801 if (WrapperKind != ISD::DELETED_NODE)
2802 Callee = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Callee);
2803 // Add extra indirection if needed.
2805 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Callee,
2806 MachinePointerInfo::getGOT(),
2807 false, false, false, 0);
2809 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2810 unsigned char OpFlags = 0;
2812 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
2813 // external symbols should go through the PLT.
2814 if (Subtarget->isTargetELF() &&
2815 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2816 OpFlags = X86II::MO_PLT;
2817 } else if (Subtarget->isPICStyleStubAny() &&
2818 (!Subtarget->getTargetTriple().isMacOSX() ||
2819 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
2820 // PC-relative references to external symbols should go through $stub,
2821 // unless we're building with the leopard linker or later, which
2822 // automatically synthesizes these stubs.
2823 OpFlags = X86II::MO_DARWIN_STUB;
2826 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2830 // Returns a chain & a flag for retval copy to use.
2831 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2832 SmallVector<SDValue, 8> Ops;
2834 if (!IsSibcall && isTailCall) {
2835 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2836 DAG.getIntPtrConstant(0, true), InFlag, dl);
2837 InFlag = Chain.getValue(1);
2840 Ops.push_back(Chain);
2841 Ops.push_back(Callee);
2844 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2846 // Add argument registers to the end of the list so that they are known live
2848 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2849 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2850 RegsToPass[i].second.getValueType()));
2852 // Add a register mask operand representing the call-preserved registers.
2853 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2854 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
2855 assert(Mask && "Missing call preserved mask for calling convention");
2856 Ops.push_back(DAG.getRegisterMask(Mask));
2858 if (InFlag.getNode())
2859 Ops.push_back(InFlag);
2863 //// If this is the first return lowered for this function, add the regs
2864 //// to the liveout set for the function.
2865 // This isn't right, although it's probably harmless on x86; liveouts
2866 // should be computed from returns not tail calls. Consider a void
2867 // function making a tail call to a function returning int.
2868 return DAG.getNode(X86ISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
2871 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2872 InFlag = Chain.getValue(1);
2874 // Create the CALLSEQ_END node.
2875 unsigned NumBytesForCalleeToPush;
2876 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
2877 getTargetMachine().Options.GuaranteedTailCallOpt))
2878 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2879 else if (!Is64Bit && !IsTailCallConvention(CallConv) && !IsWindows &&
2880 SR == StackStructReturn)
2881 // If this is a call to a struct-return function, the callee
2882 // pops the hidden struct pointer, so we have to push it back.
2883 // This is common for Darwin/X86, Linux & Mingw32 targets.
2884 // For MSVC Win32 targets, the caller pops the hidden struct pointer.
2885 NumBytesForCalleeToPush = 4;
2887 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2889 // Returns a flag for retval copy to use.
2891 Chain = DAG.getCALLSEQ_END(Chain,
2892 DAG.getIntPtrConstant(NumBytes, true),
2893 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2896 InFlag = Chain.getValue(1);
2899 // Handle result values, copying them out of physregs into vregs that we
2901 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2902 Ins, dl, DAG, InVals);
2905 //===----------------------------------------------------------------------===//
2906 // Fast Calling Convention (tail call) implementation
2907 //===----------------------------------------------------------------------===//
2909 // Like std call, callee cleans arguments, convention except that ECX is
2910 // reserved for storing the tail called function address. Only 2 registers are
2911 // free for argument passing (inreg). Tail call optimization is performed
2913 // * tailcallopt is enabled
2914 // * caller/callee are fastcc
2915 // On X86_64 architecture with GOT-style position independent code only local
2916 // (within module) calls are supported at the moment.
2917 // To keep the stack aligned according to platform abi the function
2918 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2919 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2920 // If a tail called function callee has more arguments than the caller the
2921 // caller needs to make sure that there is room to move the RETADDR to. This is
2922 // achieved by reserving an area the size of the argument delta right after the
2923 // original REtADDR, but before the saved framepointer or the spilled registers
2924 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2936 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2937 /// for a 16 byte align requirement.
2939 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2940 SelectionDAG& DAG) const {
2941 MachineFunction &MF = DAG.getMachineFunction();
2942 const TargetMachine &TM = MF.getTarget();
2943 const X86RegisterInfo *RegInfo =
2944 static_cast<const X86RegisterInfo*>(TM.getRegisterInfo());
2945 const TargetFrameLowering &TFI = *TM.getFrameLowering();
2946 unsigned StackAlignment = TFI.getStackAlignment();
2947 uint64_t AlignMask = StackAlignment - 1;
2948 int64_t Offset = StackSize;
2949 unsigned SlotSize = RegInfo->getSlotSize();
2950 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2951 // Number smaller than 12 so just add the difference.
2952 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2954 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2955 Offset = ((~AlignMask) & Offset) + StackAlignment +
2956 (StackAlignment-SlotSize);
2961 /// MatchingStackOffset - Return true if the given stack call argument is
2962 /// already available in the same position (relatively) of the caller's
2963 /// incoming argument stack.
2965 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2966 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2967 const X86InstrInfo *TII) {
2968 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2970 if (Arg.getOpcode() == ISD::CopyFromReg) {
2971 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2972 if (!TargetRegisterInfo::isVirtualRegister(VR))
2974 MachineInstr *Def = MRI->getVRegDef(VR);
2977 if (!Flags.isByVal()) {
2978 if (!TII->isLoadFromStackSlot(Def, FI))
2981 unsigned Opcode = Def->getOpcode();
2982 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2983 Def->getOperand(1).isFI()) {
2984 FI = Def->getOperand(1).getIndex();
2985 Bytes = Flags.getByValSize();
2989 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2990 if (Flags.isByVal())
2991 // ByVal argument is passed in as a pointer but it's now being
2992 // dereferenced. e.g.
2993 // define @foo(%struct.X* %A) {
2994 // tail call @bar(%struct.X* byval %A)
2997 SDValue Ptr = Ld->getBasePtr();
2998 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
3001 FI = FINode->getIndex();
3002 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
3003 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
3004 FI = FINode->getIndex();
3005 Bytes = Flags.getByValSize();
3009 assert(FI != INT_MAX);
3010 if (!MFI->isFixedObjectIndex(FI))
3012 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
3015 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
3016 /// for tail call optimization. Targets which want to do tail call
3017 /// optimization should implement this function.
3019 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
3020 CallingConv::ID CalleeCC,
3022 bool isCalleeStructRet,
3023 bool isCallerStructRet,
3025 const SmallVectorImpl<ISD::OutputArg> &Outs,
3026 const SmallVectorImpl<SDValue> &OutVals,
3027 const SmallVectorImpl<ISD::InputArg> &Ins,
3028 SelectionDAG &DAG) const {
3029 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
3032 // If -tailcallopt is specified, make fastcc functions tail-callable.
3033 const MachineFunction &MF = DAG.getMachineFunction();
3034 const Function *CallerF = MF.getFunction();
3036 // If the function return type is x86_fp80 and the callee return type is not,
3037 // then the FP_EXTEND of the call result is not a nop. It's not safe to
3038 // perform a tailcall optimization here.
3039 if (CallerF->getReturnType()->isX86_FP80Ty() && !RetTy->isX86_FP80Ty())
3042 CallingConv::ID CallerCC = CallerF->getCallingConv();
3043 bool CCMatch = CallerCC == CalleeCC;
3044 bool IsCalleeWin64 = Subtarget->isCallingConvWin64(CalleeCC);
3045 bool IsCallerWin64 = Subtarget->isCallingConvWin64(CallerCC);
3047 if (getTargetMachine().Options.GuaranteedTailCallOpt) {
3048 if (IsTailCallConvention(CalleeCC) && CCMatch)
3053 // Look for obvious safe cases to perform tail call optimization that do not
3054 // require ABI changes. This is what gcc calls sibcall.
3056 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
3057 // emit a special epilogue.
3058 const X86RegisterInfo *RegInfo =
3059 static_cast<const X86RegisterInfo*>(getTargetMachine().getRegisterInfo());
3060 if (RegInfo->needsStackRealignment(MF))
3063 // Also avoid sibcall optimization if either caller or callee uses struct
3064 // return semantics.
3065 if (isCalleeStructRet || isCallerStructRet)
3068 // An stdcall caller is expected to clean up its arguments; the callee
3069 // isn't going to do that.
3070 if (!CCMatch && CallerCC == CallingConv::X86_StdCall)
3073 // Do not sibcall optimize vararg calls unless all arguments are passed via
3075 if (isVarArg && !Outs.empty()) {
3077 // Optimizing for varargs on Win64 is unlikely to be safe without
3078 // additional testing.
3079 if (IsCalleeWin64 || IsCallerWin64)
3082 SmallVector<CCValAssign, 16> ArgLocs;
3083 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
3084 getTargetMachine(), ArgLocs, *DAG.getContext());
3086 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3087 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
3088 if (!ArgLocs[i].isRegLoc())
3092 // If the call result is in ST0 / ST1, it needs to be popped off the x87
3093 // stack. Therefore, if it's not used by the call it is not safe to optimize
3094 // this into a sibcall.
3095 bool Unused = false;
3096 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
3103 SmallVector<CCValAssign, 16> RVLocs;
3104 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(),
3105 getTargetMachine(), RVLocs, *DAG.getContext());
3106 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
3107 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3108 CCValAssign &VA = RVLocs[i];
3109 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
3114 // If the calling conventions do not match, then we'd better make sure the
3115 // results are returned in the same way as what the caller expects.
3117 SmallVector<CCValAssign, 16> RVLocs1;
3118 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
3119 getTargetMachine(), RVLocs1, *DAG.getContext());
3120 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
3122 SmallVector<CCValAssign, 16> RVLocs2;
3123 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
3124 getTargetMachine(), RVLocs2, *DAG.getContext());
3125 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
3127 if (RVLocs1.size() != RVLocs2.size())
3129 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
3130 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
3132 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
3134 if (RVLocs1[i].isRegLoc()) {
3135 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
3138 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
3144 // If the callee takes no arguments then go on to check the results of the
3146 if (!Outs.empty()) {
3147 // Check if stack adjustment is needed. For now, do not do this if any
3148 // argument is passed on the stack.
3149 SmallVector<CCValAssign, 16> ArgLocs;
3150 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
3151 getTargetMachine(), ArgLocs, *DAG.getContext());
3153 // Allocate shadow area for Win64
3155 CCInfo.AllocateStack(32, 8);
3157 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3158 if (CCInfo.getNextStackOffset()) {
3159 MachineFunction &MF = DAG.getMachineFunction();
3160 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
3163 // Check if the arguments are already laid out in the right way as
3164 // the caller's fixed stack objects.
3165 MachineFrameInfo *MFI = MF.getFrameInfo();
3166 const MachineRegisterInfo *MRI = &MF.getRegInfo();
3167 const X86InstrInfo *TII =
3168 ((const X86TargetMachine&)getTargetMachine()).getInstrInfo();
3169 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3170 CCValAssign &VA = ArgLocs[i];
3171 SDValue Arg = OutVals[i];
3172 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3173 if (VA.getLocInfo() == CCValAssign::Indirect)
3175 if (!VA.isRegLoc()) {
3176 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
3183 // If the tailcall address may be in a register, then make sure it's
3184 // possible to register allocate for it. In 32-bit, the call address can
3185 // only target EAX, EDX, or ECX since the tail call must be scheduled after
3186 // callee-saved registers are restored. These happen to be the same
3187 // registers used to pass 'inreg' arguments so watch out for those.
3188 if (!Subtarget->is64Bit() &&
3189 ((!isa<GlobalAddressSDNode>(Callee) &&
3190 !isa<ExternalSymbolSDNode>(Callee)) ||
3191 getTargetMachine().getRelocationModel() == Reloc::PIC_)) {
3192 unsigned NumInRegs = 0;
3193 // In PIC we need an extra register to formulate the address computation
3195 unsigned MaxInRegs =
3196 (getTargetMachine().getRelocationModel() == Reloc::PIC_) ? 2 : 3;
3198 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3199 CCValAssign &VA = ArgLocs[i];
3202 unsigned Reg = VA.getLocReg();
3205 case X86::EAX: case X86::EDX: case X86::ECX:
3206 if (++NumInRegs == MaxInRegs)
3218 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
3219 const TargetLibraryInfo *libInfo) const {
3220 return X86::createFastISel(funcInfo, libInfo);
3223 //===----------------------------------------------------------------------===//
3224 // Other Lowering Hooks
3225 //===----------------------------------------------------------------------===//
3227 static bool MayFoldLoad(SDValue Op) {
3228 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
3231 static bool MayFoldIntoStore(SDValue Op) {
3232 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
3235 static bool isTargetShuffle(unsigned Opcode) {
3237 default: return false;
3238 case X86ISD::PSHUFD:
3239 case X86ISD::PSHUFHW:
3240 case X86ISD::PSHUFLW:
3242 case X86ISD::PALIGNR:
3243 case X86ISD::MOVLHPS:
3244 case X86ISD::MOVLHPD:
3245 case X86ISD::MOVHLPS:
3246 case X86ISD::MOVLPS:
3247 case X86ISD::MOVLPD:
3248 case X86ISD::MOVSHDUP:
3249 case X86ISD::MOVSLDUP:
3250 case X86ISD::MOVDDUP:
3253 case X86ISD::UNPCKL:
3254 case X86ISD::UNPCKH:
3255 case X86ISD::VPERMILP:
3256 case X86ISD::VPERM2X128:
3257 case X86ISD::VPERMI:
3262 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3263 SDValue V1, SelectionDAG &DAG) {
3265 default: llvm_unreachable("Unknown x86 shuffle node");
3266 case X86ISD::MOVSHDUP:
3267 case X86ISD::MOVSLDUP:
3268 case X86ISD::MOVDDUP:
3269 return DAG.getNode(Opc, dl, VT, V1);
3273 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3274 SDValue V1, unsigned TargetMask,
3275 SelectionDAG &DAG) {
3277 default: llvm_unreachable("Unknown x86 shuffle node");
3278 case X86ISD::PSHUFD:
3279 case X86ISD::PSHUFHW:
3280 case X86ISD::PSHUFLW:
3281 case X86ISD::VPERMILP:
3282 case X86ISD::VPERMI:
3283 return DAG.getNode(Opc, dl, VT, V1, DAG.getConstant(TargetMask, MVT::i8));
3287 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3288 SDValue V1, SDValue V2, unsigned TargetMask,
3289 SelectionDAG &DAG) {
3291 default: llvm_unreachable("Unknown x86 shuffle node");
3292 case X86ISD::PALIGNR:
3294 case X86ISD::VPERM2X128:
3295 return DAG.getNode(Opc, dl, VT, V1, V2,
3296 DAG.getConstant(TargetMask, MVT::i8));
3300 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3301 SDValue V1, SDValue V2, SelectionDAG &DAG) {
3303 default: llvm_unreachable("Unknown x86 shuffle node");
3304 case X86ISD::MOVLHPS:
3305 case X86ISD::MOVLHPD:
3306 case X86ISD::MOVHLPS:
3307 case X86ISD::MOVLPS:
3308 case X86ISD::MOVLPD:
3311 case X86ISD::UNPCKL:
3312 case X86ISD::UNPCKH:
3313 return DAG.getNode(Opc, dl, VT, V1, V2);
3317 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
3318 MachineFunction &MF = DAG.getMachineFunction();
3319 const X86RegisterInfo *RegInfo =
3320 static_cast<const X86RegisterInfo*>(getTargetMachine().getRegisterInfo());
3321 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
3322 int ReturnAddrIndex = FuncInfo->getRAIndex();
3324 if (ReturnAddrIndex == 0) {
3325 // Set up a frame object for the return address.
3326 unsigned SlotSize = RegInfo->getSlotSize();
3327 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize,
3330 FuncInfo->setRAIndex(ReturnAddrIndex);
3333 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
3336 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
3337 bool hasSymbolicDisplacement) {
3338 // Offset should fit into 32 bit immediate field.
3339 if (!isInt<32>(Offset))
3342 // If we don't have a symbolic displacement - we don't have any extra
3344 if (!hasSymbolicDisplacement)
3347 // FIXME: Some tweaks might be needed for medium code model.
3348 if (M != CodeModel::Small && M != CodeModel::Kernel)
3351 // For small code model we assume that latest object is 16MB before end of 31
3352 // bits boundary. We may also accept pretty large negative constants knowing
3353 // that all objects are in the positive half of address space.
3354 if (M == CodeModel::Small && Offset < 16*1024*1024)
3357 // For kernel code model we know that all object resist in the negative half
3358 // of 32bits address space. We may not accept negative offsets, since they may
3359 // be just off and we may accept pretty large positive ones.
3360 if (M == CodeModel::Kernel && Offset > 0)
3366 /// isCalleePop - Determines whether the callee is required to pop its
3367 /// own arguments. Callee pop is necessary to support tail calls.
3368 bool X86::isCalleePop(CallingConv::ID CallingConv,
3369 bool is64Bit, bool IsVarArg, bool TailCallOpt) {
3373 switch (CallingConv) {
3376 case CallingConv::X86_StdCall:
3378 case CallingConv::X86_FastCall:
3380 case CallingConv::X86_ThisCall:
3382 case CallingConv::Fast:
3384 case CallingConv::GHC:
3386 case CallingConv::HiPE:
3391 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
3392 /// specific condition code, returning the condition code and the LHS/RHS of the
3393 /// comparison to make.
3394 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
3395 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
3397 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
3398 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
3399 // X > -1 -> X == 0, jump !sign.
3400 RHS = DAG.getConstant(0, RHS.getValueType());
3401 return X86::COND_NS;
3403 if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
3404 // X < 0 -> X == 0, jump on sign.
3407 if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
3409 RHS = DAG.getConstant(0, RHS.getValueType());
3410 return X86::COND_LE;
3414 switch (SetCCOpcode) {
3415 default: llvm_unreachable("Invalid integer condition!");
3416 case ISD::SETEQ: return X86::COND_E;
3417 case ISD::SETGT: return X86::COND_G;
3418 case ISD::SETGE: return X86::COND_GE;
3419 case ISD::SETLT: return X86::COND_L;
3420 case ISD::SETLE: return X86::COND_LE;
3421 case ISD::SETNE: return X86::COND_NE;
3422 case ISD::SETULT: return X86::COND_B;
3423 case ISD::SETUGT: return X86::COND_A;
3424 case ISD::SETULE: return X86::COND_BE;
3425 case ISD::SETUGE: return X86::COND_AE;
3429 // First determine if it is required or is profitable to flip the operands.
3431 // If LHS is a foldable load, but RHS is not, flip the condition.
3432 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
3433 !ISD::isNON_EXTLoad(RHS.getNode())) {
3434 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
3435 std::swap(LHS, RHS);
3438 switch (SetCCOpcode) {
3444 std::swap(LHS, RHS);
3448 // On a floating point condition, the flags are set as follows:
3450 // 0 | 0 | 0 | X > Y
3451 // 0 | 0 | 1 | X < Y
3452 // 1 | 0 | 0 | X == Y
3453 // 1 | 1 | 1 | unordered
3454 switch (SetCCOpcode) {
3455 default: llvm_unreachable("Condcode should be pre-legalized away");
3457 case ISD::SETEQ: return X86::COND_E;
3458 case ISD::SETOLT: // flipped
3460 case ISD::SETGT: return X86::COND_A;
3461 case ISD::SETOLE: // flipped
3463 case ISD::SETGE: return X86::COND_AE;
3464 case ISD::SETUGT: // flipped
3466 case ISD::SETLT: return X86::COND_B;
3467 case ISD::SETUGE: // flipped
3469 case ISD::SETLE: return X86::COND_BE;
3471 case ISD::SETNE: return X86::COND_NE;
3472 case ISD::SETUO: return X86::COND_P;
3473 case ISD::SETO: return X86::COND_NP;
3475 case ISD::SETUNE: return X86::COND_INVALID;
3479 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
3480 /// code. Current x86 isa includes the following FP cmov instructions:
3481 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
3482 static bool hasFPCMov(unsigned X86CC) {
3498 /// isFPImmLegal - Returns true if the target can instruction select the
3499 /// specified FP immediate natively. If false, the legalizer will
3500 /// materialize the FP immediate as a load from a constant pool.
3501 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3502 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
3503 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
3509 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
3510 /// the specified range (L, H].
3511 static bool isUndefOrInRange(int Val, int Low, int Hi) {
3512 return (Val < 0) || (Val >= Low && Val < Hi);
3515 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
3516 /// specified value.
3517 static bool isUndefOrEqual(int Val, int CmpVal) {
3518 return (Val < 0 || Val == CmpVal);
3521 /// isSequentialOrUndefInRange - Return true if every element in Mask, beginning
3522 /// from position Pos and ending in Pos+Size, falls within the specified
3523 /// sequential range (L, L+Pos]. or is undef.
3524 static bool isSequentialOrUndefInRange(ArrayRef<int> Mask,
3525 unsigned Pos, unsigned Size, int Low) {
3526 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
3527 if (!isUndefOrEqual(Mask[i], Low))
3532 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
3533 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
3534 /// the second operand.
3535 static bool isPSHUFDMask(ArrayRef<int> Mask, EVT VT) {
3536 if (VT == MVT::v4f32 || VT == MVT::v4i32 )
3537 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
3538 if (VT == MVT::v2f64 || VT == MVT::v2i64)
3539 return (Mask[0] < 2 && Mask[1] < 2);
3543 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
3544 /// is suitable for input to PSHUFHW.
3545 static bool isPSHUFHWMask(ArrayRef<int> Mask, EVT VT, bool HasInt256) {
3546 if (VT != MVT::v8i16 && (!HasInt256 || VT != MVT::v16i16))
3549 // Lower quadword copied in order or undef.
3550 if (!isSequentialOrUndefInRange(Mask, 0, 4, 0))
3553 // Upper quadword shuffled.
3554 for (unsigned i = 4; i != 8; ++i)
3555 if (!isUndefOrInRange(Mask[i], 4, 8))
3558 if (VT == MVT::v16i16) {
3559 // Lower quadword copied in order or undef.
3560 if (!isSequentialOrUndefInRange(Mask, 8, 4, 8))
3563 // Upper quadword shuffled.
3564 for (unsigned i = 12; i != 16; ++i)
3565 if (!isUndefOrInRange(Mask[i], 12, 16))
3572 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
3573 /// is suitable for input to PSHUFLW.
3574 static bool isPSHUFLWMask(ArrayRef<int> Mask, EVT VT, bool HasInt256) {
3575 if (VT != MVT::v8i16 && (!HasInt256 || VT != MVT::v16i16))
3578 // Upper quadword copied in order.
3579 if (!isSequentialOrUndefInRange(Mask, 4, 4, 4))
3582 // Lower quadword shuffled.
3583 for (unsigned i = 0; i != 4; ++i)
3584 if (!isUndefOrInRange(Mask[i], 0, 4))
3587 if (VT == MVT::v16i16) {
3588 // Upper quadword copied in order.
3589 if (!isSequentialOrUndefInRange(Mask, 12, 4, 12))
3592 // Lower quadword shuffled.
3593 for (unsigned i = 8; i != 12; ++i)
3594 if (!isUndefOrInRange(Mask[i], 8, 12))
3601 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
3602 /// is suitable for input to PALIGNR.
3603 static bool isPALIGNRMask(ArrayRef<int> Mask, EVT VT,
3604 const X86Subtarget *Subtarget) {
3605 if ((VT.is128BitVector() && !Subtarget->hasSSSE3()) ||
3606 (VT.is256BitVector() && !Subtarget->hasInt256()))
3609 unsigned NumElts = VT.getVectorNumElements();
3610 unsigned NumLanes = VT.getSizeInBits()/128;
3611 unsigned NumLaneElts = NumElts/NumLanes;
3613 // Do not handle 64-bit element shuffles with palignr.
3614 if (NumLaneElts == 2)
3617 for (unsigned l = 0; l != NumElts; l+=NumLaneElts) {
3619 for (i = 0; i != NumLaneElts; ++i) {
3624 // Lane is all undef, go to next lane
3625 if (i == NumLaneElts)
3628 int Start = Mask[i+l];
3630 // Make sure its in this lane in one of the sources
3631 if (!isUndefOrInRange(Start, l, l+NumLaneElts) &&
3632 !isUndefOrInRange(Start, l+NumElts, l+NumElts+NumLaneElts))
3635 // If not lane 0, then we must match lane 0
3636 if (l != 0 && Mask[i] >= 0 && !isUndefOrEqual(Start, Mask[i]+l))
3639 // Correct second source to be contiguous with first source
3640 if (Start >= (int)NumElts)
3641 Start -= NumElts - NumLaneElts;
3643 // Make sure we're shifting in the right direction.
3644 if (Start <= (int)(i+l))
3649 // Check the rest of the elements to see if they are consecutive.
3650 for (++i; i != NumLaneElts; ++i) {
3651 int Idx = Mask[i+l];
3653 // Make sure its in this lane
3654 if (!isUndefOrInRange(Idx, l, l+NumLaneElts) &&
3655 !isUndefOrInRange(Idx, l+NumElts, l+NumElts+NumLaneElts))
3658 // If not lane 0, then we must match lane 0
3659 if (l != 0 && Mask[i] >= 0 && !isUndefOrEqual(Idx, Mask[i]+l))
3662 if (Idx >= (int)NumElts)
3663 Idx -= NumElts - NumLaneElts;
3665 if (!isUndefOrEqual(Idx, Start+i))
3674 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3675 /// the two vector operands have swapped position.
3676 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask,
3677 unsigned NumElems) {
3678 for (unsigned i = 0; i != NumElems; ++i) {
3682 else if (idx < (int)NumElems)
3683 Mask[i] = idx + NumElems;
3685 Mask[i] = idx - NumElems;
3689 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
3690 /// specifies a shuffle of elements that is suitable for input to 128/256-bit
3691 /// SHUFPS and SHUFPD. If Commuted is true, then it checks for sources to be
3692 /// reverse of what x86 shuffles want.
3693 static bool isSHUFPMask(ArrayRef<int> Mask, EVT VT, bool HasFp256,
3694 bool Commuted = false) {
3695 if (!HasFp256 && VT.is256BitVector())
3698 unsigned NumElems = VT.getVectorNumElements();
3699 unsigned NumLanes = VT.getSizeInBits()/128;
3700 unsigned NumLaneElems = NumElems/NumLanes;
3702 if (NumLaneElems != 2 && NumLaneElems != 4)
3705 // VSHUFPSY divides the resulting vector into 4 chunks.
3706 // The sources are also splitted into 4 chunks, and each destination
3707 // chunk must come from a different source chunk.
3709 // SRC1 => X7 X6 X5 X4 X3 X2 X1 X0
3710 // SRC2 => Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y9
3712 // DST => Y7..Y4, Y7..Y4, X7..X4, X7..X4,
3713 // Y3..Y0, Y3..Y0, X3..X0, X3..X0
3715 // VSHUFPDY divides the resulting vector into 4 chunks.
3716 // The sources are also splitted into 4 chunks, and each destination
3717 // chunk must come from a different source chunk.
3719 // SRC1 => X3 X2 X1 X0
3720 // SRC2 => Y3 Y2 Y1 Y0
3722 // DST => Y3..Y2, X3..X2, Y1..Y0, X1..X0
3724 unsigned HalfLaneElems = NumLaneElems/2;
3725 for (unsigned l = 0; l != NumElems; l += NumLaneElems) {
3726 for (unsigned i = 0; i != NumLaneElems; ++i) {
3727 int Idx = Mask[i+l];
3728 unsigned RngStart = l + ((Commuted == (i<HalfLaneElems)) ? NumElems : 0);
3729 if (!isUndefOrInRange(Idx, RngStart, RngStart+NumLaneElems))
3731 // For VSHUFPSY, the mask of the second half must be the same as the
3732 // first but with the appropriate offsets. This works in the same way as
3733 // VPERMILPS works with masks.
3734 if (NumElems != 8 || l == 0 || Mask[i] < 0)
3736 if (!isUndefOrEqual(Idx, Mask[i]+l))
3744 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
3745 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
3746 static bool isMOVHLPSMask(ArrayRef<int> Mask, EVT VT) {
3747 if (!VT.is128BitVector())
3750 unsigned NumElems = VT.getVectorNumElements();
3755 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
3756 return isUndefOrEqual(Mask[0], 6) &&
3757 isUndefOrEqual(Mask[1], 7) &&
3758 isUndefOrEqual(Mask[2], 2) &&
3759 isUndefOrEqual(Mask[3], 3);
3762 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
3763 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
3765 static bool isMOVHLPS_v_undef_Mask(ArrayRef<int> Mask, EVT VT) {
3766 if (!VT.is128BitVector())
3769 unsigned NumElems = VT.getVectorNumElements();
3774 return isUndefOrEqual(Mask[0], 2) &&
3775 isUndefOrEqual(Mask[1], 3) &&
3776 isUndefOrEqual(Mask[2], 2) &&
3777 isUndefOrEqual(Mask[3], 3);
3780 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
3781 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
3782 static bool isMOVLPMask(ArrayRef<int> Mask, EVT VT) {
3783 if (!VT.is128BitVector())
3786 unsigned NumElems = VT.getVectorNumElements();
3788 if (NumElems != 2 && NumElems != 4)
3791 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3792 if (!isUndefOrEqual(Mask[i], i + NumElems))
3795 for (unsigned i = NumElems/2, e = NumElems; i != e; ++i)
3796 if (!isUndefOrEqual(Mask[i], i))
3802 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
3803 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
3804 static bool isMOVLHPSMask(ArrayRef<int> Mask, EVT VT) {
3805 if (!VT.is128BitVector())
3808 unsigned NumElems = VT.getVectorNumElements();
3810 if (NumElems != 2 && NumElems != 4)
3813 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3814 if (!isUndefOrEqual(Mask[i], i))
3817 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3818 if (!isUndefOrEqual(Mask[i + e], i + NumElems))
3825 // Some special combinations that can be optimized.
3828 SDValue Compact8x32ShuffleNode(ShuffleVectorSDNode *SVOp,
3829 SelectionDAG &DAG) {
3830 MVT VT = SVOp->getValueType(0).getSimpleVT();
3833 if (VT != MVT::v8i32 && VT != MVT::v8f32)
3836 ArrayRef<int> Mask = SVOp->getMask();
3838 // These are the special masks that may be optimized.
3839 static const int MaskToOptimizeEven[] = {0, 8, 2, 10, 4, 12, 6, 14};
3840 static const int MaskToOptimizeOdd[] = {1, 9, 3, 11, 5, 13, 7, 15};
3841 bool MatchEvenMask = true;
3842 bool MatchOddMask = true;
3843 for (int i=0; i<8; ++i) {
3844 if (!isUndefOrEqual(Mask[i], MaskToOptimizeEven[i]))
3845 MatchEvenMask = false;
3846 if (!isUndefOrEqual(Mask[i], MaskToOptimizeOdd[i]))
3847 MatchOddMask = false;
3850 if (!MatchEvenMask && !MatchOddMask)
3853 SDValue UndefNode = DAG.getNode(ISD::UNDEF, dl, VT);
3855 SDValue Op0 = SVOp->getOperand(0);
3856 SDValue Op1 = SVOp->getOperand(1);
3858 if (MatchEvenMask) {
3859 // Shift the second operand right to 32 bits.
3860 static const int ShiftRightMask[] = {-1, 0, -1, 2, -1, 4, -1, 6 };
3861 Op1 = DAG.getVectorShuffle(VT, dl, Op1, UndefNode, ShiftRightMask);
3863 // Shift the first operand left to 32 bits.
3864 static const int ShiftLeftMask[] = {1, -1, 3, -1, 5, -1, 7, -1 };
3865 Op0 = DAG.getVectorShuffle(VT, dl, Op0, UndefNode, ShiftLeftMask);
3867 static const int BlendMask[] = {0, 9, 2, 11, 4, 13, 6, 15};
3868 return DAG.getVectorShuffle(VT, dl, Op0, Op1, BlendMask);
3871 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
3872 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
3873 static bool isUNPCKLMask(ArrayRef<int> Mask, EVT VT,
3874 bool HasInt256, bool V2IsSplat = false) {
3875 unsigned NumElts = VT.getVectorNumElements();
3877 assert((VT.is128BitVector() || VT.is256BitVector()) &&
3878 "Unsupported vector type for unpckh");
3880 if (VT.is256BitVector() && NumElts != 4 && NumElts != 8 &&
3881 (!HasInt256 || (NumElts != 16 && NumElts != 32)))
3884 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
3885 // independently on 128-bit lanes.
3886 unsigned NumLanes = VT.getSizeInBits()/128;
3887 unsigned NumLaneElts = NumElts/NumLanes;
3889 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
3890 for (unsigned i = 0, j = l; i != NumLaneElts; i += 2, ++j) {
3891 int BitI = Mask[l+i];
3892 int BitI1 = Mask[l+i+1];
3893 if (!isUndefOrEqual(BitI, j))
3896 if (!isUndefOrEqual(BitI1, NumElts))
3899 if (!isUndefOrEqual(BitI1, j + NumElts))
3908 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
3909 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
3910 static bool isUNPCKHMask(ArrayRef<int> Mask, EVT VT,
3911 bool HasInt256, bool V2IsSplat = false) {
3912 unsigned NumElts = VT.getVectorNumElements();
3914 assert((VT.is128BitVector() || VT.is256BitVector()) &&
3915 "Unsupported vector type for unpckh");
3917 if (VT.is256BitVector() && NumElts != 4 && NumElts != 8 &&
3918 (!HasInt256 || (NumElts != 16 && NumElts != 32)))
3921 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
3922 // independently on 128-bit lanes.
3923 unsigned NumLanes = VT.getSizeInBits()/128;
3924 unsigned NumLaneElts = NumElts/NumLanes;
3926 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
3927 for (unsigned i = 0, j = l+NumLaneElts/2; i != NumLaneElts; i += 2, ++j) {
3928 int BitI = Mask[l+i];
3929 int BitI1 = Mask[l+i+1];
3930 if (!isUndefOrEqual(BitI, j))
3933 if (isUndefOrEqual(BitI1, NumElts))
3936 if (!isUndefOrEqual(BitI1, j+NumElts))
3944 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
3945 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
3947 static bool isUNPCKL_v_undef_Mask(ArrayRef<int> Mask, EVT VT, bool HasInt256) {
3948 unsigned NumElts = VT.getVectorNumElements();
3949 bool Is256BitVec = VT.is256BitVector();
3951 assert((VT.is128BitVector() || VT.is256BitVector()) &&
3952 "Unsupported vector type for unpckh");
3954 if (Is256BitVec && NumElts != 4 && NumElts != 8 &&
3955 (!HasInt256 || (NumElts != 16 && NumElts != 32)))
3958 // For 256-bit i64/f64, use MOVDDUPY instead, so reject the matching pattern
3959 // FIXME: Need a better way to get rid of this, there's no latency difference
3960 // between UNPCKLPD and MOVDDUP, the later should always be checked first and
3961 // the former later. We should also remove the "_undef" special mask.
3962 if (NumElts == 4 && Is256BitVec)
3965 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
3966 // independently on 128-bit lanes.
3967 unsigned NumLanes = VT.getSizeInBits()/128;
3968 unsigned NumLaneElts = NumElts/NumLanes;
3970 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
3971 for (unsigned i = 0, j = l; i != NumLaneElts; i += 2, ++j) {
3972 int BitI = Mask[l+i];
3973 int BitI1 = Mask[l+i+1];
3975 if (!isUndefOrEqual(BitI, j))
3977 if (!isUndefOrEqual(BitI1, j))
3985 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
3986 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
3988 static bool isUNPCKH_v_undef_Mask(ArrayRef<int> Mask, EVT VT, bool HasInt256) {
3989 unsigned NumElts = VT.getVectorNumElements();
3991 assert((VT.is128BitVector() || VT.is256BitVector()) &&
3992 "Unsupported vector type for unpckh");
3994 if (VT.is256BitVector() && NumElts != 4 && NumElts != 8 &&
3995 (!HasInt256 || (NumElts != 16 && NumElts != 32)))
3998 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
3999 // independently on 128-bit lanes.
4000 unsigned NumLanes = VT.getSizeInBits()/128;
4001 unsigned NumLaneElts = NumElts/NumLanes;
4003 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
4004 for (unsigned i = 0, j = l+NumLaneElts/2; i != NumLaneElts; i += 2, ++j) {
4005 int BitI = Mask[l+i];
4006 int BitI1 = Mask[l+i+1];
4007 if (!isUndefOrEqual(BitI, j))
4009 if (!isUndefOrEqual(BitI1, j))
4016 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
4017 /// specifies a shuffle of elements that is suitable for input to MOVSS,
4018 /// MOVSD, and MOVD, i.e. setting the lowest element.
4019 static bool isMOVLMask(ArrayRef<int> Mask, EVT VT) {
4020 if (VT.getVectorElementType().getSizeInBits() < 32)
4022 if (!VT.is128BitVector())
4025 unsigned NumElts = VT.getVectorNumElements();
4027 if (!isUndefOrEqual(Mask[0], NumElts))
4030 for (unsigned i = 1; i != NumElts; ++i)
4031 if (!isUndefOrEqual(Mask[i], i))
4037 /// isVPERM2X128Mask - Match 256-bit shuffles where the elements are considered
4038 /// as permutations between 128-bit chunks or halves. As an example: this
4040 /// vector_shuffle <4, 5, 6, 7, 12, 13, 14, 15>
4041 /// The first half comes from the second half of V1 and the second half from the
4042 /// the second half of V2.
4043 static bool isVPERM2X128Mask(ArrayRef<int> Mask, EVT VT, bool HasFp256) {
4044 if (!HasFp256 || !VT.is256BitVector())
4047 // The shuffle result is divided into half A and half B. In total the two
4048 // sources have 4 halves, namely: C, D, E, F. The final values of A and
4049 // B must come from C, D, E or F.
4050 unsigned HalfSize = VT.getVectorNumElements()/2;
4051 bool MatchA = false, MatchB = false;
4053 // Check if A comes from one of C, D, E, F.
4054 for (unsigned Half = 0; Half != 4; ++Half) {
4055 if (isSequentialOrUndefInRange(Mask, 0, HalfSize, Half*HalfSize)) {
4061 // Check if B comes from one of C, D, E, F.
4062 for (unsigned Half = 0; Half != 4; ++Half) {
4063 if (isSequentialOrUndefInRange(Mask, HalfSize, HalfSize, Half*HalfSize)) {
4069 return MatchA && MatchB;
4072 /// getShuffleVPERM2X128Immediate - Return the appropriate immediate to shuffle
4073 /// the specified VECTOR_MASK mask with VPERM2F128/VPERM2I128 instructions.
4074 static unsigned getShuffleVPERM2X128Immediate(ShuffleVectorSDNode *SVOp) {
4075 MVT VT = SVOp->getValueType(0).getSimpleVT();
4077 unsigned HalfSize = VT.getVectorNumElements()/2;
4079 unsigned FstHalf = 0, SndHalf = 0;
4080 for (unsigned i = 0; i < HalfSize; ++i) {
4081 if (SVOp->getMaskElt(i) > 0) {
4082 FstHalf = SVOp->getMaskElt(i)/HalfSize;
4086 for (unsigned i = HalfSize; i < HalfSize*2; ++i) {
4087 if (SVOp->getMaskElt(i) > 0) {
4088 SndHalf = SVOp->getMaskElt(i)/HalfSize;
4093 return (FstHalf | (SndHalf << 4));
4096 /// isVPERMILPMask - Return true if the specified VECTOR_SHUFFLE operand
4097 /// specifies a shuffle of elements that is suitable for input to VPERMILPD*.
4098 /// Note that VPERMIL mask matching is different depending whether theunderlying
4099 /// type is 32 or 64. In the VPERMILPS the high half of the mask should point
4100 /// to the same elements of the low, but to the higher half of the source.
4101 /// In VPERMILPD the two lanes could be shuffled independently of each other
4102 /// with the same restriction that lanes can't be crossed. Also handles PSHUFDY.
4103 static bool isVPERMILPMask(ArrayRef<int> Mask, EVT VT, bool HasFp256) {
4107 unsigned NumElts = VT.getVectorNumElements();
4108 // Only match 256-bit with 32/64-bit types
4109 if (!VT.is256BitVector() || (NumElts != 4 && NumElts != 8))
4112 unsigned NumLanes = VT.getSizeInBits()/128;
4113 unsigned LaneSize = NumElts/NumLanes;
4114 for (unsigned l = 0; l != NumElts; l += LaneSize) {
4115 for (unsigned i = 0; i != LaneSize; ++i) {
4116 if (!isUndefOrInRange(Mask[i+l], l, l+LaneSize))
4118 if (NumElts != 8 || l == 0)
4120 // VPERMILPS handling
4123 if (!isUndefOrEqual(Mask[i+l], Mask[i]+l))
4131 /// isCommutedMOVLMask - Returns true if the shuffle mask is except the reverse
4132 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
4133 /// element of vector 2 and the other elements to come from vector 1 in order.
4134 static bool isCommutedMOVLMask(ArrayRef<int> Mask, EVT VT,
4135 bool V2IsSplat = false, bool V2IsUndef = false) {
4136 if (!VT.is128BitVector())
4139 unsigned NumOps = VT.getVectorNumElements();
4140 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
4143 if (!isUndefOrEqual(Mask[0], 0))
4146 for (unsigned i = 1; i != NumOps; ++i)
4147 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
4148 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
4149 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
4155 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
4156 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
4157 /// Masks to match: <1, 1, 3, 3> or <1, 1, 3, 3, 5, 5, 7, 7>
4158 static bool isMOVSHDUPMask(ArrayRef<int> Mask, EVT VT,
4159 const X86Subtarget *Subtarget) {
4160 if (!Subtarget->hasSSE3())
4163 unsigned NumElems = VT.getVectorNumElements();
4165 if ((VT.is128BitVector() && NumElems != 4) ||
4166 (VT.is256BitVector() && NumElems != 8))
4169 // "i+1" is the value the indexed mask element must have
4170 for (unsigned i = 0; i != NumElems; i += 2)
4171 if (!isUndefOrEqual(Mask[i], i+1) ||
4172 !isUndefOrEqual(Mask[i+1], i+1))
4178 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
4179 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
4180 /// Masks to match: <0, 0, 2, 2> or <0, 0, 2, 2, 4, 4, 6, 6>
4181 static bool isMOVSLDUPMask(ArrayRef<int> Mask, EVT VT,
4182 const X86Subtarget *Subtarget) {
4183 if (!Subtarget->hasSSE3())
4186 unsigned NumElems = VT.getVectorNumElements();
4188 if ((VT.is128BitVector() && NumElems != 4) ||
4189 (VT.is256BitVector() && NumElems != 8))
4192 // "i" is the value the indexed mask element must have
4193 for (unsigned i = 0; i != NumElems; i += 2)
4194 if (!isUndefOrEqual(Mask[i], i) ||
4195 !isUndefOrEqual(Mask[i+1], i))
4201 /// isMOVDDUPYMask - Return true if the specified VECTOR_SHUFFLE operand
4202 /// specifies a shuffle of elements that is suitable for input to 256-bit
4203 /// version of MOVDDUP.
4204 static bool isMOVDDUPYMask(ArrayRef<int> Mask, EVT VT, bool HasFp256) {
4205 if (!HasFp256 || !VT.is256BitVector())
4208 unsigned NumElts = VT.getVectorNumElements();
4212 for (unsigned i = 0; i != NumElts/2; ++i)
4213 if (!isUndefOrEqual(Mask[i], 0))
4215 for (unsigned i = NumElts/2; i != NumElts; ++i)
4216 if (!isUndefOrEqual(Mask[i], NumElts/2))
4221 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
4222 /// specifies a shuffle of elements that is suitable for input to 128-bit
4223 /// version of MOVDDUP.
4224 static bool isMOVDDUPMask(ArrayRef<int> Mask, EVT VT) {
4225 if (!VT.is128BitVector())
4228 unsigned e = VT.getVectorNumElements() / 2;
4229 for (unsigned i = 0; i != e; ++i)
4230 if (!isUndefOrEqual(Mask[i], i))
4232 for (unsigned i = 0; i != e; ++i)
4233 if (!isUndefOrEqual(Mask[e+i], i))
4238 /// isVEXTRACTIndex - Return true if the specified
4239 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
4240 /// suitable for instruction that extract 128 or 256 bit vectors
4241 static bool isVEXTRACTIndex(SDNode *N, unsigned vecWidth) {
4242 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4243 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4246 // The index should be aligned on a vecWidth-bit boundary.
4248 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4250 MVT VT = N->getValueType(0).getSimpleVT();
4251 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4252 bool Result = (Index * ElSize) % vecWidth == 0;
4257 /// isVINSERTIndex - Return true if the specified INSERT_SUBVECTOR
4258 /// operand specifies a subvector insert that is suitable for input to
4259 /// insertion of 128 or 256-bit subvectors
4260 static bool isVINSERTIndex(SDNode *N, unsigned vecWidth) {
4261 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4262 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4264 // The index should be aligned on a vecWidth-bit boundary.
4266 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4268 MVT VT = N->getValueType(0).getSimpleVT();
4269 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4270 bool Result = (Index * ElSize) % vecWidth == 0;
4275 bool X86::isVINSERT128Index(SDNode *N) {
4276 return isVINSERTIndex(N, 128);
4279 bool X86::isVINSERT256Index(SDNode *N) {
4280 return isVINSERTIndex(N, 256);
4283 bool X86::isVEXTRACT128Index(SDNode *N) {
4284 return isVEXTRACTIndex(N, 128);
4287 bool X86::isVEXTRACT256Index(SDNode *N) {
4288 return isVEXTRACTIndex(N, 256);
4291 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
4292 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
4293 /// Handles 128-bit and 256-bit.
4294 static unsigned getShuffleSHUFImmediate(ShuffleVectorSDNode *N) {
4295 MVT VT = N->getValueType(0).getSimpleVT();
4297 assert((VT.is128BitVector() || VT.is256BitVector()) &&
4298 "Unsupported vector type for PSHUF/SHUFP");
4300 // Handle 128 and 256-bit vector lengths. AVX defines PSHUF/SHUFP to operate
4301 // independently on 128-bit lanes.
4302 unsigned NumElts = VT.getVectorNumElements();
4303 unsigned NumLanes = VT.getSizeInBits()/128;
4304 unsigned NumLaneElts = NumElts/NumLanes;
4306 assert((NumLaneElts == 2 || NumLaneElts == 4) &&
4307 "Only supports 2 or 4 elements per lane");
4309 unsigned Shift = (NumLaneElts == 4) ? 1 : 0;
4311 for (unsigned i = 0; i != NumElts; ++i) {
4312 int Elt = N->getMaskElt(i);
4313 if (Elt < 0) continue;
4314 Elt &= NumLaneElts - 1;
4315 unsigned ShAmt = (i << Shift) % 8;
4316 Mask |= Elt << ShAmt;
4322 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
4323 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
4324 static unsigned getShufflePSHUFHWImmediate(ShuffleVectorSDNode *N) {
4325 MVT VT = N->getValueType(0).getSimpleVT();
4327 assert((VT == MVT::v8i16 || VT == MVT::v16i16) &&
4328 "Unsupported vector type for PSHUFHW");
4330 unsigned NumElts = VT.getVectorNumElements();
4333 for (unsigned l = 0; l != NumElts; l += 8) {
4334 // 8 nodes per lane, but we only care about the last 4.
4335 for (unsigned i = 0; i < 4; ++i) {
4336 int Elt = N->getMaskElt(l+i+4);
4337 if (Elt < 0) continue;
4338 Elt &= 0x3; // only 2-bits.
4339 Mask |= Elt << (i * 2);
4346 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
4347 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
4348 static unsigned getShufflePSHUFLWImmediate(ShuffleVectorSDNode *N) {
4349 MVT VT = N->getValueType(0).getSimpleVT();
4351 assert((VT == MVT::v8i16 || VT == MVT::v16i16) &&
4352 "Unsupported vector type for PSHUFHW");
4354 unsigned NumElts = VT.getVectorNumElements();
4357 for (unsigned l = 0; l != NumElts; l += 8) {
4358 // 8 nodes per lane, but we only care about the first 4.
4359 for (unsigned i = 0; i < 4; ++i) {
4360 int Elt = N->getMaskElt(l+i);
4361 if (Elt < 0) continue;
4362 Elt &= 0x3; // only 2-bits
4363 Mask |= Elt << (i * 2);
4370 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
4371 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
4372 static unsigned getShufflePALIGNRImmediate(ShuffleVectorSDNode *SVOp) {
4373 MVT VT = SVOp->getValueType(0).getSimpleVT();
4374 unsigned EltSize = VT.getVectorElementType().getSizeInBits() >> 3;
4376 unsigned NumElts = VT.getVectorNumElements();
4377 unsigned NumLanes = VT.getSizeInBits()/128;
4378 unsigned NumLaneElts = NumElts/NumLanes;
4382 for (i = 0; i != NumElts; ++i) {
4383 Val = SVOp->getMaskElt(i);
4387 if (Val >= (int)NumElts)
4388 Val -= NumElts - NumLaneElts;
4390 assert(Val - i > 0 && "PALIGNR imm should be positive");
4391 return (Val - i) * EltSize;
4394 static unsigned getExtractVEXTRACTImmediate(SDNode *N, unsigned vecWidth) {
4395 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4396 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4397 llvm_unreachable("Illegal extract subvector for VEXTRACT");
4400 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4402 MVT VecVT = N->getOperand(0).getValueType().getSimpleVT();
4403 MVT ElVT = VecVT.getVectorElementType();
4405 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4406 return Index / NumElemsPerChunk;
4409 static unsigned getInsertVINSERTImmediate(SDNode *N, unsigned vecWidth) {
4410 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4411 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4412 llvm_unreachable("Illegal insert subvector for VINSERT");
4415 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4417 MVT VecVT = N->getValueType(0).getSimpleVT();
4418 MVT ElVT = VecVT.getVectorElementType();
4420 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4421 return Index / NumElemsPerChunk;
4424 /// getExtractVEXTRACT128Immediate - Return the appropriate immediate
4425 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF128
4426 /// and VINSERTI128 instructions.
4427 unsigned X86::getExtractVEXTRACT128Immediate(SDNode *N) {
4428 return getExtractVEXTRACTImmediate(N, 128);
4431 /// getExtractVEXTRACT256Immediate - Return the appropriate immediate
4432 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF64x4
4433 /// and VINSERTI64x4 instructions.
4434 unsigned X86::getExtractVEXTRACT256Immediate(SDNode *N) {
4435 return getExtractVEXTRACTImmediate(N, 256);
4438 /// getInsertVINSERT128Immediate - Return the appropriate immediate
4439 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF128
4440 /// and VINSERTI128 instructions.
4441 unsigned X86::getInsertVINSERT128Immediate(SDNode *N) {
4442 return getInsertVINSERTImmediate(N, 128);
4445 /// getInsertVINSERT256Immediate - Return the appropriate immediate
4446 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF46x4
4447 /// and VINSERTI64x4 instructions.
4448 unsigned X86::getInsertVINSERT256Immediate(SDNode *N) {
4449 return getInsertVINSERTImmediate(N, 256);
4452 /// getShuffleCLImmediate - Return the appropriate immediate to shuffle
4453 /// the specified VECTOR_SHUFFLE mask with VPERMQ and VPERMPD instructions.
4454 /// Handles 256-bit.
4455 static unsigned getShuffleCLImmediate(ShuffleVectorSDNode *N) {
4456 MVT VT = N->getValueType(0).getSimpleVT();
4458 unsigned NumElts = VT.getVectorNumElements();
4460 assert((VT.is256BitVector() && NumElts == 4) &&
4461 "Unsupported vector type for VPERMQ/VPERMPD");
4464 for (unsigned i = 0; i != NumElts; ++i) {
4465 int Elt = N->getMaskElt(i);
4468 Mask |= Elt << (i*2);
4473 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
4475 bool X86::isZeroNode(SDValue Elt) {
4476 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Elt))
4477 return CN->isNullValue();
4478 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Elt))
4479 return CFP->getValueAPF().isPosZero();
4483 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
4484 /// their permute mask.
4485 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
4486 SelectionDAG &DAG) {
4487 MVT VT = SVOp->getValueType(0).getSimpleVT();
4488 unsigned NumElems = VT.getVectorNumElements();
4489 SmallVector<int, 8> MaskVec;
4491 for (unsigned i = 0; i != NumElems; ++i) {
4492 int Idx = SVOp->getMaskElt(i);
4494 if (Idx < (int)NumElems)
4499 MaskVec.push_back(Idx);
4501 return DAG.getVectorShuffle(VT, SDLoc(SVOp), SVOp->getOperand(1),
4502 SVOp->getOperand(0), &MaskVec[0]);
4505 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
4506 /// match movhlps. The lower half elements should come from upper half of
4507 /// V1 (and in order), and the upper half elements should come from the upper
4508 /// half of V2 (and in order).
4509 static bool ShouldXformToMOVHLPS(ArrayRef<int> Mask, EVT VT) {
4510 if (!VT.is128BitVector())
4512 if (VT.getVectorNumElements() != 4)
4514 for (unsigned i = 0, e = 2; i != e; ++i)
4515 if (!isUndefOrEqual(Mask[i], i+2))
4517 for (unsigned i = 2; i != 4; ++i)
4518 if (!isUndefOrEqual(Mask[i], i+4))
4523 /// isScalarLoadToVector - Returns true if the node is a scalar load that
4524 /// is promoted to a vector. It also returns the LoadSDNode by reference if
4526 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
4527 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
4529 N = N->getOperand(0).getNode();
4530 if (!ISD::isNON_EXTLoad(N))
4533 *LD = cast<LoadSDNode>(N);
4537 // Test whether the given value is a vector value which will be legalized
4539 static bool WillBeConstantPoolLoad(SDNode *N) {
4540 if (N->getOpcode() != ISD::BUILD_VECTOR)
4543 // Check for any non-constant elements.
4544 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
4545 switch (N->getOperand(i).getNode()->getOpcode()) {
4547 case ISD::ConstantFP:
4554 // Vectors of all-zeros and all-ones are materialized with special
4555 // instructions rather than being loaded.
4556 return !ISD::isBuildVectorAllZeros(N) &&
4557 !ISD::isBuildVectorAllOnes(N);
4560 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
4561 /// match movlp{s|d}. The lower half elements should come from lower half of
4562 /// V1 (and in order), and the upper half elements should come from the upper
4563 /// half of V2 (and in order). And since V1 will become the source of the
4564 /// MOVLP, it must be either a vector load or a scalar load to vector.
4565 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
4566 ArrayRef<int> Mask, EVT VT) {
4567 if (!VT.is128BitVector())
4570 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
4572 // Is V2 is a vector load, don't do this transformation. We will try to use
4573 // load folding shufps op.
4574 if (ISD::isNON_EXTLoad(V2) || WillBeConstantPoolLoad(V2))
4577 unsigned NumElems = VT.getVectorNumElements();
4579 if (NumElems != 2 && NumElems != 4)
4581 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
4582 if (!isUndefOrEqual(Mask[i], i))
4584 for (unsigned i = NumElems/2, e = NumElems; i != e; ++i)
4585 if (!isUndefOrEqual(Mask[i], i+NumElems))
4590 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
4592 static bool isSplatVector(SDNode *N) {
4593 if (N->getOpcode() != ISD::BUILD_VECTOR)
4596 SDValue SplatValue = N->getOperand(0);
4597 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
4598 if (N->getOperand(i) != SplatValue)
4603 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
4604 /// to an zero vector.
4605 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
4606 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
4607 SDValue V1 = N->getOperand(0);
4608 SDValue V2 = N->getOperand(1);
4609 unsigned NumElems = N->getValueType(0).getVectorNumElements();
4610 for (unsigned i = 0; i != NumElems; ++i) {
4611 int Idx = N->getMaskElt(i);
4612 if (Idx >= (int)NumElems) {
4613 unsigned Opc = V2.getOpcode();
4614 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
4616 if (Opc != ISD::BUILD_VECTOR ||
4617 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
4619 } else if (Idx >= 0) {
4620 unsigned Opc = V1.getOpcode();
4621 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
4623 if (Opc != ISD::BUILD_VECTOR ||
4624 !X86::isZeroNode(V1.getOperand(Idx)))
4631 /// getZeroVector - Returns a vector of specified type with all zero elements.
4633 static SDValue getZeroVector(EVT VT, const X86Subtarget *Subtarget,
4634 SelectionDAG &DAG, SDLoc dl) {
4635 assert(VT.isVector() && "Expected a vector type");
4637 // Always build SSE zero vectors as <4 x i32> bitcasted
4638 // to their dest type. This ensures they get CSE'd.
4640 if (VT.is128BitVector()) { // SSE
4641 if (Subtarget->hasSSE2()) { // SSE2
4642 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
4643 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4645 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
4646 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
4648 } else if (VT.is256BitVector()) { // AVX
4649 if (Subtarget->hasInt256()) { // AVX2
4650 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
4651 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4652 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops,
4653 array_lengthof(Ops));
4655 // 256-bit logic and arithmetic instructions in AVX are all
4656 // floating-point, no support for integer ops. Emit fp zeroed vectors.
4657 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
4658 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4659 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops,
4660 array_lengthof(Ops));
4663 llvm_unreachable("Unexpected vector type");
4665 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
4668 /// getOnesVector - Returns a vector of specified type with all bits set.
4669 /// Always build ones vectors as <4 x i32> or <8 x i32>. For 256-bit types with
4670 /// no AVX2 supprt, use two <4 x i32> inserted in a <8 x i32> appropriately.
4671 /// Then bitcast to their original type, ensuring they get CSE'd.
4672 static SDValue getOnesVector(MVT VT, bool HasInt256, SelectionDAG &DAG,
4674 assert(VT.isVector() && "Expected a vector type");
4676 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
4678 if (VT.is256BitVector()) {
4679 if (HasInt256) { // AVX2
4680 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4681 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops,
4682 array_lengthof(Ops));
4684 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4685 Vec = Concat128BitVectors(Vec, Vec, MVT::v8i32, 8, DAG, dl);
4687 } else if (VT.is128BitVector()) {
4688 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4690 llvm_unreachable("Unexpected vector type");
4692 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
4695 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
4696 /// that point to V2 points to its first element.
4697 static void NormalizeMask(SmallVectorImpl<int> &Mask, unsigned NumElems) {
4698 for (unsigned i = 0; i != NumElems; ++i) {
4699 if (Mask[i] > (int)NumElems) {
4705 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
4706 /// operation of specified width.
4707 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1,
4709 unsigned NumElems = VT.getVectorNumElements();
4710 SmallVector<int, 8> Mask;
4711 Mask.push_back(NumElems);
4712 for (unsigned i = 1; i != NumElems; ++i)
4714 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4717 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
4718 static SDValue getUnpackl(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1,
4720 unsigned NumElems = VT.getVectorNumElements();
4721 SmallVector<int, 8> Mask;
4722 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
4724 Mask.push_back(i + NumElems);
4726 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4729 /// getUnpackh - Returns a vector_shuffle node for an unpackh operation.
4730 static SDValue getUnpackh(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1,
4732 unsigned NumElems = VT.getVectorNumElements();
4733 SmallVector<int, 8> Mask;
4734 for (unsigned i = 0, Half = NumElems/2; i != Half; ++i) {
4735 Mask.push_back(i + Half);
4736 Mask.push_back(i + NumElems + Half);
4738 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4741 // PromoteSplati8i16 - All i16 and i8 vector types can't be used directly by
4742 // a generic shuffle instruction because the target has no such instructions.
4743 // Generate shuffles which repeat i16 and i8 several times until they can be
4744 // represented by v4f32 and then be manipulated by target suported shuffles.
4745 static SDValue PromoteSplati8i16(SDValue V, SelectionDAG &DAG, int &EltNo) {
4746 EVT VT = V.getValueType();
4747 int NumElems = VT.getVectorNumElements();
4750 while (NumElems > 4) {
4751 if (EltNo < NumElems/2) {
4752 V = getUnpackl(DAG, dl, VT, V, V);
4754 V = getUnpackh(DAG, dl, VT, V, V);
4755 EltNo -= NumElems/2;
4762 /// getLegalSplat - Generate a legal splat with supported x86 shuffles
4763 static SDValue getLegalSplat(SelectionDAG &DAG, SDValue V, int EltNo) {
4764 EVT VT = V.getValueType();
4767 if (VT.is128BitVector()) {
4768 V = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, V);
4769 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
4770 V = DAG.getVectorShuffle(MVT::v4f32, dl, V, DAG.getUNDEF(MVT::v4f32),
4772 } else if (VT.is256BitVector()) {
4773 // To use VPERMILPS to splat scalars, the second half of indicies must
4774 // refer to the higher part, which is a duplication of the lower one,
4775 // because VPERMILPS can only handle in-lane permutations.
4776 int SplatMask[8] = { EltNo, EltNo, EltNo, EltNo,
4777 EltNo+4, EltNo+4, EltNo+4, EltNo+4 };
4779 V = DAG.getNode(ISD::BITCAST, dl, MVT::v8f32, V);
4780 V = DAG.getVectorShuffle(MVT::v8f32, dl, V, DAG.getUNDEF(MVT::v8f32),
4783 llvm_unreachable("Vector size not supported");
4785 return DAG.getNode(ISD::BITCAST, dl, VT, V);
4788 /// PromoteSplat - Splat is promoted to target supported vector shuffles.
4789 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG) {
4790 EVT SrcVT = SV->getValueType(0);
4791 SDValue V1 = SV->getOperand(0);
4794 int EltNo = SV->getSplatIndex();
4795 int NumElems = SrcVT.getVectorNumElements();
4796 bool Is256BitVec = SrcVT.is256BitVector();
4798 assert(((SrcVT.is128BitVector() && NumElems > 4) || Is256BitVec) &&
4799 "Unknown how to promote splat for type");
4801 // Extract the 128-bit part containing the splat element and update
4802 // the splat element index when it refers to the higher register.
4804 V1 = Extract128BitVector(V1, EltNo, DAG, dl);
4805 if (EltNo >= NumElems/2)
4806 EltNo -= NumElems/2;
4809 // All i16 and i8 vector types can't be used directly by a generic shuffle
4810 // instruction because the target has no such instruction. Generate shuffles
4811 // which repeat i16 and i8 several times until they fit in i32, and then can
4812 // be manipulated by target suported shuffles.
4813 EVT EltVT = SrcVT.getVectorElementType();
4814 if (EltVT == MVT::i8 || EltVT == MVT::i16)
4815 V1 = PromoteSplati8i16(V1, DAG, EltNo);
4817 // Recreate the 256-bit vector and place the same 128-bit vector
4818 // into the low and high part. This is necessary because we want
4819 // to use VPERM* to shuffle the vectors
4821 V1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, SrcVT, V1, V1);
4824 return getLegalSplat(DAG, V1, EltNo);
4827 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
4828 /// vector of zero or undef vector. This produces a shuffle where the low
4829 /// element of V2 is swizzled into the zero/undef vector, landing at element
4830 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
4831 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
4833 const X86Subtarget *Subtarget,
4834 SelectionDAG &DAG) {
4835 EVT VT = V2.getValueType();
4837 ? getZeroVector(VT, Subtarget, DAG, SDLoc(V2)) : DAG.getUNDEF(VT);
4838 unsigned NumElems = VT.getVectorNumElements();
4839 SmallVector<int, 16> MaskVec;
4840 for (unsigned i = 0; i != NumElems; ++i)
4841 // If this is the insertion idx, put the low elt of V2 here.
4842 MaskVec.push_back(i == Idx ? NumElems : i);
4843 return DAG.getVectorShuffle(VT, SDLoc(V2), V1, V2, &MaskVec[0]);
4846 /// getTargetShuffleMask - Calculates the shuffle mask corresponding to the
4847 /// target specific opcode. Returns true if the Mask could be calculated.
4848 /// Sets IsUnary to true if only uses one source.
4849 static bool getTargetShuffleMask(SDNode *N, MVT VT,
4850 SmallVectorImpl<int> &Mask, bool &IsUnary) {
4851 unsigned NumElems = VT.getVectorNumElements();
4855 switch(N->getOpcode()) {
4857 ImmN = N->getOperand(N->getNumOperands()-1);
4858 DecodeSHUFPMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4860 case X86ISD::UNPCKH:
4861 DecodeUNPCKHMask(VT, Mask);
4863 case X86ISD::UNPCKL:
4864 DecodeUNPCKLMask(VT, Mask);
4866 case X86ISD::MOVHLPS:
4867 DecodeMOVHLPSMask(NumElems, Mask);
4869 case X86ISD::MOVLHPS:
4870 DecodeMOVLHPSMask(NumElems, Mask);
4872 case X86ISD::PALIGNR:
4873 ImmN = N->getOperand(N->getNumOperands()-1);
4874 DecodePALIGNRMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4876 case X86ISD::PSHUFD:
4877 case X86ISD::VPERMILP:
4878 ImmN = N->getOperand(N->getNumOperands()-1);
4879 DecodePSHUFMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4882 case X86ISD::PSHUFHW:
4883 ImmN = N->getOperand(N->getNumOperands()-1);
4884 DecodePSHUFHWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4887 case X86ISD::PSHUFLW:
4888 ImmN = N->getOperand(N->getNumOperands()-1);
4889 DecodePSHUFLWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4892 case X86ISD::VPERMI:
4893 ImmN = N->getOperand(N->getNumOperands()-1);
4894 DecodeVPERMMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4898 case X86ISD::MOVSD: {
4899 // The index 0 always comes from the first element of the second source,
4900 // this is why MOVSS and MOVSD are used in the first place. The other
4901 // elements come from the other positions of the first source vector
4902 Mask.push_back(NumElems);
4903 for (unsigned i = 1; i != NumElems; ++i) {
4908 case X86ISD::VPERM2X128:
4909 ImmN = N->getOperand(N->getNumOperands()-1);
4910 DecodeVPERM2X128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4911 if (Mask.empty()) return false;
4913 case X86ISD::MOVDDUP:
4914 case X86ISD::MOVLHPD:
4915 case X86ISD::MOVLPD:
4916 case X86ISD::MOVLPS:
4917 case X86ISD::MOVSHDUP:
4918 case X86ISD::MOVSLDUP:
4919 // Not yet implemented
4921 default: llvm_unreachable("unknown target shuffle node");
4927 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
4928 /// element of the result of the vector shuffle.
4929 static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG,
4932 return SDValue(); // Limit search depth.
4934 SDValue V = SDValue(N, 0);
4935 EVT VT = V.getValueType();
4936 unsigned Opcode = V.getOpcode();
4938 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
4939 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
4940 int Elt = SV->getMaskElt(Index);
4943 return DAG.getUNDEF(VT.getVectorElementType());
4945 unsigned NumElems = VT.getVectorNumElements();
4946 SDValue NewV = (Elt < (int)NumElems) ? SV->getOperand(0)
4947 : SV->getOperand(1);
4948 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG, Depth+1);
4951 // Recurse into target specific vector shuffles to find scalars.
4952 if (isTargetShuffle(Opcode)) {
4953 MVT ShufVT = V.getValueType().getSimpleVT();
4954 unsigned NumElems = ShufVT.getVectorNumElements();
4955 SmallVector<int, 16> ShuffleMask;
4958 if (!getTargetShuffleMask(N, ShufVT, ShuffleMask, IsUnary))
4961 int Elt = ShuffleMask[Index];
4963 return DAG.getUNDEF(ShufVT.getVectorElementType());
4965 SDValue NewV = (Elt < (int)NumElems) ? N->getOperand(0)
4967 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG,
4971 // Actual nodes that may contain scalar elements
4972 if (Opcode == ISD::BITCAST) {
4973 V = V.getOperand(0);
4974 EVT SrcVT = V.getValueType();
4975 unsigned NumElems = VT.getVectorNumElements();
4977 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
4981 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
4982 return (Index == 0) ? V.getOperand(0)
4983 : DAG.getUNDEF(VT.getVectorElementType());
4985 if (V.getOpcode() == ISD::BUILD_VECTOR)
4986 return V.getOperand(Index);
4991 /// getNumOfConsecutiveZeros - Return the number of elements of a vector
4992 /// shuffle operation which come from a consecutively from a zero. The
4993 /// search can start in two different directions, from left or right.
4994 /// We count undefs as zeros until PreferredNum is reached.
4995 static unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp,
4996 unsigned NumElems, bool ZerosFromLeft,
4998 unsigned PreferredNum = -1U) {
4999 unsigned NumZeros = 0;
5000 for (unsigned i = 0; i != NumElems; ++i) {
5001 unsigned Index = ZerosFromLeft ? i : NumElems - i - 1;
5002 SDValue Elt = getShuffleScalarElt(SVOp, Index, DAG, 0);
5006 if (X86::isZeroNode(Elt))
5008 else if (Elt.getOpcode() == ISD::UNDEF) // Undef as zero up to PreferredNum.
5009 NumZeros = std::min(NumZeros + 1, PreferredNum);
5017 /// isShuffleMaskConsecutive - Check if the shuffle mask indicies [MaskI, MaskE)
5018 /// correspond consecutively to elements from one of the vector operands,
5019 /// starting from its index OpIdx. Also tell OpNum which source vector operand.
5021 bool isShuffleMaskConsecutive(ShuffleVectorSDNode *SVOp,
5022 unsigned MaskI, unsigned MaskE, unsigned OpIdx,
5023 unsigned NumElems, unsigned &OpNum) {
5024 bool SeenV1 = false;
5025 bool SeenV2 = false;
5027 for (unsigned i = MaskI; i != MaskE; ++i, ++OpIdx) {
5028 int Idx = SVOp->getMaskElt(i);
5029 // Ignore undef indicies
5033 if (Idx < (int)NumElems)
5038 // Only accept consecutive elements from the same vector
5039 if ((Idx % NumElems != OpIdx) || (SeenV1 && SeenV2))
5043 OpNum = SeenV1 ? 0 : 1;
5047 /// isVectorShiftRight - Returns true if the shuffle can be implemented as a
5048 /// logical left shift of a vector.
5049 static bool isVectorShiftRight(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
5050 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
5051 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
5052 unsigned NumZeros = getNumOfConsecutiveZeros(
5053 SVOp, NumElems, false /* check zeros from right */, DAG,
5054 SVOp->getMaskElt(0));
5060 // Considering the elements in the mask that are not consecutive zeros,
5061 // check if they consecutively come from only one of the source vectors.
5063 // V1 = {X, A, B, C} 0
5065 // vector_shuffle V1, V2 <1, 2, 3, X>
5067 if (!isShuffleMaskConsecutive(SVOp,
5068 0, // Mask Start Index
5069 NumElems-NumZeros, // Mask End Index(exclusive)
5070 NumZeros, // Where to start looking in the src vector
5071 NumElems, // Number of elements in vector
5072 OpSrc)) // Which source operand ?
5077 ShVal = SVOp->getOperand(OpSrc);
5081 /// isVectorShiftLeft - Returns true if the shuffle can be implemented as a
5082 /// logical left shift of a vector.
5083 static bool isVectorShiftLeft(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
5084 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
5085 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
5086 unsigned NumZeros = getNumOfConsecutiveZeros(
5087 SVOp, NumElems, true /* check zeros from left */, DAG,
5088 NumElems - SVOp->getMaskElt(NumElems - 1) - 1);
5094 // Considering the elements in the mask that are not consecutive zeros,
5095 // check if they consecutively come from only one of the source vectors.
5097 // 0 { A, B, X, X } = V2
5099 // vector_shuffle V1, V2 <X, X, 4, 5>
5101 if (!isShuffleMaskConsecutive(SVOp,
5102 NumZeros, // Mask Start Index
5103 NumElems, // Mask End Index(exclusive)
5104 0, // Where to start looking in the src vector
5105 NumElems, // Number of elements in vector
5106 OpSrc)) // Which source operand ?
5111 ShVal = SVOp->getOperand(OpSrc);
5115 /// isVectorShift - Returns true if the shuffle can be implemented as a
5116 /// logical left or right shift of a vector.
5117 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
5118 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
5119 // Although the logic below support any bitwidth size, there are no
5120 // shift instructions which handle more than 128-bit vectors.
5121 if (!SVOp->getValueType(0).is128BitVector())
5124 if (isVectorShiftLeft(SVOp, DAG, isLeft, ShVal, ShAmt) ||
5125 isVectorShiftRight(SVOp, DAG, isLeft, ShVal, ShAmt))
5131 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
5133 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
5134 unsigned NumNonZero, unsigned NumZero,
5136 const X86Subtarget* Subtarget,
5137 const TargetLowering &TLI) {
5144 for (unsigned i = 0; i < 16; ++i) {
5145 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
5146 if (ThisIsNonZero && First) {
5148 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
5150 V = DAG.getUNDEF(MVT::v8i16);
5155 SDValue ThisElt(0, 0), LastElt(0, 0);
5156 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
5157 if (LastIsNonZero) {
5158 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
5159 MVT::i16, Op.getOperand(i-1));
5161 if (ThisIsNonZero) {
5162 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
5163 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
5164 ThisElt, DAG.getConstant(8, MVT::i8));
5166 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
5170 if (ThisElt.getNode())
5171 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
5172 DAG.getIntPtrConstant(i/2));
5176 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V);
5179 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
5181 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
5182 unsigned NumNonZero, unsigned NumZero,
5184 const X86Subtarget* Subtarget,
5185 const TargetLowering &TLI) {
5192 for (unsigned i = 0; i < 8; ++i) {
5193 bool isNonZero = (NonZeros & (1 << i)) != 0;
5197 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
5199 V = DAG.getUNDEF(MVT::v8i16);
5202 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
5203 MVT::v8i16, V, Op.getOperand(i),
5204 DAG.getIntPtrConstant(i));
5211 /// getVShift - Return a vector logical shift node.
5213 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
5214 unsigned NumBits, SelectionDAG &DAG,
5215 const TargetLowering &TLI, SDLoc dl) {
5216 assert(VT.is128BitVector() && "Unknown type for VShift");
5217 EVT ShVT = MVT::v2i64;
5218 unsigned Opc = isLeft ? X86ISD::VSHLDQ : X86ISD::VSRLDQ;
5219 SrcOp = DAG.getNode(ISD::BITCAST, dl, ShVT, SrcOp);
5220 return DAG.getNode(ISD::BITCAST, dl, VT,
5221 DAG.getNode(Opc, dl, ShVT, SrcOp,
5222 DAG.getConstant(NumBits,
5223 TLI.getScalarShiftAmountTy(SrcOp.getValueType()))));
5227 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, SDLoc dl,
5228 SelectionDAG &DAG) const {
5230 // Check if the scalar load can be widened into a vector load. And if
5231 // the address is "base + cst" see if the cst can be "absorbed" into
5232 // the shuffle mask.
5233 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
5234 SDValue Ptr = LD->getBasePtr();
5235 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
5237 EVT PVT = LD->getValueType(0);
5238 if (PVT != MVT::i32 && PVT != MVT::f32)
5243 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
5244 FI = FINode->getIndex();
5246 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
5247 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
5248 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
5249 Offset = Ptr.getConstantOperandVal(1);
5250 Ptr = Ptr.getOperand(0);
5255 // FIXME: 256-bit vector instructions don't require a strict alignment,
5256 // improve this code to support it better.
5257 unsigned RequiredAlign = VT.getSizeInBits()/8;
5258 SDValue Chain = LD->getChain();
5259 // Make sure the stack object alignment is at least 16 or 32.
5260 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5261 if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
5262 if (MFI->isFixedObjectIndex(FI)) {
5263 // Can't change the alignment. FIXME: It's possible to compute
5264 // the exact stack offset and reference FI + adjust offset instead.
5265 // If someone *really* cares about this. That's the way to implement it.
5268 MFI->setObjectAlignment(FI, RequiredAlign);
5272 // (Offset % 16 or 32) must be multiple of 4. Then address is then
5273 // Ptr + (Offset & ~15).
5276 if ((Offset % RequiredAlign) & 3)
5278 int64_t StartOffset = Offset & ~(RequiredAlign-1);
5280 Ptr = DAG.getNode(ISD::ADD, SDLoc(Ptr), Ptr.getValueType(),
5281 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
5283 int EltNo = (Offset - StartOffset) >> 2;
5284 unsigned NumElems = VT.getVectorNumElements();
5286 EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
5287 SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
5288 LD->getPointerInfo().getWithOffset(StartOffset),
5289 false, false, false, 0);
5291 SmallVector<int, 8> Mask;
5292 for (unsigned i = 0; i != NumElems; ++i)
5293 Mask.push_back(EltNo);
5295 return DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), &Mask[0]);
5301 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
5302 /// vector of type 'VT', see if the elements can be replaced by a single large
5303 /// load which has the same value as a build_vector whose operands are 'elts'.
5305 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
5307 /// FIXME: we'd also like to handle the case where the last elements are zero
5308 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
5309 /// There's even a handy isZeroNode for that purpose.
5310 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
5311 SDLoc &DL, SelectionDAG &DAG) {
5312 EVT EltVT = VT.getVectorElementType();
5313 unsigned NumElems = Elts.size();
5315 LoadSDNode *LDBase = NULL;
5316 unsigned LastLoadedElt = -1U;
5318 // For each element in the initializer, see if we've found a load or an undef.
5319 // If we don't find an initial load element, or later load elements are
5320 // non-consecutive, bail out.
5321 for (unsigned i = 0; i < NumElems; ++i) {
5322 SDValue Elt = Elts[i];
5324 if (!Elt.getNode() ||
5325 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
5328 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
5330 LDBase = cast<LoadSDNode>(Elt.getNode());
5334 if (Elt.getOpcode() == ISD::UNDEF)
5337 LoadSDNode *LD = cast<LoadSDNode>(Elt);
5338 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
5343 // If we have found an entire vector of loads and undefs, then return a large
5344 // load of the entire vector width starting at the base pointer. If we found
5345 // consecutive loads for the low half, generate a vzext_load node.
5346 if (LastLoadedElt == NumElems - 1) {
5347 SDValue NewLd = SDValue();
5348 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
5349 NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
5350 LDBase->getPointerInfo(),
5351 LDBase->isVolatile(), LDBase->isNonTemporal(),
5352 LDBase->isInvariant(), 0);
5353 NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
5354 LDBase->getPointerInfo(),
5355 LDBase->isVolatile(), LDBase->isNonTemporal(),
5356 LDBase->isInvariant(), LDBase->getAlignment());
5358 if (LDBase->hasAnyUseOfValue(1)) {
5359 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5361 SDValue(NewLd.getNode(), 1));
5362 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5363 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5364 SDValue(NewLd.getNode(), 1));
5369 if (NumElems == 4 && LastLoadedElt == 1 &&
5370 DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
5371 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
5372 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
5374 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys, Ops,
5375 array_lengthof(Ops), MVT::i64,
5376 LDBase->getPointerInfo(),
5377 LDBase->getAlignment(),
5378 false/*isVolatile*/, true/*ReadMem*/,
5381 // Make sure the newly-created LOAD is in the same position as LDBase in
5382 // terms of dependency. We create a TokenFactor for LDBase and ResNode, and
5383 // update uses of LDBase's output chain to use the TokenFactor.
5384 if (LDBase->hasAnyUseOfValue(1)) {
5385 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5386 SDValue(LDBase, 1), SDValue(ResNode.getNode(), 1));
5387 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5388 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5389 SDValue(ResNode.getNode(), 1));
5392 return DAG.getNode(ISD::BITCAST, DL, VT, ResNode);
5397 /// LowerVectorBroadcast - Attempt to use the vbroadcast instruction
5398 /// to generate a splat value for the following cases:
5399 /// 1. A splat BUILD_VECTOR which uses a single scalar load, or a constant.
5400 /// 2. A splat shuffle which uses a scalar_to_vector node which comes from
5401 /// a scalar load, or a constant.
5402 /// The VBROADCAST node is returned when a pattern is found,
5403 /// or SDValue() otherwise.
5405 X86TargetLowering::LowerVectorBroadcast(SDValue Op, SelectionDAG &DAG) const {
5406 if (!Subtarget->hasFp256())
5409 MVT VT = Op.getValueType().getSimpleVT();
5412 assert((VT.is128BitVector() || VT.is256BitVector()) &&
5413 "Unsupported vector type for broadcast.");
5418 switch (Op.getOpcode()) {
5420 // Unknown pattern found.
5423 case ISD::BUILD_VECTOR: {
5424 // The BUILD_VECTOR node must be a splat.
5425 if (!isSplatVector(Op.getNode()))
5428 Ld = Op.getOperand(0);
5429 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5430 Ld.getOpcode() == ISD::ConstantFP);
5432 // The suspected load node has several users. Make sure that all
5433 // of its users are from the BUILD_VECTOR node.
5434 // Constants may have multiple users.
5435 if (!ConstSplatVal && !Ld->hasNUsesOfValue(VT.getVectorNumElements(), 0))
5440 case ISD::VECTOR_SHUFFLE: {
5441 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5443 // Shuffles must have a splat mask where the first element is
5445 if ((!SVOp->isSplat()) || SVOp->getMaskElt(0) != 0)
5448 SDValue Sc = Op.getOperand(0);
5449 if (Sc.getOpcode() != ISD::SCALAR_TO_VECTOR &&
5450 Sc.getOpcode() != ISD::BUILD_VECTOR) {
5452 if (!Subtarget->hasInt256())
5455 // Use the register form of the broadcast instruction available on AVX2.
5456 if (VT.is256BitVector())
5457 Sc = Extract128BitVector(Sc, 0, DAG, dl);
5458 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Sc);
5461 Ld = Sc.getOperand(0);
5462 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5463 Ld.getOpcode() == ISD::ConstantFP);
5465 // The scalar_to_vector node and the suspected
5466 // load node must have exactly one user.
5467 // Constants may have multiple users.
5468 if (!ConstSplatVal && (!Sc.hasOneUse() || !Ld.hasOneUse()))
5474 bool Is256 = VT.is256BitVector();
5476 // Handle the broadcasting a single constant scalar from the constant pool
5477 // into a vector. On Sandybridge it is still better to load a constant vector
5478 // from the constant pool and not to broadcast it from a scalar.
5479 if (ConstSplatVal && Subtarget->hasInt256()) {
5480 EVT CVT = Ld.getValueType();
5481 assert(!CVT.isVector() && "Must not broadcast a vector type");
5482 unsigned ScalarSize = CVT.getSizeInBits();
5484 if (ScalarSize == 32 || (Is256 && ScalarSize == 64)) {
5485 const Constant *C = 0;
5486 if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld))
5487 C = CI->getConstantIntValue();
5488 else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld))
5489 C = CF->getConstantFPValue();
5491 assert(C && "Invalid constant type");
5493 SDValue CP = DAG.getConstantPool(C, getPointerTy());
5494 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
5495 Ld = DAG.getLoad(CVT, dl, DAG.getEntryNode(), CP,
5496 MachinePointerInfo::getConstantPool(),
5497 false, false, false, Alignment);
5499 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5503 bool IsLoad = ISD::isNormalLoad(Ld.getNode());
5504 unsigned ScalarSize = Ld.getValueType().getSizeInBits();
5506 // Handle AVX2 in-register broadcasts.
5507 if (!IsLoad && Subtarget->hasInt256() &&
5508 (ScalarSize == 32 || (Is256 && ScalarSize == 64)))
5509 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5511 // The scalar source must be a normal load.
5515 if (ScalarSize == 32 || (Is256 && ScalarSize == 64))
5516 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5518 // The integer check is needed for the 64-bit into 128-bit so it doesn't match
5519 // double since there is no vbroadcastsd xmm
5520 if (Subtarget->hasInt256() && Ld.getValueType().isInteger()) {
5521 if (ScalarSize == 8 || ScalarSize == 16 || ScalarSize == 64)
5522 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5525 // Unsupported broadcast.
5530 X86TargetLowering::buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) const {
5531 EVT VT = Op.getValueType();
5533 // Skip if insert_vec_elt is not supported.
5534 if (!isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT))
5538 unsigned NumElems = Op.getNumOperands();
5542 SmallVector<unsigned, 4> InsertIndices;
5543 SmallVector<int, 8> Mask(NumElems, -1);
5545 for (unsigned i = 0; i != NumElems; ++i) {
5546 unsigned Opc = Op.getOperand(i).getOpcode();
5548 if (Opc == ISD::UNDEF)
5551 if (Opc != ISD::EXTRACT_VECTOR_ELT) {
5552 // Quit if more than 1 elements need inserting.
5553 if (InsertIndices.size() > 1)
5556 InsertIndices.push_back(i);
5560 SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0);
5561 SDValue ExtIdx = Op.getOperand(i).getOperand(1);
5563 // Quit if extracted from vector of different type.
5564 if (ExtractedFromVec.getValueType() != VT)
5567 // Quit if non-constant index.
5568 if (!isa<ConstantSDNode>(ExtIdx))
5571 if (VecIn1.getNode() == 0)
5572 VecIn1 = ExtractedFromVec;
5573 else if (VecIn1 != ExtractedFromVec) {
5574 if (VecIn2.getNode() == 0)
5575 VecIn2 = ExtractedFromVec;
5576 else if (VecIn2 != ExtractedFromVec)
5577 // Quit if more than 2 vectors to shuffle
5581 unsigned Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue();
5583 if (ExtractedFromVec == VecIn1)
5585 else if (ExtractedFromVec == VecIn2)
5586 Mask[i] = Idx + NumElems;
5589 if (VecIn1.getNode() == 0)
5592 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
5593 SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, &Mask[0]);
5594 for (unsigned i = 0, e = InsertIndices.size(); i != e; ++i) {
5595 unsigned Idx = InsertIndices[i];
5596 NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx),
5597 DAG.getIntPtrConstant(Idx));
5603 // Lower BUILD_VECTOR operation for v8i1 and v16i1 types.
5605 X86TargetLowering::LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const {
5607 EVT VT = Op.getValueType();
5608 assert((VT.getVectorElementType() == MVT::i1) && (VT.getSizeInBits() <= 16) &&
5609 "Unexpected type in LowerBUILD_VECTORvXi1!");
5612 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5613 SDValue Cst = DAG.getTargetConstant(0, MVT::i1);
5614 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
5615 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
5616 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT,
5617 Ops, VT.getVectorNumElements());
5620 if (ISD::isBuildVectorAllOnes(Op.getNode())) {
5621 SDValue Cst = DAG.getTargetConstant(1, MVT::i1);
5622 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
5623 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
5624 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT,
5625 Ops, VT.getVectorNumElements());
5628 bool AllContants = true;
5629 uint64_t Immediate = 0;
5630 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5631 SDValue In = Op.getOperand(idx);
5632 if (In.getOpcode() == ISD::UNDEF)
5634 if (!isa<ConstantSDNode>(In)) {
5635 AllContants = false;
5638 if (cast<ConstantSDNode>(In)->getZExtValue())
5639 Immediate |= (1ULL << idx);
5643 SDValue FullMask = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1,
5644 DAG.getConstant(Immediate, MVT::i16));
5645 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, FullMask,
5646 DAG.getIntPtrConstant(0));
5649 if (!isSplatVector(Op.getNode()))
5650 llvm_unreachable("Unsupported predicate operation");
5652 SDValue In = Op.getOperand(0);
5653 SDValue EFLAGS, X86CC;
5654 if (In.getOpcode() == ISD::SETCC) {
5655 SDValue Op0 = In.getOperand(0);
5656 SDValue Op1 = In.getOperand(1);
5657 ISD::CondCode CC = cast<CondCodeSDNode>(In.getOperand(2))->get();
5658 bool isFP = Op1.getValueType().isFloatingPoint();
5659 unsigned X86CCVal = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
5661 assert(X86CCVal != X86::COND_INVALID && "Unsupported predicate operation");
5663 X86CC = DAG.getConstant(X86CCVal, MVT::i8);
5664 EFLAGS = EmitCmp(Op0, Op1, X86CCVal, DAG);
5665 EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
5666 } else if (In.getOpcode() == X86ISD::SETCC) {
5667 X86CC = In.getOperand(0);
5668 EFLAGS = In.getOperand(1);
5677 // res = allOnes ### CMOVNE -1, %res
5680 MVT InVT = In.getValueType().getSimpleVT();
5681 SDValue Bit1 = DAG.getNode(ISD::AND, dl, InVT, In, DAG.getConstant(1, InVT));
5682 EFLAGS = EmitTest(Bit1, X86::COND_NE, DAG);
5683 X86CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5686 if (VT == MVT::v16i1) {
5687 SDValue Cst1 = DAG.getConstant(-1, MVT::i16);
5688 SDValue Cst0 = DAG.getConstant(0, MVT::i16);
5689 SDValue CmovOp = DAG.getNode(X86ISD::CMOV, dl, MVT::i16,
5690 Cst0, Cst1, X86CC, EFLAGS);
5691 return DAG.getNode(ISD::BITCAST, dl, VT, CmovOp);
5694 if (VT == MVT::v8i1) {
5695 SDValue Cst1 = DAG.getConstant(-1, MVT::i32);
5696 SDValue Cst0 = DAG.getConstant(0, MVT::i32);
5697 SDValue CmovOp = DAG.getNode(X86ISD::CMOV, dl, MVT::i32,
5698 Cst0, Cst1, X86CC, EFLAGS);
5699 CmovOp = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, CmovOp);
5700 return DAG.getNode(ISD::BITCAST, dl, VT, CmovOp);
5702 llvm_unreachable("Unsupported predicate operation");
5706 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5709 MVT VT = Op.getValueType().getSimpleVT();
5710 MVT ExtVT = VT.getVectorElementType();
5711 unsigned NumElems = Op.getNumOperands();
5713 // Generate vectors for predicate vectors.
5714 if (VT.getScalarType() == MVT::i1 && Subtarget->hasAVX512())
5715 return LowerBUILD_VECTORvXi1(Op, DAG);
5717 // Vectors containing all zeros can be matched by pxor and xorps later
5718 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5719 // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
5720 // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
5721 if (VT == MVT::v4i32 || VT == MVT::v8i32)
5724 return getZeroVector(VT, Subtarget, DAG, dl);
5727 // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
5728 // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use
5729 // vpcmpeqd on 256-bit vectors.
5730 if (Subtarget->hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) {
5731 if (VT == MVT::v4i32 || (VT == MVT::v8i32 && Subtarget->hasInt256()))
5734 return getOnesVector(VT, Subtarget->hasInt256(), DAG, dl);
5737 SDValue Broadcast = LowerVectorBroadcast(Op, DAG);
5738 if (Broadcast.getNode())
5741 unsigned EVTBits = ExtVT.getSizeInBits();
5743 unsigned NumZero = 0;
5744 unsigned NumNonZero = 0;
5745 unsigned NonZeros = 0;
5746 bool IsAllConstants = true;
5747 SmallSet<SDValue, 8> Values;
5748 for (unsigned i = 0; i < NumElems; ++i) {
5749 SDValue Elt = Op.getOperand(i);
5750 if (Elt.getOpcode() == ISD::UNDEF)
5753 if (Elt.getOpcode() != ISD::Constant &&
5754 Elt.getOpcode() != ISD::ConstantFP)
5755 IsAllConstants = false;
5756 if (X86::isZeroNode(Elt))
5759 NonZeros |= (1 << i);
5764 // All undef vector. Return an UNDEF. All zero vectors were handled above.
5765 if (NumNonZero == 0)
5766 return DAG.getUNDEF(VT);
5768 // Special case for single non-zero, non-undef, element.
5769 if (NumNonZero == 1) {
5770 unsigned Idx = countTrailingZeros(NonZeros);
5771 SDValue Item = Op.getOperand(Idx);
5773 // If this is an insertion of an i64 value on x86-32, and if the top bits of
5774 // the value are obviously zero, truncate the value to i32 and do the
5775 // insertion that way. Only do this if the value is non-constant or if the
5776 // value is a constant being inserted into element 0. It is cheaper to do
5777 // a constant pool load than it is to do a movd + shuffle.
5778 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
5779 (!IsAllConstants || Idx == 0)) {
5780 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
5782 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
5783 EVT VecVT = MVT::v4i32;
5784 unsigned VecElts = 4;
5786 // Truncate the value (which may itself be a constant) to i32, and
5787 // convert it to a vector with movd (S2V+shuffle to zero extend).
5788 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
5789 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
5790 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
5792 // Now we have our 32-bit value zero extended in the low element of
5793 // a vector. If Idx != 0, swizzle it into place.
5795 SmallVector<int, 4> Mask;
5796 Mask.push_back(Idx);
5797 for (unsigned i = 1; i != VecElts; ++i)
5799 Item = DAG.getVectorShuffle(VecVT, dl, Item, DAG.getUNDEF(VecVT),
5802 return DAG.getNode(ISD::BITCAST, dl, VT, Item);
5806 // If we have a constant or non-constant insertion into the low element of
5807 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
5808 // the rest of the elements. This will be matched as movd/movq/movss/movsd
5809 // depending on what the source datatype is.
5812 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5814 if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
5815 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
5816 if (VT.is256BitVector()) {
5817 SDValue ZeroVec = getZeroVector(VT, Subtarget, DAG, dl);
5818 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, ZeroVec,
5819 Item, DAG.getIntPtrConstant(0));
5821 assert(VT.is128BitVector() && "Expected an SSE value type!");
5822 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5823 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
5824 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
5827 if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
5828 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
5829 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
5830 if (VT.is256BitVector()) {
5831 SDValue ZeroVec = getZeroVector(MVT::v8i32, Subtarget, DAG, dl);
5832 Item = Insert128BitVector(ZeroVec, Item, 0, DAG, dl);
5834 assert(VT.is128BitVector() && "Expected an SSE value type!");
5835 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
5837 return DAG.getNode(ISD::BITCAST, dl, VT, Item);
5841 // Is it a vector logical left shift?
5842 if (NumElems == 2 && Idx == 1 &&
5843 X86::isZeroNode(Op.getOperand(0)) &&
5844 !X86::isZeroNode(Op.getOperand(1))) {
5845 unsigned NumBits = VT.getSizeInBits();
5846 return getVShift(true, VT,
5847 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5848 VT, Op.getOperand(1)),
5849 NumBits/2, DAG, *this, dl);
5852 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
5855 // Otherwise, if this is a vector with i32 or f32 elements, and the element
5856 // is a non-constant being inserted into an element other than the low one,
5857 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
5858 // movd/movss) to move this into the low element, then shuffle it into
5860 if (EVTBits == 32) {
5861 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5863 // Turn it into a shuffle of zero and zero-extended scalar to vector.
5864 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0, Subtarget, DAG);
5865 SmallVector<int, 8> MaskVec;
5866 for (unsigned i = 0; i != NumElems; ++i)
5867 MaskVec.push_back(i == Idx ? 0 : 1);
5868 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
5872 // Splat is obviously ok. Let legalizer expand it to a shuffle.
5873 if (Values.size() == 1) {
5874 if (EVTBits == 32) {
5875 // Instead of a shuffle like this:
5876 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
5877 // Check if it's possible to issue this instead.
5878 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
5879 unsigned Idx = countTrailingZeros(NonZeros);
5880 SDValue Item = Op.getOperand(Idx);
5881 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
5882 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
5887 // A vector full of immediates; various special cases are already
5888 // handled, so this is best done with a single constant-pool load.
5892 // For AVX-length vectors, build the individual 128-bit pieces and use
5893 // shuffles to put them in place.
5894 if (VT.is256BitVector()) {
5895 SmallVector<SDValue, 32> V;
5896 for (unsigned i = 0; i != NumElems; ++i)
5897 V.push_back(Op.getOperand(i));
5899 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
5901 // Build both the lower and upper subvector.
5902 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT, &V[0], NumElems/2);
5903 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT, &V[NumElems / 2],
5906 // Recreate the wider vector with the lower and upper part.
5907 return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
5910 // Let legalizer expand 2-wide build_vectors.
5911 if (EVTBits == 64) {
5912 if (NumNonZero == 1) {
5913 // One half is zero or undef.
5914 unsigned Idx = countTrailingZeros(NonZeros);
5915 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
5916 Op.getOperand(Idx));
5917 return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG);
5922 // If element VT is < 32 bits, convert it to inserts into a zero vector.
5923 if (EVTBits == 8 && NumElems == 16) {
5924 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
5926 if (V.getNode()) return V;
5929 if (EVTBits == 16 && NumElems == 8) {
5930 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
5932 if (V.getNode()) return V;
5935 // If element VT is == 32 bits, turn it into a number of shuffles.
5936 SmallVector<SDValue, 8> V(NumElems);
5937 if (NumElems == 4 && NumZero > 0) {
5938 for (unsigned i = 0; i < 4; ++i) {
5939 bool isZero = !(NonZeros & (1 << i));
5941 V[i] = getZeroVector(VT, Subtarget, DAG, dl);
5943 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
5946 for (unsigned i = 0; i < 2; ++i) {
5947 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
5950 V[i] = V[i*2]; // Must be a zero vector.
5953 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
5956 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
5959 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
5964 bool Reverse1 = (NonZeros & 0x3) == 2;
5965 bool Reverse2 = ((NonZeros & (0x3 << 2)) >> 2) == 2;
5969 static_cast<int>(Reverse2 ? NumElems+1 : NumElems),
5970 static_cast<int>(Reverse2 ? NumElems : NumElems+1)
5972 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
5975 if (Values.size() > 1 && VT.is128BitVector()) {
5976 // Check for a build vector of consecutive loads.
5977 for (unsigned i = 0; i < NumElems; ++i)
5978 V[i] = Op.getOperand(i);
5980 // Check for elements which are consecutive loads.
5981 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
5985 // Check for a build vector from mostly shuffle plus few inserting.
5986 SDValue Sh = buildFromShuffleMostly(Op, DAG);
5990 // For SSE 4.1, use insertps to put the high elements into the low element.
5991 if (getSubtarget()->hasSSE41()) {
5993 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
5994 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
5996 Result = DAG.getUNDEF(VT);
5998 for (unsigned i = 1; i < NumElems; ++i) {
5999 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
6000 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
6001 Op.getOperand(i), DAG.getIntPtrConstant(i));
6006 // Otherwise, expand into a number of unpckl*, start by extending each of
6007 // our (non-undef) elements to the full vector width with the element in the
6008 // bottom slot of the vector (which generates no code for SSE).
6009 for (unsigned i = 0; i < NumElems; ++i) {
6010 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
6011 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6013 V[i] = DAG.getUNDEF(VT);
6016 // Next, we iteratively mix elements, e.g. for v4f32:
6017 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
6018 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
6019 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
6020 unsigned EltStride = NumElems >> 1;
6021 while (EltStride != 0) {
6022 for (unsigned i = 0; i < EltStride; ++i) {
6023 // If V[i+EltStride] is undef and this is the first round of mixing,
6024 // then it is safe to just drop this shuffle: V[i] is already in the
6025 // right place, the one element (since it's the first round) being
6026 // inserted as undef can be dropped. This isn't safe for successive
6027 // rounds because they will permute elements within both vectors.
6028 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
6029 EltStride == NumElems/2)
6032 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
6041 // LowerAVXCONCAT_VECTORS - 256-bit AVX can use the vinsertf128 instruction
6042 // to create 256-bit vectors from two other 128-bit ones.
6043 static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
6045 MVT ResVT = Op.getValueType().getSimpleVT();
6047 assert((ResVT.is256BitVector() ||
6048 ResVT.is512BitVector()) && "Value type must be 256-/512-bit wide");
6050 SDValue V1 = Op.getOperand(0);
6051 SDValue V2 = Op.getOperand(1);
6052 unsigned NumElems = ResVT.getVectorNumElements();
6053 if(ResVT.is256BitVector())
6054 return Concat128BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6056 return Concat256BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6059 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
6060 assert(Op.getNumOperands() == 2);
6062 // AVX/AVX-512 can use the vinsertf128 instruction to create 256-bit vectors
6063 // from two other 128-bit ones.
6064 return LowerAVXCONCAT_VECTORS(Op, DAG);
6067 // Try to lower a shuffle node into a simple blend instruction.
6069 LowerVECTOR_SHUFFLEtoBlend(ShuffleVectorSDNode *SVOp,
6070 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
6071 SDValue V1 = SVOp->getOperand(0);
6072 SDValue V2 = SVOp->getOperand(1);
6074 MVT VT = SVOp->getValueType(0).getSimpleVT();
6075 MVT EltVT = VT.getVectorElementType();
6076 unsigned NumElems = VT.getVectorNumElements();
6078 if (!Subtarget->hasSSE41() || EltVT == MVT::i8)
6080 if (!Subtarget->hasInt256() && VT == MVT::v16i16)
6083 // Check the mask for BLEND and build the value.
6084 unsigned MaskValue = 0;
6085 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
6086 unsigned NumLanes = (NumElems-1)/8 + 1;
6087 unsigned NumElemsInLane = NumElems / NumLanes;
6089 // Blend for v16i16 should be symetric for the both lanes.
6090 for (unsigned i = 0; i < NumElemsInLane; ++i) {
6092 int SndLaneEltIdx = (NumLanes == 2) ?
6093 SVOp->getMaskElt(i + NumElemsInLane) : -1;
6094 int EltIdx = SVOp->getMaskElt(i);
6096 if ((EltIdx < 0 || EltIdx == (int)i) &&
6097 (SndLaneEltIdx < 0 || SndLaneEltIdx == (int)(i + NumElemsInLane)))
6100 if (((unsigned)EltIdx == (i + NumElems)) &&
6101 (SndLaneEltIdx < 0 ||
6102 (unsigned)SndLaneEltIdx == i + NumElems + NumElemsInLane))
6103 MaskValue |= (1<<i);
6108 // Convert i32 vectors to floating point if it is not AVX2.
6109 // AVX2 introduced VPBLENDD instruction for 128 and 256-bit vectors.
6111 if (EltVT == MVT::i64 || (EltVT == MVT::i32 && !Subtarget->hasInt256())) {
6112 BlendVT = MVT::getVectorVT(MVT::getFloatingPointVT(EltVT.getSizeInBits()),
6114 V1 = DAG.getNode(ISD::BITCAST, dl, VT, V1);
6115 V2 = DAG.getNode(ISD::BITCAST, dl, VT, V2);
6118 SDValue Ret = DAG.getNode(X86ISD::BLENDI, dl, BlendVT, V1, V2,
6119 DAG.getConstant(MaskValue, MVT::i32));
6120 return DAG.getNode(ISD::BITCAST, dl, VT, Ret);
6123 // v8i16 shuffles - Prefer shuffles in the following order:
6124 // 1. [all] pshuflw, pshufhw, optional move
6125 // 2. [ssse3] 1 x pshufb
6126 // 3. [ssse3] 2 x pshufb + 1 x por
6127 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
6129 LowerVECTOR_SHUFFLEv8i16(SDValue Op, const X86Subtarget *Subtarget,
6130 SelectionDAG &DAG) {
6131 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
6132 SDValue V1 = SVOp->getOperand(0);
6133 SDValue V2 = SVOp->getOperand(1);
6135 SmallVector<int, 8> MaskVals;
6137 // Determine if more than 1 of the words in each of the low and high quadwords
6138 // of the result come from the same quadword of one of the two inputs. Undef
6139 // mask values count as coming from any quadword, for better codegen.
6140 unsigned LoQuad[] = { 0, 0, 0, 0 };
6141 unsigned HiQuad[] = { 0, 0, 0, 0 };
6142 std::bitset<4> InputQuads;
6143 for (unsigned i = 0; i < 8; ++i) {
6144 unsigned *Quad = i < 4 ? LoQuad : HiQuad;
6145 int EltIdx = SVOp->getMaskElt(i);
6146 MaskVals.push_back(EltIdx);
6155 InputQuads.set(EltIdx / 4);
6158 int BestLoQuad = -1;
6159 unsigned MaxQuad = 1;
6160 for (unsigned i = 0; i < 4; ++i) {
6161 if (LoQuad[i] > MaxQuad) {
6163 MaxQuad = LoQuad[i];
6167 int BestHiQuad = -1;
6169 for (unsigned i = 0; i < 4; ++i) {
6170 if (HiQuad[i] > MaxQuad) {
6172 MaxQuad = HiQuad[i];
6176 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
6177 // of the two input vectors, shuffle them into one input vector so only a
6178 // single pshufb instruction is necessary. If There are more than 2 input
6179 // quads, disable the next transformation since it does not help SSSE3.
6180 bool V1Used = InputQuads[0] || InputQuads[1];
6181 bool V2Used = InputQuads[2] || InputQuads[3];
6182 if (Subtarget->hasSSSE3()) {
6183 if (InputQuads.count() == 2 && V1Used && V2Used) {
6184 BestLoQuad = InputQuads[0] ? 0 : 1;
6185 BestHiQuad = InputQuads[2] ? 2 : 3;
6187 if (InputQuads.count() > 2) {
6193 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
6194 // the shuffle mask. If a quad is scored as -1, that means that it contains
6195 // words from all 4 input quadwords.
6197 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
6199 BestLoQuad < 0 ? 0 : BestLoQuad,
6200 BestHiQuad < 0 ? 1 : BestHiQuad
6202 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
6203 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V1),
6204 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V2), &MaskV[0]);
6205 NewV = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, NewV);
6207 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
6208 // source words for the shuffle, to aid later transformations.
6209 bool AllWordsInNewV = true;
6210 bool InOrder[2] = { true, true };
6211 for (unsigned i = 0; i != 8; ++i) {
6212 int idx = MaskVals[i];
6214 InOrder[i/4] = false;
6215 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
6217 AllWordsInNewV = false;
6221 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
6222 if (AllWordsInNewV) {
6223 for (int i = 0; i != 8; ++i) {
6224 int idx = MaskVals[i];
6227 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
6228 if ((idx != i) && idx < 4)
6230 if ((idx != i) && idx > 3)
6239 // If we've eliminated the use of V2, and the new mask is a pshuflw or
6240 // pshufhw, that's as cheap as it gets. Return the new shuffle.
6241 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
6242 unsigned Opc = pshufhw ? X86ISD::PSHUFHW : X86ISD::PSHUFLW;
6243 unsigned TargetMask = 0;
6244 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
6245 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
6246 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(NewV.getNode());
6247 TargetMask = pshufhw ? getShufflePSHUFHWImmediate(SVOp):
6248 getShufflePSHUFLWImmediate(SVOp);
6249 V1 = NewV.getOperand(0);
6250 return getTargetShuffleNode(Opc, dl, MVT::v8i16, V1, TargetMask, DAG);
6254 // Promote splats to a larger type which usually leads to more efficient code.
6255 // FIXME: Is this true if pshufb is available?
6256 if (SVOp->isSplat())
6257 return PromoteSplat(SVOp, DAG);
6259 // If we have SSSE3, and all words of the result are from 1 input vector,
6260 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
6261 // is present, fall back to case 4.
6262 if (Subtarget->hasSSSE3()) {
6263 SmallVector<SDValue,16> pshufbMask;
6265 // If we have elements from both input vectors, set the high bit of the
6266 // shuffle mask element to zero out elements that come from V2 in the V1
6267 // mask, and elements that come from V1 in the V2 mask, so that the two
6268 // results can be OR'd together.
6269 bool TwoInputs = V1Used && V2Used;
6270 for (unsigned i = 0; i != 8; ++i) {
6271 int EltIdx = MaskVals[i] * 2;
6272 int Idx0 = (TwoInputs && (EltIdx >= 16)) ? 0x80 : EltIdx;
6273 int Idx1 = (TwoInputs && (EltIdx >= 16)) ? 0x80 : EltIdx+1;
6274 pshufbMask.push_back(DAG.getConstant(Idx0, MVT::i8));
6275 pshufbMask.push_back(DAG.getConstant(Idx1, MVT::i8));
6277 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V1);
6278 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
6279 DAG.getNode(ISD::BUILD_VECTOR, dl,
6280 MVT::v16i8, &pshufbMask[0], 16));
6282 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
6284 // Calculate the shuffle mask for the second input, shuffle it, and
6285 // OR it with the first shuffled input.
6287 for (unsigned i = 0; i != 8; ++i) {
6288 int EltIdx = MaskVals[i] * 2;
6289 int Idx0 = (EltIdx < 16) ? 0x80 : EltIdx - 16;
6290 int Idx1 = (EltIdx < 16) ? 0x80 : EltIdx - 15;
6291 pshufbMask.push_back(DAG.getConstant(Idx0, MVT::i8));
6292 pshufbMask.push_back(DAG.getConstant(Idx1, MVT::i8));
6294 V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V2);
6295 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
6296 DAG.getNode(ISD::BUILD_VECTOR, dl,
6297 MVT::v16i8, &pshufbMask[0], 16));
6298 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
6299 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
6302 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
6303 // and update MaskVals with new element order.
6304 std::bitset<8> InOrder;
6305 if (BestLoQuad >= 0) {
6306 int MaskV[] = { -1, -1, -1, -1, 4, 5, 6, 7 };
6307 for (int i = 0; i != 4; ++i) {
6308 int idx = MaskVals[i];
6311 } else if ((idx / 4) == BestLoQuad) {
6316 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
6319 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3()) {
6320 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(NewV.getNode());
6321 NewV = getTargetShuffleNode(X86ISD::PSHUFLW, dl, MVT::v8i16,
6323 getShufflePSHUFLWImmediate(SVOp), DAG);
6327 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
6328 // and update MaskVals with the new element order.
6329 if (BestHiQuad >= 0) {
6330 int MaskV[] = { 0, 1, 2, 3, -1, -1, -1, -1 };
6331 for (unsigned i = 4; i != 8; ++i) {
6332 int idx = MaskVals[i];
6335 } else if ((idx / 4) == BestHiQuad) {
6336 MaskV[i] = (idx & 3) + 4;
6340 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
6343 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3()) {
6344 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(NewV.getNode());
6345 NewV = getTargetShuffleNode(X86ISD::PSHUFHW, dl, MVT::v8i16,
6347 getShufflePSHUFHWImmediate(SVOp), DAG);
6351 // In case BestHi & BestLo were both -1, which means each quadword has a word
6352 // from each of the four input quadwords, calculate the InOrder bitvector now
6353 // before falling through to the insert/extract cleanup.
6354 if (BestLoQuad == -1 && BestHiQuad == -1) {
6356 for (int i = 0; i != 8; ++i)
6357 if (MaskVals[i] < 0 || MaskVals[i] == i)
6361 // The other elements are put in the right place using pextrw and pinsrw.
6362 for (unsigned i = 0; i != 8; ++i) {
6365 int EltIdx = MaskVals[i];
6368 SDValue ExtOp = (EltIdx < 8) ?
6369 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
6370 DAG.getIntPtrConstant(EltIdx)) :
6371 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
6372 DAG.getIntPtrConstant(EltIdx - 8));
6373 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
6374 DAG.getIntPtrConstant(i));
6379 // v16i8 shuffles - Prefer shuffles in the following order:
6380 // 1. [ssse3] 1 x pshufb
6381 // 2. [ssse3] 2 x pshufb + 1 x por
6382 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
6384 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
6386 const X86TargetLowering &TLI) {
6387 SDValue V1 = SVOp->getOperand(0);
6388 SDValue V2 = SVOp->getOperand(1);
6390 ArrayRef<int> MaskVals = SVOp->getMask();
6392 // Promote splats to a larger type which usually leads to more efficient code.
6393 // FIXME: Is this true if pshufb is available?
6394 if (SVOp->isSplat())
6395 return PromoteSplat(SVOp, DAG);
6397 // If we have SSSE3, case 1 is generated when all result bytes come from
6398 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
6399 // present, fall back to case 3.
6401 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
6402 if (TLI.getSubtarget()->hasSSSE3()) {
6403 SmallVector<SDValue,16> pshufbMask;
6405 // If all result elements are from one input vector, then only translate
6406 // undef mask values to 0x80 (zero out result) in the pshufb mask.
6408 // Otherwise, we have elements from both input vectors, and must zero out
6409 // elements that come from V2 in the first mask, and V1 in the second mask
6410 // so that we can OR them together.
6411 for (unsigned i = 0; i != 16; ++i) {
6412 int EltIdx = MaskVals[i];
6413 if (EltIdx < 0 || EltIdx >= 16)
6415 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
6417 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
6418 DAG.getNode(ISD::BUILD_VECTOR, dl,
6419 MVT::v16i8, &pshufbMask[0], 16));
6421 // As PSHUFB will zero elements with negative indices, it's safe to ignore
6422 // the 2nd operand if it's undefined or zero.
6423 if (V2.getOpcode() == ISD::UNDEF ||
6424 ISD::isBuildVectorAllZeros(V2.getNode()))
6427 // Calculate the shuffle mask for the second input, shuffle it, and
6428 // OR it with the first shuffled input.
6430 for (unsigned i = 0; i != 16; ++i) {
6431 int EltIdx = MaskVals[i];
6432 EltIdx = (EltIdx < 16) ? 0x80 : EltIdx - 16;
6433 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
6435 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
6436 DAG.getNode(ISD::BUILD_VECTOR, dl,
6437 MVT::v16i8, &pshufbMask[0], 16));
6438 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
6441 // No SSSE3 - Calculate in place words and then fix all out of place words
6442 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
6443 // the 16 different words that comprise the two doublequadword input vectors.
6444 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
6445 V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V2);
6447 for (int i = 0; i != 8; ++i) {
6448 int Elt0 = MaskVals[i*2];
6449 int Elt1 = MaskVals[i*2+1];
6451 // This word of the result is all undef, skip it.
6452 if (Elt0 < 0 && Elt1 < 0)
6455 // This word of the result is already in the correct place, skip it.
6456 if ((Elt0 == i*2) && (Elt1 == i*2+1))
6459 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
6460 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
6463 // If Elt0 and Elt1 are defined, are consecutive, and can be load
6464 // using a single extract together, load it and store it.
6465 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
6466 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
6467 DAG.getIntPtrConstant(Elt1 / 2));
6468 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
6469 DAG.getIntPtrConstant(i));
6473 // If Elt1 is defined, extract it from the appropriate source. If the
6474 // source byte is not also odd, shift the extracted word left 8 bits
6475 // otherwise clear the bottom 8 bits if we need to do an or.
6477 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
6478 DAG.getIntPtrConstant(Elt1 / 2));
6479 if ((Elt1 & 1) == 0)
6480 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
6482 TLI.getShiftAmountTy(InsElt.getValueType())));
6484 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
6485 DAG.getConstant(0xFF00, MVT::i16));
6487 // If Elt0 is defined, extract it from the appropriate source. If the
6488 // source byte is not also even, shift the extracted word right 8 bits. If
6489 // Elt1 was also defined, OR the extracted values together before
6490 // inserting them in the result.
6492 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
6493 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
6494 if ((Elt0 & 1) != 0)
6495 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
6497 TLI.getShiftAmountTy(InsElt0.getValueType())));
6499 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
6500 DAG.getConstant(0x00FF, MVT::i16));
6501 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
6504 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
6505 DAG.getIntPtrConstant(i));
6507 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, NewV);
6510 // v32i8 shuffles - Translate to VPSHUFB if possible.
6512 SDValue LowerVECTOR_SHUFFLEv32i8(ShuffleVectorSDNode *SVOp,
6513 const X86Subtarget *Subtarget,
6514 SelectionDAG &DAG) {
6515 MVT VT = SVOp->getValueType(0).getSimpleVT();
6516 SDValue V1 = SVOp->getOperand(0);
6517 SDValue V2 = SVOp->getOperand(1);
6519 SmallVector<int, 32> MaskVals(SVOp->getMask().begin(), SVOp->getMask().end());
6521 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
6522 bool V1IsAllZero = ISD::isBuildVectorAllZeros(V1.getNode());
6523 bool V2IsAllZero = ISD::isBuildVectorAllZeros(V2.getNode());
6525 // VPSHUFB may be generated if
6526 // (1) one of input vector is undefined or zeroinitializer.
6527 // The mask value 0x80 puts 0 in the corresponding slot of the vector.
6528 // And (2) the mask indexes don't cross the 128-bit lane.
6529 if (VT != MVT::v32i8 || !Subtarget->hasInt256() ||
6530 (!V2IsUndef && !V2IsAllZero && !V1IsAllZero))
6533 if (V1IsAllZero && !V2IsAllZero) {
6534 CommuteVectorShuffleMask(MaskVals, 32);
6537 SmallVector<SDValue, 32> pshufbMask;
6538 for (unsigned i = 0; i != 32; i++) {
6539 int EltIdx = MaskVals[i];
6540 if (EltIdx < 0 || EltIdx >= 32)
6543 if ((EltIdx >= 16 && i < 16) || (EltIdx < 16 && i >= 16))
6544 // Cross lane is not allowed.
6548 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
6550 return DAG.getNode(X86ISD::PSHUFB, dl, MVT::v32i8, V1,
6551 DAG.getNode(ISD::BUILD_VECTOR, dl,
6552 MVT::v32i8, &pshufbMask[0], 32));
6555 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
6556 /// ones, or rewriting v4i32 / v4f32 as 2 wide ones if possible. This can be
6557 /// done when every pair / quad of shuffle mask elements point to elements in
6558 /// the right sequence. e.g.
6559 /// vector_shuffle X, Y, <2, 3, | 10, 11, | 0, 1, | 14, 15>
6561 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
6562 SelectionDAG &DAG) {
6563 MVT VT = SVOp->getValueType(0).getSimpleVT();
6565 unsigned NumElems = VT.getVectorNumElements();
6568 switch (VT.SimpleTy) {
6569 default: llvm_unreachable("Unexpected!");
6570 case MVT::v4f32: NewVT = MVT::v2f64; Scale = 2; break;
6571 case MVT::v4i32: NewVT = MVT::v2i64; Scale = 2; break;
6572 case MVT::v8i16: NewVT = MVT::v4i32; Scale = 2; break;
6573 case MVT::v16i8: NewVT = MVT::v4i32; Scale = 4; break;
6574 case MVT::v16i16: NewVT = MVT::v8i32; Scale = 2; break;
6575 case MVT::v32i8: NewVT = MVT::v8i32; Scale = 4; break;
6578 SmallVector<int, 8> MaskVec;
6579 for (unsigned i = 0; i != NumElems; i += Scale) {
6581 for (unsigned j = 0; j != Scale; ++j) {
6582 int EltIdx = SVOp->getMaskElt(i+j);
6586 StartIdx = (EltIdx / Scale);
6587 if (EltIdx != (int)(StartIdx*Scale + j))
6590 MaskVec.push_back(StartIdx);
6593 SDValue V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, SVOp->getOperand(0));
6594 SDValue V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, SVOp->getOperand(1));
6595 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
6598 /// getVZextMovL - Return a zero-extending vector move low node.
6600 static SDValue getVZextMovL(MVT VT, EVT OpVT,
6601 SDValue SrcOp, SelectionDAG &DAG,
6602 const X86Subtarget *Subtarget, SDLoc dl) {
6603 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
6604 LoadSDNode *LD = NULL;
6605 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
6606 LD = dyn_cast<LoadSDNode>(SrcOp);
6608 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
6610 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
6611 if ((ExtVT != MVT::i64 || Subtarget->is64Bit()) &&
6612 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
6613 SrcOp.getOperand(0).getOpcode() == ISD::BITCAST &&
6614 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
6616 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
6617 return DAG.getNode(ISD::BITCAST, dl, VT,
6618 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
6619 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6627 return DAG.getNode(ISD::BITCAST, dl, VT,
6628 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
6629 DAG.getNode(ISD::BITCAST, dl,
6633 /// LowerVECTOR_SHUFFLE_256 - Handle all 256-bit wide vectors shuffles
6634 /// which could not be matched by any known target speficic shuffle
6636 LowerVECTOR_SHUFFLE_256(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
6638 SDValue NewOp = Compact8x32ShuffleNode(SVOp, DAG);
6639 if (NewOp.getNode())
6642 MVT VT = SVOp->getValueType(0).getSimpleVT();
6644 unsigned NumElems = VT.getVectorNumElements();
6645 unsigned NumLaneElems = NumElems / 2;
6648 MVT EltVT = VT.getVectorElementType();
6649 MVT NVT = MVT::getVectorVT(EltVT, NumLaneElems);
6652 SmallVector<int, 16> Mask;
6653 for (unsigned l = 0; l < 2; ++l) {
6654 // Build a shuffle mask for the output, discovering on the fly which
6655 // input vectors to use as shuffle operands (recorded in InputUsed).
6656 // If building a suitable shuffle vector proves too hard, then bail
6657 // out with UseBuildVector set.
6658 bool UseBuildVector = false;
6659 int InputUsed[2] = { -1, -1 }; // Not yet discovered.
6660 unsigned LaneStart = l * NumLaneElems;
6661 for (unsigned i = 0; i != NumLaneElems; ++i) {
6662 // The mask element. This indexes into the input.
6663 int Idx = SVOp->getMaskElt(i+LaneStart);
6665 // the mask element does not index into any input vector.
6670 // The input vector this mask element indexes into.
6671 int Input = Idx / NumLaneElems;
6673 // Turn the index into an offset from the start of the input vector.
6674 Idx -= Input * NumLaneElems;
6676 // Find or create a shuffle vector operand to hold this input.
6678 for (OpNo = 0; OpNo < array_lengthof(InputUsed); ++OpNo) {
6679 if (InputUsed[OpNo] == Input)
6680 // This input vector is already an operand.
6682 if (InputUsed[OpNo] < 0) {
6683 // Create a new operand for this input vector.
6684 InputUsed[OpNo] = Input;
6689 if (OpNo >= array_lengthof(InputUsed)) {
6690 // More than two input vectors used! Give up on trying to create a
6691 // shuffle vector. Insert all elements into a BUILD_VECTOR instead.
6692 UseBuildVector = true;
6696 // Add the mask index for the new shuffle vector.
6697 Mask.push_back(Idx + OpNo * NumLaneElems);
6700 if (UseBuildVector) {
6701 SmallVector<SDValue, 16> SVOps;
6702 for (unsigned i = 0; i != NumLaneElems; ++i) {
6703 // The mask element. This indexes into the input.
6704 int Idx = SVOp->getMaskElt(i+LaneStart);
6706 SVOps.push_back(DAG.getUNDEF(EltVT));
6710 // The input vector this mask element indexes into.
6711 int Input = Idx / NumElems;
6713 // Turn the index into an offset from the start of the input vector.
6714 Idx -= Input * NumElems;
6716 // Extract the vector element by hand.
6717 SVOps.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
6718 SVOp->getOperand(Input),
6719 DAG.getIntPtrConstant(Idx)));
6722 // Construct the output using a BUILD_VECTOR.
6723 Output[l] = DAG.getNode(ISD::BUILD_VECTOR, dl, NVT, &SVOps[0],
6725 } else if (InputUsed[0] < 0) {
6726 // No input vectors were used! The result is undefined.
6727 Output[l] = DAG.getUNDEF(NVT);
6729 SDValue Op0 = Extract128BitVector(SVOp->getOperand(InputUsed[0] / 2),
6730 (InputUsed[0] % 2) * NumLaneElems,
6732 // If only one input was used, use an undefined vector for the other.
6733 SDValue Op1 = (InputUsed[1] < 0) ? DAG.getUNDEF(NVT) :
6734 Extract128BitVector(SVOp->getOperand(InputUsed[1] / 2),
6735 (InputUsed[1] % 2) * NumLaneElems, DAG, dl);
6736 // At least one input vector was used. Create a new shuffle vector.
6737 Output[l] = DAG.getVectorShuffle(NVT, dl, Op0, Op1, &Mask[0]);
6743 // Concatenate the result back
6744 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Output[0], Output[1]);
6747 /// LowerVECTOR_SHUFFLE_128v4 - Handle all 128-bit wide vectors with
6748 /// 4 elements, and match them with several different shuffle types.
6750 LowerVECTOR_SHUFFLE_128v4(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
6751 SDValue V1 = SVOp->getOperand(0);
6752 SDValue V2 = SVOp->getOperand(1);
6754 MVT VT = SVOp->getValueType(0).getSimpleVT();
6756 assert(VT.is128BitVector() && "Unsupported vector size");
6758 std::pair<int, int> Locs[4];
6759 int Mask1[] = { -1, -1, -1, -1 };
6760 SmallVector<int, 8> PermMask(SVOp->getMask().begin(), SVOp->getMask().end());
6764 for (unsigned i = 0; i != 4; ++i) {
6765 int Idx = PermMask[i];
6767 Locs[i] = std::make_pair(-1, -1);
6769 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
6771 Locs[i] = std::make_pair(0, NumLo);
6775 Locs[i] = std::make_pair(1, NumHi);
6777 Mask1[2+NumHi] = Idx;
6783 if (NumLo <= 2 && NumHi <= 2) {
6784 // If no more than two elements come from either vector. This can be
6785 // implemented with two shuffles. First shuffle gather the elements.
6786 // The second shuffle, which takes the first shuffle as both of its
6787 // vector operands, put the elements into the right order.
6788 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
6790 int Mask2[] = { -1, -1, -1, -1 };
6792 for (unsigned i = 0; i != 4; ++i)
6793 if (Locs[i].first != -1) {
6794 unsigned Idx = (i < 2) ? 0 : 4;
6795 Idx += Locs[i].first * 2 + Locs[i].second;
6799 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
6802 if (NumLo == 3 || NumHi == 3) {
6803 // Otherwise, we must have three elements from one vector, call it X, and
6804 // one element from the other, call it Y. First, use a shufps to build an
6805 // intermediate vector with the one element from Y and the element from X
6806 // that will be in the same half in the final destination (the indexes don't
6807 // matter). Then, use a shufps to build the final vector, taking the half
6808 // containing the element from Y from the intermediate, and the other half
6811 // Normalize it so the 3 elements come from V1.
6812 CommuteVectorShuffleMask(PermMask, 4);
6816 // Find the element from V2.
6818 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
6819 int Val = PermMask[HiIndex];
6826 Mask1[0] = PermMask[HiIndex];
6828 Mask1[2] = PermMask[HiIndex^1];
6830 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
6833 Mask1[0] = PermMask[0];
6834 Mask1[1] = PermMask[1];
6835 Mask1[2] = HiIndex & 1 ? 6 : 4;
6836 Mask1[3] = HiIndex & 1 ? 4 : 6;
6837 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
6840 Mask1[0] = HiIndex & 1 ? 2 : 0;
6841 Mask1[1] = HiIndex & 1 ? 0 : 2;
6842 Mask1[2] = PermMask[2];
6843 Mask1[3] = PermMask[3];
6848 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
6851 // Break it into (shuffle shuffle_hi, shuffle_lo).
6852 int LoMask[] = { -1, -1, -1, -1 };
6853 int HiMask[] = { -1, -1, -1, -1 };
6855 int *MaskPtr = LoMask;
6856 unsigned MaskIdx = 0;
6859 for (unsigned i = 0; i != 4; ++i) {
6866 int Idx = PermMask[i];
6868 Locs[i] = std::make_pair(-1, -1);
6869 } else if (Idx < 4) {
6870 Locs[i] = std::make_pair(MaskIdx, LoIdx);
6871 MaskPtr[LoIdx] = Idx;
6874 Locs[i] = std::make_pair(MaskIdx, HiIdx);
6875 MaskPtr[HiIdx] = Idx;
6880 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
6881 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
6882 int MaskOps[] = { -1, -1, -1, -1 };
6883 for (unsigned i = 0; i != 4; ++i)
6884 if (Locs[i].first != -1)
6885 MaskOps[i] = Locs[i].first * 4 + Locs[i].second;
6886 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
6889 static bool MayFoldVectorLoad(SDValue V) {
6890 while (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
6891 V = V.getOperand(0);
6893 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
6894 V = V.getOperand(0);
6895 if (V.hasOneUse() && V.getOpcode() == ISD::BUILD_VECTOR &&
6896 V.getNumOperands() == 2 && V.getOperand(1).getOpcode() == ISD::UNDEF)
6897 // BUILD_VECTOR (load), undef
6898 V = V.getOperand(0);
6900 return MayFoldLoad(V);
6904 SDValue getMOVDDup(SDValue &Op, SDLoc &dl, SDValue V1, SelectionDAG &DAG) {
6905 EVT VT = Op.getValueType();
6907 // Canonizalize to v2f64.
6908 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, V1);
6909 return DAG.getNode(ISD::BITCAST, dl, VT,
6910 getTargetShuffleNode(X86ISD::MOVDDUP, dl, MVT::v2f64,
6915 SDValue getMOVLowToHigh(SDValue &Op, SDLoc &dl, SelectionDAG &DAG,
6917 SDValue V1 = Op.getOperand(0);
6918 SDValue V2 = Op.getOperand(1);
6919 EVT VT = Op.getValueType();
6921 assert(VT != MVT::v2i64 && "unsupported shuffle type");
6923 if (HasSSE2 && VT == MVT::v2f64)
6924 return getTargetShuffleNode(X86ISD::MOVLHPD, dl, VT, V1, V2, DAG);
6926 // v4f32 or v4i32: canonizalized to v4f32 (which is legal for SSE1)
6927 return DAG.getNode(ISD::BITCAST, dl, VT,
6928 getTargetShuffleNode(X86ISD::MOVLHPS, dl, MVT::v4f32,
6929 DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, V1),
6930 DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, V2), DAG));
6934 SDValue getMOVHighToLow(SDValue &Op, SDLoc &dl, SelectionDAG &DAG) {
6935 SDValue V1 = Op.getOperand(0);
6936 SDValue V2 = Op.getOperand(1);
6937 EVT VT = Op.getValueType();
6939 assert((VT == MVT::v4i32 || VT == MVT::v4f32) &&
6940 "unsupported shuffle type");
6942 if (V2.getOpcode() == ISD::UNDEF)
6946 return getTargetShuffleNode(X86ISD::MOVHLPS, dl, VT, V1, V2, DAG);
6950 SDValue getMOVLP(SDValue &Op, SDLoc &dl, SelectionDAG &DAG, bool HasSSE2) {
6951 SDValue V1 = Op.getOperand(0);
6952 SDValue V2 = Op.getOperand(1);
6953 EVT VT = Op.getValueType();
6954 unsigned NumElems = VT.getVectorNumElements();
6956 // Use MOVLPS and MOVLPD in case V1 or V2 are loads. During isel, the second
6957 // operand of these instructions is only memory, so check if there's a
6958 // potencial load folding here, otherwise use SHUFPS or MOVSD to match the
6960 bool CanFoldLoad = false;
6962 // Trivial case, when V2 comes from a load.
6963 if (MayFoldVectorLoad(V2))
6966 // When V1 is a load, it can be folded later into a store in isel, example:
6967 // (store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)), addr:$src1)
6969 // (MOVLPSmr addr:$src1, VR128:$src2)
6970 // So, recognize this potential and also use MOVLPS or MOVLPD
6971 else if (MayFoldVectorLoad(V1) && MayFoldIntoStore(Op))
6974 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
6976 if (HasSSE2 && NumElems == 2)
6977 return getTargetShuffleNode(X86ISD::MOVLPD, dl, VT, V1, V2, DAG);
6980 // If we don't care about the second element, proceed to use movss.
6981 if (SVOp->getMaskElt(1) != -1)
6982 return getTargetShuffleNode(X86ISD::MOVLPS, dl, VT, V1, V2, DAG);
6985 // movl and movlp will both match v2i64, but v2i64 is never matched by
6986 // movl earlier because we make it strict to avoid messing with the movlp load
6987 // folding logic (see the code above getMOVLP call). Match it here then,
6988 // this is horrible, but will stay like this until we move all shuffle
6989 // matching to x86 specific nodes. Note that for the 1st condition all
6990 // types are matched with movsd.
6992 // FIXME: isMOVLMask should be checked and matched before getMOVLP,
6993 // as to remove this logic from here, as much as possible
6994 if (NumElems == 2 || !isMOVLMask(SVOp->getMask(), VT))
6995 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
6996 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
6999 assert(VT != MVT::v4i32 && "unsupported shuffle type");
7001 // Invert the operand order and use SHUFPS to match it.
7002 return getTargetShuffleNode(X86ISD::SHUFP, dl, VT, V2, V1,
7003 getShuffleSHUFImmediate(SVOp), DAG);
7006 // Reduce a vector shuffle to zext.
7008 X86TargetLowering::LowerVectorIntExtend(SDValue Op, SelectionDAG &DAG) const {
7009 // PMOVZX is only available from SSE41.
7010 if (!Subtarget->hasSSE41())
7013 EVT VT = Op.getValueType();
7015 // Only AVX2 support 256-bit vector integer extending.
7016 if (!Subtarget->hasInt256() && VT.is256BitVector())
7019 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7021 SDValue V1 = Op.getOperand(0);
7022 SDValue V2 = Op.getOperand(1);
7023 unsigned NumElems = VT.getVectorNumElements();
7025 // Extending is an unary operation and the element type of the source vector
7026 // won't be equal to or larger than i64.
7027 if (V2.getOpcode() != ISD::UNDEF || !VT.isInteger() ||
7028 VT.getVectorElementType() == MVT::i64)
7031 // Find the expansion ratio, e.g. expanding from i8 to i32 has a ratio of 4.
7032 unsigned Shift = 1; // Start from 2, i.e. 1 << 1.
7033 while ((1U << Shift) < NumElems) {
7034 if (SVOp->getMaskElt(1U << Shift) == 1)
7037 // The maximal ratio is 8, i.e. from i8 to i64.
7042 // Check the shuffle mask.
7043 unsigned Mask = (1U << Shift) - 1;
7044 for (unsigned i = 0; i != NumElems; ++i) {
7045 int EltIdx = SVOp->getMaskElt(i);
7046 if ((i & Mask) != 0 && EltIdx != -1)
7048 if ((i & Mask) == 0 && (unsigned)EltIdx != (i >> Shift))
7052 LLVMContext *Context = DAG.getContext();
7053 unsigned NBits = VT.getVectorElementType().getSizeInBits() << Shift;
7054 EVT NeVT = EVT::getIntegerVT(*Context, NBits);
7055 EVT NVT = EVT::getVectorVT(*Context, NeVT, NumElems >> Shift);
7057 if (!isTypeLegal(NVT))
7060 // Simplify the operand as it's prepared to be fed into shuffle.
7061 unsigned SignificantBits = NVT.getSizeInBits() >> Shift;
7062 if (V1.getOpcode() == ISD::BITCAST &&
7063 V1.getOperand(0).getOpcode() == ISD::SCALAR_TO_VECTOR &&
7064 V1.getOperand(0).getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
7066 .getOperand(0).getValueType().getSizeInBits() == SignificantBits) {
7067 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast x)
7068 SDValue V = V1.getOperand(0).getOperand(0).getOperand(0);
7069 ConstantSDNode *CIdx =
7070 dyn_cast<ConstantSDNode>(V1.getOperand(0).getOperand(0).getOperand(1));
7071 // If it's foldable, i.e. normal load with single use, we will let code
7072 // selection to fold it. Otherwise, we will short the conversion sequence.
7073 if (CIdx && CIdx->getZExtValue() == 0 &&
7074 (!ISD::isNormalLoad(V.getNode()) || !V.hasOneUse())) {
7075 if (V.getValueSizeInBits() > V1.getValueSizeInBits()) {
7076 // The "ext_vec_elt" node is wider than the result node.
7077 // In this case we should extract subvector from V.
7078 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast (extract_subvector x)).
7079 unsigned Ratio = V.getValueSizeInBits() / V1.getValueSizeInBits();
7080 EVT FullVT = V.getValueType();
7081 EVT SubVecVT = EVT::getVectorVT(*Context,
7082 FullVT.getVectorElementType(),
7083 FullVT.getVectorNumElements()/Ratio);
7084 V = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVecVT, V,
7085 DAG.getIntPtrConstant(0));
7087 V1 = DAG.getNode(ISD::BITCAST, DL, V1.getValueType(), V);
7091 return DAG.getNode(ISD::BITCAST, DL, VT,
7092 DAG.getNode(X86ISD::VZEXT, DL, NVT, V1));
7096 X86TargetLowering::NormalizeVectorShuffle(SDValue Op, SelectionDAG &DAG) const {
7097 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7098 MVT VT = Op.getValueType().getSimpleVT();
7100 SDValue V1 = Op.getOperand(0);
7101 SDValue V2 = Op.getOperand(1);
7103 if (isZeroShuffle(SVOp))
7104 return getZeroVector(VT, Subtarget, DAG, dl);
7106 // Handle splat operations
7107 if (SVOp->isSplat()) {
7108 // Use vbroadcast whenever the splat comes from a foldable load
7109 SDValue Broadcast = LowerVectorBroadcast(Op, DAG);
7110 if (Broadcast.getNode())
7114 // Check integer expanding shuffles.
7115 SDValue NewOp = LowerVectorIntExtend(Op, DAG);
7116 if (NewOp.getNode())
7119 // If the shuffle can be profitably rewritten as a narrower shuffle, then
7121 if (VT == MVT::v8i16 || VT == MVT::v16i8 ||
7122 VT == MVT::v16i16 || VT == MVT::v32i8) {
7123 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG);
7124 if (NewOp.getNode())
7125 return DAG.getNode(ISD::BITCAST, dl, VT, NewOp);
7126 } else if ((VT == MVT::v4i32 ||
7127 (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
7128 // FIXME: Figure out a cleaner way to do this.
7129 // Try to make use of movq to zero out the top part.
7130 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
7131 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG);
7132 if (NewOp.getNode()) {
7133 MVT NewVT = NewOp.getValueType().getSimpleVT();
7134 if (isCommutedMOVLMask(cast<ShuffleVectorSDNode>(NewOp)->getMask(),
7135 NewVT, true, false))
7136 return getVZextMovL(VT, NewVT, NewOp.getOperand(0),
7137 DAG, Subtarget, dl);
7139 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
7140 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG);
7141 if (NewOp.getNode()) {
7142 MVT NewVT = NewOp.getValueType().getSimpleVT();
7143 if (isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)->getMask(), NewVT))
7144 return getVZextMovL(VT, NewVT, NewOp.getOperand(1),
7145 DAG, Subtarget, dl);
7153 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
7154 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7155 SDValue V1 = Op.getOperand(0);
7156 SDValue V2 = Op.getOperand(1);
7157 MVT VT = Op.getValueType().getSimpleVT();
7159 unsigned NumElems = VT.getVectorNumElements();
7160 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
7161 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
7162 bool V1IsSplat = false;
7163 bool V2IsSplat = false;
7164 bool HasSSE2 = Subtarget->hasSSE2();
7165 bool HasFp256 = Subtarget->hasFp256();
7166 bool HasInt256 = Subtarget->hasInt256();
7167 MachineFunction &MF = DAG.getMachineFunction();
7168 bool OptForSize = MF.getFunction()->getAttributes().
7169 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize);
7171 assert(VT.getSizeInBits() != 64 && "Can't lower MMX shuffles");
7173 if (V1IsUndef && V2IsUndef)
7174 return DAG.getUNDEF(VT);
7176 assert(!V1IsUndef && "Op 1 of shuffle should not be undef");
7178 // Vector shuffle lowering takes 3 steps:
7180 // 1) Normalize the input vectors. Here splats, zeroed vectors, profitable
7181 // narrowing and commutation of operands should be handled.
7182 // 2) Matching of shuffles with known shuffle masks to x86 target specific
7184 // 3) Rewriting of unmatched masks into new generic shuffle operations,
7185 // so the shuffle can be broken into other shuffles and the legalizer can
7186 // try the lowering again.
7188 // The general idea is that no vector_shuffle operation should be left to
7189 // be matched during isel, all of them must be converted to a target specific
7192 // Normalize the input vectors. Here splats, zeroed vectors, profitable
7193 // narrowing and commutation of operands should be handled. The actual code
7194 // doesn't include all of those, work in progress...
7195 SDValue NewOp = NormalizeVectorShuffle(Op, DAG);
7196 if (NewOp.getNode())
7199 SmallVector<int, 8> M(SVOp->getMask().begin(), SVOp->getMask().end());
7201 // NOTE: isPSHUFDMask can also match both masks below (unpckl_undef and
7202 // unpckh_undef). Only use pshufd if speed is more important than size.
7203 if (OptForSize && isUNPCKL_v_undef_Mask(M, VT, HasInt256))
7204 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V1, DAG);
7205 if (OptForSize && isUNPCKH_v_undef_Mask(M, VT, HasInt256))
7206 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V1, DAG);
7208 if (isMOVDDUPMask(M, VT) && Subtarget->hasSSE3() &&
7209 V2IsUndef && MayFoldVectorLoad(V1))
7210 return getMOVDDup(Op, dl, V1, DAG);
7212 if (isMOVHLPS_v_undef_Mask(M, VT))
7213 return getMOVHighToLow(Op, dl, DAG);
7215 // Use to match splats
7216 if (HasSSE2 && isUNPCKHMask(M, VT, HasInt256) && V2IsUndef &&
7217 (VT == MVT::v2f64 || VT == MVT::v2i64))
7218 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V1, DAG);
7220 if (isPSHUFDMask(M, VT)) {
7221 // The actual implementation will match the mask in the if above and then
7222 // during isel it can match several different instructions, not only pshufd
7223 // as its name says, sad but true, emulate the behavior for now...
7224 if (isMOVDDUPMask(M, VT) && ((VT == MVT::v4f32 || VT == MVT::v2i64)))
7225 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V1, DAG);
7227 unsigned TargetMask = getShuffleSHUFImmediate(SVOp);
7229 if (HasSSE2 && (VT == MVT::v4f32 || VT == MVT::v4i32))
7230 return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1, TargetMask, DAG);
7232 if (HasFp256 && (VT == MVT::v4f32 || VT == MVT::v2f64))
7233 return getTargetShuffleNode(X86ISD::VPERMILP, dl, VT, V1, TargetMask,
7236 return getTargetShuffleNode(X86ISD::SHUFP, dl, VT, V1, V1,
7240 if (isPALIGNRMask(M, VT, Subtarget))
7241 return getTargetShuffleNode(X86ISD::PALIGNR, dl, VT, V1, V2,
7242 getShufflePALIGNRImmediate(SVOp),
7245 // Check if this can be converted into a logical shift.
7246 bool isLeft = false;
7249 bool isShift = HasSSE2 && isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
7250 if (isShift && ShVal.hasOneUse()) {
7251 // If the shifted value has multiple uses, it may be cheaper to use
7252 // v_set0 + movlhps or movhlps, etc.
7253 MVT EltVT = VT.getVectorElementType();
7254 ShAmt *= EltVT.getSizeInBits();
7255 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
7258 if (isMOVLMask(M, VT)) {
7259 if (ISD::isBuildVectorAllZeros(V1.getNode()))
7260 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
7261 if (!isMOVLPMask(M, VT)) {
7262 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
7263 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
7265 if (VT == MVT::v4i32 || VT == MVT::v4f32)
7266 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
7270 // FIXME: fold these into legal mask.
7271 if (isMOVLHPSMask(M, VT) && !isUNPCKLMask(M, VT, HasInt256))
7272 return getMOVLowToHigh(Op, dl, DAG, HasSSE2);
7274 if (isMOVHLPSMask(M, VT))
7275 return getMOVHighToLow(Op, dl, DAG);
7277 if (V2IsUndef && isMOVSHDUPMask(M, VT, Subtarget))
7278 return getTargetShuffleNode(X86ISD::MOVSHDUP, dl, VT, V1, DAG);
7280 if (V2IsUndef && isMOVSLDUPMask(M, VT, Subtarget))
7281 return getTargetShuffleNode(X86ISD::MOVSLDUP, dl, VT, V1, DAG);
7283 if (isMOVLPMask(M, VT))
7284 return getMOVLP(Op, dl, DAG, HasSSE2);
7286 if (ShouldXformToMOVHLPS(M, VT) ||
7287 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), M, VT))
7288 return CommuteVectorShuffle(SVOp, DAG);
7291 // No better options. Use a vshldq / vsrldq.
7292 MVT EltVT = VT.getVectorElementType();
7293 ShAmt *= EltVT.getSizeInBits();
7294 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
7297 bool Commuted = false;
7298 // FIXME: This should also accept a bitcast of a splat? Be careful, not
7299 // 1,1,1,1 -> v8i16 though.
7300 V1IsSplat = isSplatVector(V1.getNode());
7301 V2IsSplat = isSplatVector(V2.getNode());
7303 // Canonicalize the splat or undef, if present, to be on the RHS.
7304 if (!V2IsUndef && V1IsSplat && !V2IsSplat) {
7305 CommuteVectorShuffleMask(M, NumElems);
7307 std::swap(V1IsSplat, V2IsSplat);
7311 if (isCommutedMOVLMask(M, VT, V2IsSplat, V2IsUndef)) {
7312 // Shuffling low element of v1 into undef, just return v1.
7315 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
7316 // the instruction selector will not match, so get a canonical MOVL with
7317 // swapped operands to undo the commute.
7318 return getMOVL(DAG, dl, VT, V2, V1);
7321 if (isUNPCKLMask(M, VT, HasInt256))
7322 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V2, DAG);
7324 if (isUNPCKHMask(M, VT, HasInt256))
7325 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V2, DAG);
7328 // Normalize mask so all entries that point to V2 points to its first
7329 // element then try to match unpck{h|l} again. If match, return a
7330 // new vector_shuffle with the corrected mask.p
7331 SmallVector<int, 8> NewMask(M.begin(), M.end());
7332 NormalizeMask(NewMask, NumElems);
7333 if (isUNPCKLMask(NewMask, VT, HasInt256, true))
7334 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V2, DAG);
7335 if (isUNPCKHMask(NewMask, VT, HasInt256, true))
7336 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V2, DAG);
7340 // Commute is back and try unpck* again.
7341 // FIXME: this seems wrong.
7342 CommuteVectorShuffleMask(M, NumElems);
7344 std::swap(V1IsSplat, V2IsSplat);
7347 if (isUNPCKLMask(M, VT, HasInt256))
7348 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V2, DAG);
7350 if (isUNPCKHMask(M, VT, HasInt256))
7351 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V2, DAG);
7354 // Normalize the node to match x86 shuffle ops if needed
7355 if (!V2IsUndef && (isSHUFPMask(M, VT, HasFp256, /* Commuted */ true)))
7356 return CommuteVectorShuffle(SVOp, DAG);
7358 // The checks below are all present in isShuffleMaskLegal, but they are
7359 // inlined here right now to enable us to directly emit target specific
7360 // nodes, and remove one by one until they don't return Op anymore.
7362 if (ShuffleVectorSDNode::isSplatMask(&M[0], VT) &&
7363 SVOp->getSplatIndex() == 0 && V2IsUndef) {
7364 if (VT == MVT::v2f64 || VT == MVT::v2i64)
7365 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V1, DAG);
7368 if (isPSHUFHWMask(M, VT, HasInt256))
7369 return getTargetShuffleNode(X86ISD::PSHUFHW, dl, VT, V1,
7370 getShufflePSHUFHWImmediate(SVOp),
7373 if (isPSHUFLWMask(M, VT, HasInt256))
7374 return getTargetShuffleNode(X86ISD::PSHUFLW, dl, VT, V1,
7375 getShufflePSHUFLWImmediate(SVOp),
7378 if (isSHUFPMask(M, VT, HasFp256))
7379 return getTargetShuffleNode(X86ISD::SHUFP, dl, VT, V1, V2,
7380 getShuffleSHUFImmediate(SVOp), DAG);
7382 if (isUNPCKL_v_undef_Mask(M, VT, HasInt256))
7383 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V1, DAG);
7384 if (isUNPCKH_v_undef_Mask(M, VT, HasInt256))
7385 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V1, DAG);
7387 //===--------------------------------------------------------------------===//
7388 // Generate target specific nodes for 128 or 256-bit shuffles only
7389 // supported in the AVX instruction set.
7392 // Handle VMOVDDUPY permutations
7393 if (V2IsUndef && isMOVDDUPYMask(M, VT, HasFp256))
7394 return getTargetShuffleNode(X86ISD::MOVDDUP, dl, VT, V1, DAG);
7396 // Handle VPERMILPS/D* permutations
7397 if (isVPERMILPMask(M, VT, HasFp256)) {
7398 if (HasInt256 && VT == MVT::v8i32)
7399 return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1,
7400 getShuffleSHUFImmediate(SVOp), DAG);
7401 return getTargetShuffleNode(X86ISD::VPERMILP, dl, VT, V1,
7402 getShuffleSHUFImmediate(SVOp), DAG);
7405 // Handle VPERM2F128/VPERM2I128 permutations
7406 if (isVPERM2X128Mask(M, VT, HasFp256))
7407 return getTargetShuffleNode(X86ISD::VPERM2X128, dl, VT, V1,
7408 V2, getShuffleVPERM2X128Immediate(SVOp), DAG);
7410 SDValue BlendOp = LowerVECTOR_SHUFFLEtoBlend(SVOp, Subtarget, DAG);
7411 if (BlendOp.getNode())
7414 if (V2IsUndef && HasInt256 && (VT == MVT::v8i32 || VT == MVT::v8f32)) {
7415 SmallVector<SDValue, 8> permclMask;
7416 for (unsigned i = 0; i != 8; ++i) {
7417 permclMask.push_back(DAG.getConstant((M[i]>=0) ? M[i] : 0, MVT::i32));
7419 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32,
7421 // Bitcast is for VPERMPS since mask is v8i32 but node takes v8f32
7422 return DAG.getNode(X86ISD::VPERMV, dl, VT,
7423 DAG.getNode(ISD::BITCAST, dl, VT, Mask), V1);
7426 if (V2IsUndef && HasInt256 && (VT == MVT::v4i64 || VT == MVT::v4f64))
7427 return getTargetShuffleNode(X86ISD::VPERMI, dl, VT, V1,
7428 getShuffleCLImmediate(SVOp), DAG);
7430 //===--------------------------------------------------------------------===//
7431 // Since no target specific shuffle was selected for this generic one,
7432 // lower it into other known shuffles. FIXME: this isn't true yet, but
7433 // this is the plan.
7436 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
7437 if (VT == MVT::v8i16) {
7438 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(Op, Subtarget, DAG);
7439 if (NewOp.getNode())
7443 if (VT == MVT::v16i8) {
7444 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
7445 if (NewOp.getNode())
7449 if (VT == MVT::v32i8) {
7450 SDValue NewOp = LowerVECTOR_SHUFFLEv32i8(SVOp, Subtarget, DAG);
7451 if (NewOp.getNode())
7455 // Handle all 128-bit wide vectors with 4 elements, and match them with
7456 // several different shuffle types.
7457 if (NumElems == 4 && VT.is128BitVector())
7458 return LowerVECTOR_SHUFFLE_128v4(SVOp, DAG);
7460 // Handle general 256-bit shuffles
7461 if (VT.is256BitVector())
7462 return LowerVECTOR_SHUFFLE_256(SVOp, DAG);
7467 static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
7468 MVT VT = Op.getValueType().getSimpleVT();
7471 if (!Op.getOperand(0).getValueType().getSimpleVT().is128BitVector())
7474 if (VT.getSizeInBits() == 8) {
7475 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
7476 Op.getOperand(0), Op.getOperand(1));
7477 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
7478 DAG.getValueType(VT));
7479 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
7482 if (VT.getSizeInBits() == 16) {
7483 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
7484 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
7486 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
7487 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
7488 DAG.getNode(ISD::BITCAST, dl,
7492 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
7493 Op.getOperand(0), Op.getOperand(1));
7494 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
7495 DAG.getValueType(VT));
7496 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
7499 if (VT == MVT::f32) {
7500 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
7501 // the result back to FR32 register. It's only worth matching if the
7502 // result has a single use which is a store or a bitcast to i32. And in
7503 // the case of a store, it's not worth it if the index is a constant 0,
7504 // because a MOVSSmr can be used instead, which is smaller and faster.
7505 if (!Op.hasOneUse())
7507 SDNode *User = *Op.getNode()->use_begin();
7508 if ((User->getOpcode() != ISD::STORE ||
7509 (isa<ConstantSDNode>(Op.getOperand(1)) &&
7510 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
7511 (User->getOpcode() != ISD::BITCAST ||
7512 User->getValueType(0) != MVT::i32))
7514 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
7515 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32,
7518 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, Extract);
7521 if (VT == MVT::i32 || VT == MVT::i64) {
7522 // ExtractPS/pextrq works with constant index.
7523 if (isa<ConstantSDNode>(Op.getOperand(1)))
7530 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
7531 SelectionDAG &DAG) const {
7533 if (!isa<ConstantSDNode>(Op.getOperand(1)))
7536 SDValue Vec = Op.getOperand(0);
7537 MVT VecVT = Vec.getValueType().getSimpleVT();
7539 // If this is a 256-bit vector result, first extract the 128-bit vector and
7540 // then extract the element from the 128-bit vector.
7541 if (VecVT.is256BitVector() || VecVT.is512BitVector()) {
7542 SDValue Idx = Op.getOperand(1);
7543 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
7545 // Get the 128-bit vector.
7546 Vec = Extract128BitVector(Vec, IdxVal, DAG, dl);
7547 EVT EltVT = VecVT.getVectorElementType();
7549 unsigned ElemsPerChunk = 128 / EltVT.getSizeInBits();
7551 //if (IdxVal >= NumElems/2)
7552 // IdxVal -= NumElems/2;
7553 IdxVal -= (IdxVal/ElemsPerChunk)*ElemsPerChunk;
7554 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
7555 DAG.getConstant(IdxVal, MVT::i32));
7558 assert(VecVT.is128BitVector() && "Unexpected vector length");
7560 if (Subtarget->hasSSE41()) {
7561 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
7566 MVT VT = Op.getValueType().getSimpleVT();
7567 // TODO: handle v16i8.
7568 if (VT.getSizeInBits() == 16) {
7569 SDValue Vec = Op.getOperand(0);
7570 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
7572 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
7573 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
7574 DAG.getNode(ISD::BITCAST, dl,
7577 // Transform it so it match pextrw which produces a 32-bit result.
7578 MVT EltVT = MVT::i32;
7579 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
7580 Op.getOperand(0), Op.getOperand(1));
7581 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
7582 DAG.getValueType(VT));
7583 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
7586 if (VT.getSizeInBits() == 32) {
7587 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
7591 // SHUFPS the element to the lowest double word, then movss.
7592 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
7593 MVT VVT = Op.getOperand(0).getValueType().getSimpleVT();
7594 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
7595 DAG.getUNDEF(VVT), Mask);
7596 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
7597 DAG.getIntPtrConstant(0));
7600 if (VT.getSizeInBits() == 64) {
7601 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
7602 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
7603 // to match extract_elt for f64.
7604 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
7608 // UNPCKHPD the element to the lowest double word, then movsd.
7609 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
7610 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
7611 int Mask[2] = { 1, -1 };
7612 MVT VVT = Op.getOperand(0).getValueType().getSimpleVT();
7613 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
7614 DAG.getUNDEF(VVT), Mask);
7615 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
7616 DAG.getIntPtrConstant(0));
7622 static SDValue LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
7623 MVT VT = Op.getValueType().getSimpleVT();
7624 MVT EltVT = VT.getVectorElementType();
7627 SDValue N0 = Op.getOperand(0);
7628 SDValue N1 = Op.getOperand(1);
7629 SDValue N2 = Op.getOperand(2);
7631 if (!VT.is128BitVector())
7634 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
7635 isa<ConstantSDNode>(N2)) {
7637 if (VT == MVT::v8i16)
7638 Opc = X86ISD::PINSRW;
7639 else if (VT == MVT::v16i8)
7640 Opc = X86ISD::PINSRB;
7642 Opc = X86ISD::PINSRB;
7644 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
7646 if (N1.getValueType() != MVT::i32)
7647 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
7648 if (N2.getValueType() != MVT::i32)
7649 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
7650 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
7653 if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
7654 // Bits [7:6] of the constant are the source select. This will always be
7655 // zero here. The DAG Combiner may combine an extract_elt index into these
7656 // bits. For example (insert (extract, 3), 2) could be matched by putting
7657 // the '3' into bits [7:6] of X86ISD::INSERTPS.
7658 // Bits [5:4] of the constant are the destination select. This is the
7659 // value of the incoming immediate.
7660 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
7661 // combine either bitwise AND or insert of float 0.0 to set these bits.
7662 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
7663 // Create this as a scalar to vector..
7664 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
7665 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
7668 if ((EltVT == MVT::i32 || EltVT == MVT::i64) && isa<ConstantSDNode>(N2)) {
7669 // PINSR* works with constant index.
7676 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
7677 MVT VT = Op.getValueType().getSimpleVT();
7678 MVT EltVT = VT.getVectorElementType();
7681 SDValue N0 = Op.getOperand(0);
7682 SDValue N1 = Op.getOperand(1);
7683 SDValue N2 = Op.getOperand(2);
7685 // If this is a 256-bit vector result, first extract the 128-bit vector,
7686 // insert the element into the extracted half and then place it back.
7687 if (VT.is256BitVector() || VT.is512BitVector()) {
7688 if (!isa<ConstantSDNode>(N2))
7691 // Get the desired 128-bit vector half.
7692 unsigned IdxVal = cast<ConstantSDNode>(N2)->getZExtValue();
7693 SDValue V = Extract128BitVector(N0, IdxVal, DAG, dl);
7695 // Insert the element into the desired half.
7696 unsigned NumEltsIn128 = 128/EltVT.getSizeInBits();
7697 unsigned IdxIn128 = IdxVal - (IdxVal/NumEltsIn128) * NumEltsIn128;
7699 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1,
7700 DAG.getConstant(IdxIn128, MVT::i32));
7702 // Insert the changed part back to the 256-bit vector
7703 return Insert128BitVector(N0, V, IdxVal, DAG, dl);
7706 if (Subtarget->hasSSE41())
7707 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
7709 if (EltVT == MVT::i8)
7712 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
7713 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
7714 // as its second argument.
7715 if (N1.getValueType() != MVT::i32)
7716 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
7717 if (N2.getValueType() != MVT::i32)
7718 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
7719 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
7724 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
7725 LLVMContext *Context = DAG.getContext();
7727 MVT OpVT = Op.getValueType().getSimpleVT();
7729 // If this is a 256-bit vector result, first insert into a 128-bit
7730 // vector and then insert into the 256-bit vector.
7731 if (!OpVT.is128BitVector()) {
7732 // Insert into a 128-bit vector.
7733 unsigned SizeFactor = OpVT.getSizeInBits()/128;
7734 EVT VT128 = EVT::getVectorVT(*Context,
7735 OpVT.getVectorElementType(),
7736 OpVT.getVectorNumElements() / SizeFactor);
7738 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
7740 // Insert the 128-bit vector.
7741 return Insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl);
7744 if (OpVT == MVT::v1i64 &&
7745 Op.getOperand(0).getValueType() == MVT::i64)
7746 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
7748 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
7749 assert(OpVT.is128BitVector() && "Expected an SSE type!");
7750 return DAG.getNode(ISD::BITCAST, dl, OpVT,
7751 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,AnyExt));
7754 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
7755 // a simple subregister reference or explicit instructions to grab
7756 // upper bits of a vector.
7757 static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
7758 SelectionDAG &DAG) {
7760 SDValue In = Op.getOperand(0);
7761 SDValue Idx = Op.getOperand(1);
7762 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
7763 EVT ResVT = Op.getValueType();
7764 EVT InVT = In.getValueType();
7766 if (Subtarget->hasFp256()) {
7767 if (ResVT.is128BitVector() &&
7768 (InVT.is256BitVector() || InVT.is512BitVector()) &&
7769 isa<ConstantSDNode>(Idx)) {
7770 return Extract128BitVector(In, IdxVal, DAG, dl);
7772 if (ResVT.is256BitVector() && InVT.is512BitVector() &&
7773 isa<ConstantSDNode>(Idx)) {
7774 return Extract256BitVector(In, IdxVal, DAG, dl);
7780 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
7781 // simple superregister reference or explicit instructions to insert
7782 // the upper bits of a vector.
7783 static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
7784 SelectionDAG &DAG) {
7785 if (Subtarget->hasFp256()) {
7786 SDLoc dl(Op.getNode());
7787 SDValue Vec = Op.getNode()->getOperand(0);
7788 SDValue SubVec = Op.getNode()->getOperand(1);
7789 SDValue Idx = Op.getNode()->getOperand(2);
7791 if ((Op.getNode()->getValueType(0).is256BitVector() ||
7792 Op.getNode()->getValueType(0).is512BitVector()) &&
7793 SubVec.getNode()->getValueType(0).is128BitVector() &&
7794 isa<ConstantSDNode>(Idx)) {
7795 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
7796 return Insert128BitVector(Vec, SubVec, IdxVal, DAG, dl);
7799 if (Op.getNode()->getValueType(0).is512BitVector() &&
7800 SubVec.getNode()->getValueType(0).is256BitVector() &&
7801 isa<ConstantSDNode>(Idx)) {
7802 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
7803 return Insert256BitVector(Vec, SubVec, IdxVal, DAG, dl);
7809 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
7810 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
7811 // one of the above mentioned nodes. It has to be wrapped because otherwise
7812 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
7813 // be used to form addressing mode. These wrapped nodes will be selected
7816 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
7817 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
7819 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
7821 unsigned char OpFlag = 0;
7822 unsigned WrapperKind = X86ISD::Wrapper;
7823 CodeModel::Model M = getTargetMachine().getCodeModel();
7825 if (Subtarget->isPICStyleRIPRel() &&
7826 (M == CodeModel::Small || M == CodeModel::Kernel))
7827 WrapperKind = X86ISD::WrapperRIP;
7828 else if (Subtarget->isPICStyleGOT())
7829 OpFlag = X86II::MO_GOTOFF;
7830 else if (Subtarget->isPICStyleStubPIC())
7831 OpFlag = X86II::MO_PIC_BASE_OFFSET;
7833 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
7835 CP->getOffset(), OpFlag);
7837 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
7838 // With PIC, the address is actually $g + Offset.
7840 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
7841 DAG.getNode(X86ISD::GlobalBaseReg,
7842 SDLoc(), getPointerTy()),
7849 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
7850 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
7852 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
7854 unsigned char OpFlag = 0;
7855 unsigned WrapperKind = X86ISD::Wrapper;
7856 CodeModel::Model M = getTargetMachine().getCodeModel();
7858 if (Subtarget->isPICStyleRIPRel() &&
7859 (M == CodeModel::Small || M == CodeModel::Kernel))
7860 WrapperKind = X86ISD::WrapperRIP;
7861 else if (Subtarget->isPICStyleGOT())
7862 OpFlag = X86II::MO_GOTOFF;
7863 else if (Subtarget->isPICStyleStubPIC())
7864 OpFlag = X86II::MO_PIC_BASE_OFFSET;
7866 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
7869 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
7871 // With PIC, the address is actually $g + Offset.
7873 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
7874 DAG.getNode(X86ISD::GlobalBaseReg,
7875 SDLoc(), getPointerTy()),
7882 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
7883 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
7885 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
7887 unsigned char OpFlag = 0;
7888 unsigned WrapperKind = X86ISD::Wrapper;
7889 CodeModel::Model M = getTargetMachine().getCodeModel();
7891 if (Subtarget->isPICStyleRIPRel() &&
7892 (M == CodeModel::Small || M == CodeModel::Kernel)) {
7893 if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF())
7894 OpFlag = X86II::MO_GOTPCREL;
7895 WrapperKind = X86ISD::WrapperRIP;
7896 } else if (Subtarget->isPICStyleGOT()) {
7897 OpFlag = X86II::MO_GOT;
7898 } else if (Subtarget->isPICStyleStubPIC()) {
7899 OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE;
7900 } else if (Subtarget->isPICStyleStubNoDynamic()) {
7901 OpFlag = X86II::MO_DARWIN_NONLAZY;
7904 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
7907 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
7909 // With PIC, the address is actually $g + Offset.
7910 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
7911 !Subtarget->is64Bit()) {
7912 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
7913 DAG.getNode(X86ISD::GlobalBaseReg,
7914 SDLoc(), getPointerTy()),
7918 // For symbols that require a load from a stub to get the address, emit the
7920 if (isGlobalStubReference(OpFlag))
7921 Result = DAG.getLoad(getPointerTy(), DL, DAG.getEntryNode(), Result,
7922 MachinePointerInfo::getGOT(), false, false, false, 0);
7928 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
7929 // Create the TargetBlockAddressAddress node.
7930 unsigned char OpFlags =
7931 Subtarget->ClassifyBlockAddressReference();
7932 CodeModel::Model M = getTargetMachine().getCodeModel();
7933 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
7934 int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset();
7936 SDValue Result = DAG.getTargetBlockAddress(BA, getPointerTy(), Offset,
7939 if (Subtarget->isPICStyleRIPRel() &&
7940 (M == CodeModel::Small || M == CodeModel::Kernel))
7941 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
7943 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
7945 // With PIC, the address is actually $g + Offset.
7946 if (isGlobalRelativeToPICBase(OpFlags)) {
7947 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
7948 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
7956 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
7957 int64_t Offset, SelectionDAG &DAG) const {
7958 // Create the TargetGlobalAddress node, folding in the constant
7959 // offset if it is legal.
7960 unsigned char OpFlags =
7961 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
7962 CodeModel::Model M = getTargetMachine().getCodeModel();
7964 if (OpFlags == X86II::MO_NO_FLAG &&
7965 X86::isOffsetSuitableForCodeModel(Offset, M)) {
7966 // A direct static reference to a global.
7967 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
7970 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
7973 if (Subtarget->isPICStyleRIPRel() &&
7974 (M == CodeModel::Small || M == CodeModel::Kernel))
7975 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
7977 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
7979 // With PIC, the address is actually $g + Offset.
7980 if (isGlobalRelativeToPICBase(OpFlags)) {
7981 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
7982 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
7986 // For globals that require a load from a stub to get the address, emit the
7988 if (isGlobalStubReference(OpFlags))
7989 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
7990 MachinePointerInfo::getGOT(), false, false, false, 0);
7992 // If there was a non-zero offset that we didn't fold, create an explicit
7995 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
7996 DAG.getConstant(Offset, getPointerTy()));
8002 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
8003 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
8004 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
8005 return LowerGlobalAddress(GV, SDLoc(Op), Offset, DAG);
8009 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
8010 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
8011 unsigned char OperandFlags, bool LocalDynamic = false) {
8012 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8013 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
8015 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
8016 GA->getValueType(0),
8020 X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR
8024 SDValue Ops[] = { Chain, TGA, *InFlag };
8025 Chain = DAG.getNode(CallType, dl, NodeTys, Ops, array_lengthof(Ops));
8027 SDValue Ops[] = { Chain, TGA };
8028 Chain = DAG.getNode(CallType, dl, NodeTys, Ops, array_lengthof(Ops));
8031 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
8032 MFI->setAdjustsStack(true);
8034 SDValue Flag = Chain.getValue(1);
8035 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
8038 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
8040 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
8043 SDLoc dl(GA); // ? function entry point might be better
8044 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
8045 DAG.getNode(X86ISD::GlobalBaseReg,
8046 SDLoc(), PtrVT), InFlag);
8047 InFlag = Chain.getValue(1);
8049 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
8052 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
8054 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
8056 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
8057 X86::RAX, X86II::MO_TLSGD);
8060 static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA,
8066 // Get the start address of the TLS block for this module.
8067 X86MachineFunctionInfo* MFI = DAG.getMachineFunction()
8068 .getInfo<X86MachineFunctionInfo>();
8069 MFI->incNumLocalDynamicTLSAccesses();
8073 Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT, X86::RAX,
8074 X86II::MO_TLSLD, /*LocalDynamic=*/true);
8077 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
8078 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), InFlag);
8079 InFlag = Chain.getValue(1);
8080 Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX,
8081 X86II::MO_TLSLDM, /*LocalDynamic=*/true);
8084 // Note: the CleanupLocalDynamicTLSPass will remove redundant computations
8088 unsigned char OperandFlags = X86II::MO_DTPOFF;
8089 unsigned WrapperKind = X86ISD::Wrapper;
8090 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
8091 GA->getValueType(0),
8092 GA->getOffset(), OperandFlags);
8093 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
8095 // Add x@dtpoff with the base.
8096 return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base);
8099 // Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model.
8100 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
8101 const EVT PtrVT, TLSModel::Model model,
8102 bool is64Bit, bool isPIC) {
8105 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
8106 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
8107 is64Bit ? 257 : 256));
8109 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
8110 DAG.getIntPtrConstant(0),
8111 MachinePointerInfo(Ptr),
8112 false, false, false, 0);
8114 unsigned char OperandFlags = 0;
8115 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
8117 unsigned WrapperKind = X86ISD::Wrapper;
8118 if (model == TLSModel::LocalExec) {
8119 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
8120 } else if (model == TLSModel::InitialExec) {
8122 OperandFlags = X86II::MO_GOTTPOFF;
8123 WrapperKind = X86ISD::WrapperRIP;
8125 OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF;
8128 llvm_unreachable("Unexpected model");
8131 // emit "addl x@ntpoff,%eax" (local exec)
8132 // or "addl x@indntpoff,%eax" (initial exec)
8133 // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic)
8134 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
8135 GA->getValueType(0),
8136 GA->getOffset(), OperandFlags);
8137 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
8139 if (model == TLSModel::InitialExec) {
8140 if (isPIC && !is64Bit) {
8141 Offset = DAG.getNode(ISD::ADD, dl, PtrVT,
8142 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
8146 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
8147 MachinePointerInfo::getGOT(), false, false, false,
8151 // The address of the thread local variable is the add of the thread
8152 // pointer with the offset of the variable.
8153 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
8157 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
8159 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
8160 const GlobalValue *GV = GA->getGlobal();
8162 if (Subtarget->isTargetELF()) {
8163 TLSModel::Model model = getTargetMachine().getTLSModel(GV);
8166 case TLSModel::GeneralDynamic:
8167 if (Subtarget->is64Bit())
8168 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
8169 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
8170 case TLSModel::LocalDynamic:
8171 return LowerToTLSLocalDynamicModel(GA, DAG, getPointerTy(),
8172 Subtarget->is64Bit());
8173 case TLSModel::InitialExec:
8174 case TLSModel::LocalExec:
8175 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
8176 Subtarget->is64Bit(),
8177 getTargetMachine().getRelocationModel() == Reloc::PIC_);
8179 llvm_unreachable("Unknown TLS model.");
8182 if (Subtarget->isTargetDarwin()) {
8183 // Darwin only has one model of TLS. Lower to that.
8184 unsigned char OpFlag = 0;
8185 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
8186 X86ISD::WrapperRIP : X86ISD::Wrapper;
8188 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
8190 bool PIC32 = (getTargetMachine().getRelocationModel() == Reloc::PIC_) &&
8191 !Subtarget->is64Bit();
8193 OpFlag = X86II::MO_TLVP_PIC_BASE;
8195 OpFlag = X86II::MO_TLVP;
8197 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
8198 GA->getValueType(0),
8199 GA->getOffset(), OpFlag);
8200 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
8202 // With PIC32, the address is actually $g + Offset.
8204 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
8205 DAG.getNode(X86ISD::GlobalBaseReg,
8206 SDLoc(), getPointerTy()),
8209 // Lowering the machine isd will make sure everything is in the right
8211 SDValue Chain = DAG.getEntryNode();
8212 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
8213 SDValue Args[] = { Chain, Offset };
8214 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args, 2);
8216 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
8217 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8218 MFI->setAdjustsStack(true);
8220 // And our return value (tls address) is in the standard call return value
8222 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
8223 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy(),
8227 if (Subtarget->isTargetWindows() || Subtarget->isTargetMingw()) {
8228 // Just use the implicit TLS architecture
8229 // Need to generate someting similar to:
8230 // mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage
8232 // mov ecx, dword [rel _tls_index]: Load index (from C runtime)
8233 // mov rcx, qword [rdx+rcx*8]
8234 // mov eax, .tls$:tlsvar
8235 // [rax+rcx] contains the address
8236 // Windows 64bit: gs:0x58
8237 // Windows 32bit: fs:__tls_array
8239 // If GV is an alias then use the aliasee for determining
8240 // thread-localness.
8241 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
8242 GV = GA->resolveAliasedGlobal(false);
8244 SDValue Chain = DAG.getEntryNode();
8246 // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or
8247 // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly
8248 // use its literal value of 0x2C.
8249 Value *Ptr = Constant::getNullValue(Subtarget->is64Bit()
8250 ? Type::getInt8PtrTy(*DAG.getContext(),
8252 : Type::getInt32PtrTy(*DAG.getContext(),
8255 SDValue TlsArray = Subtarget->is64Bit() ? DAG.getIntPtrConstant(0x58) :
8256 (Subtarget->isTargetMingw() ? DAG.getIntPtrConstant(0x2C) :
8257 DAG.getExternalSymbol("_tls_array", getPointerTy()));
8259 SDValue ThreadPointer = DAG.getLoad(getPointerTy(), dl, Chain, TlsArray,
8260 MachinePointerInfo(Ptr),
8261 false, false, false, 0);
8263 // Load the _tls_index variable
8264 SDValue IDX = DAG.getExternalSymbol("_tls_index", getPointerTy());
8265 if (Subtarget->is64Bit())
8266 IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, getPointerTy(), Chain,
8267 IDX, MachinePointerInfo(), MVT::i32,
8270 IDX = DAG.getLoad(getPointerTy(), dl, Chain, IDX, MachinePointerInfo(),
8271 false, false, false, 0);
8273 SDValue Scale = DAG.getConstant(Log2_64_Ceil(TD->getPointerSize()),
8275 IDX = DAG.getNode(ISD::SHL, dl, getPointerTy(), IDX, Scale);
8277 SDValue res = DAG.getNode(ISD::ADD, dl, getPointerTy(), ThreadPointer, IDX);
8278 res = DAG.getLoad(getPointerTy(), dl, Chain, res, MachinePointerInfo(),
8279 false, false, false, 0);
8281 // Get the offset of start of .tls section
8282 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
8283 GA->getValueType(0),
8284 GA->getOffset(), X86II::MO_SECREL);
8285 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), TGA);
8287 // The address of the thread local variable is the add of the thread
8288 // pointer with the offset of the variable.
8289 return DAG.getNode(ISD::ADD, dl, getPointerTy(), res, Offset);
8292 llvm_unreachable("TLS not implemented for this target.");
8295 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values
8296 /// and take a 2 x i32 value to shift plus a shift amount.
8297 SDValue X86TargetLowering::LowerShiftParts(SDValue Op, SelectionDAG &DAG) const{
8298 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
8299 EVT VT = Op.getValueType();
8300 unsigned VTBits = VT.getSizeInBits();
8302 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
8303 SDValue ShOpLo = Op.getOperand(0);
8304 SDValue ShOpHi = Op.getOperand(1);
8305 SDValue ShAmt = Op.getOperand(2);
8306 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
8307 DAG.getConstant(VTBits - 1, MVT::i8))
8308 : DAG.getConstant(0, VT);
8311 if (Op.getOpcode() == ISD::SHL_PARTS) {
8312 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
8313 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
8315 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
8316 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
8319 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
8320 DAG.getConstant(VTBits, MVT::i8));
8321 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
8322 AndNode, DAG.getConstant(0, MVT::i8));
8325 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
8326 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
8327 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
8329 if (Op.getOpcode() == ISD::SHL_PARTS) {
8330 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
8331 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
8333 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
8334 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
8337 SDValue Ops[2] = { Lo, Hi };
8338 return DAG.getMergeValues(Ops, array_lengthof(Ops), dl);
8341 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
8342 SelectionDAG &DAG) const {
8343 EVT SrcVT = Op.getOperand(0).getValueType();
8345 if (SrcVT.isVector())
8348 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
8349 "Unknown SINT_TO_FP to lower!");
8351 // These are really Legal; return the operand so the caller accepts it as
8353 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
8355 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
8356 Subtarget->is64Bit()) {
8361 unsigned Size = SrcVT.getSizeInBits()/8;
8362 MachineFunction &MF = DAG.getMachineFunction();
8363 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
8364 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
8365 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
8367 MachinePointerInfo::getFixedStack(SSFI),
8369 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
8372 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
8374 SelectionDAG &DAG) const {
8378 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
8380 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
8382 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
8384 unsigned ByteSize = SrcVT.getSizeInBits()/8;
8386 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
8387 MachineMemOperand *MMO;
8389 int SSFI = FI->getIndex();
8391 DAG.getMachineFunction()
8392 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
8393 MachineMemOperand::MOLoad, ByteSize, ByteSize);
8395 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
8396 StackSlot = StackSlot.getOperand(1);
8398 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
8399 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
8401 Tys, Ops, array_lengthof(Ops),
8405 Chain = Result.getValue(1);
8406 SDValue InFlag = Result.getValue(2);
8408 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
8409 // shouldn't be necessary except that RFP cannot be live across
8410 // multiple blocks. When stackifier is fixed, they can be uncoupled.
8411 MachineFunction &MF = DAG.getMachineFunction();
8412 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
8413 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
8414 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
8415 Tys = DAG.getVTList(MVT::Other);
8417 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
8419 MachineMemOperand *MMO =
8420 DAG.getMachineFunction()
8421 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
8422 MachineMemOperand::MOStore, SSFISize, SSFISize);
8424 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
8425 Ops, array_lengthof(Ops),
8426 Op.getValueType(), MMO);
8427 Result = DAG.getLoad(Op.getValueType(), DL, Chain, StackSlot,
8428 MachinePointerInfo::getFixedStack(SSFI),
8429 false, false, false, 0);
8435 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
8436 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
8437 SelectionDAG &DAG) const {
8438 // This algorithm is not obvious. Here it is what we're trying to output:
8441 punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U }
8442 subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 }
8446 pshufd $0x4e, %xmm0, %xmm1
8452 LLVMContext *Context = DAG.getContext();
8454 // Build some magic constants.
8455 static const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 };
8456 Constant *C0 = ConstantDataVector::get(*Context, CV0);
8457 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
8459 SmallVector<Constant*,2> CV1;
8461 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
8462 APInt(64, 0x4330000000000000ULL))));
8464 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
8465 APInt(64, 0x4530000000000000ULL))));
8466 Constant *C1 = ConstantVector::get(CV1);
8467 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
8469 // Load the 64-bit value into an XMM register.
8470 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
8472 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
8473 MachinePointerInfo::getConstantPool(),
8474 false, false, false, 16);
8475 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32,
8476 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, XR1),
8479 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
8480 MachinePointerInfo::getConstantPool(),
8481 false, false, false, 16);
8482 SDValue XR2F = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Unpck1);
8483 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
8486 if (Subtarget->hasSSE3()) {
8487 // FIXME: The 'haddpd' instruction may be slower than 'movhlps + addsd'.
8488 Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub);
8490 SDValue S2F = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Sub);
8491 SDValue Shuffle = getTargetShuffleNode(X86ISD::PSHUFD, dl, MVT::v4i32,
8493 Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64,
8494 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Shuffle),
8498 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result,
8499 DAG.getIntPtrConstant(0));
8502 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
8503 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
8504 SelectionDAG &DAG) const {
8506 // FP constant to bias correct the final result.
8507 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
8510 // Load the 32-bit value into an XMM register.
8511 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
8514 // Zero out the upper parts of the register.
8515 Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG);
8517 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
8518 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Load),
8519 DAG.getIntPtrConstant(0));
8521 // Or the load with the bias.
8522 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
8523 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
8524 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
8526 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
8527 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
8528 MVT::v2f64, Bias)));
8529 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
8530 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or),
8531 DAG.getIntPtrConstant(0));
8533 // Subtract the bias.
8534 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
8536 // Handle final rounding.
8537 EVT DestVT = Op.getValueType();
8539 if (DestVT.bitsLT(MVT::f64))
8540 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
8541 DAG.getIntPtrConstant(0));
8542 if (DestVT.bitsGT(MVT::f64))
8543 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
8545 // Handle final rounding.
8549 SDValue X86TargetLowering::lowerUINT_TO_FP_vec(SDValue Op,
8550 SelectionDAG &DAG) const {
8551 SDValue N0 = Op.getOperand(0);
8552 EVT SVT = N0.getValueType();
8555 assert((SVT == MVT::v4i8 || SVT == MVT::v4i16 ||
8556 SVT == MVT::v8i8 || SVT == MVT::v8i16) &&
8557 "Custom UINT_TO_FP is not supported!");
8559 EVT NVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
8560 SVT.getVectorNumElements());
8561 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
8562 DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0));
8565 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
8566 SelectionDAG &DAG) const {
8567 SDValue N0 = Op.getOperand(0);
8570 if (Op.getValueType().isVector())
8571 return lowerUINT_TO_FP_vec(Op, DAG);
8573 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
8574 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
8575 // the optimization here.
8576 if (DAG.SignBitIsZero(N0))
8577 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
8579 EVT SrcVT = N0.getValueType();
8580 EVT DstVT = Op.getValueType();
8581 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
8582 return LowerUINT_TO_FP_i64(Op, DAG);
8583 if (SrcVT == MVT::i32 && X86ScalarSSEf64)
8584 return LowerUINT_TO_FP_i32(Op, DAG);
8585 if (Subtarget->is64Bit() && SrcVT == MVT::i64 && DstVT == MVT::f32)
8588 // Make a 64-bit buffer, and use it to build an FILD.
8589 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
8590 if (SrcVT == MVT::i32) {
8591 SDValue WordOff = DAG.getConstant(4, getPointerTy());
8592 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
8593 getPointerTy(), StackSlot, WordOff);
8594 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
8595 StackSlot, MachinePointerInfo(),
8597 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
8598 OffsetSlot, MachinePointerInfo(),
8600 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
8604 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
8605 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
8606 StackSlot, MachinePointerInfo(),
8608 // For i64 source, we need to add the appropriate power of 2 if the input
8609 // was negative. This is the same as the optimization in
8610 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
8611 // we must be careful to do the computation in x87 extended precision, not
8612 // in SSE. (The generic code can't know it's OK to do this, or how to.)
8613 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
8614 MachineMemOperand *MMO =
8615 DAG.getMachineFunction()
8616 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
8617 MachineMemOperand::MOLoad, 8, 8);
8619 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
8620 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
8621 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops,
8622 array_lengthof(Ops), MVT::i64, MMO);
8624 APInt FF(32, 0x5F800000ULL);
8626 // Check whether the sign bit is set.
8627 SDValue SignSet = DAG.getSetCC(dl,
8628 getSetCCResultType(*DAG.getContext(), MVT::i64),
8629 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
8632 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
8633 SDValue FudgePtr = DAG.getConstantPool(
8634 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
8637 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
8638 SDValue Zero = DAG.getIntPtrConstant(0);
8639 SDValue Four = DAG.getIntPtrConstant(4);
8640 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
8642 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
8644 // Load the value out, extending it from f32 to f80.
8645 // FIXME: Avoid the extend by constructing the right constant pool?
8646 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
8647 FudgePtr, MachinePointerInfo::getConstantPool(),
8648 MVT::f32, false, false, 4);
8649 // Extend everything to 80 bits to force it to be done on x87.
8650 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
8651 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
8654 std::pair<SDValue,SDValue>
8655 X86TargetLowering:: FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
8656 bool IsSigned, bool IsReplace) const {
8659 EVT DstTy = Op.getValueType();
8661 if (!IsSigned && !isIntegerTypeFTOL(DstTy)) {
8662 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
8666 assert(DstTy.getSimpleVT() <= MVT::i64 &&
8667 DstTy.getSimpleVT() >= MVT::i16 &&
8668 "Unknown FP_TO_INT to lower!");
8670 // These are really Legal.
8671 if (DstTy == MVT::i32 &&
8672 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
8673 return std::make_pair(SDValue(), SDValue());
8674 if (Subtarget->is64Bit() &&
8675 DstTy == MVT::i64 &&
8676 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
8677 return std::make_pair(SDValue(), SDValue());
8679 // We lower FP->int64 either into FISTP64 followed by a load from a temporary
8680 // stack slot, or into the FTOL runtime function.
8681 MachineFunction &MF = DAG.getMachineFunction();
8682 unsigned MemSize = DstTy.getSizeInBits()/8;
8683 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
8684 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
8687 if (!IsSigned && isIntegerTypeFTOL(DstTy))
8688 Opc = X86ISD::WIN_FTOL;
8690 switch (DstTy.getSimpleVT().SimpleTy) {
8691 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
8692 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
8693 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
8694 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
8697 SDValue Chain = DAG.getEntryNode();
8698 SDValue Value = Op.getOperand(0);
8699 EVT TheVT = Op.getOperand(0).getValueType();
8700 // FIXME This causes a redundant load/store if the SSE-class value is already
8701 // in memory, such as if it is on the callstack.
8702 if (isScalarFPTypeInSSEReg(TheVT)) {
8703 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
8704 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
8705 MachinePointerInfo::getFixedStack(SSFI),
8707 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
8709 Chain, StackSlot, DAG.getValueType(TheVT)
8712 MachineMemOperand *MMO =
8713 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
8714 MachineMemOperand::MOLoad, MemSize, MemSize);
8715 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops,
8716 array_lengthof(Ops), DstTy, MMO);
8717 Chain = Value.getValue(1);
8718 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
8719 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
8722 MachineMemOperand *MMO =
8723 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
8724 MachineMemOperand::MOStore, MemSize, MemSize);
8726 if (Opc != X86ISD::WIN_FTOL) {
8727 // Build the FP_TO_INT*_IN_MEM
8728 SDValue Ops[] = { Chain, Value, StackSlot };
8729 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
8730 Ops, array_lengthof(Ops), DstTy,
8732 return std::make_pair(FIST, StackSlot);
8734 SDValue ftol = DAG.getNode(X86ISD::WIN_FTOL, DL,
8735 DAG.getVTList(MVT::Other, MVT::Glue),
8737 SDValue eax = DAG.getCopyFromReg(ftol, DL, X86::EAX,
8738 MVT::i32, ftol.getValue(1));
8739 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), DL, X86::EDX,
8740 MVT::i32, eax.getValue(2));
8741 SDValue Ops[] = { eax, edx };
8742 SDValue pair = IsReplace
8743 ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops, array_lengthof(Ops))
8744 : DAG.getMergeValues(Ops, array_lengthof(Ops), DL);
8745 return std::make_pair(pair, SDValue());
8749 static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
8750 const X86Subtarget *Subtarget) {
8751 MVT VT = Op->getValueType(0).getSimpleVT();
8752 SDValue In = Op->getOperand(0);
8753 MVT InVT = In.getValueType().getSimpleVT();
8756 // Optimize vectors in AVX mode:
8759 // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32.
8760 // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32.
8761 // Concat upper and lower parts.
8764 // Use vpunpckldq for 4 lower elements v4i32 -> v2i64.
8765 // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64.
8766 // Concat upper and lower parts.
8769 if (((VT != MVT::v8i32) || (InVT != MVT::v8i16)) &&
8770 ((VT != MVT::v4i64) || (InVT != MVT::v4i32)))
8773 if (Subtarget->hasInt256())
8774 return DAG.getNode(X86ISD::VZEXT_MOVL, dl, VT, In);
8776 SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl);
8777 SDValue Undef = DAG.getUNDEF(InVT);
8778 bool NeedZero = Op.getOpcode() == ISD::ZERO_EXTEND;
8779 SDValue OpLo = getUnpackl(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
8780 SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
8782 MVT HVT = MVT::getVectorVT(VT.getVectorElementType(),
8783 VT.getVectorNumElements()/2);
8785 OpLo = DAG.getNode(ISD::BITCAST, dl, HVT, OpLo);
8786 OpHi = DAG.getNode(ISD::BITCAST, dl, HVT, OpHi);
8788 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
8791 SDValue X86TargetLowering::LowerANY_EXTEND(SDValue Op,
8792 SelectionDAG &DAG) const {
8793 if (Subtarget->hasFp256()) {
8794 SDValue Res = LowerAVXExtend(Op, DAG, Subtarget);
8801 SDValue X86TargetLowering::LowerZERO_EXTEND(SDValue Op,
8802 SelectionDAG &DAG) const {
8804 MVT VT = Op.getValueType().getSimpleVT();
8805 SDValue In = Op.getOperand(0);
8806 MVT SVT = In.getValueType().getSimpleVT();
8808 if (Subtarget->hasFp256()) {
8809 SDValue Res = LowerAVXExtend(Op, DAG, Subtarget);
8814 if (!VT.is256BitVector() || !SVT.is128BitVector() ||
8815 VT.getVectorNumElements() != SVT.getVectorNumElements())
8818 assert(Subtarget->hasFp256() && "256-bit vector is observed without AVX!");
8820 // AVX2 has better support of integer extending.
8821 if (Subtarget->hasInt256())
8822 return DAG.getNode(X86ISD::VZEXT, DL, VT, In);
8824 SDValue Lo = DAG.getNode(X86ISD::VZEXT, DL, MVT::v4i32, In);
8825 static const int Mask[] = {4, 5, 6, 7, -1, -1, -1, -1};
8826 SDValue Hi = DAG.getNode(X86ISD::VZEXT, DL, MVT::v4i32,
8827 DAG.getVectorShuffle(MVT::v8i16, DL, In,
8828 DAG.getUNDEF(MVT::v8i16),
8831 return DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v8i32, Lo, Hi);
8834 SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
8836 MVT VT = Op.getValueType().getSimpleVT();
8837 SDValue In = Op.getOperand(0);
8838 MVT SVT = In.getValueType().getSimpleVT();
8840 if ((VT == MVT::v4i32) && (SVT == MVT::v4i64)) {
8841 // On AVX2, v4i64 -> v4i32 becomes VPERMD.
8842 if (Subtarget->hasInt256()) {
8843 static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1};
8844 In = DAG.getNode(ISD::BITCAST, DL, MVT::v8i32, In);
8845 In = DAG.getVectorShuffle(MVT::v8i32, DL, In, DAG.getUNDEF(MVT::v8i32),
8847 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, In,
8848 DAG.getIntPtrConstant(0));
8851 // On AVX, v4i64 -> v4i32 becomes a sequence that uses PSHUFD and MOVLHPS.
8852 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
8853 DAG.getIntPtrConstant(0));
8854 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
8855 DAG.getIntPtrConstant(2));
8857 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpLo);
8858 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpHi);
8861 static const int ShufMask1[] = {0, 2, 0, 0};
8862 SDValue Undef = DAG.getUNDEF(VT);
8863 OpLo = DAG.getVectorShuffle(VT, DL, OpLo, Undef, ShufMask1);
8864 OpHi = DAG.getVectorShuffle(VT, DL, OpHi, Undef, ShufMask1);
8866 // The MOVLHPS mask:
8867 static const int ShufMask2[] = {0, 1, 4, 5};
8868 return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask2);
8871 if ((VT == MVT::v8i16) && (SVT == MVT::v8i32)) {
8872 // On AVX2, v8i32 -> v8i16 becomed PSHUFB.
8873 if (Subtarget->hasInt256()) {
8874 In = DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, In);
8876 SmallVector<SDValue,32> pshufbMask;
8877 for (unsigned i = 0; i < 2; ++i) {
8878 pshufbMask.push_back(DAG.getConstant(0x0, MVT::i8));
8879 pshufbMask.push_back(DAG.getConstant(0x1, MVT::i8));
8880 pshufbMask.push_back(DAG.getConstant(0x4, MVT::i8));
8881 pshufbMask.push_back(DAG.getConstant(0x5, MVT::i8));
8882 pshufbMask.push_back(DAG.getConstant(0x8, MVT::i8));
8883 pshufbMask.push_back(DAG.getConstant(0x9, MVT::i8));
8884 pshufbMask.push_back(DAG.getConstant(0xc, MVT::i8));
8885 pshufbMask.push_back(DAG.getConstant(0xd, MVT::i8));
8886 for (unsigned j = 0; j < 8; ++j)
8887 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
8889 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8,
8890 &pshufbMask[0], 32);
8891 In = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8, In, BV);
8892 In = DAG.getNode(ISD::BITCAST, DL, MVT::v4i64, In);
8894 static const int ShufMask[] = {0, 2, -1, -1};
8895 In = DAG.getVectorShuffle(MVT::v4i64, DL, In, DAG.getUNDEF(MVT::v4i64),
8897 In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
8898 DAG.getIntPtrConstant(0));
8899 return DAG.getNode(ISD::BITCAST, DL, VT, In);
8902 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
8903 DAG.getIntPtrConstant(0));
8905 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
8906 DAG.getIntPtrConstant(4));
8908 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, OpLo);
8909 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, OpHi);
8912 static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13,
8913 -1, -1, -1, -1, -1, -1, -1, -1};
8915 SDValue Undef = DAG.getUNDEF(MVT::v16i8);
8916 OpLo = DAG.getVectorShuffle(MVT::v16i8, DL, OpLo, Undef, ShufMask1);
8917 OpHi = DAG.getVectorShuffle(MVT::v16i8, DL, OpHi, Undef, ShufMask1);
8919 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpLo);
8920 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpHi);
8922 // The MOVLHPS Mask:
8923 static const int ShufMask2[] = {0, 1, 4, 5};
8924 SDValue res = DAG.getVectorShuffle(MVT::v4i32, DL, OpLo, OpHi, ShufMask2);
8925 return DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, res);
8928 // Handle truncation of V256 to V128 using shuffles.
8929 if (!VT.is128BitVector() || !SVT.is256BitVector())
8932 assert(VT.getVectorNumElements() != SVT.getVectorNumElements() &&
8934 assert(Subtarget->hasFp256() && "256-bit vector without AVX!");
8936 unsigned NumElems = VT.getVectorNumElements();
8937 EVT NVT = EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(),
8940 SmallVector<int, 16> MaskVec(NumElems * 2, -1);
8941 // Prepare truncation shuffle mask
8942 for (unsigned i = 0; i != NumElems; ++i)
8944 SDValue V = DAG.getVectorShuffle(NVT, DL,
8945 DAG.getNode(ISD::BITCAST, DL, NVT, In),
8946 DAG.getUNDEF(NVT), &MaskVec[0]);
8947 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V,
8948 DAG.getIntPtrConstant(0));
8951 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
8952 SelectionDAG &DAG) const {
8953 MVT VT = Op.getValueType().getSimpleVT();
8954 if (VT.isVector()) {
8955 if (VT == MVT::v8i16)
8956 return DAG.getNode(ISD::TRUNCATE, SDLoc(Op), VT,
8957 DAG.getNode(ISD::FP_TO_SINT, SDLoc(Op),
8958 MVT::v8i32, Op.getOperand(0)));
8962 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
8963 /*IsSigned=*/ true, /*IsReplace=*/ false);
8964 SDValue FIST = Vals.first, StackSlot = Vals.second;
8965 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
8966 if (FIST.getNode() == 0) return Op;
8968 if (StackSlot.getNode())
8970 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
8971 FIST, StackSlot, MachinePointerInfo(),
8972 false, false, false, 0);
8974 // The node is the result.
8978 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
8979 SelectionDAG &DAG) const {
8980 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
8981 /*IsSigned=*/ false, /*IsReplace=*/ false);
8982 SDValue FIST = Vals.first, StackSlot = Vals.second;
8983 assert(FIST.getNode() && "Unexpected failure");
8985 if (StackSlot.getNode())
8987 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
8988 FIST, StackSlot, MachinePointerInfo(),
8989 false, false, false, 0);
8991 // The node is the result.
8995 static SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) {
8997 MVT VT = Op.getValueType().getSimpleVT();
8998 SDValue In = Op.getOperand(0);
8999 MVT SVT = In.getValueType().getSimpleVT();
9001 assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!");
9003 return DAG.getNode(X86ISD::VFPEXT, DL, VT,
9004 DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32,
9005 In, DAG.getUNDEF(SVT)));
9008 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) const {
9009 LLVMContext *Context = DAG.getContext();
9011 MVT VT = Op.getValueType().getSimpleVT();
9013 unsigned NumElts = VT == MVT::f64 ? 2 : 4;
9014 if (VT.isVector()) {
9015 EltVT = VT.getVectorElementType();
9016 NumElts = VT.getVectorNumElements();
9019 if (EltVT == MVT::f64)
9020 C = ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
9021 APInt(64, ~(1ULL << 63))));
9023 C = ConstantFP::get(*Context, APFloat(APFloat::IEEEsingle,
9024 APInt(32, ~(1U << 31))));
9025 C = ConstantVector::getSplat(NumElts, C);
9026 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy());
9027 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
9028 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
9029 MachinePointerInfo::getConstantPool(),
9030 false, false, false, Alignment);
9031 if (VT.isVector()) {
9032 MVT ANDVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
9033 return DAG.getNode(ISD::BITCAST, dl, VT,
9034 DAG.getNode(ISD::AND, dl, ANDVT,
9035 DAG.getNode(ISD::BITCAST, dl, ANDVT,
9037 DAG.getNode(ISD::BITCAST, dl, ANDVT, Mask)));
9039 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
9042 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
9043 LLVMContext *Context = DAG.getContext();
9045 MVT VT = Op.getValueType().getSimpleVT();
9047 unsigned NumElts = VT == MVT::f64 ? 2 : 4;
9048 if (VT.isVector()) {
9049 EltVT = VT.getVectorElementType();
9050 NumElts = VT.getVectorNumElements();
9053 if (EltVT == MVT::f64)
9054 C = ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
9055 APInt(64, 1ULL << 63)));
9057 C = ConstantFP::get(*Context, APFloat(APFloat::IEEEsingle,
9058 APInt(32, 1U << 31)));
9059 C = ConstantVector::getSplat(NumElts, C);
9060 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy());
9061 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
9062 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
9063 MachinePointerInfo::getConstantPool(),
9064 false, false, false, Alignment);
9065 if (VT.isVector()) {
9066 MVT XORVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
9067 return DAG.getNode(ISD::BITCAST, dl, VT,
9068 DAG.getNode(ISD::XOR, dl, XORVT,
9069 DAG.getNode(ISD::BITCAST, dl, XORVT,
9071 DAG.getNode(ISD::BITCAST, dl, XORVT, Mask)));
9074 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
9077 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
9078 LLVMContext *Context = DAG.getContext();
9079 SDValue Op0 = Op.getOperand(0);
9080 SDValue Op1 = Op.getOperand(1);
9082 MVT VT = Op.getValueType().getSimpleVT();
9083 MVT SrcVT = Op1.getValueType().getSimpleVT();
9085 // If second operand is smaller, extend it first.
9086 if (SrcVT.bitsLT(VT)) {
9087 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
9090 // And if it is bigger, shrink it first.
9091 if (SrcVT.bitsGT(VT)) {
9092 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
9096 // At this point the operands and the result should have the same
9097 // type, and that won't be f80 since that is not custom lowered.
9099 // First get the sign bit of second operand.
9100 SmallVector<Constant*,4> CV;
9101 if (SrcVT == MVT::f64) {
9102 const fltSemantics &Sem = APFloat::IEEEdouble;
9103 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(64, 1ULL << 63))));
9104 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(64, 0))));
9106 const fltSemantics &Sem = APFloat::IEEEsingle;
9107 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 1U << 31))));
9108 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 0))));
9109 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 0))));
9110 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 0))));
9112 Constant *C = ConstantVector::get(CV);
9113 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
9114 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
9115 MachinePointerInfo::getConstantPool(),
9116 false, false, false, 16);
9117 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
9119 // Shift sign bit right or left if the two operands have different types.
9120 if (SrcVT.bitsGT(VT)) {
9121 // Op0 is MVT::f32, Op1 is MVT::f64.
9122 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
9123 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
9124 DAG.getConstant(32, MVT::i32));
9125 SignBit = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, SignBit);
9126 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
9127 DAG.getIntPtrConstant(0));
9130 // Clear first operand sign bit.
9132 if (VT == MVT::f64) {
9133 const fltSemantics &Sem = APFloat::IEEEdouble;
9134 CV.push_back(ConstantFP::get(*Context, APFloat(Sem,
9135 APInt(64, ~(1ULL << 63)))));
9136 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(64, 0))));
9138 const fltSemantics &Sem = APFloat::IEEEsingle;
9139 CV.push_back(ConstantFP::get(*Context, APFloat(Sem,
9140 APInt(32, ~(1U << 31)))));
9141 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 0))));
9142 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 0))));
9143 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 0))));
9145 C = ConstantVector::get(CV);
9146 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
9147 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
9148 MachinePointerInfo::getConstantPool(),
9149 false, false, false, 16);
9150 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
9152 // Or the value with the sign bit.
9153 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
9156 static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) {
9157 SDValue N0 = Op.getOperand(0);
9159 MVT VT = Op.getValueType().getSimpleVT();
9161 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
9162 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
9163 DAG.getConstant(1, VT));
9164 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, VT));
9167 // LowerVectorAllZeroTest - Check whether an OR'd tree is PTEST-able.
9169 SDValue X86TargetLowering::LowerVectorAllZeroTest(SDValue Op,
9170 SelectionDAG &DAG) const {
9171 assert(Op.getOpcode() == ISD::OR && "Only check OR'd tree.");
9173 if (!Subtarget->hasSSE41())
9176 if (!Op->hasOneUse())
9179 SDNode *N = Op.getNode();
9182 SmallVector<SDValue, 8> Opnds;
9183 DenseMap<SDValue, unsigned> VecInMap;
9184 EVT VT = MVT::Other;
9186 // Recognize a special case where a vector is casted into wide integer to
9188 Opnds.push_back(N->getOperand(0));
9189 Opnds.push_back(N->getOperand(1));
9191 for (unsigned Slot = 0, e = Opnds.size(); Slot < e; ++Slot) {
9192 SmallVectorImpl<SDValue>::const_iterator I = Opnds.begin() + Slot;
9193 // BFS traverse all OR'd operands.
9194 if (I->getOpcode() == ISD::OR) {
9195 Opnds.push_back(I->getOperand(0));
9196 Opnds.push_back(I->getOperand(1));
9197 // Re-evaluate the number of nodes to be traversed.
9198 e += 2; // 2 more nodes (LHS and RHS) are pushed.
9202 // Quit if a non-EXTRACT_VECTOR_ELT
9203 if (I->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
9206 // Quit if without a constant index.
9207 SDValue Idx = I->getOperand(1);
9208 if (!isa<ConstantSDNode>(Idx))
9211 SDValue ExtractedFromVec = I->getOperand(0);
9212 DenseMap<SDValue, unsigned>::iterator M = VecInMap.find(ExtractedFromVec);
9213 if (M == VecInMap.end()) {
9214 VT = ExtractedFromVec.getValueType();
9215 // Quit if not 128/256-bit vector.
9216 if (!VT.is128BitVector() && !VT.is256BitVector())
9218 // Quit if not the same type.
9219 if (VecInMap.begin() != VecInMap.end() &&
9220 VT != VecInMap.begin()->first.getValueType())
9222 M = VecInMap.insert(std::make_pair(ExtractedFromVec, 0)).first;
9224 M->second |= 1U << cast<ConstantSDNode>(Idx)->getZExtValue();
9227 assert((VT.is128BitVector() || VT.is256BitVector()) &&
9228 "Not extracted from 128-/256-bit vector.");
9230 unsigned FullMask = (1U << VT.getVectorNumElements()) - 1U;
9231 SmallVector<SDValue, 8> VecIns;
9233 for (DenseMap<SDValue, unsigned>::const_iterator
9234 I = VecInMap.begin(), E = VecInMap.end(); I != E; ++I) {
9235 // Quit if not all elements are used.
9236 if (I->second != FullMask)
9238 VecIns.push_back(I->first);
9241 EVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
9243 // Cast all vectors into TestVT for PTEST.
9244 for (unsigned i = 0, e = VecIns.size(); i < e; ++i)
9245 VecIns[i] = DAG.getNode(ISD::BITCAST, DL, TestVT, VecIns[i]);
9247 // If more than one full vectors are evaluated, OR them first before PTEST.
9248 for (unsigned Slot = 0, e = VecIns.size(); e - Slot > 1; Slot += 2, e += 1) {
9249 // Each iteration will OR 2 nodes and append the result until there is only
9250 // 1 node left, i.e. the final OR'd value of all vectors.
9251 SDValue LHS = VecIns[Slot];
9252 SDValue RHS = VecIns[Slot + 1];
9253 VecIns.push_back(DAG.getNode(ISD::OR, DL, TestVT, LHS, RHS));
9256 return DAG.getNode(X86ISD::PTEST, DL, MVT::i32,
9257 VecIns.back(), VecIns.back());
9260 /// Emit nodes that will be selected as "test Op0,Op0", or something
9262 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
9263 SelectionDAG &DAG) const {
9266 // CF and OF aren't always set the way we want. Determine which
9267 // of these we need.
9268 bool NeedCF = false;
9269 bool NeedOF = false;
9272 case X86::COND_A: case X86::COND_AE:
9273 case X86::COND_B: case X86::COND_BE:
9276 case X86::COND_G: case X86::COND_GE:
9277 case X86::COND_L: case X86::COND_LE:
9278 case X86::COND_O: case X86::COND_NO:
9283 // See if we can use the EFLAGS value from the operand instead of
9284 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
9285 // we prove that the arithmetic won't overflow, we can't use OF or CF.
9286 if (Op.getResNo() != 0 || NeedOF || NeedCF)
9287 // Emit a CMP with 0, which is the TEST pattern.
9288 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
9289 DAG.getConstant(0, Op.getValueType()));
9291 unsigned Opcode = 0;
9292 unsigned NumOperands = 0;
9294 // Truncate operations may prevent the merge of the SETCC instruction
9295 // and the arithmetic intruction before it. Attempt to truncate the operands
9296 // of the arithmetic instruction and use a reduced bit-width instruction.
9297 bool NeedTruncation = false;
9298 SDValue ArithOp = Op;
9299 if (Op->getOpcode() == ISD::TRUNCATE && Op->hasOneUse()) {
9300 SDValue Arith = Op->getOperand(0);
9301 // Both the trunc and the arithmetic op need to have one user each.
9302 if (Arith->hasOneUse())
9303 switch (Arith.getOpcode()) {
9310 NeedTruncation = true;
9316 // NOTICE: In the code below we use ArithOp to hold the arithmetic operation
9317 // which may be the result of a CAST. We use the variable 'Op', which is the
9318 // non-casted variable when we check for possible users.
9319 switch (ArithOp.getOpcode()) {
9321 // Due to an isel shortcoming, be conservative if this add is likely to be
9322 // selected as part of a load-modify-store instruction. When the root node
9323 // in a match is a store, isel doesn't know how to remap non-chain non-flag
9324 // uses of other nodes in the match, such as the ADD in this case. This
9325 // leads to the ADD being left around and reselected, with the result being
9326 // two adds in the output. Alas, even if none our users are stores, that
9327 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
9328 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
9329 // climbing the DAG back to the root, and it doesn't seem to be worth the
9331 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
9332 UE = Op.getNode()->use_end(); UI != UE; ++UI)
9333 if (UI->getOpcode() != ISD::CopyToReg &&
9334 UI->getOpcode() != ISD::SETCC &&
9335 UI->getOpcode() != ISD::STORE)
9338 if (ConstantSDNode *C =
9339 dyn_cast<ConstantSDNode>(ArithOp.getNode()->getOperand(1))) {
9340 // An add of one will be selected as an INC.
9341 if (C->getAPIntValue() == 1) {
9342 Opcode = X86ISD::INC;
9347 // An add of negative one (subtract of one) will be selected as a DEC.
9348 if (C->getAPIntValue().isAllOnesValue()) {
9349 Opcode = X86ISD::DEC;
9355 // Otherwise use a regular EFLAGS-setting add.
9356 Opcode = X86ISD::ADD;
9360 // If the primary and result isn't used, don't bother using X86ISD::AND,
9361 // because a TEST instruction will be better.
9362 bool NonFlagUse = false;
9363 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
9364 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
9366 unsigned UOpNo = UI.getOperandNo();
9367 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
9368 // Look pass truncate.
9369 UOpNo = User->use_begin().getOperandNo();
9370 User = *User->use_begin();
9373 if (User->getOpcode() != ISD::BRCOND &&
9374 User->getOpcode() != ISD::SETCC &&
9375 !(User->getOpcode() == ISD::SELECT && UOpNo == 0)) {
9388 // Due to the ISEL shortcoming noted above, be conservative if this op is
9389 // likely to be selected as part of a load-modify-store instruction.
9390 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
9391 UE = Op.getNode()->use_end(); UI != UE; ++UI)
9392 if (UI->getOpcode() == ISD::STORE)
9395 // Otherwise use a regular EFLAGS-setting instruction.
9396 switch (ArithOp.getOpcode()) {
9397 default: llvm_unreachable("unexpected operator!");
9398 case ISD::SUB: Opcode = X86ISD::SUB; break;
9399 case ISD::XOR: Opcode = X86ISD::XOR; break;
9400 case ISD::AND: Opcode = X86ISD::AND; break;
9402 if (!NeedTruncation && (X86CC == X86::COND_E || X86CC == X86::COND_NE)) {
9403 SDValue EFLAGS = LowerVectorAllZeroTest(Op, DAG);
9404 if (EFLAGS.getNode())
9407 Opcode = X86ISD::OR;
9421 return SDValue(Op.getNode(), 1);
9427 // If we found that truncation is beneficial, perform the truncation and
9429 if (NeedTruncation) {
9430 EVT VT = Op.getValueType();
9431 SDValue WideVal = Op->getOperand(0);
9432 EVT WideVT = WideVal.getValueType();
9433 unsigned ConvertedOp = 0;
9434 // Use a target machine opcode to prevent further DAGCombine
9435 // optimizations that may separate the arithmetic operations
9436 // from the setcc node.
9437 switch (WideVal.getOpcode()) {
9439 case ISD::ADD: ConvertedOp = X86ISD::ADD; break;
9440 case ISD::SUB: ConvertedOp = X86ISD::SUB; break;
9441 case ISD::AND: ConvertedOp = X86ISD::AND; break;
9442 case ISD::OR: ConvertedOp = X86ISD::OR; break;
9443 case ISD::XOR: ConvertedOp = X86ISD::XOR; break;
9447 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9448 if (TLI.isOperationLegal(WideVal.getOpcode(), WideVT)) {
9449 SDValue V0 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(0));
9450 SDValue V1 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(1));
9451 Op = DAG.getNode(ConvertedOp, dl, VT, V0, V1);
9457 // Emit a CMP with 0, which is the TEST pattern.
9458 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
9459 DAG.getConstant(0, Op.getValueType()));
9461 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
9462 SmallVector<SDValue, 4> Ops;
9463 for (unsigned i = 0; i != NumOperands; ++i)
9464 Ops.push_back(Op.getOperand(i));
9466 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
9467 DAG.ReplaceAllUsesWith(Op, New);
9468 return SDValue(New.getNode(), 1);
9471 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
9473 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
9474 SelectionDAG &DAG) const {
9475 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
9476 if (C->getAPIntValue() == 0)
9477 return EmitTest(Op0, X86CC, DAG);
9480 if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 ||
9481 Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) {
9482 // Use SUB instead of CMP to enable CSE between SUB and CMP.
9483 SDVTList VTs = DAG.getVTList(Op0.getValueType(), MVT::i32);
9484 SDValue Sub = DAG.getNode(X86ISD::SUB, dl, VTs,
9486 return SDValue(Sub.getNode(), 1);
9488 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
9491 /// Convert a comparison if required by the subtarget.
9492 SDValue X86TargetLowering::ConvertCmpIfNecessary(SDValue Cmp,
9493 SelectionDAG &DAG) const {
9494 // If the subtarget does not support the FUCOMI instruction, floating-point
9495 // comparisons have to be converted.
9496 if (Subtarget->hasCMov() ||
9497 Cmp.getOpcode() != X86ISD::CMP ||
9498 !Cmp.getOperand(0).getValueType().isFloatingPoint() ||
9499 !Cmp.getOperand(1).getValueType().isFloatingPoint())
9502 // The instruction selector will select an FUCOM instruction instead of
9503 // FUCOMI, which writes the comparison result to FPSW instead of EFLAGS. Hence
9504 // build an SDNode sequence that transfers the result from FPSW into EFLAGS:
9505 // (X86sahf (trunc (srl (X86fp_stsw (trunc (X86cmp ...)), 8))))
9507 SDValue TruncFPSW = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, Cmp);
9508 SDValue FNStSW = DAG.getNode(X86ISD::FNSTSW16r, dl, MVT::i16, TruncFPSW);
9509 SDValue Srl = DAG.getNode(ISD::SRL, dl, MVT::i16, FNStSW,
9510 DAG.getConstant(8, MVT::i8));
9511 SDValue TruncSrl = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Srl);
9512 return DAG.getNode(X86ISD::SAHF, dl, MVT::i32, TruncSrl);
9515 static bool isAllOnes(SDValue V) {
9516 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
9517 return C && C->isAllOnesValue();
9520 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
9521 /// if it's possible.
9522 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
9523 SDLoc dl, SelectionDAG &DAG) const {
9524 SDValue Op0 = And.getOperand(0);
9525 SDValue Op1 = And.getOperand(1);
9526 if (Op0.getOpcode() == ISD::TRUNCATE)
9527 Op0 = Op0.getOperand(0);
9528 if (Op1.getOpcode() == ISD::TRUNCATE)
9529 Op1 = Op1.getOperand(0);
9532 if (Op1.getOpcode() == ISD::SHL)
9533 std::swap(Op0, Op1);
9534 if (Op0.getOpcode() == ISD::SHL) {
9535 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
9536 if (And00C->getZExtValue() == 1) {
9537 // If we looked past a truncate, check that it's only truncating away
9539 unsigned BitWidth = Op0.getValueSizeInBits();
9540 unsigned AndBitWidth = And.getValueSizeInBits();
9541 if (BitWidth > AndBitWidth) {
9543 DAG.ComputeMaskedBits(Op0, Zeros, Ones);
9544 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
9548 RHS = Op0.getOperand(1);
9550 } else if (Op1.getOpcode() == ISD::Constant) {
9551 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
9552 uint64_t AndRHSVal = AndRHS->getZExtValue();
9553 SDValue AndLHS = Op0;
9555 if (AndRHSVal == 1 && AndLHS.getOpcode() == ISD::SRL) {
9556 LHS = AndLHS.getOperand(0);
9557 RHS = AndLHS.getOperand(1);
9560 // Use BT if the immediate can't be encoded in a TEST instruction.
9561 if (!isUInt<32>(AndRHSVal) && isPowerOf2_64(AndRHSVal)) {
9563 RHS = DAG.getConstant(Log2_64_Ceil(AndRHSVal), LHS.getValueType());
9567 if (LHS.getNode()) {
9568 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
9569 // instruction. Since the shift amount is in-range-or-undefined, we know
9570 // that doing a bittest on the i32 value is ok. We extend to i32 because
9571 // the encoding for the i16 version is larger than the i32 version.
9572 // Also promote i16 to i32 for performance / code size reason.
9573 if (LHS.getValueType() == MVT::i8 ||
9574 LHS.getValueType() == MVT::i16)
9575 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
9577 // If the operand types disagree, extend the shift amount to match. Since
9578 // BT ignores high bits (like shifts) we can use anyextend.
9579 if (LHS.getValueType() != RHS.getValueType())
9580 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
9582 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
9583 X86::CondCode Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
9584 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
9585 DAG.getConstant(Cond, MVT::i8), BT);
9591 /// \brief - Turns an ISD::CondCode into a value suitable for SSE floating point
9593 static int translateX86FSETCC(ISD::CondCode SetCCOpcode, SDValue &Op0,
9598 // SSE Condition code mapping:
9607 switch (SetCCOpcode) {
9608 default: llvm_unreachable("Unexpected SETCC condition");
9610 case ISD::SETEQ: SSECC = 0; break;
9612 case ISD::SETGT: Swap = true; // Fallthrough
9614 case ISD::SETOLT: SSECC = 1; break;
9616 case ISD::SETGE: Swap = true; // Fallthrough
9618 case ISD::SETOLE: SSECC = 2; break;
9619 case ISD::SETUO: SSECC = 3; break;
9621 case ISD::SETNE: SSECC = 4; break;
9622 case ISD::SETULE: Swap = true; // Fallthrough
9623 case ISD::SETUGE: SSECC = 5; break;
9624 case ISD::SETULT: Swap = true; // Fallthrough
9625 case ISD::SETUGT: SSECC = 6; break;
9626 case ISD::SETO: SSECC = 7; break;
9628 case ISD::SETONE: SSECC = 8; break;
9631 std::swap(Op0, Op1);
9636 // Lower256IntVSETCC - Break a VSETCC 256-bit integer VSETCC into two new 128
9637 // ones, and then concatenate the result back.
9638 static SDValue Lower256IntVSETCC(SDValue Op, SelectionDAG &DAG) {
9639 MVT VT = Op.getValueType().getSimpleVT();
9641 assert(VT.is256BitVector() && Op.getOpcode() == ISD::SETCC &&
9642 "Unsupported value type for operation");
9644 unsigned NumElems = VT.getVectorNumElements();
9646 SDValue CC = Op.getOperand(2);
9648 // Extract the LHS vectors
9649 SDValue LHS = Op.getOperand(0);
9650 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
9651 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
9653 // Extract the RHS vectors
9654 SDValue RHS = Op.getOperand(1);
9655 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
9656 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
9658 // Issue the operation on the smaller types and concatenate the result back
9659 MVT EltVT = VT.getVectorElementType();
9660 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
9661 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
9662 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC),
9663 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
9666 static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
9667 SelectionDAG &DAG) {
9669 SDValue Op0 = Op.getOperand(0);
9670 SDValue Op1 = Op.getOperand(1);
9671 SDValue CC = Op.getOperand(2);
9672 MVT VT = Op.getValueType().getSimpleVT();
9673 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
9674 bool isFP = Op.getOperand(1).getValueType().getSimpleVT().isFloatingPoint();
9679 MVT EltVT = Op0.getValueType().getVectorElementType().getSimpleVT();
9680 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
9683 unsigned SSECC = translateX86FSETCC(SetCCOpcode, Op0, Op1);
9685 // In the two special cases we can't handle, emit two comparisons.
9688 unsigned CombineOpc;
9689 if (SetCCOpcode == ISD::SETUEQ) {
9690 CC0 = 3; CC1 = 0; CombineOpc = ISD::OR;
9692 assert(SetCCOpcode == ISD::SETONE);
9693 CC0 = 7; CC1 = 4; CombineOpc = ISD::AND;
9696 SDValue Cmp0 = DAG.getNode(X86ISD::CMPP, dl, VT, Op0, Op1,
9697 DAG.getConstant(CC0, MVT::i8));
9698 SDValue Cmp1 = DAG.getNode(X86ISD::CMPP, dl, VT, Op0, Op1,
9699 DAG.getConstant(CC1, MVT::i8));
9700 return DAG.getNode(CombineOpc, dl, VT, Cmp0, Cmp1);
9702 // Handle all other FP comparisons here.
9703 return DAG.getNode(X86ISD::CMPP, dl, VT, Op0, Op1,
9704 DAG.getConstant(SSECC, MVT::i8));
9707 // Break 256-bit integer vector compare into smaller ones.
9708 if (VT.is256BitVector() && !Subtarget->hasInt256())
9709 return Lower256IntVSETCC(Op, DAG);
9711 // We are handling one of the integer comparisons here. Since SSE only has
9712 // GT and EQ comparisons for integer, swapping operands and multiple
9713 // operations may be required for some comparisons.
9715 bool Swap = false, Invert = false, FlipSigns = false, MinMax = false;
9717 switch (SetCCOpcode) {
9718 default: llvm_unreachable("Unexpected SETCC condition");
9719 case ISD::SETNE: Invert = true;
9720 case ISD::SETEQ: Opc = X86ISD::PCMPEQ; break;
9721 case ISD::SETLT: Swap = true;
9722 case ISD::SETGT: Opc = X86ISD::PCMPGT; break;
9723 case ISD::SETGE: Swap = true;
9724 case ISD::SETLE: Opc = X86ISD::PCMPGT; Invert = true; break;
9725 case ISD::SETULT: Swap = true;
9726 case ISD::SETUGT: Opc = X86ISD::PCMPGT; FlipSigns = true; break;
9727 case ISD::SETUGE: Swap = true;
9728 case ISD::SETULE: Opc = X86ISD::PCMPGT; FlipSigns = true; Invert = true; break;
9731 // Special case: Use min/max operations for SETULE/SETUGE
9732 MVT VET = VT.getVectorElementType();
9734 (Subtarget->hasSSE41() && (VET >= MVT::i8 && VET <= MVT::i32))
9735 || (Subtarget->hasSSE2() && (VET == MVT::i8));
9738 switch (SetCCOpcode) {
9740 case ISD::SETULE: Opc = X86ISD::UMIN; MinMax = true; break;
9741 case ISD::SETUGE: Opc = X86ISD::UMAX; MinMax = true; break;
9744 if (MinMax) { Swap = false; Invert = false; FlipSigns = false; }
9748 std::swap(Op0, Op1);
9750 // Check that the operation in question is available (most are plain SSE2,
9751 // but PCMPGTQ and PCMPEQQ have different requirements).
9752 if (VT == MVT::v2i64) {
9753 if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42()) {
9754 assert(Subtarget->hasSSE2() && "Don't know how to lower!");
9756 // First cast everything to the right type.
9757 Op0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op0);
9758 Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op1);
9760 // Since SSE has no unsigned integer comparisons, we need to flip the sign
9761 // bits of the inputs before performing those operations. The lower
9762 // compare is always unsigned.
9765 SB = DAG.getConstant(0x80000000U, MVT::v4i32);
9767 SDValue Sign = DAG.getConstant(0x80000000U, MVT::i32);
9768 SDValue Zero = DAG.getConstant(0x00000000U, MVT::i32);
9769 SB = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
9770 Sign, Zero, Sign, Zero);
9772 Op0 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op0, SB);
9773 Op1 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op1, SB);
9775 // Emulate PCMPGTQ with (hi1 > hi2) | ((hi1 == hi2) & (lo1 > lo2))
9776 SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1);
9777 SDValue EQ = DAG.getNode(X86ISD::PCMPEQ, dl, MVT::v4i32, Op0, Op1);
9779 // Create masks for only the low parts/high parts of the 64 bit integers.
9780 static const int MaskHi[] = { 1, 1, 3, 3 };
9781 static const int MaskLo[] = { 0, 0, 2, 2 };
9782 SDValue EQHi = DAG.getVectorShuffle(MVT::v4i32, dl, EQ, EQ, MaskHi);
9783 SDValue GTLo = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskLo);
9784 SDValue GTHi = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi);
9786 SDValue Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, EQHi, GTLo);
9787 Result = DAG.getNode(ISD::OR, dl, MVT::v4i32, Result, GTHi);
9790 Result = DAG.getNOT(dl, Result, MVT::v4i32);
9792 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
9795 if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) {
9796 // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with
9797 // pcmpeqd + pshufd + pand.
9798 assert(Subtarget->hasSSE2() && !FlipSigns && "Don't know how to lower!");
9800 // First cast everything to the right type.
9801 Op0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op0);
9802 Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op1);
9805 SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1);
9807 // Make sure the lower and upper halves are both all-ones.
9808 static const int Mask[] = { 1, 0, 3, 2 };
9809 SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask);
9810 Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf);
9813 Result = DAG.getNOT(dl, Result, MVT::v4i32);
9815 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
9819 // Since SSE has no unsigned integer comparisons, we need to flip the sign
9820 // bits of the inputs before performing those operations.
9822 EVT EltVT = VT.getVectorElementType();
9823 SDValue SB = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()), VT);
9824 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SB);
9825 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SB);
9828 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
9830 // If the logical-not of the result is required, perform that now.
9832 Result = DAG.getNOT(dl, Result, VT);
9835 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Op0, Result);
9840 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
9842 MVT VT = Op.getValueType().getSimpleVT();
9844 if (VT.isVector()) return LowerVSETCC(Op, Subtarget, DAG);
9846 assert(VT == MVT::i8 && "SetCC type must be 8-bit integer");
9847 SDValue Op0 = Op.getOperand(0);
9848 SDValue Op1 = Op.getOperand(1);
9850 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
9852 // Optimize to BT if possible.
9853 // Lower (X & (1 << N)) == 0 to BT(X, N).
9854 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
9855 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
9856 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
9857 Op1.getOpcode() == ISD::Constant &&
9858 cast<ConstantSDNode>(Op1)->isNullValue() &&
9859 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
9860 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
9861 if (NewSetCC.getNode())
9865 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
9867 if (Op1.getOpcode() == ISD::Constant &&
9868 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
9869 cast<ConstantSDNode>(Op1)->isNullValue()) &&
9870 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
9872 // If the input is a setcc, then reuse the input setcc or use a new one with
9873 // the inverted condition.
9874 if (Op0.getOpcode() == X86ISD::SETCC) {
9875 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
9876 bool Invert = (CC == ISD::SETNE) ^
9877 cast<ConstantSDNode>(Op1)->isNullValue();
9878 if (!Invert) return Op0;
9880 CCode = X86::GetOppositeBranchCondition(CCode);
9881 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
9882 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
9886 bool isFP = Op1.getValueType().getSimpleVT().isFloatingPoint();
9887 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
9888 if (X86CC == X86::COND_INVALID)
9891 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, DAG);
9892 EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
9893 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
9894 DAG.getConstant(X86CC, MVT::i8), EFLAGS);
9897 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
9898 static bool isX86LogicalCmp(SDValue Op) {
9899 unsigned Opc = Op.getNode()->getOpcode();
9900 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI ||
9901 Opc == X86ISD::SAHF)
9903 if (Op.getResNo() == 1 &&
9904 (Opc == X86ISD::ADD ||
9905 Opc == X86ISD::SUB ||
9906 Opc == X86ISD::ADC ||
9907 Opc == X86ISD::SBB ||
9908 Opc == X86ISD::SMUL ||
9909 Opc == X86ISD::UMUL ||
9910 Opc == X86ISD::INC ||
9911 Opc == X86ISD::DEC ||
9912 Opc == X86ISD::OR ||
9913 Opc == X86ISD::XOR ||
9914 Opc == X86ISD::AND))
9917 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
9923 static bool isZero(SDValue V) {
9924 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
9925 return C && C->isNullValue();
9928 static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) {
9929 if (V.getOpcode() != ISD::TRUNCATE)
9932 SDValue VOp0 = V.getOperand(0);
9933 unsigned InBits = VOp0.getValueSizeInBits();
9934 unsigned Bits = V.getValueSizeInBits();
9935 return DAG.MaskedValueIsZero(VOp0, APInt::getHighBitsSet(InBits,InBits-Bits));
9938 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
9939 bool addTest = true;
9940 SDValue Cond = Op.getOperand(0);
9941 SDValue Op1 = Op.getOperand(1);
9942 SDValue Op2 = Op.getOperand(2);
9944 EVT VT = Op1.getValueType();
9947 // Lower fp selects into a CMP/AND/ANDN/OR sequence when the necessary SSE ops
9948 // are available. Otherwise fp cmovs get lowered into a less efficient branch
9949 // sequence later on.
9950 if (Cond.getOpcode() == ISD::SETCC &&
9951 ((Subtarget->hasSSE2() && (VT == MVT::f32 || VT == MVT::f64)) ||
9952 (Subtarget->hasSSE1() && VT == MVT::f32)) &&
9953 VT == Cond.getOperand(0).getValueType() && Cond->hasOneUse()) {
9954 SDValue CondOp0 = Cond.getOperand(0), CondOp1 = Cond.getOperand(1);
9955 int SSECC = translateX86FSETCC(
9956 cast<CondCodeSDNode>(Cond.getOperand(2))->get(), CondOp0, CondOp1);
9959 unsigned Opcode = VT == MVT::f32 ? X86ISD::FSETCCss : X86ISD::FSETCCsd;
9960 SDValue Cmp = DAG.getNode(Opcode, DL, VT, CondOp0, CondOp1,
9961 DAG.getConstant(SSECC, MVT::i8));
9962 SDValue AndN = DAG.getNode(X86ISD::FANDN, DL, VT, Cmp, Op2);
9963 SDValue And = DAG.getNode(X86ISD::FAND, DL, VT, Cmp, Op1);
9964 return DAG.getNode(X86ISD::FOR, DL, VT, AndN, And);
9968 if (Cond.getOpcode() == ISD::SETCC) {
9969 SDValue NewCond = LowerSETCC(Cond, DAG);
9970 if (NewCond.getNode())
9974 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
9975 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
9976 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
9977 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
9978 if (Cond.getOpcode() == X86ISD::SETCC &&
9979 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
9980 isZero(Cond.getOperand(1).getOperand(1))) {
9981 SDValue Cmp = Cond.getOperand(1);
9983 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
9985 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
9986 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
9987 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
9989 SDValue CmpOp0 = Cmp.getOperand(0);
9990 // Apply further optimizations for special cases
9991 // (select (x != 0), -1, 0) -> neg & sbb
9992 // (select (x == 0), 0, -1) -> neg & sbb
9993 if (ConstantSDNode *YC = dyn_cast<ConstantSDNode>(Y))
9994 if (YC->isNullValue() &&
9995 (isAllOnes(Op1) == (CondCode == X86::COND_NE))) {
9996 SDVTList VTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32);
9997 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, VTs,
9998 DAG.getConstant(0, CmpOp0.getValueType()),
10000 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
10001 DAG.getConstant(X86::COND_B, MVT::i8),
10002 SDValue(Neg.getNode(), 1));
10006 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
10007 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
10008 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
10010 SDValue Res = // Res = 0 or -1.
10011 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
10012 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
10014 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
10015 Res = DAG.getNOT(DL, Res, Res.getValueType());
10017 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
10018 if (N2C == 0 || !N2C->isNullValue())
10019 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
10024 // Look past (and (setcc_carry (cmp ...)), 1).
10025 if (Cond.getOpcode() == ISD::AND &&
10026 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
10027 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
10028 if (C && C->getAPIntValue() == 1)
10029 Cond = Cond.getOperand(0);
10032 // If condition flag is set by a X86ISD::CMP, then use it as the condition
10033 // setting operand in place of the X86ISD::SETCC.
10034 unsigned CondOpcode = Cond.getOpcode();
10035 if (CondOpcode == X86ISD::SETCC ||
10036 CondOpcode == X86ISD::SETCC_CARRY) {
10037 CC = Cond.getOperand(0);
10039 SDValue Cmp = Cond.getOperand(1);
10040 unsigned Opc = Cmp.getOpcode();
10041 MVT VT = Op.getValueType().getSimpleVT();
10043 bool IllegalFPCMov = false;
10044 if (VT.isFloatingPoint() && !VT.isVector() &&
10045 !isScalarFPTypeInSSEReg(VT)) // FPStack?
10046 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
10048 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
10049 Opc == X86ISD::BT) { // FIXME
10053 } else if (CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
10054 CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
10055 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
10056 Cond.getOperand(0).getValueType() != MVT::i8)) {
10057 SDValue LHS = Cond.getOperand(0);
10058 SDValue RHS = Cond.getOperand(1);
10059 unsigned X86Opcode;
10062 switch (CondOpcode) {
10063 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
10064 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
10065 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
10066 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
10067 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
10068 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
10069 default: llvm_unreachable("unexpected overflowing operator");
10071 if (CondOpcode == ISD::UMULO)
10072 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
10075 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
10077 SDValue X86Op = DAG.getNode(X86Opcode, DL, VTs, LHS, RHS);
10079 if (CondOpcode == ISD::UMULO)
10080 Cond = X86Op.getValue(2);
10082 Cond = X86Op.getValue(1);
10084 CC = DAG.getConstant(X86Cond, MVT::i8);
10089 // Look pass the truncate if the high bits are known zero.
10090 if (isTruncWithZeroHighBitsInput(Cond, DAG))
10091 Cond = Cond.getOperand(0);
10093 // We know the result of AND is compared against zero. Try to match
10095 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
10096 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
10097 if (NewSetCC.getNode()) {
10098 CC = NewSetCC.getOperand(0);
10099 Cond = NewSetCC.getOperand(1);
10106 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
10107 Cond = EmitTest(Cond, X86::COND_NE, DAG);
10110 // a < b ? -1 : 0 -> RES = ~setcc_carry
10111 // a < b ? 0 : -1 -> RES = setcc_carry
10112 // a >= b ? -1 : 0 -> RES = setcc_carry
10113 // a >= b ? 0 : -1 -> RES = ~setcc_carry
10114 if (Cond.getOpcode() == X86ISD::SUB) {
10115 Cond = ConvertCmpIfNecessary(Cond, DAG);
10116 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
10118 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
10119 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
10120 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
10121 DAG.getConstant(X86::COND_B, MVT::i8), Cond);
10122 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
10123 return DAG.getNOT(DL, Res, Res.getValueType());
10128 // X86 doesn't have an i8 cmov. If both operands are the result of a truncate
10129 // widen the cmov and push the truncate through. This avoids introducing a new
10130 // branch during isel and doesn't add any extensions.
10131 if (Op.getValueType() == MVT::i8 &&
10132 Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) {
10133 SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0);
10134 if (T1.getValueType() == T2.getValueType() &&
10135 // Blacklist CopyFromReg to avoid partial register stalls.
10136 T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){
10137 SDVTList VTs = DAG.getVTList(T1.getValueType(), MVT::Glue);
10138 SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, VTs, T2, T1, CC, Cond);
10139 return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov);
10143 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
10144 // condition is true.
10145 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
10146 SDValue Ops[] = { Op2, Op1, CC, Cond };
10147 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops, array_lengthof(Ops));
10150 SDValue X86TargetLowering::LowerSIGN_EXTEND(SDValue Op,
10151 SelectionDAG &DAG) const {
10152 MVT VT = Op->getValueType(0).getSimpleVT();
10153 SDValue In = Op->getOperand(0);
10154 MVT InVT = In.getValueType().getSimpleVT();
10157 if ((VT != MVT::v4i64 || InVT != MVT::v4i32) &&
10158 (VT != MVT::v8i32 || InVT != MVT::v8i16))
10161 if (Subtarget->hasInt256())
10162 return DAG.getNode(X86ISD::VSEXT_MOVL, dl, VT, In);
10164 // Optimize vectors in AVX mode
10165 // Sign extend v8i16 to v8i32 and
10168 // Divide input vector into two parts
10169 // for v4i32 the shuffle mask will be { 0, 1, -1, -1} {2, 3, -1, -1}
10170 // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32
10171 // concat the vectors to original VT
10173 unsigned NumElems = InVT.getVectorNumElements();
10174 SDValue Undef = DAG.getUNDEF(InVT);
10176 SmallVector<int,8> ShufMask1(NumElems, -1);
10177 for (unsigned i = 0; i != NumElems/2; ++i)
10180 SDValue OpLo = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask1[0]);
10182 SmallVector<int,8> ShufMask2(NumElems, -1);
10183 for (unsigned i = 0; i != NumElems/2; ++i)
10184 ShufMask2[i] = i + NumElems/2;
10186 SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask2[0]);
10188 MVT HalfVT = MVT::getVectorVT(VT.getScalarType(),
10189 VT.getVectorNumElements()/2);
10191 OpLo = DAG.getNode(X86ISD::VSEXT_MOVL, dl, HalfVT, OpLo);
10192 OpHi = DAG.getNode(X86ISD::VSEXT_MOVL, dl, HalfVT, OpHi);
10194 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
10197 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
10198 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
10199 // from the AND / OR.
10200 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
10201 Opc = Op.getOpcode();
10202 if (Opc != ISD::OR && Opc != ISD::AND)
10204 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
10205 Op.getOperand(0).hasOneUse() &&
10206 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
10207 Op.getOperand(1).hasOneUse());
10210 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
10211 // 1 and that the SETCC node has a single use.
10212 static bool isXor1OfSetCC(SDValue Op) {
10213 if (Op.getOpcode() != ISD::XOR)
10215 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
10216 if (N1C && N1C->getAPIntValue() == 1) {
10217 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
10218 Op.getOperand(0).hasOneUse();
10223 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
10224 bool addTest = true;
10225 SDValue Chain = Op.getOperand(0);
10226 SDValue Cond = Op.getOperand(1);
10227 SDValue Dest = Op.getOperand(2);
10230 bool Inverted = false;
10232 if (Cond.getOpcode() == ISD::SETCC) {
10233 // Check for setcc([su]{add,sub,mul}o == 0).
10234 if (cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETEQ &&
10235 isa<ConstantSDNode>(Cond.getOperand(1)) &&
10236 cast<ConstantSDNode>(Cond.getOperand(1))->isNullValue() &&
10237 Cond.getOperand(0).getResNo() == 1 &&
10238 (Cond.getOperand(0).getOpcode() == ISD::SADDO ||
10239 Cond.getOperand(0).getOpcode() == ISD::UADDO ||
10240 Cond.getOperand(0).getOpcode() == ISD::SSUBO ||
10241 Cond.getOperand(0).getOpcode() == ISD::USUBO ||
10242 Cond.getOperand(0).getOpcode() == ISD::SMULO ||
10243 Cond.getOperand(0).getOpcode() == ISD::UMULO)) {
10245 Cond = Cond.getOperand(0);
10247 SDValue NewCond = LowerSETCC(Cond, DAG);
10248 if (NewCond.getNode())
10253 // FIXME: LowerXALUO doesn't handle these!!
10254 else if (Cond.getOpcode() == X86ISD::ADD ||
10255 Cond.getOpcode() == X86ISD::SUB ||
10256 Cond.getOpcode() == X86ISD::SMUL ||
10257 Cond.getOpcode() == X86ISD::UMUL)
10258 Cond = LowerXALUO(Cond, DAG);
10261 // Look pass (and (setcc_carry (cmp ...)), 1).
10262 if (Cond.getOpcode() == ISD::AND &&
10263 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
10264 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
10265 if (C && C->getAPIntValue() == 1)
10266 Cond = Cond.getOperand(0);
10269 // If condition flag is set by a X86ISD::CMP, then use it as the condition
10270 // setting operand in place of the X86ISD::SETCC.
10271 unsigned CondOpcode = Cond.getOpcode();
10272 if (CondOpcode == X86ISD::SETCC ||
10273 CondOpcode == X86ISD::SETCC_CARRY) {
10274 CC = Cond.getOperand(0);
10276 SDValue Cmp = Cond.getOperand(1);
10277 unsigned Opc = Cmp.getOpcode();
10278 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
10279 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
10283 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
10287 // These can only come from an arithmetic instruction with overflow,
10288 // e.g. SADDO, UADDO.
10289 Cond = Cond.getNode()->getOperand(1);
10295 CondOpcode = Cond.getOpcode();
10296 if (CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
10297 CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
10298 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
10299 Cond.getOperand(0).getValueType() != MVT::i8)) {
10300 SDValue LHS = Cond.getOperand(0);
10301 SDValue RHS = Cond.getOperand(1);
10302 unsigned X86Opcode;
10305 switch (CondOpcode) {
10306 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
10307 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
10308 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
10309 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
10310 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
10311 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
10312 default: llvm_unreachable("unexpected overflowing operator");
10315 X86Cond = X86::GetOppositeBranchCondition((X86::CondCode)X86Cond);
10316 if (CondOpcode == ISD::UMULO)
10317 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
10320 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
10322 SDValue X86Op = DAG.getNode(X86Opcode, dl, VTs, LHS, RHS);
10324 if (CondOpcode == ISD::UMULO)
10325 Cond = X86Op.getValue(2);
10327 Cond = X86Op.getValue(1);
10329 CC = DAG.getConstant(X86Cond, MVT::i8);
10333 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
10334 SDValue Cmp = Cond.getOperand(0).getOperand(1);
10335 if (CondOpc == ISD::OR) {
10336 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
10337 // two branches instead of an explicit OR instruction with a
10339 if (Cmp == Cond.getOperand(1).getOperand(1) &&
10340 isX86LogicalCmp(Cmp)) {
10341 CC = Cond.getOperand(0).getOperand(0);
10342 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
10343 Chain, Dest, CC, Cmp);
10344 CC = Cond.getOperand(1).getOperand(0);
10348 } else { // ISD::AND
10349 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
10350 // two branches instead of an explicit AND instruction with a
10351 // separate test. However, we only do this if this block doesn't
10352 // have a fall-through edge, because this requires an explicit
10353 // jmp when the condition is false.
10354 if (Cmp == Cond.getOperand(1).getOperand(1) &&
10355 isX86LogicalCmp(Cmp) &&
10356 Op.getNode()->hasOneUse()) {
10357 X86::CondCode CCode =
10358 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
10359 CCode = X86::GetOppositeBranchCondition(CCode);
10360 CC = DAG.getConstant(CCode, MVT::i8);
10361 SDNode *User = *Op.getNode()->use_begin();
10362 // Look for an unconditional branch following this conditional branch.
10363 // We need this because we need to reverse the successors in order
10364 // to implement FCMP_OEQ.
10365 if (User->getOpcode() == ISD::BR) {
10366 SDValue FalseBB = User->getOperand(1);
10368 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
10369 assert(NewBR == User);
10373 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
10374 Chain, Dest, CC, Cmp);
10375 X86::CondCode CCode =
10376 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
10377 CCode = X86::GetOppositeBranchCondition(CCode);
10378 CC = DAG.getConstant(CCode, MVT::i8);
10384 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
10385 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
10386 // It should be transformed during dag combiner except when the condition
10387 // is set by a arithmetics with overflow node.
10388 X86::CondCode CCode =
10389 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
10390 CCode = X86::GetOppositeBranchCondition(CCode);
10391 CC = DAG.getConstant(CCode, MVT::i8);
10392 Cond = Cond.getOperand(0).getOperand(1);
10394 } else if (Cond.getOpcode() == ISD::SETCC &&
10395 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETOEQ) {
10396 // For FCMP_OEQ, we can emit
10397 // two branches instead of an explicit AND instruction with a
10398 // separate test. However, we only do this if this block doesn't
10399 // have a fall-through edge, because this requires an explicit
10400 // jmp when the condition is false.
10401 if (Op.getNode()->hasOneUse()) {
10402 SDNode *User = *Op.getNode()->use_begin();
10403 // Look for an unconditional branch following this conditional branch.
10404 // We need this because we need to reverse the successors in order
10405 // to implement FCMP_OEQ.
10406 if (User->getOpcode() == ISD::BR) {
10407 SDValue FalseBB = User->getOperand(1);
10409 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
10410 assert(NewBR == User);
10414 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
10415 Cond.getOperand(0), Cond.getOperand(1));
10416 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
10417 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
10418 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
10419 Chain, Dest, CC, Cmp);
10420 CC = DAG.getConstant(X86::COND_P, MVT::i8);
10425 } else if (Cond.getOpcode() == ISD::SETCC &&
10426 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETUNE) {
10427 // For FCMP_UNE, we can emit
10428 // two branches instead of an explicit AND instruction with a
10429 // separate test. However, we only do this if this block doesn't
10430 // have a fall-through edge, because this requires an explicit
10431 // jmp when the condition is false.
10432 if (Op.getNode()->hasOneUse()) {
10433 SDNode *User = *Op.getNode()->use_begin();
10434 // Look for an unconditional branch following this conditional branch.
10435 // We need this because we need to reverse the successors in order
10436 // to implement FCMP_UNE.
10437 if (User->getOpcode() == ISD::BR) {
10438 SDValue FalseBB = User->getOperand(1);
10440 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
10441 assert(NewBR == User);
10444 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
10445 Cond.getOperand(0), Cond.getOperand(1));
10446 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
10447 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
10448 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
10449 Chain, Dest, CC, Cmp);
10450 CC = DAG.getConstant(X86::COND_NP, MVT::i8);
10460 // Look pass the truncate if the high bits are known zero.
10461 if (isTruncWithZeroHighBitsInput(Cond, DAG))
10462 Cond = Cond.getOperand(0);
10464 // We know the result of AND is compared against zero. Try to match
10466 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
10467 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
10468 if (NewSetCC.getNode()) {
10469 CC = NewSetCC.getOperand(0);
10470 Cond = NewSetCC.getOperand(1);
10477 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
10478 Cond = EmitTest(Cond, X86::COND_NE, DAG);
10480 Cond = ConvertCmpIfNecessary(Cond, DAG);
10481 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
10482 Chain, Dest, CC, Cond);
10485 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
10486 // Calls to _alloca is needed to probe the stack when allocating more than 4k
10487 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
10488 // that the guard pages used by the OS virtual memory manager are allocated in
10489 // correct sequence.
10491 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
10492 SelectionDAG &DAG) const {
10493 assert((Subtarget->isTargetCygMing() || Subtarget->isTargetWindows() ||
10494 getTargetMachine().Options.EnableSegmentedStacks) &&
10495 "This should be used only on Windows targets or when segmented stacks "
10497 assert(!Subtarget->isTargetEnvMacho() && "Not implemented");
10501 SDValue Chain = Op.getOperand(0);
10502 SDValue Size = Op.getOperand(1);
10503 // FIXME: Ensure alignment here
10505 bool Is64Bit = Subtarget->is64Bit();
10506 EVT SPTy = Is64Bit ? MVT::i64 : MVT::i32;
10508 if (getTargetMachine().Options.EnableSegmentedStacks) {
10509 MachineFunction &MF = DAG.getMachineFunction();
10510 MachineRegisterInfo &MRI = MF.getRegInfo();
10513 // The 64 bit implementation of segmented stacks needs to clobber both r10
10514 // r11. This makes it impossible to use it along with nested parameters.
10515 const Function *F = MF.getFunction();
10517 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
10519 if (I->hasNestAttr())
10520 report_fatal_error("Cannot use segmented stacks with functions that "
10521 "have nested arguments.");
10524 const TargetRegisterClass *AddrRegClass =
10525 getRegClassFor(Subtarget->is64Bit() ? MVT::i64:MVT::i32);
10526 unsigned Vreg = MRI.createVirtualRegister(AddrRegClass);
10527 Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size);
10528 SDValue Value = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain,
10529 DAG.getRegister(Vreg, SPTy));
10530 SDValue Ops1[2] = { Value, Chain };
10531 return DAG.getMergeValues(Ops1, 2, dl);
10534 unsigned Reg = (Subtarget->is64Bit() ? X86::RAX : X86::EAX);
10536 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
10537 Flag = Chain.getValue(1);
10538 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
10540 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
10541 Flag = Chain.getValue(1);
10543 const X86RegisterInfo *RegInfo =
10544 static_cast<const X86RegisterInfo*>(getTargetMachine().getRegisterInfo());
10545 Chain = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
10548 SDValue Ops1[2] = { Chain.getValue(0), Chain };
10549 return DAG.getMergeValues(Ops1, 2, dl);
10553 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
10554 MachineFunction &MF = DAG.getMachineFunction();
10555 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
10557 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
10560 if (!Subtarget->is64Bit() || Subtarget->isTargetWin64()) {
10561 // vastart just stores the address of the VarArgsFrameIndex slot into the
10562 // memory location argument.
10563 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
10565 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
10566 MachinePointerInfo(SV), false, false, 0);
10570 // gp_offset (0 - 6 * 8)
10571 // fp_offset (48 - 48 + 8 * 16)
10572 // overflow_arg_area (point to parameters coming in memory).
10574 SmallVector<SDValue, 8> MemOps;
10575 SDValue FIN = Op.getOperand(1);
10577 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
10578 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
10580 FIN, MachinePointerInfo(SV), false, false, 0);
10581 MemOps.push_back(Store);
10584 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
10585 FIN, DAG.getIntPtrConstant(4));
10586 Store = DAG.getStore(Op.getOperand(0), DL,
10587 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
10589 FIN, MachinePointerInfo(SV, 4), false, false, 0);
10590 MemOps.push_back(Store);
10592 // Store ptr to overflow_arg_area
10593 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
10594 FIN, DAG.getIntPtrConstant(4));
10595 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
10597 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
10598 MachinePointerInfo(SV, 8),
10600 MemOps.push_back(Store);
10602 // Store ptr to reg_save_area.
10603 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
10604 FIN, DAG.getIntPtrConstant(8));
10605 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
10607 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN,
10608 MachinePointerInfo(SV, 16), false, false, 0);
10609 MemOps.push_back(Store);
10610 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
10611 &MemOps[0], MemOps.size());
10614 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
10615 assert(Subtarget->is64Bit() &&
10616 "LowerVAARG only handles 64-bit va_arg!");
10617 assert((Subtarget->isTargetLinux() ||
10618 Subtarget->isTargetDarwin()) &&
10619 "Unhandled target in LowerVAARG");
10620 assert(Op.getNode()->getNumOperands() == 4);
10621 SDValue Chain = Op.getOperand(0);
10622 SDValue SrcPtr = Op.getOperand(1);
10623 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
10624 unsigned Align = Op.getConstantOperandVal(3);
10627 EVT ArgVT = Op.getNode()->getValueType(0);
10628 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
10629 uint32_t ArgSize = getDataLayout()->getTypeAllocSize(ArgTy);
10632 // Decide which area this value should be read from.
10633 // TODO: Implement the AMD64 ABI in its entirety. This simple
10634 // selection mechanism works only for the basic types.
10635 if (ArgVT == MVT::f80) {
10636 llvm_unreachable("va_arg for f80 not yet implemented");
10637 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
10638 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
10639 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
10640 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
10642 llvm_unreachable("Unhandled argument type in LowerVAARG");
10645 if (ArgMode == 2) {
10646 // Sanity Check: Make sure using fp_offset makes sense.
10647 assert(!getTargetMachine().Options.UseSoftFloat &&
10648 !(DAG.getMachineFunction()
10649 .getFunction()->getAttributes()
10650 .hasAttribute(AttributeSet::FunctionIndex,
10651 Attribute::NoImplicitFloat)) &&
10652 Subtarget->hasSSE1());
10655 // Insert VAARG_64 node into the DAG
10656 // VAARG_64 returns two values: Variable Argument Address, Chain
10657 SmallVector<SDValue, 11> InstOps;
10658 InstOps.push_back(Chain);
10659 InstOps.push_back(SrcPtr);
10660 InstOps.push_back(DAG.getConstant(ArgSize, MVT::i32));
10661 InstOps.push_back(DAG.getConstant(ArgMode, MVT::i8));
10662 InstOps.push_back(DAG.getConstant(Align, MVT::i32));
10663 SDVTList VTs = DAG.getVTList(getPointerTy(), MVT::Other);
10664 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
10665 VTs, &InstOps[0], InstOps.size(),
10667 MachinePointerInfo(SV),
10669 /*Volatile=*/false,
10671 /*WriteMem=*/true);
10672 Chain = VAARG.getValue(1);
10674 // Load the next argument and return it
10675 return DAG.getLoad(ArgVT, dl,
10678 MachinePointerInfo(),
10679 false, false, false, 0);
10682 static SDValue LowerVACOPY(SDValue Op, const X86Subtarget *Subtarget,
10683 SelectionDAG &DAG) {
10684 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
10685 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
10686 SDValue Chain = Op.getOperand(0);
10687 SDValue DstPtr = Op.getOperand(1);
10688 SDValue SrcPtr = Op.getOperand(2);
10689 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
10690 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
10693 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
10694 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
10696 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
10699 // getTargetVShiftNode - Handle vector element shifts where the shift amount
10700 // may or may not be a constant. Takes immediate version of shift as input.
10701 static SDValue getTargetVShiftNode(unsigned Opc, SDLoc dl, EVT VT,
10702 SDValue SrcOp, SDValue ShAmt,
10703 SelectionDAG &DAG) {
10704 assert(ShAmt.getValueType() == MVT::i32 && "ShAmt is not i32");
10706 if (isa<ConstantSDNode>(ShAmt)) {
10707 // Constant may be a TargetConstant. Use a regular constant.
10708 uint32_t ShiftAmt = cast<ConstantSDNode>(ShAmt)->getZExtValue();
10710 default: llvm_unreachable("Unknown target vector shift node");
10711 case X86ISD::VSHLI:
10712 case X86ISD::VSRLI:
10713 case X86ISD::VSRAI:
10714 return DAG.getNode(Opc, dl, VT, SrcOp,
10715 DAG.getConstant(ShiftAmt, MVT::i32));
10719 // Change opcode to non-immediate version
10721 default: llvm_unreachable("Unknown target vector shift node");
10722 case X86ISD::VSHLI: Opc = X86ISD::VSHL; break;
10723 case X86ISD::VSRLI: Opc = X86ISD::VSRL; break;
10724 case X86ISD::VSRAI: Opc = X86ISD::VSRA; break;
10727 // Need to build a vector containing shift amount
10728 // Shift amount is 32-bits, but SSE instructions read 64-bit, so fill with 0
10731 ShOps[1] = DAG.getConstant(0, MVT::i32);
10732 ShOps[2] = ShOps[3] = DAG.getUNDEF(MVT::i32);
10733 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, &ShOps[0], 4);
10735 // The return type has to be a 128-bit type with the same element
10736 // type as the input type.
10737 MVT EltVT = VT.getVectorElementType().getSimpleVT();
10738 EVT ShVT = MVT::getVectorVT(EltVT, 128/EltVT.getSizeInBits());
10740 ShAmt = DAG.getNode(ISD::BITCAST, dl, ShVT, ShAmt);
10741 return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt);
10744 static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
10746 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
10748 default: return SDValue(); // Don't custom lower most intrinsics.
10749 // Comparison intrinsics.
10750 case Intrinsic::x86_sse_comieq_ss:
10751 case Intrinsic::x86_sse_comilt_ss:
10752 case Intrinsic::x86_sse_comile_ss:
10753 case Intrinsic::x86_sse_comigt_ss:
10754 case Intrinsic::x86_sse_comige_ss:
10755 case Intrinsic::x86_sse_comineq_ss:
10756 case Intrinsic::x86_sse_ucomieq_ss:
10757 case Intrinsic::x86_sse_ucomilt_ss:
10758 case Intrinsic::x86_sse_ucomile_ss:
10759 case Intrinsic::x86_sse_ucomigt_ss:
10760 case Intrinsic::x86_sse_ucomige_ss:
10761 case Intrinsic::x86_sse_ucomineq_ss:
10762 case Intrinsic::x86_sse2_comieq_sd:
10763 case Intrinsic::x86_sse2_comilt_sd:
10764 case Intrinsic::x86_sse2_comile_sd:
10765 case Intrinsic::x86_sse2_comigt_sd:
10766 case Intrinsic::x86_sse2_comige_sd:
10767 case Intrinsic::x86_sse2_comineq_sd:
10768 case Intrinsic::x86_sse2_ucomieq_sd:
10769 case Intrinsic::x86_sse2_ucomilt_sd:
10770 case Intrinsic::x86_sse2_ucomile_sd:
10771 case Intrinsic::x86_sse2_ucomigt_sd:
10772 case Intrinsic::x86_sse2_ucomige_sd:
10773 case Intrinsic::x86_sse2_ucomineq_sd: {
10777 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
10778 case Intrinsic::x86_sse_comieq_ss:
10779 case Intrinsic::x86_sse2_comieq_sd:
10780 Opc = X86ISD::COMI;
10783 case Intrinsic::x86_sse_comilt_ss:
10784 case Intrinsic::x86_sse2_comilt_sd:
10785 Opc = X86ISD::COMI;
10788 case Intrinsic::x86_sse_comile_ss:
10789 case Intrinsic::x86_sse2_comile_sd:
10790 Opc = X86ISD::COMI;
10793 case Intrinsic::x86_sse_comigt_ss:
10794 case Intrinsic::x86_sse2_comigt_sd:
10795 Opc = X86ISD::COMI;
10798 case Intrinsic::x86_sse_comige_ss:
10799 case Intrinsic::x86_sse2_comige_sd:
10800 Opc = X86ISD::COMI;
10803 case Intrinsic::x86_sse_comineq_ss:
10804 case Intrinsic::x86_sse2_comineq_sd:
10805 Opc = X86ISD::COMI;
10808 case Intrinsic::x86_sse_ucomieq_ss:
10809 case Intrinsic::x86_sse2_ucomieq_sd:
10810 Opc = X86ISD::UCOMI;
10813 case Intrinsic::x86_sse_ucomilt_ss:
10814 case Intrinsic::x86_sse2_ucomilt_sd:
10815 Opc = X86ISD::UCOMI;
10818 case Intrinsic::x86_sse_ucomile_ss:
10819 case Intrinsic::x86_sse2_ucomile_sd:
10820 Opc = X86ISD::UCOMI;
10823 case Intrinsic::x86_sse_ucomigt_ss:
10824 case Intrinsic::x86_sse2_ucomigt_sd:
10825 Opc = X86ISD::UCOMI;
10828 case Intrinsic::x86_sse_ucomige_ss:
10829 case Intrinsic::x86_sse2_ucomige_sd:
10830 Opc = X86ISD::UCOMI;
10833 case Intrinsic::x86_sse_ucomineq_ss:
10834 case Intrinsic::x86_sse2_ucomineq_sd:
10835 Opc = X86ISD::UCOMI;
10840 SDValue LHS = Op.getOperand(1);
10841 SDValue RHS = Op.getOperand(2);
10842 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
10843 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
10844 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
10845 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
10846 DAG.getConstant(X86CC, MVT::i8), Cond);
10847 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
10850 // Arithmetic intrinsics.
10851 case Intrinsic::x86_sse2_pmulu_dq:
10852 case Intrinsic::x86_avx2_pmulu_dq:
10853 return DAG.getNode(X86ISD::PMULUDQ, dl, Op.getValueType(),
10854 Op.getOperand(1), Op.getOperand(2));
10856 // SSE2/AVX2 sub with unsigned saturation intrinsics
10857 case Intrinsic::x86_sse2_psubus_b:
10858 case Intrinsic::x86_sse2_psubus_w:
10859 case Intrinsic::x86_avx2_psubus_b:
10860 case Intrinsic::x86_avx2_psubus_w:
10861 return DAG.getNode(X86ISD::SUBUS, dl, Op.getValueType(),
10862 Op.getOperand(1), Op.getOperand(2));
10864 // SSE3/AVX horizontal add/sub intrinsics
10865 case Intrinsic::x86_sse3_hadd_ps:
10866 case Intrinsic::x86_sse3_hadd_pd:
10867 case Intrinsic::x86_avx_hadd_ps_256:
10868 case Intrinsic::x86_avx_hadd_pd_256:
10869 case Intrinsic::x86_sse3_hsub_ps:
10870 case Intrinsic::x86_sse3_hsub_pd:
10871 case Intrinsic::x86_avx_hsub_ps_256:
10872 case Intrinsic::x86_avx_hsub_pd_256:
10873 case Intrinsic::x86_ssse3_phadd_w_128:
10874 case Intrinsic::x86_ssse3_phadd_d_128:
10875 case Intrinsic::x86_avx2_phadd_w:
10876 case Intrinsic::x86_avx2_phadd_d:
10877 case Intrinsic::x86_ssse3_phsub_w_128:
10878 case Intrinsic::x86_ssse3_phsub_d_128:
10879 case Intrinsic::x86_avx2_phsub_w:
10880 case Intrinsic::x86_avx2_phsub_d: {
10883 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
10884 case Intrinsic::x86_sse3_hadd_ps:
10885 case Intrinsic::x86_sse3_hadd_pd:
10886 case Intrinsic::x86_avx_hadd_ps_256:
10887 case Intrinsic::x86_avx_hadd_pd_256:
10888 Opcode = X86ISD::FHADD;
10890 case Intrinsic::x86_sse3_hsub_ps:
10891 case Intrinsic::x86_sse3_hsub_pd:
10892 case Intrinsic::x86_avx_hsub_ps_256:
10893 case Intrinsic::x86_avx_hsub_pd_256:
10894 Opcode = X86ISD::FHSUB;
10896 case Intrinsic::x86_ssse3_phadd_w_128:
10897 case Intrinsic::x86_ssse3_phadd_d_128:
10898 case Intrinsic::x86_avx2_phadd_w:
10899 case Intrinsic::x86_avx2_phadd_d:
10900 Opcode = X86ISD::HADD;
10902 case Intrinsic::x86_ssse3_phsub_w_128:
10903 case Intrinsic::x86_ssse3_phsub_d_128:
10904 case Intrinsic::x86_avx2_phsub_w:
10905 case Intrinsic::x86_avx2_phsub_d:
10906 Opcode = X86ISD::HSUB;
10909 return DAG.getNode(Opcode, dl, Op.getValueType(),
10910 Op.getOperand(1), Op.getOperand(2));
10913 // SSE2/SSE41/AVX2 integer max/min intrinsics.
10914 case Intrinsic::x86_sse2_pmaxu_b:
10915 case Intrinsic::x86_sse41_pmaxuw:
10916 case Intrinsic::x86_sse41_pmaxud:
10917 case Intrinsic::x86_avx2_pmaxu_b:
10918 case Intrinsic::x86_avx2_pmaxu_w:
10919 case Intrinsic::x86_avx2_pmaxu_d:
10920 case Intrinsic::x86_sse2_pminu_b:
10921 case Intrinsic::x86_sse41_pminuw:
10922 case Intrinsic::x86_sse41_pminud:
10923 case Intrinsic::x86_avx2_pminu_b:
10924 case Intrinsic::x86_avx2_pminu_w:
10925 case Intrinsic::x86_avx2_pminu_d:
10926 case Intrinsic::x86_sse41_pmaxsb:
10927 case Intrinsic::x86_sse2_pmaxs_w:
10928 case Intrinsic::x86_sse41_pmaxsd:
10929 case Intrinsic::x86_avx2_pmaxs_b:
10930 case Intrinsic::x86_avx2_pmaxs_w:
10931 case Intrinsic::x86_avx2_pmaxs_d:
10932 case Intrinsic::x86_sse41_pminsb:
10933 case Intrinsic::x86_sse2_pmins_w:
10934 case Intrinsic::x86_sse41_pminsd:
10935 case Intrinsic::x86_avx2_pmins_b:
10936 case Intrinsic::x86_avx2_pmins_w:
10937 case Intrinsic::x86_avx2_pmins_d: {
10940 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
10941 case Intrinsic::x86_sse2_pmaxu_b:
10942 case Intrinsic::x86_sse41_pmaxuw:
10943 case Intrinsic::x86_sse41_pmaxud:
10944 case Intrinsic::x86_avx2_pmaxu_b:
10945 case Intrinsic::x86_avx2_pmaxu_w:
10946 case Intrinsic::x86_avx2_pmaxu_d:
10947 Opcode = X86ISD::UMAX;
10949 case Intrinsic::x86_sse2_pminu_b:
10950 case Intrinsic::x86_sse41_pminuw:
10951 case Intrinsic::x86_sse41_pminud:
10952 case Intrinsic::x86_avx2_pminu_b:
10953 case Intrinsic::x86_avx2_pminu_w:
10954 case Intrinsic::x86_avx2_pminu_d:
10955 Opcode = X86ISD::UMIN;
10957 case Intrinsic::x86_sse41_pmaxsb:
10958 case Intrinsic::x86_sse2_pmaxs_w:
10959 case Intrinsic::x86_sse41_pmaxsd:
10960 case Intrinsic::x86_avx2_pmaxs_b:
10961 case Intrinsic::x86_avx2_pmaxs_w:
10962 case Intrinsic::x86_avx2_pmaxs_d:
10963 Opcode = X86ISD::SMAX;
10965 case Intrinsic::x86_sse41_pminsb:
10966 case Intrinsic::x86_sse2_pmins_w:
10967 case Intrinsic::x86_sse41_pminsd:
10968 case Intrinsic::x86_avx2_pmins_b:
10969 case Intrinsic::x86_avx2_pmins_w:
10970 case Intrinsic::x86_avx2_pmins_d:
10971 Opcode = X86ISD::SMIN;
10974 return DAG.getNode(Opcode, dl, Op.getValueType(),
10975 Op.getOperand(1), Op.getOperand(2));
10978 // SSE/SSE2/AVX floating point max/min intrinsics.
10979 case Intrinsic::x86_sse_max_ps:
10980 case Intrinsic::x86_sse2_max_pd:
10981 case Intrinsic::x86_avx_max_ps_256:
10982 case Intrinsic::x86_avx_max_pd_256:
10983 case Intrinsic::x86_sse_min_ps:
10984 case Intrinsic::x86_sse2_min_pd:
10985 case Intrinsic::x86_avx_min_ps_256:
10986 case Intrinsic::x86_avx_min_pd_256: {
10989 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
10990 case Intrinsic::x86_sse_max_ps:
10991 case Intrinsic::x86_sse2_max_pd:
10992 case Intrinsic::x86_avx_max_ps_256:
10993 case Intrinsic::x86_avx_max_pd_256:
10994 Opcode = X86ISD::FMAX;
10996 case Intrinsic::x86_sse_min_ps:
10997 case Intrinsic::x86_sse2_min_pd:
10998 case Intrinsic::x86_avx_min_ps_256:
10999 case Intrinsic::x86_avx_min_pd_256:
11000 Opcode = X86ISD::FMIN;
11003 return DAG.getNode(Opcode, dl, Op.getValueType(),
11004 Op.getOperand(1), Op.getOperand(2));
11007 // AVX2 variable shift intrinsics
11008 case Intrinsic::x86_avx2_psllv_d:
11009 case Intrinsic::x86_avx2_psllv_q:
11010 case Intrinsic::x86_avx2_psllv_d_256:
11011 case Intrinsic::x86_avx2_psllv_q_256:
11012 case Intrinsic::x86_avx2_psrlv_d:
11013 case Intrinsic::x86_avx2_psrlv_q:
11014 case Intrinsic::x86_avx2_psrlv_d_256:
11015 case Intrinsic::x86_avx2_psrlv_q_256:
11016 case Intrinsic::x86_avx2_psrav_d:
11017 case Intrinsic::x86_avx2_psrav_d_256: {
11020 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
11021 case Intrinsic::x86_avx2_psllv_d:
11022 case Intrinsic::x86_avx2_psllv_q:
11023 case Intrinsic::x86_avx2_psllv_d_256:
11024 case Intrinsic::x86_avx2_psllv_q_256:
11027 case Intrinsic::x86_avx2_psrlv_d:
11028 case Intrinsic::x86_avx2_psrlv_q:
11029 case Intrinsic::x86_avx2_psrlv_d_256:
11030 case Intrinsic::x86_avx2_psrlv_q_256:
11033 case Intrinsic::x86_avx2_psrav_d:
11034 case Intrinsic::x86_avx2_psrav_d_256:
11038 return DAG.getNode(Opcode, dl, Op.getValueType(),
11039 Op.getOperand(1), Op.getOperand(2));
11042 case Intrinsic::x86_ssse3_pshuf_b_128:
11043 case Intrinsic::x86_avx2_pshuf_b:
11044 return DAG.getNode(X86ISD::PSHUFB, dl, Op.getValueType(),
11045 Op.getOperand(1), Op.getOperand(2));
11047 case Intrinsic::x86_ssse3_psign_b_128:
11048 case Intrinsic::x86_ssse3_psign_w_128:
11049 case Intrinsic::x86_ssse3_psign_d_128:
11050 case Intrinsic::x86_avx2_psign_b:
11051 case Intrinsic::x86_avx2_psign_w:
11052 case Intrinsic::x86_avx2_psign_d:
11053 return DAG.getNode(X86ISD::PSIGN, dl, Op.getValueType(),
11054 Op.getOperand(1), Op.getOperand(2));
11056 case Intrinsic::x86_sse41_insertps:
11057 return DAG.getNode(X86ISD::INSERTPS, dl, Op.getValueType(),
11058 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
11060 case Intrinsic::x86_avx_vperm2f128_ps_256:
11061 case Intrinsic::x86_avx_vperm2f128_pd_256:
11062 case Intrinsic::x86_avx_vperm2f128_si_256:
11063 case Intrinsic::x86_avx2_vperm2i128:
11064 return DAG.getNode(X86ISD::VPERM2X128, dl, Op.getValueType(),
11065 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
11067 case Intrinsic::x86_avx2_permd:
11068 case Intrinsic::x86_avx2_permps:
11069 // Operands intentionally swapped. Mask is last operand to intrinsic,
11070 // but second operand for node/intruction.
11071 return DAG.getNode(X86ISD::VPERMV, dl, Op.getValueType(),
11072 Op.getOperand(2), Op.getOperand(1));
11074 case Intrinsic::x86_sse_sqrt_ps:
11075 case Intrinsic::x86_sse2_sqrt_pd:
11076 case Intrinsic::x86_avx_sqrt_ps_256:
11077 case Intrinsic::x86_avx_sqrt_pd_256:
11078 return DAG.getNode(ISD::FSQRT, dl, Op.getValueType(), Op.getOperand(1));
11080 // ptest and testp intrinsics. The intrinsic these come from are designed to
11081 // return an integer value, not just an instruction so lower it to the ptest
11082 // or testp pattern and a setcc for the result.
11083 case Intrinsic::x86_sse41_ptestz:
11084 case Intrinsic::x86_sse41_ptestc:
11085 case Intrinsic::x86_sse41_ptestnzc:
11086 case Intrinsic::x86_avx_ptestz_256:
11087 case Intrinsic::x86_avx_ptestc_256:
11088 case Intrinsic::x86_avx_ptestnzc_256:
11089 case Intrinsic::x86_avx_vtestz_ps:
11090 case Intrinsic::x86_avx_vtestc_ps:
11091 case Intrinsic::x86_avx_vtestnzc_ps:
11092 case Intrinsic::x86_avx_vtestz_pd:
11093 case Intrinsic::x86_avx_vtestc_pd:
11094 case Intrinsic::x86_avx_vtestnzc_pd:
11095 case Intrinsic::x86_avx_vtestz_ps_256:
11096 case Intrinsic::x86_avx_vtestc_ps_256:
11097 case Intrinsic::x86_avx_vtestnzc_ps_256:
11098 case Intrinsic::x86_avx_vtestz_pd_256:
11099 case Intrinsic::x86_avx_vtestc_pd_256:
11100 case Intrinsic::x86_avx_vtestnzc_pd_256: {
11101 bool IsTestPacked = false;
11104 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
11105 case Intrinsic::x86_avx_vtestz_ps:
11106 case Intrinsic::x86_avx_vtestz_pd:
11107 case Intrinsic::x86_avx_vtestz_ps_256:
11108 case Intrinsic::x86_avx_vtestz_pd_256:
11109 IsTestPacked = true; // Fallthrough
11110 case Intrinsic::x86_sse41_ptestz:
11111 case Intrinsic::x86_avx_ptestz_256:
11113 X86CC = X86::COND_E;
11115 case Intrinsic::x86_avx_vtestc_ps:
11116 case Intrinsic::x86_avx_vtestc_pd:
11117 case Intrinsic::x86_avx_vtestc_ps_256:
11118 case Intrinsic::x86_avx_vtestc_pd_256:
11119 IsTestPacked = true; // Fallthrough
11120 case Intrinsic::x86_sse41_ptestc:
11121 case Intrinsic::x86_avx_ptestc_256:
11123 X86CC = X86::COND_B;
11125 case Intrinsic::x86_avx_vtestnzc_ps:
11126 case Intrinsic::x86_avx_vtestnzc_pd:
11127 case Intrinsic::x86_avx_vtestnzc_ps_256:
11128 case Intrinsic::x86_avx_vtestnzc_pd_256:
11129 IsTestPacked = true; // Fallthrough
11130 case Intrinsic::x86_sse41_ptestnzc:
11131 case Intrinsic::x86_avx_ptestnzc_256:
11133 X86CC = X86::COND_A;
11137 SDValue LHS = Op.getOperand(1);
11138 SDValue RHS = Op.getOperand(2);
11139 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
11140 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
11141 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
11142 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
11143 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
11146 // SSE/AVX shift intrinsics
11147 case Intrinsic::x86_sse2_psll_w:
11148 case Intrinsic::x86_sse2_psll_d:
11149 case Intrinsic::x86_sse2_psll_q:
11150 case Intrinsic::x86_avx2_psll_w:
11151 case Intrinsic::x86_avx2_psll_d:
11152 case Intrinsic::x86_avx2_psll_q:
11153 case Intrinsic::x86_sse2_psrl_w:
11154 case Intrinsic::x86_sse2_psrl_d:
11155 case Intrinsic::x86_sse2_psrl_q:
11156 case Intrinsic::x86_avx2_psrl_w:
11157 case Intrinsic::x86_avx2_psrl_d:
11158 case Intrinsic::x86_avx2_psrl_q:
11159 case Intrinsic::x86_sse2_psra_w:
11160 case Intrinsic::x86_sse2_psra_d:
11161 case Intrinsic::x86_avx2_psra_w:
11162 case Intrinsic::x86_avx2_psra_d: {
11165 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
11166 case Intrinsic::x86_sse2_psll_w:
11167 case Intrinsic::x86_sse2_psll_d:
11168 case Intrinsic::x86_sse2_psll_q:
11169 case Intrinsic::x86_avx2_psll_w:
11170 case Intrinsic::x86_avx2_psll_d:
11171 case Intrinsic::x86_avx2_psll_q:
11172 Opcode = X86ISD::VSHL;
11174 case Intrinsic::x86_sse2_psrl_w:
11175 case Intrinsic::x86_sse2_psrl_d:
11176 case Intrinsic::x86_sse2_psrl_q:
11177 case Intrinsic::x86_avx2_psrl_w:
11178 case Intrinsic::x86_avx2_psrl_d:
11179 case Intrinsic::x86_avx2_psrl_q:
11180 Opcode = X86ISD::VSRL;
11182 case Intrinsic::x86_sse2_psra_w:
11183 case Intrinsic::x86_sse2_psra_d:
11184 case Intrinsic::x86_avx2_psra_w:
11185 case Intrinsic::x86_avx2_psra_d:
11186 Opcode = X86ISD::VSRA;
11189 return DAG.getNode(Opcode, dl, Op.getValueType(),
11190 Op.getOperand(1), Op.getOperand(2));
11193 // SSE/AVX immediate shift intrinsics
11194 case Intrinsic::x86_sse2_pslli_w:
11195 case Intrinsic::x86_sse2_pslli_d:
11196 case Intrinsic::x86_sse2_pslli_q:
11197 case Intrinsic::x86_avx2_pslli_w:
11198 case Intrinsic::x86_avx2_pslli_d:
11199 case Intrinsic::x86_avx2_pslli_q:
11200 case Intrinsic::x86_sse2_psrli_w:
11201 case Intrinsic::x86_sse2_psrli_d:
11202 case Intrinsic::x86_sse2_psrli_q:
11203 case Intrinsic::x86_avx2_psrli_w:
11204 case Intrinsic::x86_avx2_psrli_d:
11205 case Intrinsic::x86_avx2_psrli_q:
11206 case Intrinsic::x86_sse2_psrai_w:
11207 case Intrinsic::x86_sse2_psrai_d:
11208 case Intrinsic::x86_avx2_psrai_w:
11209 case Intrinsic::x86_avx2_psrai_d: {
11212 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
11213 case Intrinsic::x86_sse2_pslli_w:
11214 case Intrinsic::x86_sse2_pslli_d:
11215 case Intrinsic::x86_sse2_pslli_q:
11216 case Intrinsic::x86_avx2_pslli_w:
11217 case Intrinsic::x86_avx2_pslli_d:
11218 case Intrinsic::x86_avx2_pslli_q:
11219 Opcode = X86ISD::VSHLI;
11221 case Intrinsic::x86_sse2_psrli_w:
11222 case Intrinsic::x86_sse2_psrli_d:
11223 case Intrinsic::x86_sse2_psrli_q:
11224 case Intrinsic::x86_avx2_psrli_w:
11225 case Intrinsic::x86_avx2_psrli_d:
11226 case Intrinsic::x86_avx2_psrli_q:
11227 Opcode = X86ISD::VSRLI;
11229 case Intrinsic::x86_sse2_psrai_w:
11230 case Intrinsic::x86_sse2_psrai_d:
11231 case Intrinsic::x86_avx2_psrai_w:
11232 case Intrinsic::x86_avx2_psrai_d:
11233 Opcode = X86ISD::VSRAI;
11236 return getTargetVShiftNode(Opcode, dl, Op.getValueType(),
11237 Op.getOperand(1), Op.getOperand(2), DAG);
11240 case Intrinsic::x86_sse42_pcmpistria128:
11241 case Intrinsic::x86_sse42_pcmpestria128:
11242 case Intrinsic::x86_sse42_pcmpistric128:
11243 case Intrinsic::x86_sse42_pcmpestric128:
11244 case Intrinsic::x86_sse42_pcmpistrio128:
11245 case Intrinsic::x86_sse42_pcmpestrio128:
11246 case Intrinsic::x86_sse42_pcmpistris128:
11247 case Intrinsic::x86_sse42_pcmpestris128:
11248 case Intrinsic::x86_sse42_pcmpistriz128:
11249 case Intrinsic::x86_sse42_pcmpestriz128: {
11253 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
11254 case Intrinsic::x86_sse42_pcmpistria128:
11255 Opcode = X86ISD::PCMPISTRI;
11256 X86CC = X86::COND_A;
11258 case Intrinsic::x86_sse42_pcmpestria128:
11259 Opcode = X86ISD::PCMPESTRI;
11260 X86CC = X86::COND_A;
11262 case Intrinsic::x86_sse42_pcmpistric128:
11263 Opcode = X86ISD::PCMPISTRI;
11264 X86CC = X86::COND_B;
11266 case Intrinsic::x86_sse42_pcmpestric128:
11267 Opcode = X86ISD::PCMPESTRI;
11268 X86CC = X86::COND_B;
11270 case Intrinsic::x86_sse42_pcmpistrio128:
11271 Opcode = X86ISD::PCMPISTRI;
11272 X86CC = X86::COND_O;
11274 case Intrinsic::x86_sse42_pcmpestrio128:
11275 Opcode = X86ISD::PCMPESTRI;
11276 X86CC = X86::COND_O;
11278 case Intrinsic::x86_sse42_pcmpistris128:
11279 Opcode = X86ISD::PCMPISTRI;
11280 X86CC = X86::COND_S;
11282 case Intrinsic::x86_sse42_pcmpestris128:
11283 Opcode = X86ISD::PCMPESTRI;
11284 X86CC = X86::COND_S;
11286 case Intrinsic::x86_sse42_pcmpistriz128:
11287 Opcode = X86ISD::PCMPISTRI;
11288 X86CC = X86::COND_E;
11290 case Intrinsic::x86_sse42_pcmpestriz128:
11291 Opcode = X86ISD::PCMPESTRI;
11292 X86CC = X86::COND_E;
11295 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
11296 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
11297 SDValue PCMP = DAG.getNode(Opcode, dl, VTs, NewOps.data(), NewOps.size());
11298 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
11299 DAG.getConstant(X86CC, MVT::i8),
11300 SDValue(PCMP.getNode(), 1));
11301 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
11304 case Intrinsic::x86_sse42_pcmpistri128:
11305 case Intrinsic::x86_sse42_pcmpestri128: {
11307 if (IntNo == Intrinsic::x86_sse42_pcmpistri128)
11308 Opcode = X86ISD::PCMPISTRI;
11310 Opcode = X86ISD::PCMPESTRI;
11312 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
11313 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
11314 return DAG.getNode(Opcode, dl, VTs, NewOps.data(), NewOps.size());
11316 case Intrinsic::x86_fma_vfmadd_ps:
11317 case Intrinsic::x86_fma_vfmadd_pd:
11318 case Intrinsic::x86_fma_vfmsub_ps:
11319 case Intrinsic::x86_fma_vfmsub_pd:
11320 case Intrinsic::x86_fma_vfnmadd_ps:
11321 case Intrinsic::x86_fma_vfnmadd_pd:
11322 case Intrinsic::x86_fma_vfnmsub_ps:
11323 case Intrinsic::x86_fma_vfnmsub_pd:
11324 case Intrinsic::x86_fma_vfmaddsub_ps:
11325 case Intrinsic::x86_fma_vfmaddsub_pd:
11326 case Intrinsic::x86_fma_vfmsubadd_ps:
11327 case Intrinsic::x86_fma_vfmsubadd_pd:
11328 case Intrinsic::x86_fma_vfmadd_ps_256:
11329 case Intrinsic::x86_fma_vfmadd_pd_256:
11330 case Intrinsic::x86_fma_vfmsub_ps_256:
11331 case Intrinsic::x86_fma_vfmsub_pd_256:
11332 case Intrinsic::x86_fma_vfnmadd_ps_256:
11333 case Intrinsic::x86_fma_vfnmadd_pd_256:
11334 case Intrinsic::x86_fma_vfnmsub_ps_256:
11335 case Intrinsic::x86_fma_vfnmsub_pd_256:
11336 case Intrinsic::x86_fma_vfmaddsub_ps_256:
11337 case Intrinsic::x86_fma_vfmaddsub_pd_256:
11338 case Intrinsic::x86_fma_vfmsubadd_ps_256:
11339 case Intrinsic::x86_fma_vfmsubadd_pd_256: {
11342 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
11343 case Intrinsic::x86_fma_vfmadd_ps:
11344 case Intrinsic::x86_fma_vfmadd_pd:
11345 case Intrinsic::x86_fma_vfmadd_ps_256:
11346 case Intrinsic::x86_fma_vfmadd_pd_256:
11347 Opc = X86ISD::FMADD;
11349 case Intrinsic::x86_fma_vfmsub_ps:
11350 case Intrinsic::x86_fma_vfmsub_pd:
11351 case Intrinsic::x86_fma_vfmsub_ps_256:
11352 case Intrinsic::x86_fma_vfmsub_pd_256:
11353 Opc = X86ISD::FMSUB;
11355 case Intrinsic::x86_fma_vfnmadd_ps:
11356 case Intrinsic::x86_fma_vfnmadd_pd:
11357 case Intrinsic::x86_fma_vfnmadd_ps_256:
11358 case Intrinsic::x86_fma_vfnmadd_pd_256:
11359 Opc = X86ISD::FNMADD;
11361 case Intrinsic::x86_fma_vfnmsub_ps:
11362 case Intrinsic::x86_fma_vfnmsub_pd:
11363 case Intrinsic::x86_fma_vfnmsub_ps_256:
11364 case Intrinsic::x86_fma_vfnmsub_pd_256:
11365 Opc = X86ISD::FNMSUB;
11367 case Intrinsic::x86_fma_vfmaddsub_ps:
11368 case Intrinsic::x86_fma_vfmaddsub_pd:
11369 case Intrinsic::x86_fma_vfmaddsub_ps_256:
11370 case Intrinsic::x86_fma_vfmaddsub_pd_256:
11371 Opc = X86ISD::FMADDSUB;
11373 case Intrinsic::x86_fma_vfmsubadd_ps:
11374 case Intrinsic::x86_fma_vfmsubadd_pd:
11375 case Intrinsic::x86_fma_vfmsubadd_ps_256:
11376 case Intrinsic::x86_fma_vfmsubadd_pd_256:
11377 Opc = X86ISD::FMSUBADD;
11381 return DAG.getNode(Opc, dl, Op.getValueType(), Op.getOperand(1),
11382 Op.getOperand(2), Op.getOperand(3));
11387 static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, SelectionDAG &DAG) {
11389 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11391 default: return SDValue(); // Don't custom lower most intrinsics.
11393 // RDRAND/RDSEED intrinsics.
11394 case Intrinsic::x86_rdrand_16:
11395 case Intrinsic::x86_rdrand_32:
11396 case Intrinsic::x86_rdrand_64:
11397 case Intrinsic::x86_rdseed_16:
11398 case Intrinsic::x86_rdseed_32:
11399 case Intrinsic::x86_rdseed_64: {
11400 unsigned Opcode = (IntNo == Intrinsic::x86_rdseed_16 ||
11401 IntNo == Intrinsic::x86_rdseed_32 ||
11402 IntNo == Intrinsic::x86_rdseed_64) ? X86ISD::RDSEED :
11404 // Emit the node with the right value type.
11405 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other);
11406 SDValue Result = DAG.getNode(Opcode, dl, VTs, Op.getOperand(0));
11408 // If the value returned by RDRAND/RDSEED was valid (CF=1), return 1.
11409 // Otherwise return the value from Rand, which is always 0, casted to i32.
11410 SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)),
11411 DAG.getConstant(1, Op->getValueType(1)),
11412 DAG.getConstant(X86::COND_B, MVT::i32),
11413 SDValue(Result.getNode(), 1) };
11414 SDValue isValid = DAG.getNode(X86ISD::CMOV, dl,
11415 DAG.getVTList(Op->getValueType(1), MVT::Glue),
11416 Ops, array_lengthof(Ops));
11418 // Return { result, isValid, chain }.
11419 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid,
11420 SDValue(Result.getNode(), 2));
11423 // XTEST intrinsics.
11424 case Intrinsic::x86_xtest: {
11425 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
11426 SDValue InTrans = DAG.getNode(X86ISD::XTEST, dl, VTs, Op.getOperand(0));
11427 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
11428 DAG.getConstant(X86::COND_NE, MVT::i8),
11430 SDValue Ret = DAG.getNode(ISD::ZERO_EXTEND, dl, Op->getValueType(0), SetCC);
11431 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(),
11432 Ret, SDValue(InTrans.getNode(), 1));
11437 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
11438 SelectionDAG &DAG) const {
11439 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
11440 MFI->setReturnAddressIsTaken(true);
11442 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
11444 EVT PtrVT = getPointerTy();
11447 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
11448 const X86RegisterInfo *RegInfo =
11449 static_cast<const X86RegisterInfo*>(getTargetMachine().getRegisterInfo());
11450 SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), PtrVT);
11451 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
11452 DAG.getNode(ISD::ADD, dl, PtrVT,
11453 FrameAddr, Offset),
11454 MachinePointerInfo(), false, false, false, 0);
11457 // Just load the return address.
11458 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
11459 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
11460 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
11463 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
11464 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
11465 MFI->setFrameAddressIsTaken(true);
11467 EVT VT = Op.getValueType();
11468 SDLoc dl(Op); // FIXME probably not meaningful
11469 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
11470 const X86RegisterInfo *RegInfo =
11471 static_cast<const X86RegisterInfo*>(getTargetMachine().getRegisterInfo());
11472 unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
11473 assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
11474 (FrameReg == X86::EBP && VT == MVT::i32)) &&
11475 "Invalid Frame Register!");
11476 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
11478 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
11479 MachinePointerInfo(),
11480 false, false, false, 0);
11484 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
11485 SelectionDAG &DAG) const {
11486 const X86RegisterInfo *RegInfo =
11487 static_cast<const X86RegisterInfo*>(getTargetMachine().getRegisterInfo());
11488 return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize());
11491 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
11492 SDValue Chain = Op.getOperand(0);
11493 SDValue Offset = Op.getOperand(1);
11494 SDValue Handler = Op.getOperand(2);
11497 EVT PtrVT = getPointerTy();
11498 const X86RegisterInfo *RegInfo =
11499 static_cast<const X86RegisterInfo*>(getTargetMachine().getRegisterInfo());
11500 unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
11501 assert(((FrameReg == X86::RBP && PtrVT == MVT::i64) ||
11502 (FrameReg == X86::EBP && PtrVT == MVT::i32)) &&
11503 "Invalid Frame Register!");
11504 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, PtrVT);
11505 unsigned StoreAddrReg = (PtrVT == MVT::i64) ? X86::RCX : X86::ECX;
11507 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Frame,
11508 DAG.getIntPtrConstant(RegInfo->getSlotSize()));
11509 StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, StoreAddr, Offset);
11510 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
11512 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
11514 return DAG.getNode(X86ISD::EH_RETURN, dl, MVT::Other, Chain,
11515 DAG.getRegister(StoreAddrReg, PtrVT));
11518 SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
11519 SelectionDAG &DAG) const {
11521 return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL,
11522 DAG.getVTList(MVT::i32, MVT::Other),
11523 Op.getOperand(0), Op.getOperand(1));
11526 SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
11527 SelectionDAG &DAG) const {
11529 return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
11530 Op.getOperand(0), Op.getOperand(1));
11533 static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
11534 return Op.getOperand(0);
11537 SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
11538 SelectionDAG &DAG) const {
11539 SDValue Root = Op.getOperand(0);
11540 SDValue Trmp = Op.getOperand(1); // trampoline
11541 SDValue FPtr = Op.getOperand(2); // nested function
11542 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
11545 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
11546 const TargetRegisterInfo* TRI = getTargetMachine().getRegisterInfo();
11548 if (Subtarget->is64Bit()) {
11549 SDValue OutChains[6];
11551 // Large code-model.
11552 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
11553 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
11555 const unsigned char N86R10 = TRI->getEncodingValue(X86::R10) & 0x7;
11556 const unsigned char N86R11 = TRI->getEncodingValue(X86::R11) & 0x7;
11558 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
11560 // Load the pointer to the nested function into R11.
11561 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
11562 SDValue Addr = Trmp;
11563 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
11564 Addr, MachinePointerInfo(TrmpAddr),
11567 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
11568 DAG.getConstant(2, MVT::i64));
11569 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
11570 MachinePointerInfo(TrmpAddr, 2),
11573 // Load the 'nest' parameter value into R10.
11574 // R10 is specified in X86CallingConv.td
11575 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
11576 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
11577 DAG.getConstant(10, MVT::i64));
11578 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
11579 Addr, MachinePointerInfo(TrmpAddr, 10),
11582 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
11583 DAG.getConstant(12, MVT::i64));
11584 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
11585 MachinePointerInfo(TrmpAddr, 12),
11588 // Jump to the nested function.
11589 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
11590 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
11591 DAG.getConstant(20, MVT::i64));
11592 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
11593 Addr, MachinePointerInfo(TrmpAddr, 20),
11596 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
11597 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
11598 DAG.getConstant(22, MVT::i64));
11599 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
11600 MachinePointerInfo(TrmpAddr, 22),
11603 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6);
11605 const Function *Func =
11606 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
11607 CallingConv::ID CC = Func->getCallingConv();
11612 llvm_unreachable("Unsupported calling convention");
11613 case CallingConv::C:
11614 case CallingConv::X86_StdCall: {
11615 // Pass 'nest' parameter in ECX.
11616 // Must be kept in sync with X86CallingConv.td
11617 NestReg = X86::ECX;
11619 // Check that ECX wasn't needed by an 'inreg' parameter.
11620 FunctionType *FTy = Func->getFunctionType();
11621 const AttributeSet &Attrs = Func->getAttributes();
11623 if (!Attrs.isEmpty() && !Func->isVarArg()) {
11624 unsigned InRegCount = 0;
11627 for (FunctionType::param_iterator I = FTy->param_begin(),
11628 E = FTy->param_end(); I != E; ++I, ++Idx)
11629 if (Attrs.hasAttribute(Idx, Attribute::InReg))
11630 // FIXME: should only count parameters that are lowered to integers.
11631 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
11633 if (InRegCount > 2) {
11634 report_fatal_error("Nest register in use - reduce number of inreg"
11640 case CallingConv::X86_FastCall:
11641 case CallingConv::X86_ThisCall:
11642 case CallingConv::Fast:
11643 // Pass 'nest' parameter in EAX.
11644 // Must be kept in sync with X86CallingConv.td
11645 NestReg = X86::EAX;
11649 SDValue OutChains[4];
11650 SDValue Addr, Disp;
11652 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
11653 DAG.getConstant(10, MVT::i32));
11654 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
11656 // This is storing the opcode for MOV32ri.
11657 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
11658 const unsigned char N86Reg = TRI->getEncodingValue(NestReg) & 0x7;
11659 OutChains[0] = DAG.getStore(Root, dl,
11660 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
11661 Trmp, MachinePointerInfo(TrmpAddr),
11664 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
11665 DAG.getConstant(1, MVT::i32));
11666 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
11667 MachinePointerInfo(TrmpAddr, 1),
11670 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
11671 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
11672 DAG.getConstant(5, MVT::i32));
11673 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
11674 MachinePointerInfo(TrmpAddr, 5),
11677 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
11678 DAG.getConstant(6, MVT::i32));
11679 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
11680 MachinePointerInfo(TrmpAddr, 6),
11683 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4);
11687 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
11688 SelectionDAG &DAG) const {
11690 The rounding mode is in bits 11:10 of FPSR, and has the following
11692 00 Round to nearest
11697 FLT_ROUNDS, on the other hand, expects the following:
11704 To perform the conversion, we do:
11705 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
11708 MachineFunction &MF = DAG.getMachineFunction();
11709 const TargetMachine &TM = MF.getTarget();
11710 const TargetFrameLowering &TFI = *TM.getFrameLowering();
11711 unsigned StackAlignment = TFI.getStackAlignment();
11712 EVT VT = Op.getValueType();
11715 // Save FP Control Word to stack slot
11716 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
11717 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
11719 MachineMemOperand *MMO =
11720 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11721 MachineMemOperand::MOStore, 2, 2);
11723 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
11724 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
11725 DAG.getVTList(MVT::Other),
11726 Ops, array_lengthof(Ops), MVT::i16,
11729 // Load FP Control Word from stack slot
11730 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
11731 MachinePointerInfo(), false, false, false, 0);
11733 // Transform as necessary
11735 DAG.getNode(ISD::SRL, DL, MVT::i16,
11736 DAG.getNode(ISD::AND, DL, MVT::i16,
11737 CWD, DAG.getConstant(0x800, MVT::i16)),
11738 DAG.getConstant(11, MVT::i8));
11740 DAG.getNode(ISD::SRL, DL, MVT::i16,
11741 DAG.getNode(ISD::AND, DL, MVT::i16,
11742 CWD, DAG.getConstant(0x400, MVT::i16)),
11743 DAG.getConstant(9, MVT::i8));
11746 DAG.getNode(ISD::AND, DL, MVT::i16,
11747 DAG.getNode(ISD::ADD, DL, MVT::i16,
11748 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
11749 DAG.getConstant(1, MVT::i16)),
11750 DAG.getConstant(3, MVT::i16));
11752 return DAG.getNode((VT.getSizeInBits() < 16 ?
11753 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
11756 static SDValue LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
11757 EVT VT = Op.getValueType();
11759 unsigned NumBits = VT.getSizeInBits();
11762 Op = Op.getOperand(0);
11763 if (VT == MVT::i8) {
11764 // Zero extend to i32 since there is not an i8 bsr.
11766 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
11769 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
11770 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
11771 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
11773 // If src is zero (i.e. bsr sets ZF), returns NumBits.
11776 DAG.getConstant(NumBits+NumBits-1, OpVT),
11777 DAG.getConstant(X86::COND_E, MVT::i8),
11780 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
11782 // Finally xor with NumBits-1.
11783 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
11786 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
11790 static SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, SelectionDAG &DAG) {
11791 EVT VT = Op.getValueType();
11793 unsigned NumBits = VT.getSizeInBits();
11796 Op = Op.getOperand(0);
11797 if (VT == MVT::i8) {
11798 // Zero extend to i32 since there is not an i8 bsr.
11800 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
11803 // Issue a bsr (scan bits in reverse).
11804 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
11805 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
11807 // And xor with NumBits-1.
11808 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
11811 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
11815 static SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
11816 EVT VT = Op.getValueType();
11817 unsigned NumBits = VT.getSizeInBits();
11819 Op = Op.getOperand(0);
11821 // Issue a bsf (scan bits forward) which also sets EFLAGS.
11822 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
11823 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
11825 // If src is zero (i.e. bsf sets ZF), returns NumBits.
11828 DAG.getConstant(NumBits, VT),
11829 DAG.getConstant(X86::COND_E, MVT::i8),
11832 return DAG.getNode(X86ISD::CMOV, dl, VT, Ops, array_lengthof(Ops));
11835 // Lower256IntArith - Break a 256-bit integer operation into two new 128-bit
11836 // ones, and then concatenate the result back.
11837 static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) {
11838 EVT VT = Op.getValueType();
11840 assert(VT.is256BitVector() && VT.isInteger() &&
11841 "Unsupported value type for operation");
11843 unsigned NumElems = VT.getVectorNumElements();
11846 // Extract the LHS vectors
11847 SDValue LHS = Op.getOperand(0);
11848 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
11849 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
11851 // Extract the RHS vectors
11852 SDValue RHS = Op.getOperand(1);
11853 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
11854 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
11856 MVT EltVT = VT.getVectorElementType().getSimpleVT();
11857 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
11859 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
11860 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1),
11861 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
11864 static SDValue LowerADD(SDValue Op, SelectionDAG &DAG) {
11865 assert(Op.getValueType().is256BitVector() &&
11866 Op.getValueType().isInteger() &&
11867 "Only handle AVX 256-bit vector integer operation");
11868 return Lower256IntArith(Op, DAG);
11871 static SDValue LowerSUB(SDValue Op, SelectionDAG &DAG) {
11872 assert(Op.getValueType().is256BitVector() &&
11873 Op.getValueType().isInteger() &&
11874 "Only handle AVX 256-bit vector integer operation");
11875 return Lower256IntArith(Op, DAG);
11878 static SDValue LowerMUL(SDValue Op, const X86Subtarget *Subtarget,
11879 SelectionDAG &DAG) {
11881 EVT VT = Op.getValueType();
11883 // Decompose 256-bit ops into smaller 128-bit ops.
11884 if (VT.is256BitVector() && !Subtarget->hasInt256())
11885 return Lower256IntArith(Op, DAG);
11887 SDValue A = Op.getOperand(0);
11888 SDValue B = Op.getOperand(1);
11890 // Lower v4i32 mul as 2x shuffle, 2x pmuludq, 2x shuffle.
11891 if (VT == MVT::v4i32) {
11892 assert(Subtarget->hasSSE2() && !Subtarget->hasSSE41() &&
11893 "Should not custom lower when pmuldq is available!");
11895 // Extract the odd parts.
11896 static const int UnpackMask[] = { 1, -1, 3, -1 };
11897 SDValue Aodds = DAG.getVectorShuffle(VT, dl, A, A, UnpackMask);
11898 SDValue Bodds = DAG.getVectorShuffle(VT, dl, B, B, UnpackMask);
11900 // Multiply the even parts.
11901 SDValue Evens = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, A, B);
11902 // Now multiply odd parts.
11903 SDValue Odds = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, Aodds, Bodds);
11905 Evens = DAG.getNode(ISD::BITCAST, dl, VT, Evens);
11906 Odds = DAG.getNode(ISD::BITCAST, dl, VT, Odds);
11908 // Merge the two vectors back together with a shuffle. This expands into 2
11910 static const int ShufMask[] = { 0, 4, 2, 6 };
11911 return DAG.getVectorShuffle(VT, dl, Evens, Odds, ShufMask);
11914 assert((VT == MVT::v2i64 || VT == MVT::v4i64) &&
11915 "Only know how to lower V2I64/V4I64 multiply");
11917 // Ahi = psrlqi(a, 32);
11918 // Bhi = psrlqi(b, 32);
11920 // AloBlo = pmuludq(a, b);
11921 // AloBhi = pmuludq(a, Bhi);
11922 // AhiBlo = pmuludq(Ahi, b);
11924 // AloBhi = psllqi(AloBhi, 32);
11925 // AhiBlo = psllqi(AhiBlo, 32);
11926 // return AloBlo + AloBhi + AhiBlo;
11928 SDValue ShAmt = DAG.getConstant(32, MVT::i32);
11930 SDValue Ahi = DAG.getNode(X86ISD::VSRLI, dl, VT, A, ShAmt);
11931 SDValue Bhi = DAG.getNode(X86ISD::VSRLI, dl, VT, B, ShAmt);
11933 // Bit cast to 32-bit vectors for MULUDQ
11934 EVT MulVT = (VT == MVT::v2i64) ? MVT::v4i32 : MVT::v8i32;
11935 A = DAG.getNode(ISD::BITCAST, dl, MulVT, A);
11936 B = DAG.getNode(ISD::BITCAST, dl, MulVT, B);
11937 Ahi = DAG.getNode(ISD::BITCAST, dl, MulVT, Ahi);
11938 Bhi = DAG.getNode(ISD::BITCAST, dl, MulVT, Bhi);
11940 SDValue AloBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, B);
11941 SDValue AloBhi = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, Bhi);
11942 SDValue AhiBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Ahi, B);
11944 AloBhi = DAG.getNode(X86ISD::VSHLI, dl, VT, AloBhi, ShAmt);
11945 AhiBlo = DAG.getNode(X86ISD::VSHLI, dl, VT, AhiBlo, ShAmt);
11947 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
11948 return DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
11951 SDValue X86TargetLowering::LowerSDIV(SDValue Op, SelectionDAG &DAG) const {
11952 EVT VT = Op.getValueType();
11953 EVT EltTy = VT.getVectorElementType();
11954 unsigned NumElts = VT.getVectorNumElements();
11955 SDValue N0 = Op.getOperand(0);
11958 // Lower sdiv X, pow2-const.
11959 BuildVectorSDNode *C = dyn_cast<BuildVectorSDNode>(Op.getOperand(1));
11963 APInt SplatValue, SplatUndef;
11964 unsigned SplatBitSize;
11966 if (!C->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
11968 EltTy.getSizeInBits() < SplatBitSize)
11971 if ((SplatValue != 0) &&
11972 (SplatValue.isPowerOf2() || (-SplatValue).isPowerOf2())) {
11973 unsigned lg2 = SplatValue.countTrailingZeros();
11974 // Splat the sign bit.
11975 SDValue Sz = DAG.getConstant(EltTy.getSizeInBits()-1, MVT::i32);
11976 SDValue SGN = getTargetVShiftNode(X86ISD::VSRAI, dl, VT, N0, Sz, DAG);
11977 // Add (N0 < 0) ? abs2 - 1 : 0;
11978 SDValue Amt = DAG.getConstant(EltTy.getSizeInBits() - lg2, MVT::i32);
11979 SDValue SRL = getTargetVShiftNode(X86ISD::VSRLI, dl, VT, SGN, Amt, DAG);
11980 SDValue ADD = DAG.getNode(ISD::ADD, dl, VT, N0, SRL);
11981 SDValue Lg2Amt = DAG.getConstant(lg2, MVT::i32);
11982 SDValue SRA = getTargetVShiftNode(X86ISD::VSRAI, dl, VT, ADD, Lg2Amt, DAG);
11984 // If we're dividing by a positive value, we're done. Otherwise, we must
11985 // negate the result.
11986 if (SplatValue.isNonNegative())
11989 SmallVector<SDValue, 16> V(NumElts, DAG.getConstant(0, EltTy));
11990 SDValue Zero = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &V[0], NumElts);
11991 return DAG.getNode(ISD::SUB, dl, VT, Zero, SRA);
11996 static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
11997 const X86Subtarget *Subtarget) {
11998 EVT VT = Op.getValueType();
12000 SDValue R = Op.getOperand(0);
12001 SDValue Amt = Op.getOperand(1);
12003 // Optimize shl/srl/sra with constant shift amount.
12004 if (isSplatVector(Amt.getNode())) {
12005 SDValue SclrAmt = Amt->getOperand(0);
12006 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(SclrAmt)) {
12007 uint64_t ShiftAmt = C->getZExtValue();
12009 if (VT == MVT::v2i64 || VT == MVT::v4i32 || VT == MVT::v8i16 ||
12010 (Subtarget->hasInt256() &&
12011 (VT == MVT::v4i64 || VT == MVT::v8i32 || VT == MVT::v16i16))) {
12012 if (Op.getOpcode() == ISD::SHL)
12013 return DAG.getNode(X86ISD::VSHLI, dl, VT, R,
12014 DAG.getConstant(ShiftAmt, MVT::i32));
12015 if (Op.getOpcode() == ISD::SRL)
12016 return DAG.getNode(X86ISD::VSRLI, dl, VT, R,
12017 DAG.getConstant(ShiftAmt, MVT::i32));
12018 if (Op.getOpcode() == ISD::SRA && VT != MVT::v2i64 && VT != MVT::v4i64)
12019 return DAG.getNode(X86ISD::VSRAI, dl, VT, R,
12020 DAG.getConstant(ShiftAmt, MVT::i32));
12023 if (VT == MVT::v16i8) {
12024 if (Op.getOpcode() == ISD::SHL) {
12025 // Make a large shift.
12026 SDValue SHL = DAG.getNode(X86ISD::VSHLI, dl, MVT::v8i16, R,
12027 DAG.getConstant(ShiftAmt, MVT::i32));
12028 SHL = DAG.getNode(ISD::BITCAST, dl, VT, SHL);
12029 // Zero out the rightmost bits.
12030 SmallVector<SDValue, 16> V(16,
12031 DAG.getConstant(uint8_t(-1U << ShiftAmt),
12033 return DAG.getNode(ISD::AND, dl, VT, SHL,
12034 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &V[0], 16));
12036 if (Op.getOpcode() == ISD::SRL) {
12037 // Make a large shift.
12038 SDValue SRL = DAG.getNode(X86ISD::VSRLI, dl, MVT::v8i16, R,
12039 DAG.getConstant(ShiftAmt, MVT::i32));
12040 SRL = DAG.getNode(ISD::BITCAST, dl, VT, SRL);
12041 // Zero out the leftmost bits.
12042 SmallVector<SDValue, 16> V(16,
12043 DAG.getConstant(uint8_t(-1U) >> ShiftAmt,
12045 return DAG.getNode(ISD::AND, dl, VT, SRL,
12046 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &V[0], 16));
12048 if (Op.getOpcode() == ISD::SRA) {
12049 if (ShiftAmt == 7) {
12050 // R s>> 7 === R s< 0
12051 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
12052 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
12055 // R s>> a === ((R u>> a) ^ m) - m
12056 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
12057 SmallVector<SDValue, 16> V(16, DAG.getConstant(128 >> ShiftAmt,
12059 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &V[0], 16);
12060 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
12061 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
12064 llvm_unreachable("Unknown shift opcode.");
12067 if (Subtarget->hasInt256() && VT == MVT::v32i8) {
12068 if (Op.getOpcode() == ISD::SHL) {
12069 // Make a large shift.
12070 SDValue SHL = DAG.getNode(X86ISD::VSHLI, dl, MVT::v16i16, R,
12071 DAG.getConstant(ShiftAmt, MVT::i32));
12072 SHL = DAG.getNode(ISD::BITCAST, dl, VT, SHL);
12073 // Zero out the rightmost bits.
12074 SmallVector<SDValue, 32> V(32,
12075 DAG.getConstant(uint8_t(-1U << ShiftAmt),
12077 return DAG.getNode(ISD::AND, dl, VT, SHL,
12078 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &V[0], 32));
12080 if (Op.getOpcode() == ISD::SRL) {
12081 // Make a large shift.
12082 SDValue SRL = DAG.getNode(X86ISD::VSRLI, dl, MVT::v16i16, R,
12083 DAG.getConstant(ShiftAmt, MVT::i32));
12084 SRL = DAG.getNode(ISD::BITCAST, dl, VT, SRL);
12085 // Zero out the leftmost bits.
12086 SmallVector<SDValue, 32> V(32,
12087 DAG.getConstant(uint8_t(-1U) >> ShiftAmt,
12089 return DAG.getNode(ISD::AND, dl, VT, SRL,
12090 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &V[0], 32));
12092 if (Op.getOpcode() == ISD::SRA) {
12093 if (ShiftAmt == 7) {
12094 // R s>> 7 === R s< 0
12095 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
12096 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
12099 // R s>> a === ((R u>> a) ^ m) - m
12100 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
12101 SmallVector<SDValue, 32> V(32, DAG.getConstant(128 >> ShiftAmt,
12103 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &V[0], 32);
12104 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
12105 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
12108 llvm_unreachable("Unknown shift opcode.");
12113 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
12114 if (!Subtarget->is64Bit() &&
12115 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
12116 Amt.getOpcode() == ISD::BITCAST &&
12117 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
12118 Amt = Amt.getOperand(0);
12119 unsigned Ratio = Amt.getValueType().getVectorNumElements() /
12120 VT.getVectorNumElements();
12121 unsigned RatioInLog2 = Log2_32_Ceil(Ratio);
12122 uint64_t ShiftAmt = 0;
12123 for (unsigned i = 0; i != Ratio; ++i) {
12124 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i));
12128 ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2)));
12130 // Check remaining shift amounts.
12131 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
12132 uint64_t ShAmt = 0;
12133 for (unsigned j = 0; j != Ratio; ++j) {
12134 ConstantSDNode *C =
12135 dyn_cast<ConstantSDNode>(Amt.getOperand(i + j));
12139 ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2)));
12141 if (ShAmt != ShiftAmt)
12144 switch (Op.getOpcode()) {
12146 llvm_unreachable("Unknown shift opcode!");
12148 return DAG.getNode(X86ISD::VSHLI, dl, VT, R,
12149 DAG.getConstant(ShiftAmt, MVT::i32));
12151 return DAG.getNode(X86ISD::VSRLI, dl, VT, R,
12152 DAG.getConstant(ShiftAmt, MVT::i32));
12154 return DAG.getNode(X86ISD::VSRAI, dl, VT, R,
12155 DAG.getConstant(ShiftAmt, MVT::i32));
12162 static SDValue LowerScalarVariableShift(SDValue Op, SelectionDAG &DAG,
12163 const X86Subtarget* Subtarget) {
12164 EVT VT = Op.getValueType();
12166 SDValue R = Op.getOperand(0);
12167 SDValue Amt = Op.getOperand(1);
12169 if ((VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) ||
12170 VT == MVT::v4i32 || VT == MVT::v8i16 ||
12171 (Subtarget->hasInt256() &&
12172 ((VT == MVT::v4i64 && Op.getOpcode() != ISD::SRA) ||
12173 VT == MVT::v8i32 || VT == MVT::v16i16))) {
12175 EVT EltVT = VT.getVectorElementType();
12177 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
12178 unsigned NumElts = VT.getVectorNumElements();
12180 for (i = 0; i != NumElts; ++i) {
12181 if (Amt.getOperand(i).getOpcode() == ISD::UNDEF)
12185 for (j = i; j != NumElts; ++j) {
12186 SDValue Arg = Amt.getOperand(j);
12187 if (Arg.getOpcode() == ISD::UNDEF) continue;
12188 if (Arg != Amt.getOperand(i))
12191 if (i != NumElts && j == NumElts)
12192 BaseShAmt = Amt.getOperand(i);
12194 if (Amt.getOpcode() == ISD::EXTRACT_SUBVECTOR)
12195 Amt = Amt.getOperand(0);
12196 if (Amt.getOpcode() == ISD::VECTOR_SHUFFLE &&
12197 cast<ShuffleVectorSDNode>(Amt)->isSplat()) {
12198 SDValue InVec = Amt.getOperand(0);
12199 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
12200 unsigned NumElts = InVec.getValueType().getVectorNumElements();
12202 for (; i != NumElts; ++i) {
12203 SDValue Arg = InVec.getOperand(i);
12204 if (Arg.getOpcode() == ISD::UNDEF) continue;
12208 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
12209 if (ConstantSDNode *C =
12210 dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
12211 unsigned SplatIdx =
12212 cast<ShuffleVectorSDNode>(Amt)->getSplatIndex();
12213 if (C->getZExtValue() == SplatIdx)
12214 BaseShAmt = InVec.getOperand(1);
12217 if (BaseShAmt.getNode() == 0)
12218 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, Amt,
12219 DAG.getIntPtrConstant(0));
12223 if (BaseShAmt.getNode()) {
12224 if (EltVT.bitsGT(MVT::i32))
12225 BaseShAmt = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BaseShAmt);
12226 else if (EltVT.bitsLT(MVT::i32))
12227 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt);
12229 switch (Op.getOpcode()) {
12231 llvm_unreachable("Unknown shift opcode!");
12233 switch (VT.getSimpleVT().SimpleTy) {
12234 default: return SDValue();
12241 return getTargetVShiftNode(X86ISD::VSHLI, dl, VT, R, BaseShAmt, DAG);
12244 switch (VT.getSimpleVT().SimpleTy) {
12245 default: return SDValue();
12250 return getTargetVShiftNode(X86ISD::VSRAI, dl, VT, R, BaseShAmt, DAG);
12253 switch (VT.getSimpleVT().SimpleTy) {
12254 default: return SDValue();
12261 return getTargetVShiftNode(X86ISD::VSRLI, dl, VT, R, BaseShAmt, DAG);
12267 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
12268 if (!Subtarget->is64Bit() &&
12269 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
12270 Amt.getOpcode() == ISD::BITCAST &&
12271 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
12272 Amt = Amt.getOperand(0);
12273 unsigned Ratio = Amt.getValueType().getVectorNumElements() /
12274 VT.getVectorNumElements();
12275 std::vector<SDValue> Vals(Ratio);
12276 for (unsigned i = 0; i != Ratio; ++i)
12277 Vals[i] = Amt.getOperand(i);
12278 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
12279 for (unsigned j = 0; j != Ratio; ++j)
12280 if (Vals[j] != Amt.getOperand(i + j))
12283 switch (Op.getOpcode()) {
12285 llvm_unreachable("Unknown shift opcode!");
12287 return DAG.getNode(X86ISD::VSHL, dl, VT, R, Op.getOperand(1));
12289 return DAG.getNode(X86ISD::VSRL, dl, VT, R, Op.getOperand(1));
12291 return DAG.getNode(X86ISD::VSRA, dl, VT, R, Op.getOperand(1));
12298 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
12300 EVT VT = Op.getValueType();
12302 SDValue R = Op.getOperand(0);
12303 SDValue Amt = Op.getOperand(1);
12306 if (!Subtarget->hasSSE2())
12309 V = LowerScalarImmediateShift(Op, DAG, Subtarget);
12313 V = LowerScalarVariableShift(Op, DAG, Subtarget);
12317 // AVX2 has VPSLLV/VPSRAV/VPSRLV.
12318 if (Subtarget->hasInt256()) {
12319 if (Op.getOpcode() == ISD::SRL &&
12320 (VT == MVT::v2i64 || VT == MVT::v4i32 ||
12321 VT == MVT::v4i64 || VT == MVT::v8i32))
12323 if (Op.getOpcode() == ISD::SHL &&
12324 (VT == MVT::v2i64 || VT == MVT::v4i32 ||
12325 VT == MVT::v4i64 || VT == MVT::v8i32))
12327 if (Op.getOpcode() == ISD::SRA && (VT == MVT::v4i32 || VT == MVT::v8i32))
12331 // Lower SHL with variable shift amount.
12332 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
12333 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, VT));
12335 Op = DAG.getNode(ISD::ADD, dl, VT, Op, DAG.getConstant(0x3f800000U, VT));
12336 Op = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, Op);
12337 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
12338 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
12340 if (VT == MVT::v16i8 && Op->getOpcode() == ISD::SHL) {
12341 assert(Subtarget->hasSSE2() && "Need SSE2 for pslli/pcmpeq.");
12344 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(5, VT));
12345 Op = DAG.getNode(ISD::BITCAST, dl, VT, Op);
12347 // Turn 'a' into a mask suitable for VSELECT
12348 SDValue VSelM = DAG.getConstant(0x80, VT);
12349 SDValue OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
12350 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
12352 SDValue CM1 = DAG.getConstant(0x0f, VT);
12353 SDValue CM2 = DAG.getConstant(0x3f, VT);
12355 // r = VSELECT(r, psllw(r & (char16)15, 4), a);
12356 SDValue M = DAG.getNode(ISD::AND, dl, VT, R, CM1);
12357 M = getTargetVShiftNode(X86ISD::VSHLI, dl, MVT::v8i16, M,
12358 DAG.getConstant(4, MVT::i32), DAG);
12359 M = DAG.getNode(ISD::BITCAST, dl, VT, M);
12360 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, M, R);
12363 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
12364 OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
12365 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
12367 // r = VSELECT(r, psllw(r & (char16)63, 2), a);
12368 M = DAG.getNode(ISD::AND, dl, VT, R, CM2);
12369 M = getTargetVShiftNode(X86ISD::VSHLI, dl, MVT::v8i16, M,
12370 DAG.getConstant(2, MVT::i32), DAG);
12371 M = DAG.getNode(ISD::BITCAST, dl, VT, M);
12372 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, M, R);
12375 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
12376 OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
12377 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
12379 // return VSELECT(r, r+r, a);
12380 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel,
12381 DAG.getNode(ISD::ADD, dl, VT, R, R), R);
12385 // Decompose 256-bit shifts into smaller 128-bit shifts.
12386 if (VT.is256BitVector()) {
12387 unsigned NumElems = VT.getVectorNumElements();
12388 MVT EltVT = VT.getVectorElementType().getSimpleVT();
12389 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
12391 // Extract the two vectors
12392 SDValue V1 = Extract128BitVector(R, 0, DAG, dl);
12393 SDValue V2 = Extract128BitVector(R, NumElems/2, DAG, dl);
12395 // Recreate the shift amount vectors
12396 SDValue Amt1, Amt2;
12397 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
12398 // Constant shift amount
12399 SmallVector<SDValue, 4> Amt1Csts;
12400 SmallVector<SDValue, 4> Amt2Csts;
12401 for (unsigned i = 0; i != NumElems/2; ++i)
12402 Amt1Csts.push_back(Amt->getOperand(i));
12403 for (unsigned i = NumElems/2; i != NumElems; ++i)
12404 Amt2Csts.push_back(Amt->getOperand(i));
12406 Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT,
12407 &Amt1Csts[0], NumElems/2);
12408 Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT,
12409 &Amt2Csts[0], NumElems/2);
12411 // Variable shift amount
12412 Amt1 = Extract128BitVector(Amt, 0, DAG, dl);
12413 Amt2 = Extract128BitVector(Amt, NumElems/2, DAG, dl);
12416 // Issue new vector shifts for the smaller types
12417 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
12418 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
12420 // Concatenate the result back
12421 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
12427 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
12428 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
12429 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
12430 // looks for this combo and may remove the "setcc" instruction if the "setcc"
12431 // has only one use.
12432 SDNode *N = Op.getNode();
12433 SDValue LHS = N->getOperand(0);
12434 SDValue RHS = N->getOperand(1);
12435 unsigned BaseOp = 0;
12438 switch (Op.getOpcode()) {
12439 default: llvm_unreachable("Unknown ovf instruction!");
12441 // A subtract of one will be selected as a INC. Note that INC doesn't
12442 // set CF, so we can't do this for UADDO.
12443 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
12445 BaseOp = X86ISD::INC;
12446 Cond = X86::COND_O;
12449 BaseOp = X86ISD::ADD;
12450 Cond = X86::COND_O;
12453 BaseOp = X86ISD::ADD;
12454 Cond = X86::COND_B;
12457 // A subtract of one will be selected as a DEC. Note that DEC doesn't
12458 // set CF, so we can't do this for USUBO.
12459 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
12461 BaseOp = X86ISD::DEC;
12462 Cond = X86::COND_O;
12465 BaseOp = X86ISD::SUB;
12466 Cond = X86::COND_O;
12469 BaseOp = X86ISD::SUB;
12470 Cond = X86::COND_B;
12473 BaseOp = X86ISD::SMUL;
12474 Cond = X86::COND_O;
12476 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
12477 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
12479 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
12482 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
12483 DAG.getConstant(X86::COND_O, MVT::i32),
12484 SDValue(Sum.getNode(), 2));
12486 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
12490 // Also sets EFLAGS.
12491 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
12492 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
12495 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
12496 DAG.getConstant(Cond, MVT::i32),
12497 SDValue(Sum.getNode(), 1));
12499 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
12502 SDValue X86TargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
12503 SelectionDAG &DAG) const {
12505 EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
12506 EVT VT = Op.getValueType();
12508 if (!Subtarget->hasSSE2() || !VT.isVector())
12511 unsigned BitsDiff = VT.getScalarType().getSizeInBits() -
12512 ExtraVT.getScalarType().getSizeInBits();
12513 SDValue ShAmt = DAG.getConstant(BitsDiff, MVT::i32);
12515 switch (VT.getSimpleVT().SimpleTy) {
12516 default: return SDValue();
12519 if (!Subtarget->hasFp256())
12521 if (!Subtarget->hasInt256()) {
12522 // needs to be split
12523 unsigned NumElems = VT.getVectorNumElements();
12525 // Extract the LHS vectors
12526 SDValue LHS = Op.getOperand(0);
12527 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
12528 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
12530 MVT EltVT = VT.getVectorElementType().getSimpleVT();
12531 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
12533 EVT ExtraEltVT = ExtraVT.getVectorElementType();
12534 unsigned ExtraNumElems = ExtraVT.getVectorNumElements();
12535 ExtraVT = EVT::getVectorVT(*DAG.getContext(), ExtraEltVT,
12537 SDValue Extra = DAG.getValueType(ExtraVT);
12539 LHS1 = DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, Extra);
12540 LHS2 = DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, Extra);
12542 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, LHS1, LHS2);
12547 // (sext (vzext x)) -> (vsext x)
12548 SDValue Op0 = Op.getOperand(0);
12549 SDValue Op00 = Op0.getOperand(0);
12551 // Hopefully, this VECTOR_SHUFFLE is just a VZEXT.
12552 if (Op0.getOpcode() == ISD::BITCAST &&
12553 Op00.getOpcode() == ISD::VECTOR_SHUFFLE)
12554 Tmp1 = LowerVectorIntExtend(Op00, DAG);
12555 if (Tmp1.getNode()) {
12556 SDValue Tmp1Op0 = Tmp1.getOperand(0);
12557 assert(Tmp1Op0.getOpcode() == X86ISD::VZEXT &&
12558 "This optimization is invalid without a VZEXT.");
12559 return DAG.getNode(X86ISD::VSEXT, dl, VT, Tmp1Op0.getOperand(0));
12562 // If the above didn't work, then just use Shift-Left + Shift-Right.
12563 Tmp1 = getTargetVShiftNode(X86ISD::VSHLI, dl, VT, Op0, ShAmt, DAG);
12564 return getTargetVShiftNode(X86ISD::VSRAI, dl, VT, Tmp1, ShAmt, DAG);
12569 static SDValue LowerATOMIC_FENCE(SDValue Op, const X86Subtarget *Subtarget,
12570 SelectionDAG &DAG) {
12572 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
12573 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
12574 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
12575 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
12577 // The only fence that needs an instruction is a sequentially-consistent
12578 // cross-thread fence.
12579 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
12580 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
12581 // no-sse2). There isn't any reason to disable it if the target processor
12583 if (Subtarget->hasSSE2() || Subtarget->is64Bit())
12584 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
12586 SDValue Chain = Op.getOperand(0);
12587 SDValue Zero = DAG.getConstant(0, MVT::i32);
12589 DAG.getRegister(X86::ESP, MVT::i32), // Base
12590 DAG.getTargetConstant(1, MVT::i8), // Scale
12591 DAG.getRegister(0, MVT::i32), // Index
12592 DAG.getTargetConstant(0, MVT::i32), // Disp
12593 DAG.getRegister(0, MVT::i32), // Segment.
12597 SDNode *Res = DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops);
12598 return SDValue(Res, 0);
12601 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
12602 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
12605 static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget *Subtarget,
12606 SelectionDAG &DAG) {
12607 EVT T = Op.getValueType();
12611 switch(T.getSimpleVT().SimpleTy) {
12612 default: llvm_unreachable("Invalid value type!");
12613 case MVT::i8: Reg = X86::AL; size = 1; break;
12614 case MVT::i16: Reg = X86::AX; size = 2; break;
12615 case MVT::i32: Reg = X86::EAX; size = 4; break;
12617 assert(Subtarget->is64Bit() && "Node not type legal!");
12618 Reg = X86::RAX; size = 8;
12621 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
12622 Op.getOperand(2), SDValue());
12623 SDValue Ops[] = { cpIn.getValue(0),
12626 DAG.getTargetConstant(size, MVT::i8),
12627 cpIn.getValue(1) };
12628 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
12629 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
12630 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
12631 Ops, array_lengthof(Ops), T, MMO);
12633 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
12637 static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget *Subtarget,
12638 SelectionDAG &DAG) {
12639 assert(Subtarget->is64Bit() && "Result not type legalized?");
12640 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
12641 SDValue TheChain = Op.getOperand(0);
12643 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
12644 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
12645 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
12647 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
12648 DAG.getConstant(32, MVT::i8));
12650 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
12653 return DAG.getMergeValues(Ops, array_lengthof(Ops), dl);
12656 SDValue X86TargetLowering::LowerBITCAST(SDValue Op, SelectionDAG &DAG) const {
12657 EVT SrcVT = Op.getOperand(0).getValueType();
12658 EVT DstVT = Op.getValueType();
12659 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
12660 Subtarget->hasMMX() && "Unexpected custom BITCAST");
12661 assert((DstVT == MVT::i64 ||
12662 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
12663 "Unexpected custom BITCAST");
12664 // i64 <=> MMX conversions are Legal.
12665 if (SrcVT==MVT::i64 && DstVT.isVector())
12667 if (DstVT==MVT::i64 && SrcVT.isVector())
12669 // MMX <=> MMX conversions are Legal.
12670 if (SrcVT.isVector() && DstVT.isVector())
12672 // All other conversions need to be expanded.
12676 static SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
12677 SDNode *Node = Op.getNode();
12679 EVT T = Node->getValueType(0);
12680 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
12681 DAG.getConstant(0, T), Node->getOperand(2));
12682 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
12683 cast<AtomicSDNode>(Node)->getMemoryVT(),
12684 Node->getOperand(0),
12685 Node->getOperand(1), negOp,
12686 cast<AtomicSDNode>(Node)->getSrcValue(),
12687 cast<AtomicSDNode>(Node)->getAlignment(),
12688 cast<AtomicSDNode>(Node)->getOrdering(),
12689 cast<AtomicSDNode>(Node)->getSynchScope());
12692 static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) {
12693 SDNode *Node = Op.getNode();
12695 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
12697 // Convert seq_cst store -> xchg
12698 // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b)
12699 // FIXME: On 32-bit, store -> fist or movq would be more efficient
12700 // (The only way to get a 16-byte store is cmpxchg16b)
12701 // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment.
12702 if (cast<AtomicSDNode>(Node)->getOrdering() == SequentiallyConsistent ||
12703 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
12704 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
12705 cast<AtomicSDNode>(Node)->getMemoryVT(),
12706 Node->getOperand(0),
12707 Node->getOperand(1), Node->getOperand(2),
12708 cast<AtomicSDNode>(Node)->getMemOperand(),
12709 cast<AtomicSDNode>(Node)->getOrdering(),
12710 cast<AtomicSDNode>(Node)->getSynchScope());
12711 return Swap.getValue(1);
12713 // Other atomic stores have a simple pattern.
12717 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
12718 EVT VT = Op.getNode()->getValueType(0);
12720 // Let legalize expand this if it isn't a legal type yet.
12721 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
12724 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
12727 bool ExtraOp = false;
12728 switch (Op.getOpcode()) {
12729 default: llvm_unreachable("Invalid code");
12730 case ISD::ADDC: Opc = X86ISD::ADD; break;
12731 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
12732 case ISD::SUBC: Opc = X86ISD::SUB; break;
12733 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
12737 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
12739 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
12740 Op.getOperand(1), Op.getOperand(2));
12743 SDValue X86TargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
12744 assert(Subtarget->isTargetDarwin() && Subtarget->is64Bit());
12746 // For MacOSX, we want to call an alternative entry point: __sincos_stret,
12747 // which returns the values as { float, float } (in XMM0) or
12748 // { double, double } (which is returned in XMM0, XMM1).
12750 SDValue Arg = Op.getOperand(0);
12751 EVT ArgVT = Arg.getValueType();
12752 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
12755 ArgListEntry Entry;
12759 Entry.isSExt = false;
12760 Entry.isZExt = false;
12761 Args.push_back(Entry);
12763 bool isF64 = ArgVT == MVT::f64;
12764 // Only optimize x86_64 for now. i386 is a bit messy. For f32,
12765 // the small struct {f32, f32} is returned in (eax, edx). For f64,
12766 // the results are returned via SRet in memory.
12767 const char *LibcallName = isF64 ? "__sincos_stret" : "__sincosf_stret";
12768 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy());
12770 Type *RetTy = isF64
12771 ? (Type*)StructType::get(ArgTy, ArgTy, NULL)
12772 : (Type*)VectorType::get(ArgTy, 4);
12774 CallLoweringInfo CLI(DAG.getEntryNode(), RetTy,
12775 false, false, false, false, 0,
12776 CallingConv::C, /*isTaillCall=*/false,
12777 /*doesNotRet=*/false, /*isReturnValueUsed*/true,
12778 Callee, Args, DAG, dl);
12779 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
12782 // Returned in xmm0 and xmm1.
12783 return CallResult.first;
12785 // Returned in bits 0:31 and 32:64 xmm0.
12786 SDValue SinVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
12787 CallResult.first, DAG.getIntPtrConstant(0));
12788 SDValue CosVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
12789 CallResult.first, DAG.getIntPtrConstant(1));
12790 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
12791 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys, SinVal, CosVal);
12794 /// LowerOperation - Provide custom lowering hooks for some operations.
12796 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
12797 switch (Op.getOpcode()) {
12798 default: llvm_unreachable("Should not custom lower this!");
12799 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op,DAG);
12800 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, Subtarget, DAG);
12801 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op, Subtarget, DAG);
12802 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
12803 case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG);
12804 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
12805 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
12806 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
12807 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
12808 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
12809 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op,Subtarget,DAG);
12810 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, Subtarget,DAG);
12811 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
12812 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
12813 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
12814 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
12815 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
12816 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
12817 case ISD::SHL_PARTS:
12818 case ISD::SRA_PARTS:
12819 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
12820 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
12821 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
12822 case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG);
12823 case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, DAG);
12824 case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, DAG);
12825 case ISD::ANY_EXTEND: return LowerANY_EXTEND(Op, DAG);
12826 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
12827 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
12828 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
12829 case ISD::FABS: return LowerFABS(Op, DAG);
12830 case ISD::FNEG: return LowerFNEG(Op, DAG);
12831 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
12832 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
12833 case ISD::SETCC: return LowerSETCC(Op, DAG);
12834 case ISD::SELECT: return LowerSELECT(Op, DAG);
12835 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
12836 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
12837 case ISD::VASTART: return LowerVASTART(Op, DAG);
12838 case ISD::VAARG: return LowerVAARG(Op, DAG);
12839 case ISD::VACOPY: return LowerVACOPY(Op, Subtarget, DAG);
12840 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
12841 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, DAG);
12842 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
12843 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
12844 case ISD::FRAME_TO_ARGS_OFFSET:
12845 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
12846 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
12847 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
12848 case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
12849 case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
12850 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
12851 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
12852 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
12853 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
12854 case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, DAG);
12855 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
12856 case ISD::MUL: return LowerMUL(Op, Subtarget, DAG);
12859 case ISD::SHL: return LowerShift(Op, DAG);
12865 case ISD::UMULO: return LowerXALUO(Op, DAG);
12866 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, Subtarget,DAG);
12867 case ISD::BITCAST: return LowerBITCAST(Op, DAG);
12871 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
12872 case ISD::ADD: return LowerADD(Op, DAG);
12873 case ISD::SUB: return LowerSUB(Op, DAG);
12874 case ISD::SDIV: return LowerSDIV(Op, DAG);
12875 case ISD::FSINCOS: return LowerFSINCOS(Op, DAG);
12879 static void ReplaceATOMIC_LOAD(SDNode *Node,
12880 SmallVectorImpl<SDValue> &Results,
12881 SelectionDAG &DAG) {
12883 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
12885 // Convert wide load -> cmpxchg8b/cmpxchg16b
12886 // FIXME: On 32-bit, load -> fild or movq would be more efficient
12887 // (The only way to get a 16-byte load is cmpxchg16b)
12888 // FIXME: 16-byte ATOMIC_CMP_SWAP isn't actually hooked up at the moment.
12889 SDValue Zero = DAG.getConstant(0, VT);
12890 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_CMP_SWAP, dl, VT,
12891 Node->getOperand(0),
12892 Node->getOperand(1), Zero, Zero,
12893 cast<AtomicSDNode>(Node)->getMemOperand(),
12894 cast<AtomicSDNode>(Node)->getOrdering(),
12895 cast<AtomicSDNode>(Node)->getSynchScope());
12896 Results.push_back(Swap.getValue(0));
12897 Results.push_back(Swap.getValue(1));
12901 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
12902 SelectionDAG &DAG, unsigned NewOp) {
12904 assert (Node->getValueType(0) == MVT::i64 &&
12905 "Only know how to expand i64 atomics");
12907 SDValue Chain = Node->getOperand(0);
12908 SDValue In1 = Node->getOperand(1);
12909 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
12910 Node->getOperand(2), DAG.getIntPtrConstant(0));
12911 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
12912 Node->getOperand(2), DAG.getIntPtrConstant(1));
12913 SDValue Ops[] = { Chain, In1, In2L, In2H };
12914 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
12916 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, array_lengthof(Ops), MVT::i64,
12917 cast<MemSDNode>(Node)->getMemOperand());
12918 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
12919 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
12920 Results.push_back(Result.getValue(2));
12923 /// ReplaceNodeResults - Replace a node with an illegal result type
12924 /// with a new node built out of custom code.
12925 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
12926 SmallVectorImpl<SDValue>&Results,
12927 SelectionDAG &DAG) const {
12929 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12930 switch (N->getOpcode()) {
12932 llvm_unreachable("Do not know how to custom type legalize this operation!");
12933 case ISD::SIGN_EXTEND_INREG:
12938 // We don't want to expand or promote these.
12940 case ISD::FP_TO_SINT:
12941 case ISD::FP_TO_UINT: {
12942 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
12944 if (!IsSigned && !isIntegerTypeFTOL(SDValue(N, 0).getValueType()))
12947 std::pair<SDValue,SDValue> Vals =
12948 FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true);
12949 SDValue FIST = Vals.first, StackSlot = Vals.second;
12950 if (FIST.getNode() != 0) {
12951 EVT VT = N->getValueType(0);
12952 // Return a load from the stack slot.
12953 if (StackSlot.getNode() != 0)
12954 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
12955 MachinePointerInfo(),
12956 false, false, false, 0));
12958 Results.push_back(FIST);
12962 case ISD::UINT_TO_FP: {
12963 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
12964 if (N->getOperand(0).getValueType() != MVT::v2i32 ||
12965 N->getValueType(0) != MVT::v2f32)
12967 SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64,
12969 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
12971 SDValue VBias = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2f64, Bias, Bias);
12972 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn,
12973 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, VBias));
12974 Or = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or);
12975 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias);
12976 Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub));
12979 case ISD::FP_ROUND: {
12980 if (!TLI.isTypeLegal(N->getOperand(0).getValueType()))
12982 SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0));
12983 Results.push_back(V);
12986 case ISD::READCYCLECOUNTER: {
12987 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
12988 SDValue TheChain = N->getOperand(0);
12989 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
12990 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
12992 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
12994 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
12995 SDValue Ops[] = { eax, edx };
12996 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops,
12997 array_lengthof(Ops)));
12998 Results.push_back(edx.getValue(1));
13001 case ISD::ATOMIC_CMP_SWAP: {
13002 EVT T = N->getValueType(0);
13003 assert((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair");
13004 bool Regs64bit = T == MVT::i128;
13005 EVT HalfT = Regs64bit ? MVT::i64 : MVT::i32;
13006 SDValue cpInL, cpInH;
13007 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
13008 DAG.getConstant(0, HalfT));
13009 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
13010 DAG.getConstant(1, HalfT));
13011 cpInL = DAG.getCopyToReg(N->getOperand(0), dl,
13012 Regs64bit ? X86::RAX : X86::EAX,
13014 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl,
13015 Regs64bit ? X86::RDX : X86::EDX,
13016 cpInH, cpInL.getValue(1));
13017 SDValue swapInL, swapInH;
13018 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
13019 DAG.getConstant(0, HalfT));
13020 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
13021 DAG.getConstant(1, HalfT));
13022 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl,
13023 Regs64bit ? X86::RBX : X86::EBX,
13024 swapInL, cpInH.getValue(1));
13025 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl,
13026 Regs64bit ? X86::RCX : X86::ECX,
13027 swapInH, swapInL.getValue(1));
13028 SDValue Ops[] = { swapInH.getValue(0),
13030 swapInH.getValue(1) };
13031 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
13032 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
13033 unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_DAG :
13034 X86ISD::LCMPXCHG8_DAG;
13035 SDValue Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys,
13036 Ops, array_lengthof(Ops), T, MMO);
13037 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl,
13038 Regs64bit ? X86::RAX : X86::EAX,
13039 HalfT, Result.getValue(1));
13040 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl,
13041 Regs64bit ? X86::RDX : X86::EDX,
13042 HalfT, cpOutL.getValue(2));
13043 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
13044 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF, 2));
13045 Results.push_back(cpOutH.getValue(1));
13048 case ISD::ATOMIC_LOAD_ADD:
13049 case ISD::ATOMIC_LOAD_AND:
13050 case ISD::ATOMIC_LOAD_NAND:
13051 case ISD::ATOMIC_LOAD_OR:
13052 case ISD::ATOMIC_LOAD_SUB:
13053 case ISD::ATOMIC_LOAD_XOR:
13054 case ISD::ATOMIC_LOAD_MAX:
13055 case ISD::ATOMIC_LOAD_MIN:
13056 case ISD::ATOMIC_LOAD_UMAX:
13057 case ISD::ATOMIC_LOAD_UMIN:
13058 case ISD::ATOMIC_SWAP: {
13060 switch (N->getOpcode()) {
13061 default: llvm_unreachable("Unexpected opcode");
13062 case ISD::ATOMIC_LOAD_ADD:
13063 Opc = X86ISD::ATOMADD64_DAG;
13065 case ISD::ATOMIC_LOAD_AND:
13066 Opc = X86ISD::ATOMAND64_DAG;
13068 case ISD::ATOMIC_LOAD_NAND:
13069 Opc = X86ISD::ATOMNAND64_DAG;
13071 case ISD::ATOMIC_LOAD_OR:
13072 Opc = X86ISD::ATOMOR64_DAG;
13074 case ISD::ATOMIC_LOAD_SUB:
13075 Opc = X86ISD::ATOMSUB64_DAG;
13077 case ISD::ATOMIC_LOAD_XOR:
13078 Opc = X86ISD::ATOMXOR64_DAG;
13080 case ISD::ATOMIC_LOAD_MAX:
13081 Opc = X86ISD::ATOMMAX64_DAG;
13083 case ISD::ATOMIC_LOAD_MIN:
13084 Opc = X86ISD::ATOMMIN64_DAG;
13086 case ISD::ATOMIC_LOAD_UMAX:
13087 Opc = X86ISD::ATOMUMAX64_DAG;
13089 case ISD::ATOMIC_LOAD_UMIN:
13090 Opc = X86ISD::ATOMUMIN64_DAG;
13092 case ISD::ATOMIC_SWAP:
13093 Opc = X86ISD::ATOMSWAP64_DAG;
13096 ReplaceATOMIC_BINARY_64(N, Results, DAG, Opc);
13099 case ISD::ATOMIC_LOAD:
13100 ReplaceATOMIC_LOAD(N, Results, DAG);
13104 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
13106 default: return NULL;
13107 case X86ISD::BSF: return "X86ISD::BSF";
13108 case X86ISD::BSR: return "X86ISD::BSR";
13109 case X86ISD::SHLD: return "X86ISD::SHLD";
13110 case X86ISD::SHRD: return "X86ISD::SHRD";
13111 case X86ISD::FAND: return "X86ISD::FAND";
13112 case X86ISD::FANDN: return "X86ISD::FANDN";
13113 case X86ISD::FOR: return "X86ISD::FOR";
13114 case X86ISD::FXOR: return "X86ISD::FXOR";
13115 case X86ISD::FSRL: return "X86ISD::FSRL";
13116 case X86ISD::FILD: return "X86ISD::FILD";
13117 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
13118 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
13119 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
13120 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
13121 case X86ISD::FLD: return "X86ISD::FLD";
13122 case X86ISD::FST: return "X86ISD::FST";
13123 case X86ISD::CALL: return "X86ISD::CALL";
13124 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
13125 case X86ISD::BT: return "X86ISD::BT";
13126 case X86ISD::CMP: return "X86ISD::CMP";
13127 case X86ISD::COMI: return "X86ISD::COMI";
13128 case X86ISD::UCOMI: return "X86ISD::UCOMI";
13129 case X86ISD::SETCC: return "X86ISD::SETCC";
13130 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
13131 case X86ISD::FSETCCsd: return "X86ISD::FSETCCsd";
13132 case X86ISD::FSETCCss: return "X86ISD::FSETCCss";
13133 case X86ISD::CMOV: return "X86ISD::CMOV";
13134 case X86ISD::BRCOND: return "X86ISD::BRCOND";
13135 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
13136 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
13137 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
13138 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
13139 case X86ISD::Wrapper: return "X86ISD::Wrapper";
13140 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
13141 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
13142 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
13143 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
13144 case X86ISD::PINSRB: return "X86ISD::PINSRB";
13145 case X86ISD::PINSRW: return "X86ISD::PINSRW";
13146 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
13147 case X86ISD::ANDNP: return "X86ISD::ANDNP";
13148 case X86ISD::PSIGN: return "X86ISD::PSIGN";
13149 case X86ISD::BLENDV: return "X86ISD::BLENDV";
13150 case X86ISD::BLENDI: return "X86ISD::BLENDI";
13151 case X86ISD::SUBUS: return "X86ISD::SUBUS";
13152 case X86ISD::HADD: return "X86ISD::HADD";
13153 case X86ISD::HSUB: return "X86ISD::HSUB";
13154 case X86ISD::FHADD: return "X86ISD::FHADD";
13155 case X86ISD::FHSUB: return "X86ISD::FHSUB";
13156 case X86ISD::UMAX: return "X86ISD::UMAX";
13157 case X86ISD::UMIN: return "X86ISD::UMIN";
13158 case X86ISD::SMAX: return "X86ISD::SMAX";
13159 case X86ISD::SMIN: return "X86ISD::SMIN";
13160 case X86ISD::FMAX: return "X86ISD::FMAX";
13161 case X86ISD::FMIN: return "X86ISD::FMIN";
13162 case X86ISD::FMAXC: return "X86ISD::FMAXC";
13163 case X86ISD::FMINC: return "X86ISD::FMINC";
13164 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
13165 case X86ISD::FRCP: return "X86ISD::FRCP";
13166 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
13167 case X86ISD::TLSBASEADDR: return "X86ISD::TLSBASEADDR";
13168 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
13169 case X86ISD::EH_SJLJ_SETJMP: return "X86ISD::EH_SJLJ_SETJMP";
13170 case X86ISD::EH_SJLJ_LONGJMP: return "X86ISD::EH_SJLJ_LONGJMP";
13171 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
13172 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
13173 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
13174 case X86ISD::FNSTSW16r: return "X86ISD::FNSTSW16r";
13175 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
13176 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
13177 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
13178 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
13179 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
13180 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
13181 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
13182 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
13183 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
13184 case X86ISD::VSEXT_MOVL: return "X86ISD::VSEXT_MOVL";
13185 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
13186 case X86ISD::VZEXT: return "X86ISD::VZEXT";
13187 case X86ISD::VSEXT: return "X86ISD::VSEXT";
13188 case X86ISD::VFPEXT: return "X86ISD::VFPEXT";
13189 case X86ISD::VFPROUND: return "X86ISD::VFPROUND";
13190 case X86ISD::VSHLDQ: return "X86ISD::VSHLDQ";
13191 case X86ISD::VSRLDQ: return "X86ISD::VSRLDQ";
13192 case X86ISD::VSHL: return "X86ISD::VSHL";
13193 case X86ISD::VSRL: return "X86ISD::VSRL";
13194 case X86ISD::VSRA: return "X86ISD::VSRA";
13195 case X86ISD::VSHLI: return "X86ISD::VSHLI";
13196 case X86ISD::VSRLI: return "X86ISD::VSRLI";
13197 case X86ISD::VSRAI: return "X86ISD::VSRAI";
13198 case X86ISD::CMPP: return "X86ISD::CMPP";
13199 case X86ISD::PCMPEQ: return "X86ISD::PCMPEQ";
13200 case X86ISD::PCMPGT: return "X86ISD::PCMPGT";
13201 case X86ISD::ADD: return "X86ISD::ADD";
13202 case X86ISD::SUB: return "X86ISD::SUB";
13203 case X86ISD::ADC: return "X86ISD::ADC";
13204 case X86ISD::SBB: return "X86ISD::SBB";
13205 case X86ISD::SMUL: return "X86ISD::SMUL";
13206 case X86ISD::UMUL: return "X86ISD::UMUL";
13207 case X86ISD::INC: return "X86ISD::INC";
13208 case X86ISD::DEC: return "X86ISD::DEC";
13209 case X86ISD::OR: return "X86ISD::OR";
13210 case X86ISD::XOR: return "X86ISD::XOR";
13211 case X86ISD::AND: return "X86ISD::AND";
13212 case X86ISD::BLSI: return "X86ISD::BLSI";
13213 case X86ISD::BLSMSK: return "X86ISD::BLSMSK";
13214 case X86ISD::BLSR: return "X86ISD::BLSR";
13215 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
13216 case X86ISD::PTEST: return "X86ISD::PTEST";
13217 case X86ISD::TESTP: return "X86ISD::TESTP";
13218 case X86ISD::PALIGNR: return "X86ISD::PALIGNR";
13219 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
13220 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
13221 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
13222 case X86ISD::SHUFP: return "X86ISD::SHUFP";
13223 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
13224 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
13225 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
13226 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
13227 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
13228 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
13229 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
13230 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
13231 case X86ISD::MOVSD: return "X86ISD::MOVSD";
13232 case X86ISD::MOVSS: return "X86ISD::MOVSS";
13233 case X86ISD::UNPCKL: return "X86ISD::UNPCKL";
13234 case X86ISD::UNPCKH: return "X86ISD::UNPCKH";
13235 case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
13236 case X86ISD::VPERMILP: return "X86ISD::VPERMILP";
13237 case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128";
13238 case X86ISD::VPERMV: return "X86ISD::VPERMV";
13239 case X86ISD::VPERMI: return "X86ISD::VPERMI";
13240 case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
13241 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
13242 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
13243 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
13244 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
13245 case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA";
13246 case X86ISD::WIN_FTOL: return "X86ISD::WIN_FTOL";
13247 case X86ISD::SAHF: return "X86ISD::SAHF";
13248 case X86ISD::RDRAND: return "X86ISD::RDRAND";
13249 case X86ISD::RDSEED: return "X86ISD::RDSEED";
13250 case X86ISD::FMADD: return "X86ISD::FMADD";
13251 case X86ISD::FMSUB: return "X86ISD::FMSUB";
13252 case X86ISD::FNMADD: return "X86ISD::FNMADD";
13253 case X86ISD::FNMSUB: return "X86ISD::FNMSUB";
13254 case X86ISD::FMADDSUB: return "X86ISD::FMADDSUB";
13255 case X86ISD::FMSUBADD: return "X86ISD::FMSUBADD";
13256 case X86ISD::PCMPESTRI: return "X86ISD::PCMPESTRI";
13257 case X86ISD::PCMPISTRI: return "X86ISD::PCMPISTRI";
13258 case X86ISD::XTEST: return "X86ISD::XTEST";
13262 // isLegalAddressingMode - Return true if the addressing mode represented
13263 // by AM is legal for this target, for a load/store of the specified type.
13264 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
13266 // X86 supports extremely general addressing modes.
13267 CodeModel::Model M = getTargetMachine().getCodeModel();
13268 Reloc::Model R = getTargetMachine().getRelocationModel();
13270 // X86 allows a sign-extended 32-bit immediate field as a displacement.
13271 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
13276 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
13278 // If a reference to this global requires an extra load, we can't fold it.
13279 if (isGlobalStubReference(GVFlags))
13282 // If BaseGV requires a register for the PIC base, we cannot also have a
13283 // BaseReg specified.
13284 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
13287 // If lower 4G is not available, then we must use rip-relative addressing.
13288 if ((M != CodeModel::Small || R != Reloc::Static) &&
13289 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
13293 switch (AM.Scale) {
13299 // These scales always work.
13304 // These scales are formed with basereg+scalereg. Only accept if there is
13309 default: // Other stuff never works.
13316 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
13317 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
13319 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
13320 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
13321 return NumBits1 > NumBits2;
13324 bool X86TargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
13325 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
13328 if (!isTypeLegal(EVT::getEVT(Ty1)))
13331 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
13333 // Assuming the caller doesn't have a zeroext or signext return parameter,
13334 // truncation all the way down to i1 is valid.
13338 bool X86TargetLowering::isLegalICmpImmediate(int64_t Imm) const {
13339 return isInt<32>(Imm);
13342 bool X86TargetLowering::isLegalAddImmediate(int64_t Imm) const {
13343 // Can also use sub to handle negated immediates.
13344 return isInt<32>(Imm);
13347 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
13348 if (!VT1.isInteger() || !VT2.isInteger())
13350 unsigned NumBits1 = VT1.getSizeInBits();
13351 unsigned NumBits2 = VT2.getSizeInBits();
13352 return NumBits1 > NumBits2;
13355 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
13356 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
13357 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
13360 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
13361 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
13362 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
13365 bool X86TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
13366 EVT VT1 = Val.getValueType();
13367 if (isZExtFree(VT1, VT2))
13370 if (Val.getOpcode() != ISD::LOAD)
13373 if (!VT1.isSimple() || !VT1.isInteger() ||
13374 !VT2.isSimple() || !VT2.isInteger())
13377 switch (VT1.getSimpleVT().SimpleTy) {
13382 // X86 has 8, 16, and 32-bit zero-extending loads.
13390 X86TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
13391 if (!(Subtarget->hasFMA() || Subtarget->hasFMA4()))
13394 VT = VT.getScalarType();
13396 if (!VT.isSimple())
13399 switch (VT.getSimpleVT().SimpleTy) {
13410 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
13411 // i16 instructions are longer (0x66 prefix) and potentially slower.
13412 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
13415 /// isShuffleMaskLegal - Targets can use this to indicate that they only
13416 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
13417 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
13418 /// are assumed to be legal.
13420 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
13422 // Very little shuffling can be done for 64-bit vectors right now.
13423 if (VT.getSizeInBits() == 64)
13426 // FIXME: pshufb, blends, shifts.
13427 return (VT.getVectorNumElements() == 2 ||
13428 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
13429 isMOVLMask(M, VT) ||
13430 isSHUFPMask(M, VT, Subtarget->hasFp256()) ||
13431 isPSHUFDMask(M, VT) ||
13432 isPSHUFHWMask(M, VT, Subtarget->hasInt256()) ||
13433 isPSHUFLWMask(M, VT, Subtarget->hasInt256()) ||
13434 isPALIGNRMask(M, VT, Subtarget) ||
13435 isUNPCKLMask(M, VT, Subtarget->hasInt256()) ||
13436 isUNPCKHMask(M, VT, Subtarget->hasInt256()) ||
13437 isUNPCKL_v_undef_Mask(M, VT, Subtarget->hasInt256()) ||
13438 isUNPCKH_v_undef_Mask(M, VT, Subtarget->hasInt256()));
13442 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
13444 unsigned NumElts = VT.getVectorNumElements();
13445 // FIXME: This collection of masks seems suspect.
13448 if (NumElts == 4 && VT.is128BitVector()) {
13449 return (isMOVLMask(Mask, VT) ||
13450 isCommutedMOVLMask(Mask, VT, true) ||
13451 isSHUFPMask(Mask, VT, Subtarget->hasFp256()) ||
13452 isSHUFPMask(Mask, VT, Subtarget->hasFp256(), /* Commuted */ true));
13457 //===----------------------------------------------------------------------===//
13458 // X86 Scheduler Hooks
13459 //===----------------------------------------------------------------------===//
13461 /// Utility function to emit xbegin specifying the start of an RTM region.
13462 static MachineBasicBlock *EmitXBegin(MachineInstr *MI, MachineBasicBlock *MBB,
13463 const TargetInstrInfo *TII) {
13464 DebugLoc DL = MI->getDebugLoc();
13466 const BasicBlock *BB = MBB->getBasicBlock();
13467 MachineFunction::iterator I = MBB;
13470 // For the v = xbegin(), we generate
13481 MachineBasicBlock *thisMBB = MBB;
13482 MachineFunction *MF = MBB->getParent();
13483 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
13484 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
13485 MF->insert(I, mainMBB);
13486 MF->insert(I, sinkMBB);
13488 // Transfer the remainder of BB and its successor edges to sinkMBB.
13489 sinkMBB->splice(sinkMBB->begin(), MBB,
13490 llvm::next(MachineBasicBlock::iterator(MI)), MBB->end());
13491 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
13495 // # fallthrough to mainMBB
13496 // # abortion to sinkMBB
13497 BuildMI(thisMBB, DL, TII->get(X86::XBEGIN_4)).addMBB(sinkMBB);
13498 thisMBB->addSuccessor(mainMBB);
13499 thisMBB->addSuccessor(sinkMBB);
13503 BuildMI(mainMBB, DL, TII->get(X86::MOV32ri), X86::EAX).addImm(-1);
13504 mainMBB->addSuccessor(sinkMBB);
13507 // EAX is live into the sinkMBB
13508 sinkMBB->addLiveIn(X86::EAX);
13509 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
13510 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
13513 MI->eraseFromParent();
13517 // Get CMPXCHG opcode for the specified data type.
13518 static unsigned getCmpXChgOpcode(EVT VT) {
13519 switch (VT.getSimpleVT().SimpleTy) {
13520 case MVT::i8: return X86::LCMPXCHG8;
13521 case MVT::i16: return X86::LCMPXCHG16;
13522 case MVT::i32: return X86::LCMPXCHG32;
13523 case MVT::i64: return X86::LCMPXCHG64;
13527 llvm_unreachable("Invalid operand size!");
13530 // Get LOAD opcode for the specified data type.
13531 static unsigned getLoadOpcode(EVT VT) {
13532 switch (VT.getSimpleVT().SimpleTy) {
13533 case MVT::i8: return X86::MOV8rm;
13534 case MVT::i16: return X86::MOV16rm;
13535 case MVT::i32: return X86::MOV32rm;
13536 case MVT::i64: return X86::MOV64rm;
13540 llvm_unreachable("Invalid operand size!");
13543 // Get opcode of the non-atomic one from the specified atomic instruction.
13544 static unsigned getNonAtomicOpcode(unsigned Opc) {
13546 case X86::ATOMAND8: return X86::AND8rr;
13547 case X86::ATOMAND16: return X86::AND16rr;
13548 case X86::ATOMAND32: return X86::AND32rr;
13549 case X86::ATOMAND64: return X86::AND64rr;
13550 case X86::ATOMOR8: return X86::OR8rr;
13551 case X86::ATOMOR16: return X86::OR16rr;
13552 case X86::ATOMOR32: return X86::OR32rr;
13553 case X86::ATOMOR64: return X86::OR64rr;
13554 case X86::ATOMXOR8: return X86::XOR8rr;
13555 case X86::ATOMXOR16: return X86::XOR16rr;
13556 case X86::ATOMXOR32: return X86::XOR32rr;
13557 case X86::ATOMXOR64: return X86::XOR64rr;
13559 llvm_unreachable("Unhandled atomic-load-op opcode!");
13562 // Get opcode of the non-atomic one from the specified atomic instruction with
13564 static unsigned getNonAtomicOpcodeWithExtraOpc(unsigned Opc,
13565 unsigned &ExtraOpc) {
13567 case X86::ATOMNAND8: ExtraOpc = X86::NOT8r; return X86::AND8rr;
13568 case X86::ATOMNAND16: ExtraOpc = X86::NOT16r; return X86::AND16rr;
13569 case X86::ATOMNAND32: ExtraOpc = X86::NOT32r; return X86::AND32rr;
13570 case X86::ATOMNAND64: ExtraOpc = X86::NOT64r; return X86::AND64rr;
13571 case X86::ATOMMAX8: ExtraOpc = X86::CMP8rr; return X86::CMOVL32rr;
13572 case X86::ATOMMAX16: ExtraOpc = X86::CMP16rr; return X86::CMOVL16rr;
13573 case X86::ATOMMAX32: ExtraOpc = X86::CMP32rr; return X86::CMOVL32rr;
13574 case X86::ATOMMAX64: ExtraOpc = X86::CMP64rr; return X86::CMOVL64rr;
13575 case X86::ATOMMIN8: ExtraOpc = X86::CMP8rr; return X86::CMOVG32rr;
13576 case X86::ATOMMIN16: ExtraOpc = X86::CMP16rr; return X86::CMOVG16rr;
13577 case X86::ATOMMIN32: ExtraOpc = X86::CMP32rr; return X86::CMOVG32rr;
13578 case X86::ATOMMIN64: ExtraOpc = X86::CMP64rr; return X86::CMOVG64rr;
13579 case X86::ATOMUMAX8: ExtraOpc = X86::CMP8rr; return X86::CMOVB32rr;
13580 case X86::ATOMUMAX16: ExtraOpc = X86::CMP16rr; return X86::CMOVB16rr;
13581 case X86::ATOMUMAX32: ExtraOpc = X86::CMP32rr; return X86::CMOVB32rr;
13582 case X86::ATOMUMAX64: ExtraOpc = X86::CMP64rr; return X86::CMOVB64rr;
13583 case X86::ATOMUMIN8: ExtraOpc = X86::CMP8rr; return X86::CMOVA32rr;
13584 case X86::ATOMUMIN16: ExtraOpc = X86::CMP16rr; return X86::CMOVA16rr;
13585 case X86::ATOMUMIN32: ExtraOpc = X86::CMP32rr; return X86::CMOVA32rr;
13586 case X86::ATOMUMIN64: ExtraOpc = X86::CMP64rr; return X86::CMOVA64rr;
13588 llvm_unreachable("Unhandled atomic-load-op opcode!");
13591 // Get opcode of the non-atomic one from the specified atomic instruction for
13592 // 64-bit data type on 32-bit target.
13593 static unsigned getNonAtomic6432Opcode(unsigned Opc, unsigned &HiOpc) {
13595 case X86::ATOMAND6432: HiOpc = X86::AND32rr; return X86::AND32rr;
13596 case X86::ATOMOR6432: HiOpc = X86::OR32rr; return X86::OR32rr;
13597 case X86::ATOMXOR6432: HiOpc = X86::XOR32rr; return X86::XOR32rr;
13598 case X86::ATOMADD6432: HiOpc = X86::ADC32rr; return X86::ADD32rr;
13599 case X86::ATOMSUB6432: HiOpc = X86::SBB32rr; return X86::SUB32rr;
13600 case X86::ATOMSWAP6432: HiOpc = X86::MOV32rr; return X86::MOV32rr;
13601 case X86::ATOMMAX6432: HiOpc = X86::SETLr; return X86::SETLr;
13602 case X86::ATOMMIN6432: HiOpc = X86::SETGr; return X86::SETGr;
13603 case X86::ATOMUMAX6432: HiOpc = X86::SETBr; return X86::SETBr;
13604 case X86::ATOMUMIN6432: HiOpc = X86::SETAr; return X86::SETAr;
13606 llvm_unreachable("Unhandled atomic-load-op opcode!");
13609 // Get opcode of the non-atomic one from the specified atomic instruction for
13610 // 64-bit data type on 32-bit target with extra opcode.
13611 static unsigned getNonAtomic6432OpcodeWithExtraOpc(unsigned Opc,
13613 unsigned &ExtraOpc) {
13615 case X86::ATOMNAND6432:
13616 ExtraOpc = X86::NOT32r;
13617 HiOpc = X86::AND32rr;
13618 return X86::AND32rr;
13620 llvm_unreachable("Unhandled atomic-load-op opcode!");
13623 // Get pseudo CMOV opcode from the specified data type.
13624 static unsigned getPseudoCMOVOpc(EVT VT) {
13625 switch (VT.getSimpleVT().SimpleTy) {
13626 case MVT::i8: return X86::CMOV_GR8;
13627 case MVT::i16: return X86::CMOV_GR16;
13628 case MVT::i32: return X86::CMOV_GR32;
13632 llvm_unreachable("Unknown CMOV opcode!");
13635 // EmitAtomicLoadArith - emit the code sequence for pseudo atomic instructions.
13636 // They will be translated into a spin-loop or compare-exchange loop from
13639 // dst = atomic-fetch-op MI.addr, MI.val
13645 // t1 = LOAD MI.addr
13647 // t4 = phi(t1, t3 / loop)
13648 // t2 = OP MI.val, t4
13650 // LCMPXCHG [MI.addr], t2, [EAX is implicitly used & defined]
13656 MachineBasicBlock *
13657 X86TargetLowering::EmitAtomicLoadArith(MachineInstr *MI,
13658 MachineBasicBlock *MBB) const {
13659 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
13660 DebugLoc DL = MI->getDebugLoc();
13662 MachineFunction *MF = MBB->getParent();
13663 MachineRegisterInfo &MRI = MF->getRegInfo();
13665 const BasicBlock *BB = MBB->getBasicBlock();
13666 MachineFunction::iterator I = MBB;
13669 assert(MI->getNumOperands() <= X86::AddrNumOperands + 4 &&
13670 "Unexpected number of operands");
13672 assert(MI->hasOneMemOperand() &&
13673 "Expected atomic-load-op to have one memoperand");
13675 // Memory Reference
13676 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
13677 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
13679 unsigned DstReg, SrcReg;
13680 unsigned MemOpndSlot;
13682 unsigned CurOp = 0;
13684 DstReg = MI->getOperand(CurOp++).getReg();
13685 MemOpndSlot = CurOp;
13686 CurOp += X86::AddrNumOperands;
13687 SrcReg = MI->getOperand(CurOp++).getReg();
13689 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
13690 MVT::SimpleValueType VT = *RC->vt_begin();
13691 unsigned t1 = MRI.createVirtualRegister(RC);
13692 unsigned t2 = MRI.createVirtualRegister(RC);
13693 unsigned t3 = MRI.createVirtualRegister(RC);
13694 unsigned t4 = MRI.createVirtualRegister(RC);
13695 unsigned PhyReg = getX86SubSuperRegister(X86::EAX, VT);
13697 unsigned LCMPXCHGOpc = getCmpXChgOpcode(VT);
13698 unsigned LOADOpc = getLoadOpcode(VT);
13700 // For the atomic load-arith operator, we generate
13703 // t1 = LOAD [MI.addr]
13705 // t4 = phi(t1 / thisMBB, t3 / mainMBB)
13706 // t1 = OP MI.val, EAX
13708 // LCMPXCHG [MI.addr], t1, [EAX is implicitly used & defined]
13714 MachineBasicBlock *thisMBB = MBB;
13715 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
13716 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
13717 MF->insert(I, mainMBB);
13718 MF->insert(I, sinkMBB);
13720 MachineInstrBuilder MIB;
13722 // Transfer the remainder of BB and its successor edges to sinkMBB.
13723 sinkMBB->splice(sinkMBB->begin(), MBB,
13724 llvm::next(MachineBasicBlock::iterator(MI)), MBB->end());
13725 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
13728 MIB = BuildMI(thisMBB, DL, TII->get(LOADOpc), t1);
13729 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
13730 MachineOperand NewMO = MI->getOperand(MemOpndSlot + i);
13732 NewMO.setIsKill(false);
13733 MIB.addOperand(NewMO);
13735 for (MachineInstr::mmo_iterator MMOI = MMOBegin; MMOI != MMOEnd; ++MMOI) {
13736 unsigned flags = (*MMOI)->getFlags();
13737 flags = (flags & ~MachineMemOperand::MOStore) | MachineMemOperand::MOLoad;
13738 MachineMemOperand *MMO =
13739 MF->getMachineMemOperand((*MMOI)->getPointerInfo(), flags,
13740 (*MMOI)->getSize(),
13741 (*MMOI)->getBaseAlignment(),
13742 (*MMOI)->getTBAAInfo(),
13743 (*MMOI)->getRanges());
13744 MIB.addMemOperand(MMO);
13747 thisMBB->addSuccessor(mainMBB);
13750 MachineBasicBlock *origMainMBB = mainMBB;
13753 MachineInstr *Phi = BuildMI(mainMBB, DL, TII->get(X86::PHI), t4)
13754 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(mainMBB);
13756 unsigned Opc = MI->getOpcode();
13759 llvm_unreachable("Unhandled atomic-load-op opcode!");
13760 case X86::ATOMAND8:
13761 case X86::ATOMAND16:
13762 case X86::ATOMAND32:
13763 case X86::ATOMAND64:
13765 case X86::ATOMOR16:
13766 case X86::ATOMOR32:
13767 case X86::ATOMOR64:
13768 case X86::ATOMXOR8:
13769 case X86::ATOMXOR16:
13770 case X86::ATOMXOR32:
13771 case X86::ATOMXOR64: {
13772 unsigned ARITHOpc = getNonAtomicOpcode(Opc);
13773 BuildMI(mainMBB, DL, TII->get(ARITHOpc), t2).addReg(SrcReg)
13777 case X86::ATOMNAND8:
13778 case X86::ATOMNAND16:
13779 case X86::ATOMNAND32:
13780 case X86::ATOMNAND64: {
13781 unsigned Tmp = MRI.createVirtualRegister(RC);
13783 unsigned ANDOpc = getNonAtomicOpcodeWithExtraOpc(Opc, NOTOpc);
13784 BuildMI(mainMBB, DL, TII->get(ANDOpc), Tmp).addReg(SrcReg)
13786 BuildMI(mainMBB, DL, TII->get(NOTOpc), t2).addReg(Tmp);
13789 case X86::ATOMMAX8:
13790 case X86::ATOMMAX16:
13791 case X86::ATOMMAX32:
13792 case X86::ATOMMAX64:
13793 case X86::ATOMMIN8:
13794 case X86::ATOMMIN16:
13795 case X86::ATOMMIN32:
13796 case X86::ATOMMIN64:
13797 case X86::ATOMUMAX8:
13798 case X86::ATOMUMAX16:
13799 case X86::ATOMUMAX32:
13800 case X86::ATOMUMAX64:
13801 case X86::ATOMUMIN8:
13802 case X86::ATOMUMIN16:
13803 case X86::ATOMUMIN32:
13804 case X86::ATOMUMIN64: {
13806 unsigned CMOVOpc = getNonAtomicOpcodeWithExtraOpc(Opc, CMPOpc);
13808 BuildMI(mainMBB, DL, TII->get(CMPOpc))
13812 if (Subtarget->hasCMov()) {
13813 if (VT != MVT::i8) {
13815 BuildMI(mainMBB, DL, TII->get(CMOVOpc), t2)
13819 // Promote i8 to i32 to use CMOV32
13820 const TargetRegisterInfo* TRI = getTargetMachine().getRegisterInfo();
13821 const TargetRegisterClass *RC32 =
13822 TRI->getSubClassWithSubReg(getRegClassFor(MVT::i32), X86::sub_8bit);
13823 unsigned SrcReg32 = MRI.createVirtualRegister(RC32);
13824 unsigned AccReg32 = MRI.createVirtualRegister(RC32);
13825 unsigned Tmp = MRI.createVirtualRegister(RC32);
13827 unsigned Undef = MRI.createVirtualRegister(RC32);
13828 BuildMI(mainMBB, DL, TII->get(TargetOpcode::IMPLICIT_DEF), Undef);
13830 BuildMI(mainMBB, DL, TII->get(TargetOpcode::INSERT_SUBREG), SrcReg32)
13833 .addImm(X86::sub_8bit);
13834 BuildMI(mainMBB, DL, TII->get(TargetOpcode::INSERT_SUBREG), AccReg32)
13837 .addImm(X86::sub_8bit);
13839 BuildMI(mainMBB, DL, TII->get(CMOVOpc), Tmp)
13843 BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), t2)
13844 .addReg(Tmp, 0, X86::sub_8bit);
13847 // Use pseudo select and lower them.
13848 assert((VT == MVT::i8 || VT == MVT::i16 || VT == MVT::i32) &&
13849 "Invalid atomic-load-op transformation!");
13850 unsigned SelOpc = getPseudoCMOVOpc(VT);
13851 X86::CondCode CC = X86::getCondFromCMovOpc(CMOVOpc);
13852 assert(CC != X86::COND_INVALID && "Invalid atomic-load-op transformation!");
13853 MIB = BuildMI(mainMBB, DL, TII->get(SelOpc), t2)
13854 .addReg(SrcReg).addReg(t4)
13856 mainMBB = EmitLoweredSelect(MIB, mainMBB);
13857 // Replace the original PHI node as mainMBB is changed after CMOV
13859 BuildMI(*origMainMBB, Phi, DL, TII->get(X86::PHI), t4)
13860 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(mainMBB);
13861 Phi->eraseFromParent();
13867 // Copy PhyReg back from virtual register.
13868 BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), PhyReg)
13871 MIB = BuildMI(mainMBB, DL, TII->get(LCMPXCHGOpc));
13872 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
13873 MachineOperand NewMO = MI->getOperand(MemOpndSlot + i);
13875 NewMO.setIsKill(false);
13876 MIB.addOperand(NewMO);
13879 MIB.setMemRefs(MMOBegin, MMOEnd);
13881 // Copy PhyReg back to virtual register.
13882 BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), t3)
13885 BuildMI(mainMBB, DL, TII->get(X86::JNE_4)).addMBB(origMainMBB);
13887 mainMBB->addSuccessor(origMainMBB);
13888 mainMBB->addSuccessor(sinkMBB);
13891 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
13892 TII->get(TargetOpcode::COPY), DstReg)
13895 MI->eraseFromParent();
13899 // EmitAtomicLoadArith6432 - emit the code sequence for pseudo atomic
13900 // instructions. They will be translated into a spin-loop or compare-exchange
13904 // dst = atomic-fetch-op MI.addr, MI.val
13910 // t1L = LOAD [MI.addr + 0]
13911 // t1H = LOAD [MI.addr + 4]
13913 // t4L = phi(t1L, t3L / loop)
13914 // t4H = phi(t1H, t3H / loop)
13915 // t2L = OP MI.val.lo, t4L
13916 // t2H = OP MI.val.hi, t4H
13921 // LCMPXCHG8B [MI.addr], [ECX:EBX & EDX:EAX are implicitly used and EDX:EAX is implicitly defined]
13929 MachineBasicBlock *
13930 X86TargetLowering::EmitAtomicLoadArith6432(MachineInstr *MI,
13931 MachineBasicBlock *MBB) const {
13932 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
13933 DebugLoc DL = MI->getDebugLoc();
13935 MachineFunction *MF = MBB->getParent();
13936 MachineRegisterInfo &MRI = MF->getRegInfo();
13938 const BasicBlock *BB = MBB->getBasicBlock();
13939 MachineFunction::iterator I = MBB;
13942 assert(MI->getNumOperands() <= X86::AddrNumOperands + 7 &&
13943 "Unexpected number of operands");
13945 assert(MI->hasOneMemOperand() &&
13946 "Expected atomic-load-op32 to have one memoperand");
13948 // Memory Reference
13949 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
13950 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
13952 unsigned DstLoReg, DstHiReg;
13953 unsigned SrcLoReg, SrcHiReg;
13954 unsigned MemOpndSlot;
13956 unsigned CurOp = 0;
13958 DstLoReg = MI->getOperand(CurOp++).getReg();
13959 DstHiReg = MI->getOperand(CurOp++).getReg();
13960 MemOpndSlot = CurOp;
13961 CurOp += X86::AddrNumOperands;
13962 SrcLoReg = MI->getOperand(CurOp++).getReg();
13963 SrcHiReg = MI->getOperand(CurOp++).getReg();
13965 const TargetRegisterClass *RC = &X86::GR32RegClass;
13966 const TargetRegisterClass *RC8 = &X86::GR8RegClass;
13968 unsigned t1L = MRI.createVirtualRegister(RC);
13969 unsigned t1H = MRI.createVirtualRegister(RC);
13970 unsigned t2L = MRI.createVirtualRegister(RC);
13971 unsigned t2H = MRI.createVirtualRegister(RC);
13972 unsigned t3L = MRI.createVirtualRegister(RC);
13973 unsigned t3H = MRI.createVirtualRegister(RC);
13974 unsigned t4L = MRI.createVirtualRegister(RC);
13975 unsigned t4H = MRI.createVirtualRegister(RC);
13977 unsigned LCMPXCHGOpc = X86::LCMPXCHG8B;
13978 unsigned LOADOpc = X86::MOV32rm;
13980 // For the atomic load-arith operator, we generate
13983 // t1L = LOAD [MI.addr + 0]
13984 // t1H = LOAD [MI.addr + 4]
13986 // t4L = phi(t1L / thisMBB, t3L / mainMBB)
13987 // t4H = phi(t1H / thisMBB, t3H / mainMBB)
13988 // t2L = OP MI.val.lo, t4L
13989 // t2H = OP MI.val.hi, t4H
13992 // LCMPXCHG8B [MI.addr], [ECX:EBX & EDX:EAX are implicitly used and EDX:EAX is implicitly defined]
14000 MachineBasicBlock *thisMBB = MBB;
14001 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
14002 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
14003 MF->insert(I, mainMBB);
14004 MF->insert(I, sinkMBB);
14006 MachineInstrBuilder MIB;
14008 // Transfer the remainder of BB and its successor edges to sinkMBB.
14009 sinkMBB->splice(sinkMBB->begin(), MBB,
14010 llvm::next(MachineBasicBlock::iterator(MI)), MBB->end());
14011 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
14015 MIB = BuildMI(thisMBB, DL, TII->get(LOADOpc), t1L);
14016 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
14017 MachineOperand NewMO = MI->getOperand(MemOpndSlot + i);
14019 NewMO.setIsKill(false);
14020 MIB.addOperand(NewMO);
14022 for (MachineInstr::mmo_iterator MMOI = MMOBegin; MMOI != MMOEnd; ++MMOI) {
14023 unsigned flags = (*MMOI)->getFlags();
14024 flags = (flags & ~MachineMemOperand::MOStore) | MachineMemOperand::MOLoad;
14025 MachineMemOperand *MMO =
14026 MF->getMachineMemOperand((*MMOI)->getPointerInfo(), flags,
14027 (*MMOI)->getSize(),
14028 (*MMOI)->getBaseAlignment(),
14029 (*MMOI)->getTBAAInfo(),
14030 (*MMOI)->getRanges());
14031 MIB.addMemOperand(MMO);
14033 MachineInstr *LowMI = MIB;
14036 MIB = BuildMI(thisMBB, DL, TII->get(LOADOpc), t1H);
14037 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
14038 if (i == X86::AddrDisp) {
14039 MIB.addDisp(MI->getOperand(MemOpndSlot + i), 4); // 4 == sizeof(i32)
14041 MachineOperand NewMO = MI->getOperand(MemOpndSlot + i);
14043 NewMO.setIsKill(false);
14044 MIB.addOperand(NewMO);
14047 MIB.setMemRefs(LowMI->memoperands_begin(), LowMI->memoperands_end());
14049 thisMBB->addSuccessor(mainMBB);
14052 MachineBasicBlock *origMainMBB = mainMBB;
14055 MachineInstr *PhiL = BuildMI(mainMBB, DL, TII->get(X86::PHI), t4L)
14056 .addReg(t1L).addMBB(thisMBB).addReg(t3L).addMBB(mainMBB);
14057 MachineInstr *PhiH = BuildMI(mainMBB, DL, TII->get(X86::PHI), t4H)
14058 .addReg(t1H).addMBB(thisMBB).addReg(t3H).addMBB(mainMBB);
14060 unsigned Opc = MI->getOpcode();
14063 llvm_unreachable("Unhandled atomic-load-op6432 opcode!");
14064 case X86::ATOMAND6432:
14065 case X86::ATOMOR6432:
14066 case X86::ATOMXOR6432:
14067 case X86::ATOMADD6432:
14068 case X86::ATOMSUB6432: {
14070 unsigned LoOpc = getNonAtomic6432Opcode(Opc, HiOpc);
14071 BuildMI(mainMBB, DL, TII->get(LoOpc), t2L).addReg(t4L)
14073 BuildMI(mainMBB, DL, TII->get(HiOpc), t2H).addReg(t4H)
14077 case X86::ATOMNAND6432: {
14078 unsigned HiOpc, NOTOpc;
14079 unsigned LoOpc = getNonAtomic6432OpcodeWithExtraOpc(Opc, HiOpc, NOTOpc);
14080 unsigned TmpL = MRI.createVirtualRegister(RC);
14081 unsigned TmpH = MRI.createVirtualRegister(RC);
14082 BuildMI(mainMBB, DL, TII->get(LoOpc), TmpL).addReg(SrcLoReg)
14084 BuildMI(mainMBB, DL, TII->get(HiOpc), TmpH).addReg(SrcHiReg)
14086 BuildMI(mainMBB, DL, TII->get(NOTOpc), t2L).addReg(TmpL);
14087 BuildMI(mainMBB, DL, TII->get(NOTOpc), t2H).addReg(TmpH);
14090 case X86::ATOMMAX6432:
14091 case X86::ATOMMIN6432:
14092 case X86::ATOMUMAX6432:
14093 case X86::ATOMUMIN6432: {
14095 unsigned LoOpc = getNonAtomic6432Opcode(Opc, HiOpc);
14096 unsigned cL = MRI.createVirtualRegister(RC8);
14097 unsigned cH = MRI.createVirtualRegister(RC8);
14098 unsigned cL32 = MRI.createVirtualRegister(RC);
14099 unsigned cH32 = MRI.createVirtualRegister(RC);
14100 unsigned cc = MRI.createVirtualRegister(RC);
14101 // cl := cmp src_lo, lo
14102 BuildMI(mainMBB, DL, TII->get(X86::CMP32rr))
14103 .addReg(SrcLoReg).addReg(t4L);
14104 BuildMI(mainMBB, DL, TII->get(LoOpc), cL);
14105 BuildMI(mainMBB, DL, TII->get(X86::MOVZX32rr8), cL32).addReg(cL);
14106 // ch := cmp src_hi, hi
14107 BuildMI(mainMBB, DL, TII->get(X86::CMP32rr))
14108 .addReg(SrcHiReg).addReg(t4H);
14109 BuildMI(mainMBB, DL, TII->get(HiOpc), cH);
14110 BuildMI(mainMBB, DL, TII->get(X86::MOVZX32rr8), cH32).addReg(cH);
14111 // cc := if (src_hi == hi) ? cl : ch;
14112 if (Subtarget->hasCMov()) {
14113 BuildMI(mainMBB, DL, TII->get(X86::CMOVE32rr), cc)
14114 .addReg(cH32).addReg(cL32);
14116 MIB = BuildMI(mainMBB, DL, TII->get(X86::CMOV_GR32), cc)
14117 .addReg(cH32).addReg(cL32)
14118 .addImm(X86::COND_E);
14119 mainMBB = EmitLoweredSelect(MIB, mainMBB);
14121 BuildMI(mainMBB, DL, TII->get(X86::TEST32rr)).addReg(cc).addReg(cc);
14122 if (Subtarget->hasCMov()) {
14123 BuildMI(mainMBB, DL, TII->get(X86::CMOVNE32rr), t2L)
14124 .addReg(SrcLoReg).addReg(t4L);
14125 BuildMI(mainMBB, DL, TII->get(X86::CMOVNE32rr), t2H)
14126 .addReg(SrcHiReg).addReg(t4H);
14128 MIB = BuildMI(mainMBB, DL, TII->get(X86::CMOV_GR32), t2L)
14129 .addReg(SrcLoReg).addReg(t4L)
14130 .addImm(X86::COND_NE);
14131 mainMBB = EmitLoweredSelect(MIB, mainMBB);
14132 // As the lowered CMOV won't clobber EFLAGS, we could reuse it for the
14133 // 2nd CMOV lowering.
14134 mainMBB->addLiveIn(X86::EFLAGS);
14135 MIB = BuildMI(mainMBB, DL, TII->get(X86::CMOV_GR32), t2H)
14136 .addReg(SrcHiReg).addReg(t4H)
14137 .addImm(X86::COND_NE);
14138 mainMBB = EmitLoweredSelect(MIB, mainMBB);
14139 // Replace the original PHI node as mainMBB is changed after CMOV
14141 BuildMI(*origMainMBB, PhiL, DL, TII->get(X86::PHI), t4L)
14142 .addReg(t1L).addMBB(thisMBB).addReg(t3L).addMBB(mainMBB);
14143 BuildMI(*origMainMBB, PhiH, DL, TII->get(X86::PHI), t4H)
14144 .addReg(t1H).addMBB(thisMBB).addReg(t3H).addMBB(mainMBB);
14145 PhiL->eraseFromParent();
14146 PhiH->eraseFromParent();
14150 case X86::ATOMSWAP6432: {
14152 unsigned LoOpc = getNonAtomic6432Opcode(Opc, HiOpc);
14153 BuildMI(mainMBB, DL, TII->get(LoOpc), t2L).addReg(SrcLoReg);
14154 BuildMI(mainMBB, DL, TII->get(HiOpc), t2H).addReg(SrcHiReg);
14159 // Copy EDX:EAX back from HiReg:LoReg
14160 BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), X86::EAX).addReg(t4L);
14161 BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), X86::EDX).addReg(t4H);
14162 // Copy ECX:EBX from t1H:t1L
14163 BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), X86::EBX).addReg(t2L);
14164 BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), X86::ECX).addReg(t2H);
14166 MIB = BuildMI(mainMBB, DL, TII->get(LCMPXCHGOpc));
14167 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
14168 MachineOperand NewMO = MI->getOperand(MemOpndSlot + i);
14170 NewMO.setIsKill(false);
14171 MIB.addOperand(NewMO);
14173 MIB.setMemRefs(MMOBegin, MMOEnd);
14175 // Copy EDX:EAX back to t3H:t3L
14176 BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), t3L).addReg(X86::EAX);
14177 BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), t3H).addReg(X86::EDX);
14179 BuildMI(mainMBB, DL, TII->get(X86::JNE_4)).addMBB(origMainMBB);
14181 mainMBB->addSuccessor(origMainMBB);
14182 mainMBB->addSuccessor(sinkMBB);
14185 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
14186 TII->get(TargetOpcode::COPY), DstLoReg)
14188 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
14189 TII->get(TargetOpcode::COPY), DstHiReg)
14192 MI->eraseFromParent();
14196 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
14197 // or XMM0_V32I8 in AVX all of this code can be replaced with that
14198 // in the .td file.
14199 static MachineBasicBlock *EmitPCMPSTRM(MachineInstr *MI, MachineBasicBlock *BB,
14200 const TargetInstrInfo *TII) {
14202 switch (MI->getOpcode()) {
14203 default: llvm_unreachable("illegal opcode!");
14204 case X86::PCMPISTRM128REG: Opc = X86::PCMPISTRM128rr; break;
14205 case X86::VPCMPISTRM128REG: Opc = X86::VPCMPISTRM128rr; break;
14206 case X86::PCMPISTRM128MEM: Opc = X86::PCMPISTRM128rm; break;
14207 case X86::VPCMPISTRM128MEM: Opc = X86::VPCMPISTRM128rm; break;
14208 case X86::PCMPESTRM128REG: Opc = X86::PCMPESTRM128rr; break;
14209 case X86::VPCMPESTRM128REG: Opc = X86::VPCMPESTRM128rr; break;
14210 case X86::PCMPESTRM128MEM: Opc = X86::PCMPESTRM128rm; break;
14211 case X86::VPCMPESTRM128MEM: Opc = X86::VPCMPESTRM128rm; break;
14214 DebugLoc dl = MI->getDebugLoc();
14215 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
14217 unsigned NumArgs = MI->getNumOperands();
14218 for (unsigned i = 1; i < NumArgs; ++i) {
14219 MachineOperand &Op = MI->getOperand(i);
14220 if (!(Op.isReg() && Op.isImplicit()))
14221 MIB.addOperand(Op);
14223 if (MI->hasOneMemOperand())
14224 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
14226 BuildMI(*BB, MI, dl,
14227 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
14228 .addReg(X86::XMM0);
14230 MI->eraseFromParent();
14234 // FIXME: Custom handling because TableGen doesn't support multiple implicit
14235 // defs in an instruction pattern
14236 static MachineBasicBlock *EmitPCMPSTRI(MachineInstr *MI, MachineBasicBlock *BB,
14237 const TargetInstrInfo *TII) {
14239 switch (MI->getOpcode()) {
14240 default: llvm_unreachable("illegal opcode!");
14241 case X86::PCMPISTRIREG: Opc = X86::PCMPISTRIrr; break;
14242 case X86::VPCMPISTRIREG: Opc = X86::VPCMPISTRIrr; break;
14243 case X86::PCMPISTRIMEM: Opc = X86::PCMPISTRIrm; break;
14244 case X86::VPCMPISTRIMEM: Opc = X86::VPCMPISTRIrm; break;
14245 case X86::PCMPESTRIREG: Opc = X86::PCMPESTRIrr; break;
14246 case X86::VPCMPESTRIREG: Opc = X86::VPCMPESTRIrr; break;
14247 case X86::PCMPESTRIMEM: Opc = X86::PCMPESTRIrm; break;
14248 case X86::VPCMPESTRIMEM: Opc = X86::VPCMPESTRIrm; break;
14251 DebugLoc dl = MI->getDebugLoc();
14252 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
14254 unsigned NumArgs = MI->getNumOperands(); // remove the results
14255 for (unsigned i = 1; i < NumArgs; ++i) {
14256 MachineOperand &Op = MI->getOperand(i);
14257 if (!(Op.isReg() && Op.isImplicit()))
14258 MIB.addOperand(Op);
14260 if (MI->hasOneMemOperand())
14261 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
14263 BuildMI(*BB, MI, dl,
14264 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
14267 MI->eraseFromParent();
14271 static MachineBasicBlock * EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB,
14272 const TargetInstrInfo *TII,
14273 const X86Subtarget* Subtarget) {
14274 DebugLoc dl = MI->getDebugLoc();
14276 // Address into RAX/EAX, other two args into ECX, EDX.
14277 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
14278 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
14279 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
14280 for (int i = 0; i < X86::AddrNumOperands; ++i)
14281 MIB.addOperand(MI->getOperand(i));
14283 unsigned ValOps = X86::AddrNumOperands;
14284 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
14285 .addReg(MI->getOperand(ValOps).getReg());
14286 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
14287 .addReg(MI->getOperand(ValOps+1).getReg());
14289 // The instruction doesn't actually take any operands though.
14290 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
14292 MI->eraseFromParent(); // The pseudo is gone now.
14296 MachineBasicBlock *
14297 X86TargetLowering::EmitVAARG64WithCustomInserter(
14299 MachineBasicBlock *MBB) const {
14300 // Emit va_arg instruction on X86-64.
14302 // Operands to this pseudo-instruction:
14303 // 0 ) Output : destination address (reg)
14304 // 1-5) Input : va_list address (addr, i64mem)
14305 // 6 ) ArgSize : Size (in bytes) of vararg type
14306 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
14307 // 8 ) Align : Alignment of type
14308 // 9 ) EFLAGS (implicit-def)
14310 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
14311 assert(X86::AddrNumOperands == 5 && "VAARG_64 assumes 5 address operands");
14313 unsigned DestReg = MI->getOperand(0).getReg();
14314 MachineOperand &Base = MI->getOperand(1);
14315 MachineOperand &Scale = MI->getOperand(2);
14316 MachineOperand &Index = MI->getOperand(3);
14317 MachineOperand &Disp = MI->getOperand(4);
14318 MachineOperand &Segment = MI->getOperand(5);
14319 unsigned ArgSize = MI->getOperand(6).getImm();
14320 unsigned ArgMode = MI->getOperand(7).getImm();
14321 unsigned Align = MI->getOperand(8).getImm();
14323 // Memory Reference
14324 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
14325 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
14326 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
14328 // Machine Information
14329 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
14330 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
14331 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
14332 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
14333 DebugLoc DL = MI->getDebugLoc();
14335 // struct va_list {
14338 // i64 overflow_area (address)
14339 // i64 reg_save_area (address)
14341 // sizeof(va_list) = 24
14342 // alignment(va_list) = 8
14344 unsigned TotalNumIntRegs = 6;
14345 unsigned TotalNumXMMRegs = 8;
14346 bool UseGPOffset = (ArgMode == 1);
14347 bool UseFPOffset = (ArgMode == 2);
14348 unsigned MaxOffset = TotalNumIntRegs * 8 +
14349 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
14351 /* Align ArgSize to a multiple of 8 */
14352 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
14353 bool NeedsAlign = (Align > 8);
14355 MachineBasicBlock *thisMBB = MBB;
14356 MachineBasicBlock *overflowMBB;
14357 MachineBasicBlock *offsetMBB;
14358 MachineBasicBlock *endMBB;
14360 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
14361 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
14362 unsigned OffsetReg = 0;
14364 if (!UseGPOffset && !UseFPOffset) {
14365 // If we only pull from the overflow region, we don't create a branch.
14366 // We don't need to alter control flow.
14367 OffsetDestReg = 0; // unused
14368 OverflowDestReg = DestReg;
14371 overflowMBB = thisMBB;
14374 // First emit code to check if gp_offset (or fp_offset) is below the bound.
14375 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
14376 // If not, pull from overflow_area. (branch to overflowMBB)
14381 // offsetMBB overflowMBB
14386 // Registers for the PHI in endMBB
14387 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
14388 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
14390 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
14391 MachineFunction *MF = MBB->getParent();
14392 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
14393 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
14394 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
14396 MachineFunction::iterator MBBIter = MBB;
14399 // Insert the new basic blocks
14400 MF->insert(MBBIter, offsetMBB);
14401 MF->insert(MBBIter, overflowMBB);
14402 MF->insert(MBBIter, endMBB);
14404 // Transfer the remainder of MBB and its successor edges to endMBB.
14405 endMBB->splice(endMBB->begin(), thisMBB,
14406 llvm::next(MachineBasicBlock::iterator(MI)),
14408 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
14410 // Make offsetMBB and overflowMBB successors of thisMBB
14411 thisMBB->addSuccessor(offsetMBB);
14412 thisMBB->addSuccessor(overflowMBB);
14414 // endMBB is a successor of both offsetMBB and overflowMBB
14415 offsetMBB->addSuccessor(endMBB);
14416 overflowMBB->addSuccessor(endMBB);
14418 // Load the offset value into a register
14419 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
14420 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
14424 .addDisp(Disp, UseFPOffset ? 4 : 0)
14425 .addOperand(Segment)
14426 .setMemRefs(MMOBegin, MMOEnd);
14428 // Check if there is enough room left to pull this argument.
14429 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
14431 .addImm(MaxOffset + 8 - ArgSizeA8);
14433 // Branch to "overflowMBB" if offset >= max
14434 // Fall through to "offsetMBB" otherwise
14435 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
14436 .addMBB(overflowMBB);
14439 // In offsetMBB, emit code to use the reg_save_area.
14441 assert(OffsetReg != 0);
14443 // Read the reg_save_area address.
14444 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
14445 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
14450 .addOperand(Segment)
14451 .setMemRefs(MMOBegin, MMOEnd);
14453 // Zero-extend the offset
14454 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
14455 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
14458 .addImm(X86::sub_32bit);
14460 // Add the offset to the reg_save_area to get the final address.
14461 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
14462 .addReg(OffsetReg64)
14463 .addReg(RegSaveReg);
14465 // Compute the offset for the next argument
14466 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
14467 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
14469 .addImm(UseFPOffset ? 16 : 8);
14471 // Store it back into the va_list.
14472 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
14476 .addDisp(Disp, UseFPOffset ? 4 : 0)
14477 .addOperand(Segment)
14478 .addReg(NextOffsetReg)
14479 .setMemRefs(MMOBegin, MMOEnd);
14482 BuildMI(offsetMBB, DL, TII->get(X86::JMP_4))
14487 // Emit code to use overflow area
14490 // Load the overflow_area address into a register.
14491 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
14492 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
14497 .addOperand(Segment)
14498 .setMemRefs(MMOBegin, MMOEnd);
14500 // If we need to align it, do so. Otherwise, just copy the address
14501 // to OverflowDestReg.
14503 // Align the overflow address
14504 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
14505 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
14507 // aligned_addr = (addr + (align-1)) & ~(align-1)
14508 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
14509 .addReg(OverflowAddrReg)
14512 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
14514 .addImm(~(uint64_t)(Align-1));
14516 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
14517 .addReg(OverflowAddrReg);
14520 // Compute the next overflow address after this argument.
14521 // (the overflow address should be kept 8-byte aligned)
14522 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
14523 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
14524 .addReg(OverflowDestReg)
14525 .addImm(ArgSizeA8);
14527 // Store the new overflow address.
14528 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
14533 .addOperand(Segment)
14534 .addReg(NextAddrReg)
14535 .setMemRefs(MMOBegin, MMOEnd);
14537 // If we branched, emit the PHI to the front of endMBB.
14539 BuildMI(*endMBB, endMBB->begin(), DL,
14540 TII->get(X86::PHI), DestReg)
14541 .addReg(OffsetDestReg).addMBB(offsetMBB)
14542 .addReg(OverflowDestReg).addMBB(overflowMBB);
14545 // Erase the pseudo instruction
14546 MI->eraseFromParent();
14551 MachineBasicBlock *
14552 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
14554 MachineBasicBlock *MBB) const {
14555 // Emit code to save XMM registers to the stack. The ABI says that the
14556 // number of registers to save is given in %al, so it's theoretically
14557 // possible to do an indirect jump trick to avoid saving all of them,
14558 // however this code takes a simpler approach and just executes all
14559 // of the stores if %al is non-zero. It's less code, and it's probably
14560 // easier on the hardware branch predictor, and stores aren't all that
14561 // expensive anyway.
14563 // Create the new basic blocks. One block contains all the XMM stores,
14564 // and one block is the final destination regardless of whether any
14565 // stores were performed.
14566 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
14567 MachineFunction *F = MBB->getParent();
14568 MachineFunction::iterator MBBIter = MBB;
14570 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
14571 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
14572 F->insert(MBBIter, XMMSaveMBB);
14573 F->insert(MBBIter, EndMBB);
14575 // Transfer the remainder of MBB and its successor edges to EndMBB.
14576 EndMBB->splice(EndMBB->begin(), MBB,
14577 llvm::next(MachineBasicBlock::iterator(MI)),
14579 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
14581 // The original block will now fall through to the XMM save block.
14582 MBB->addSuccessor(XMMSaveMBB);
14583 // The XMMSaveMBB will fall through to the end block.
14584 XMMSaveMBB->addSuccessor(EndMBB);
14586 // Now add the instructions.
14587 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
14588 DebugLoc DL = MI->getDebugLoc();
14590 unsigned CountReg = MI->getOperand(0).getReg();
14591 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
14592 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
14594 if (!Subtarget->isTargetWin64()) {
14595 // If %al is 0, branch around the XMM save block.
14596 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
14597 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
14598 MBB->addSuccessor(EndMBB);
14601 unsigned MOVOpc = Subtarget->hasFp256() ? X86::VMOVAPSmr : X86::MOVAPSmr;
14602 // In the XMM save block, save all the XMM argument registers.
14603 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
14604 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
14605 MachineMemOperand *MMO =
14606 F->getMachineMemOperand(
14607 MachinePointerInfo::getFixedStack(RegSaveFrameIndex, Offset),
14608 MachineMemOperand::MOStore,
14609 /*Size=*/16, /*Align=*/16);
14610 BuildMI(XMMSaveMBB, DL, TII->get(MOVOpc))
14611 .addFrameIndex(RegSaveFrameIndex)
14612 .addImm(/*Scale=*/1)
14613 .addReg(/*IndexReg=*/0)
14614 .addImm(/*Disp=*/Offset)
14615 .addReg(/*Segment=*/0)
14616 .addReg(MI->getOperand(i).getReg())
14617 .addMemOperand(MMO);
14620 MI->eraseFromParent(); // The pseudo instruction is gone now.
14625 // The EFLAGS operand of SelectItr might be missing a kill marker
14626 // because there were multiple uses of EFLAGS, and ISel didn't know
14627 // which to mark. Figure out whether SelectItr should have had a
14628 // kill marker, and set it if it should. Returns the correct kill
14630 static bool checkAndUpdateEFLAGSKill(MachineBasicBlock::iterator SelectItr,
14631 MachineBasicBlock* BB,
14632 const TargetRegisterInfo* TRI) {
14633 // Scan forward through BB for a use/def of EFLAGS.
14634 MachineBasicBlock::iterator miI(llvm::next(SelectItr));
14635 for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) {
14636 const MachineInstr& mi = *miI;
14637 if (mi.readsRegister(X86::EFLAGS))
14639 if (mi.definesRegister(X86::EFLAGS))
14640 break; // Should have kill-flag - update below.
14643 // If we hit the end of the block, check whether EFLAGS is live into a
14645 if (miI == BB->end()) {
14646 for (MachineBasicBlock::succ_iterator sItr = BB->succ_begin(),
14647 sEnd = BB->succ_end();
14648 sItr != sEnd; ++sItr) {
14649 MachineBasicBlock* succ = *sItr;
14650 if (succ->isLiveIn(X86::EFLAGS))
14655 // We found a def, or hit the end of the basic block and EFLAGS wasn't live
14656 // out. SelectMI should have a kill flag on EFLAGS.
14657 SelectItr->addRegisterKilled(X86::EFLAGS, TRI);
14661 MachineBasicBlock *
14662 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
14663 MachineBasicBlock *BB) const {
14664 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
14665 DebugLoc DL = MI->getDebugLoc();
14667 // To "insert" a SELECT_CC instruction, we actually have to insert the
14668 // diamond control-flow pattern. The incoming instruction knows the
14669 // destination vreg to set, the condition code register to branch on, the
14670 // true/false values to select between, and a branch opcode to use.
14671 const BasicBlock *LLVM_BB = BB->getBasicBlock();
14672 MachineFunction::iterator It = BB;
14678 // cmpTY ccX, r1, r2
14680 // fallthrough --> copy0MBB
14681 MachineBasicBlock *thisMBB = BB;
14682 MachineFunction *F = BB->getParent();
14683 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
14684 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
14685 F->insert(It, copy0MBB);
14686 F->insert(It, sinkMBB);
14688 // If the EFLAGS register isn't dead in the terminator, then claim that it's
14689 // live into the sink and copy blocks.
14690 const TargetRegisterInfo* TRI = getTargetMachine().getRegisterInfo();
14691 if (!MI->killsRegister(X86::EFLAGS) &&
14692 !checkAndUpdateEFLAGSKill(MI, BB, TRI)) {
14693 copy0MBB->addLiveIn(X86::EFLAGS);
14694 sinkMBB->addLiveIn(X86::EFLAGS);
14697 // Transfer the remainder of BB and its successor edges to sinkMBB.
14698 sinkMBB->splice(sinkMBB->begin(), BB,
14699 llvm::next(MachineBasicBlock::iterator(MI)),
14701 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
14703 // Add the true and fallthrough blocks as its successors.
14704 BB->addSuccessor(copy0MBB);
14705 BB->addSuccessor(sinkMBB);
14707 // Create the conditional branch instruction.
14709 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
14710 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
14713 // %FalseValue = ...
14714 // # fallthrough to sinkMBB
14715 copy0MBB->addSuccessor(sinkMBB);
14718 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
14720 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
14721 TII->get(X86::PHI), MI->getOperand(0).getReg())
14722 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
14723 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
14725 MI->eraseFromParent(); // The pseudo instruction is gone now.
14729 MachineBasicBlock *
14730 X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI, MachineBasicBlock *BB,
14731 bool Is64Bit) const {
14732 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
14733 DebugLoc DL = MI->getDebugLoc();
14734 MachineFunction *MF = BB->getParent();
14735 const BasicBlock *LLVM_BB = BB->getBasicBlock();
14737 assert(getTargetMachine().Options.EnableSegmentedStacks);
14739 unsigned TlsReg = Is64Bit ? X86::FS : X86::GS;
14740 unsigned TlsOffset = Is64Bit ? 0x70 : 0x30;
14743 // ... [Till the alloca]
14744 // If stacklet is not large enough, jump to mallocMBB
14747 // Allocate by subtracting from RSP
14748 // Jump to continueMBB
14751 // Allocate by call to runtime
14755 // [rest of original BB]
14758 MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB);
14759 MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB);
14760 MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB);
14762 MachineRegisterInfo &MRI = MF->getRegInfo();
14763 const TargetRegisterClass *AddrRegClass =
14764 getRegClassFor(Is64Bit ? MVT::i64:MVT::i32);
14766 unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass),
14767 bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass),
14768 tmpSPVReg = MRI.createVirtualRegister(AddrRegClass),
14769 SPLimitVReg = MRI.createVirtualRegister(AddrRegClass),
14770 sizeVReg = MI->getOperand(1).getReg(),
14771 physSPReg = Is64Bit ? X86::RSP : X86::ESP;
14773 MachineFunction::iterator MBBIter = BB;
14776 MF->insert(MBBIter, bumpMBB);
14777 MF->insert(MBBIter, mallocMBB);
14778 MF->insert(MBBIter, continueMBB);
14780 continueMBB->splice(continueMBB->begin(), BB, llvm::next
14781 (MachineBasicBlock::iterator(MI)), BB->end());
14782 continueMBB->transferSuccessorsAndUpdatePHIs(BB);
14784 // Add code to the main basic block to check if the stack limit has been hit,
14785 // and if so, jump to mallocMBB otherwise to bumpMBB.
14786 BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg);
14787 BuildMI(BB, DL, TII->get(Is64Bit ? X86::SUB64rr:X86::SUB32rr), SPLimitVReg)
14788 .addReg(tmpSPVReg).addReg(sizeVReg);
14789 BuildMI(BB, DL, TII->get(Is64Bit ? X86::CMP64mr:X86::CMP32mr))
14790 .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg)
14791 .addReg(SPLimitVReg);
14792 BuildMI(BB, DL, TII->get(X86::JG_4)).addMBB(mallocMBB);
14794 // bumpMBB simply decreases the stack pointer, since we know the current
14795 // stacklet has enough space.
14796 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg)
14797 .addReg(SPLimitVReg);
14798 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg)
14799 .addReg(SPLimitVReg);
14800 BuildMI(bumpMBB, DL, TII->get(X86::JMP_4)).addMBB(continueMBB);
14802 // Calls into a routine in libgcc to allocate more space from the heap.
14803 const uint32_t *RegMask =
14804 getTargetMachine().getRegisterInfo()->getCallPreservedMask(CallingConv::C);
14806 BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
14808 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
14809 .addExternalSymbol("__morestack_allocate_stack_space")
14810 .addRegMask(RegMask)
14811 .addReg(X86::RDI, RegState::Implicit)
14812 .addReg(X86::RAX, RegState::ImplicitDefine);
14814 BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg)
14816 BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg);
14817 BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32))
14818 .addExternalSymbol("__morestack_allocate_stack_space")
14819 .addRegMask(RegMask)
14820 .addReg(X86::EAX, RegState::ImplicitDefine);
14824 BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg)
14827 BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg)
14828 .addReg(Is64Bit ? X86::RAX : X86::EAX);
14829 BuildMI(mallocMBB, DL, TII->get(X86::JMP_4)).addMBB(continueMBB);
14831 // Set up the CFG correctly.
14832 BB->addSuccessor(bumpMBB);
14833 BB->addSuccessor(mallocMBB);
14834 mallocMBB->addSuccessor(continueMBB);
14835 bumpMBB->addSuccessor(continueMBB);
14837 // Take care of the PHI nodes.
14838 BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI),
14839 MI->getOperand(0).getReg())
14840 .addReg(mallocPtrVReg).addMBB(mallocMBB)
14841 .addReg(bumpSPPtrVReg).addMBB(bumpMBB);
14843 // Delete the original pseudo instruction.
14844 MI->eraseFromParent();
14847 return continueMBB;
14850 MachineBasicBlock *
14851 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
14852 MachineBasicBlock *BB) const {
14853 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
14854 DebugLoc DL = MI->getDebugLoc();
14856 assert(!Subtarget->isTargetEnvMacho());
14858 // The lowering is pretty easy: we're just emitting the call to _alloca. The
14859 // non-trivial part is impdef of ESP.
14861 if (Subtarget->isTargetWin64()) {
14862 if (Subtarget->isTargetCygMing()) {
14863 // ___chkstk(Mingw64):
14864 // Clobbers R10, R11, RAX and EFLAGS.
14866 BuildMI(*BB, MI, DL, TII->get(X86::W64ALLOCA))
14867 .addExternalSymbol("___chkstk")
14868 .addReg(X86::RAX, RegState::Implicit)
14869 .addReg(X86::RSP, RegState::Implicit)
14870 .addReg(X86::RAX, RegState::Define | RegState::Implicit)
14871 .addReg(X86::RSP, RegState::Define | RegState::Implicit)
14872 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
14874 // __chkstk(MSVCRT): does not update stack pointer.
14875 // Clobbers R10, R11 and EFLAGS.
14876 BuildMI(*BB, MI, DL, TII->get(X86::W64ALLOCA))
14877 .addExternalSymbol("__chkstk")
14878 .addReg(X86::RAX, RegState::Implicit)
14879 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
14880 // RAX has the offset to be subtracted from RSP.
14881 BuildMI(*BB, MI, DL, TII->get(X86::SUB64rr), X86::RSP)
14886 const char *StackProbeSymbol =
14887 Subtarget->isTargetWindows() ? "_chkstk" : "_alloca";
14889 BuildMI(*BB, MI, DL, TII->get(X86::CALLpcrel32))
14890 .addExternalSymbol(StackProbeSymbol)
14891 .addReg(X86::EAX, RegState::Implicit)
14892 .addReg(X86::ESP, RegState::Implicit)
14893 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
14894 .addReg(X86::ESP, RegState::Define | RegState::Implicit)
14895 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
14898 MI->eraseFromParent(); // The pseudo instruction is gone now.
14902 MachineBasicBlock *
14903 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
14904 MachineBasicBlock *BB) const {
14905 // This is pretty easy. We're taking the value that we received from
14906 // our load from the relocation, sticking it in either RDI (x86-64)
14907 // or EAX and doing an indirect call. The return value will then
14908 // be in the normal return register.
14909 const X86InstrInfo *TII
14910 = static_cast<const X86InstrInfo*>(getTargetMachine().getInstrInfo());
14911 DebugLoc DL = MI->getDebugLoc();
14912 MachineFunction *F = BB->getParent();
14914 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
14915 assert(MI->getOperand(3).isGlobal() && "This should be a global");
14917 // Get a register mask for the lowered call.
14918 // FIXME: The 32-bit calls have non-standard calling conventions. Use a
14919 // proper register mask.
14920 const uint32_t *RegMask =
14921 getTargetMachine().getRegisterInfo()->getCallPreservedMask(CallingConv::C);
14922 if (Subtarget->is64Bit()) {
14923 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
14924 TII->get(X86::MOV64rm), X86::RDI)
14926 .addImm(0).addReg(0)
14927 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
14928 MI->getOperand(3).getTargetFlags())
14930 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
14931 addDirectMem(MIB, X86::RDI);
14932 MIB.addReg(X86::RAX, RegState::ImplicitDefine).addRegMask(RegMask);
14933 } else if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
14934 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
14935 TII->get(X86::MOV32rm), X86::EAX)
14937 .addImm(0).addReg(0)
14938 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
14939 MI->getOperand(3).getTargetFlags())
14941 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
14942 addDirectMem(MIB, X86::EAX);
14943 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
14945 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
14946 TII->get(X86::MOV32rm), X86::EAX)
14947 .addReg(TII->getGlobalBaseReg(F))
14948 .addImm(0).addReg(0)
14949 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
14950 MI->getOperand(3).getTargetFlags())
14952 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
14953 addDirectMem(MIB, X86::EAX);
14954 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
14957 MI->eraseFromParent(); // The pseudo instruction is gone now.
14961 MachineBasicBlock *
14962 X86TargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
14963 MachineBasicBlock *MBB) const {
14964 DebugLoc DL = MI->getDebugLoc();
14965 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
14967 MachineFunction *MF = MBB->getParent();
14968 MachineRegisterInfo &MRI = MF->getRegInfo();
14970 const BasicBlock *BB = MBB->getBasicBlock();
14971 MachineFunction::iterator I = MBB;
14974 // Memory Reference
14975 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
14976 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
14979 unsigned MemOpndSlot = 0;
14981 unsigned CurOp = 0;
14983 DstReg = MI->getOperand(CurOp++).getReg();
14984 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
14985 assert(RC->hasType(MVT::i32) && "Invalid destination!");
14986 unsigned mainDstReg = MRI.createVirtualRegister(RC);
14987 unsigned restoreDstReg = MRI.createVirtualRegister(RC);
14989 MemOpndSlot = CurOp;
14991 MVT PVT = getPointerTy();
14992 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
14993 "Invalid Pointer Size!");
14995 // For v = setjmp(buf), we generate
14998 // buf[LabelOffset] = restoreMBB
14999 // SjLjSetup restoreMBB
15005 // v = phi(main, restore)
15010 MachineBasicBlock *thisMBB = MBB;
15011 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
15012 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
15013 MachineBasicBlock *restoreMBB = MF->CreateMachineBasicBlock(BB);
15014 MF->insert(I, mainMBB);
15015 MF->insert(I, sinkMBB);
15016 MF->push_back(restoreMBB);
15018 MachineInstrBuilder MIB;
15020 // Transfer the remainder of BB and its successor edges to sinkMBB.
15021 sinkMBB->splice(sinkMBB->begin(), MBB,
15022 llvm::next(MachineBasicBlock::iterator(MI)), MBB->end());
15023 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
15026 unsigned PtrStoreOpc = 0;
15027 unsigned LabelReg = 0;
15028 const int64_t LabelOffset = 1 * PVT.getStoreSize();
15029 Reloc::Model RM = getTargetMachine().getRelocationModel();
15030 bool UseImmLabel = (getTargetMachine().getCodeModel() == CodeModel::Small) &&
15031 (RM == Reloc::Static || RM == Reloc::DynamicNoPIC);
15033 // Prepare IP either in reg or imm.
15034 if (!UseImmLabel) {
15035 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr;
15036 const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
15037 LabelReg = MRI.createVirtualRegister(PtrRC);
15038 if (Subtarget->is64Bit()) {
15039 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA64r), LabelReg)
15043 .addMBB(restoreMBB)
15046 const X86InstrInfo *XII = static_cast<const X86InstrInfo*>(TII);
15047 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA32r), LabelReg)
15048 .addReg(XII->getGlobalBaseReg(MF))
15051 .addMBB(restoreMBB, Subtarget->ClassifyBlockAddressReference())
15055 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi;
15057 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PtrStoreOpc));
15058 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
15059 if (i == X86::AddrDisp)
15060 MIB.addDisp(MI->getOperand(MemOpndSlot + i), LabelOffset);
15062 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
15065 MIB.addReg(LabelReg);
15067 MIB.addMBB(restoreMBB);
15068 MIB.setMemRefs(MMOBegin, MMOEnd);
15070 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
15071 .addMBB(restoreMBB);
15073 const X86RegisterInfo *RegInfo =
15074 static_cast<const X86RegisterInfo*>(getTargetMachine().getRegisterInfo());
15075 MIB.addRegMask(RegInfo->getNoPreservedMask());
15076 thisMBB->addSuccessor(mainMBB);
15077 thisMBB->addSuccessor(restoreMBB);
15081 BuildMI(mainMBB, DL, TII->get(X86::MOV32r0), mainDstReg);
15082 mainMBB->addSuccessor(sinkMBB);
15085 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
15086 TII->get(X86::PHI), DstReg)
15087 .addReg(mainDstReg).addMBB(mainMBB)
15088 .addReg(restoreDstReg).addMBB(restoreMBB);
15091 BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1);
15092 BuildMI(restoreMBB, DL, TII->get(X86::JMP_4)).addMBB(sinkMBB);
15093 restoreMBB->addSuccessor(sinkMBB);
15095 MI->eraseFromParent();
15099 MachineBasicBlock *
15100 X86TargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
15101 MachineBasicBlock *MBB) const {
15102 DebugLoc DL = MI->getDebugLoc();
15103 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
15105 MachineFunction *MF = MBB->getParent();
15106 MachineRegisterInfo &MRI = MF->getRegInfo();
15108 // Memory Reference
15109 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
15110 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
15112 MVT PVT = getPointerTy();
15113 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
15114 "Invalid Pointer Size!");
15116 const TargetRegisterClass *RC =
15117 (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass;
15118 unsigned Tmp = MRI.createVirtualRegister(RC);
15119 // Since FP is only updated here but NOT referenced, it's treated as GPR.
15120 const X86RegisterInfo *RegInfo =
15121 static_cast<const X86RegisterInfo*>(getTargetMachine().getRegisterInfo());
15122 unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
15123 unsigned SP = RegInfo->getStackRegister();
15125 MachineInstrBuilder MIB;
15127 const int64_t LabelOffset = 1 * PVT.getStoreSize();
15128 const int64_t SPOffset = 2 * PVT.getStoreSize();
15130 unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm;
15131 unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r;
15134 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP);
15135 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
15136 MIB.addOperand(MI->getOperand(i));
15137 MIB.setMemRefs(MMOBegin, MMOEnd);
15139 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp);
15140 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
15141 if (i == X86::AddrDisp)
15142 MIB.addDisp(MI->getOperand(i), LabelOffset);
15144 MIB.addOperand(MI->getOperand(i));
15146 MIB.setMemRefs(MMOBegin, MMOEnd);
15148 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP);
15149 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
15150 if (i == X86::AddrDisp)
15151 MIB.addDisp(MI->getOperand(i), SPOffset);
15153 MIB.addOperand(MI->getOperand(i));
15155 MIB.setMemRefs(MMOBegin, MMOEnd);
15157 BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp);
15159 MI->eraseFromParent();
15163 MachineBasicBlock *
15164 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
15165 MachineBasicBlock *BB) const {
15166 switch (MI->getOpcode()) {
15167 default: llvm_unreachable("Unexpected instr type to insert");
15168 case X86::TAILJMPd64:
15169 case X86::TAILJMPr64:
15170 case X86::TAILJMPm64:
15171 llvm_unreachable("TAILJMP64 would not be touched here.");
15172 case X86::TCRETURNdi64:
15173 case X86::TCRETURNri64:
15174 case X86::TCRETURNmi64:
15176 case X86::WIN_ALLOCA:
15177 return EmitLoweredWinAlloca(MI, BB);
15178 case X86::SEG_ALLOCA_32:
15179 return EmitLoweredSegAlloca(MI, BB, false);
15180 case X86::SEG_ALLOCA_64:
15181 return EmitLoweredSegAlloca(MI, BB, true);
15182 case X86::TLSCall_32:
15183 case X86::TLSCall_64:
15184 return EmitLoweredTLSCall(MI, BB);
15185 case X86::CMOV_GR8:
15186 case X86::CMOV_FR32:
15187 case X86::CMOV_FR64:
15188 case X86::CMOV_V4F32:
15189 case X86::CMOV_V2F64:
15190 case X86::CMOV_V2I64:
15191 case X86::CMOV_V8F32:
15192 case X86::CMOV_V4F64:
15193 case X86::CMOV_V4I64:
15194 case X86::CMOV_GR16:
15195 case X86::CMOV_GR32:
15196 case X86::CMOV_RFP32:
15197 case X86::CMOV_RFP64:
15198 case X86::CMOV_RFP80:
15199 return EmitLoweredSelect(MI, BB);
15201 case X86::FP32_TO_INT16_IN_MEM:
15202 case X86::FP32_TO_INT32_IN_MEM:
15203 case X86::FP32_TO_INT64_IN_MEM:
15204 case X86::FP64_TO_INT16_IN_MEM:
15205 case X86::FP64_TO_INT32_IN_MEM:
15206 case X86::FP64_TO_INT64_IN_MEM:
15207 case X86::FP80_TO_INT16_IN_MEM:
15208 case X86::FP80_TO_INT32_IN_MEM:
15209 case X86::FP80_TO_INT64_IN_MEM: {
15210 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
15211 DebugLoc DL = MI->getDebugLoc();
15213 // Change the floating point control register to use "round towards zero"
15214 // mode when truncating to an integer value.
15215 MachineFunction *F = BB->getParent();
15216 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
15217 addFrameReference(BuildMI(*BB, MI, DL,
15218 TII->get(X86::FNSTCW16m)), CWFrameIdx);
15220 // Load the old value of the high byte of the control word...
15222 F->getRegInfo().createVirtualRegister(&X86::GR16RegClass);
15223 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
15226 // Set the high part to be round to zero...
15227 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
15230 // Reload the modified control word now...
15231 addFrameReference(BuildMI(*BB, MI, DL,
15232 TII->get(X86::FLDCW16m)), CWFrameIdx);
15234 // Restore the memory image of control word to original value
15235 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
15238 // Get the X86 opcode to use.
15240 switch (MI->getOpcode()) {
15241 default: llvm_unreachable("illegal opcode!");
15242 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
15243 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
15244 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
15245 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
15246 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
15247 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
15248 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
15249 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
15250 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
15254 MachineOperand &Op = MI->getOperand(0);
15256 AM.BaseType = X86AddressMode::RegBase;
15257 AM.Base.Reg = Op.getReg();
15259 AM.BaseType = X86AddressMode::FrameIndexBase;
15260 AM.Base.FrameIndex = Op.getIndex();
15262 Op = MI->getOperand(1);
15264 AM.Scale = Op.getImm();
15265 Op = MI->getOperand(2);
15267 AM.IndexReg = Op.getImm();
15268 Op = MI->getOperand(3);
15269 if (Op.isGlobal()) {
15270 AM.GV = Op.getGlobal();
15272 AM.Disp = Op.getImm();
15274 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
15275 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
15277 // Reload the original control word now.
15278 addFrameReference(BuildMI(*BB, MI, DL,
15279 TII->get(X86::FLDCW16m)), CWFrameIdx);
15281 MI->eraseFromParent(); // The pseudo instruction is gone now.
15284 // String/text processing lowering.
15285 case X86::PCMPISTRM128REG:
15286 case X86::VPCMPISTRM128REG:
15287 case X86::PCMPISTRM128MEM:
15288 case X86::VPCMPISTRM128MEM:
15289 case X86::PCMPESTRM128REG:
15290 case X86::VPCMPESTRM128REG:
15291 case X86::PCMPESTRM128MEM:
15292 case X86::VPCMPESTRM128MEM:
15293 assert(Subtarget->hasSSE42() &&
15294 "Target must have SSE4.2 or AVX features enabled");
15295 return EmitPCMPSTRM(MI, BB, getTargetMachine().getInstrInfo());
15297 // String/text processing lowering.
15298 case X86::PCMPISTRIREG:
15299 case X86::VPCMPISTRIREG:
15300 case X86::PCMPISTRIMEM:
15301 case X86::VPCMPISTRIMEM:
15302 case X86::PCMPESTRIREG:
15303 case X86::VPCMPESTRIREG:
15304 case X86::PCMPESTRIMEM:
15305 case X86::VPCMPESTRIMEM:
15306 assert(Subtarget->hasSSE42() &&
15307 "Target must have SSE4.2 or AVX features enabled");
15308 return EmitPCMPSTRI(MI, BB, getTargetMachine().getInstrInfo());
15310 // Thread synchronization.
15312 return EmitMonitor(MI, BB, getTargetMachine().getInstrInfo(), Subtarget);
15316 return EmitXBegin(MI, BB, getTargetMachine().getInstrInfo());
15318 // Atomic Lowering.
15319 case X86::ATOMAND8:
15320 case X86::ATOMAND16:
15321 case X86::ATOMAND32:
15322 case X86::ATOMAND64:
15325 case X86::ATOMOR16:
15326 case X86::ATOMOR32:
15327 case X86::ATOMOR64:
15329 case X86::ATOMXOR16:
15330 case X86::ATOMXOR8:
15331 case X86::ATOMXOR32:
15332 case X86::ATOMXOR64:
15334 case X86::ATOMNAND8:
15335 case X86::ATOMNAND16:
15336 case X86::ATOMNAND32:
15337 case X86::ATOMNAND64:
15339 case X86::ATOMMAX8:
15340 case X86::ATOMMAX16:
15341 case X86::ATOMMAX32:
15342 case X86::ATOMMAX64:
15344 case X86::ATOMMIN8:
15345 case X86::ATOMMIN16:
15346 case X86::ATOMMIN32:
15347 case X86::ATOMMIN64:
15349 case X86::ATOMUMAX8:
15350 case X86::ATOMUMAX16:
15351 case X86::ATOMUMAX32:
15352 case X86::ATOMUMAX64:
15354 case X86::ATOMUMIN8:
15355 case X86::ATOMUMIN16:
15356 case X86::ATOMUMIN32:
15357 case X86::ATOMUMIN64:
15358 return EmitAtomicLoadArith(MI, BB);
15360 // This group does 64-bit operations on a 32-bit host.
15361 case X86::ATOMAND6432:
15362 case X86::ATOMOR6432:
15363 case X86::ATOMXOR6432:
15364 case X86::ATOMNAND6432:
15365 case X86::ATOMADD6432:
15366 case X86::ATOMSUB6432:
15367 case X86::ATOMMAX6432:
15368 case X86::ATOMMIN6432:
15369 case X86::ATOMUMAX6432:
15370 case X86::ATOMUMIN6432:
15371 case X86::ATOMSWAP6432:
15372 return EmitAtomicLoadArith6432(MI, BB);
15374 case X86::VASTART_SAVE_XMM_REGS:
15375 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
15377 case X86::VAARG_64:
15378 return EmitVAARG64WithCustomInserter(MI, BB);
15380 case X86::EH_SjLj_SetJmp32:
15381 case X86::EH_SjLj_SetJmp64:
15382 return emitEHSjLjSetJmp(MI, BB);
15384 case X86::EH_SjLj_LongJmp32:
15385 case X86::EH_SjLj_LongJmp64:
15386 return emitEHSjLjLongJmp(MI, BB);
15390 //===----------------------------------------------------------------------===//
15391 // X86 Optimization Hooks
15392 //===----------------------------------------------------------------------===//
15394 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
15397 const SelectionDAG &DAG,
15398 unsigned Depth) const {
15399 unsigned BitWidth = KnownZero.getBitWidth();
15400 unsigned Opc = Op.getOpcode();
15401 assert((Opc >= ISD::BUILTIN_OP_END ||
15402 Opc == ISD::INTRINSIC_WO_CHAIN ||
15403 Opc == ISD::INTRINSIC_W_CHAIN ||
15404 Opc == ISD::INTRINSIC_VOID) &&
15405 "Should use MaskedValueIsZero if you don't know whether Op"
15406 " is a target node!");
15408 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
15422 // These nodes' second result is a boolean.
15423 if (Op.getResNo() == 0)
15426 case X86ISD::SETCC:
15427 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
15429 case ISD::INTRINSIC_WO_CHAIN: {
15430 unsigned IntId = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
15431 unsigned NumLoBits = 0;
15434 case Intrinsic::x86_sse_movmsk_ps:
15435 case Intrinsic::x86_avx_movmsk_ps_256:
15436 case Intrinsic::x86_sse2_movmsk_pd:
15437 case Intrinsic::x86_avx_movmsk_pd_256:
15438 case Intrinsic::x86_mmx_pmovmskb:
15439 case Intrinsic::x86_sse2_pmovmskb_128:
15440 case Intrinsic::x86_avx2_pmovmskb: {
15441 // High bits of movmskp{s|d}, pmovmskb are known zero.
15443 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
15444 case Intrinsic::x86_sse_movmsk_ps: NumLoBits = 4; break;
15445 case Intrinsic::x86_avx_movmsk_ps_256: NumLoBits = 8; break;
15446 case Intrinsic::x86_sse2_movmsk_pd: NumLoBits = 2; break;
15447 case Intrinsic::x86_avx_movmsk_pd_256: NumLoBits = 4; break;
15448 case Intrinsic::x86_mmx_pmovmskb: NumLoBits = 8; break;
15449 case Intrinsic::x86_sse2_pmovmskb_128: NumLoBits = 16; break;
15450 case Intrinsic::x86_avx2_pmovmskb: NumLoBits = 32; break;
15452 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - NumLoBits);
15461 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
15462 unsigned Depth) const {
15463 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
15464 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
15465 return Op.getValueType().getScalarType().getSizeInBits();
15471 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
15472 /// node is a GlobalAddress + offset.
15473 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
15474 const GlobalValue* &GA,
15475 int64_t &Offset) const {
15476 if (N->getOpcode() == X86ISD::Wrapper) {
15477 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
15478 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
15479 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
15483 return TargetLowering::isGAPlusOffset(N, GA, Offset);
15486 /// isShuffleHigh128VectorInsertLow - Checks whether the shuffle node is the
15487 /// same as extracting the high 128-bit part of 256-bit vector and then
15488 /// inserting the result into the low part of a new 256-bit vector
15489 static bool isShuffleHigh128VectorInsertLow(ShuffleVectorSDNode *SVOp) {
15490 EVT VT = SVOp->getValueType(0);
15491 unsigned NumElems = VT.getVectorNumElements();
15493 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
15494 for (unsigned i = 0, j = NumElems/2; i != NumElems/2; ++i, ++j)
15495 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
15496 SVOp->getMaskElt(j) >= 0)
15502 /// isShuffleLow128VectorInsertHigh - Checks whether the shuffle node is the
15503 /// same as extracting the low 128-bit part of 256-bit vector and then
15504 /// inserting the result into the high part of a new 256-bit vector
15505 static bool isShuffleLow128VectorInsertHigh(ShuffleVectorSDNode *SVOp) {
15506 EVT VT = SVOp->getValueType(0);
15507 unsigned NumElems = VT.getVectorNumElements();
15509 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
15510 for (unsigned i = NumElems/2, j = 0; i != NumElems; ++i, ++j)
15511 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
15512 SVOp->getMaskElt(j) >= 0)
15518 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
15519 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
15520 TargetLowering::DAGCombinerInfo &DCI,
15521 const X86Subtarget* Subtarget) {
15523 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
15524 SDValue V1 = SVOp->getOperand(0);
15525 SDValue V2 = SVOp->getOperand(1);
15526 EVT VT = SVOp->getValueType(0);
15527 unsigned NumElems = VT.getVectorNumElements();
15529 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
15530 V2.getOpcode() == ISD::CONCAT_VECTORS) {
15534 // V UNDEF BUILD_VECTOR UNDEF
15536 // CONCAT_VECTOR CONCAT_VECTOR
15539 // RESULT: V + zero extended
15541 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
15542 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
15543 V1.getOperand(1).getOpcode() != ISD::UNDEF)
15546 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
15549 // To match the shuffle mask, the first half of the mask should
15550 // be exactly the first vector, and all the rest a splat with the
15551 // first element of the second one.
15552 for (unsigned i = 0; i != NumElems/2; ++i)
15553 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
15554 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
15557 // If V1 is coming from a vector load then just fold to a VZEXT_LOAD.
15558 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(V1.getOperand(0))) {
15559 if (Ld->hasNUsesOfValue(1, 0)) {
15560 SDVTList Tys = DAG.getVTList(MVT::v4i64, MVT::Other);
15561 SDValue Ops[] = { Ld->getChain(), Ld->getBasePtr() };
15563 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops,
15564 array_lengthof(Ops),
15566 Ld->getPointerInfo(),
15567 Ld->getAlignment(),
15568 false/*isVolatile*/, true/*ReadMem*/,
15569 false/*WriteMem*/);
15571 // Make sure the newly-created LOAD is in the same position as Ld in
15572 // terms of dependency. We create a TokenFactor for Ld and ResNode,
15573 // and update uses of Ld's output chain to use the TokenFactor.
15574 if (Ld->hasAnyUseOfValue(1)) {
15575 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
15576 SDValue(Ld, 1), SDValue(ResNode.getNode(), 1));
15577 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), NewChain);
15578 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(Ld, 1),
15579 SDValue(ResNode.getNode(), 1));
15582 return DAG.getNode(ISD::BITCAST, dl, VT, ResNode);
15586 // Emit a zeroed vector and insert the desired subvector on its
15588 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
15589 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0), 0, DAG, dl);
15590 return DCI.CombineTo(N, InsV);
15593 //===--------------------------------------------------------------------===//
15594 // Combine some shuffles into subvector extracts and inserts:
15597 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
15598 if (isShuffleHigh128VectorInsertLow(SVOp)) {
15599 SDValue V = Extract128BitVector(V1, NumElems/2, DAG, dl);
15600 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, 0, DAG, dl);
15601 return DCI.CombineTo(N, InsV);
15604 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
15605 if (isShuffleLow128VectorInsertHigh(SVOp)) {
15606 SDValue V = Extract128BitVector(V1, 0, DAG, dl);
15607 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, NumElems/2, DAG, dl);
15608 return DCI.CombineTo(N, InsV);
15614 /// PerformShuffleCombine - Performs several different shuffle combines.
15615 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
15616 TargetLowering::DAGCombinerInfo &DCI,
15617 const X86Subtarget *Subtarget) {
15619 EVT VT = N->getValueType(0);
15621 // Don't create instructions with illegal types after legalize types has run.
15622 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15623 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
15626 // Combine 256-bit vector shuffles. This is only profitable when in AVX mode
15627 if (Subtarget->hasFp256() && VT.is256BitVector() &&
15628 N->getOpcode() == ISD::VECTOR_SHUFFLE)
15629 return PerformShuffleCombine256(N, DAG, DCI, Subtarget);
15631 // Only handle 128 wide vector from here on.
15632 if (!VT.is128BitVector())
15635 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
15636 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
15637 // consecutive, non-overlapping, and in the right order.
15638 SmallVector<SDValue, 16> Elts;
15639 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
15640 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
15642 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
15645 /// PerformTruncateCombine - Converts truncate operation to
15646 /// a sequence of vector shuffle operations.
15647 /// It is possible when we truncate 256-bit vector to 128-bit vector
15648 static SDValue PerformTruncateCombine(SDNode *N, SelectionDAG &DAG,
15649 TargetLowering::DAGCombinerInfo &DCI,
15650 const X86Subtarget *Subtarget) {
15654 /// XFormVExtractWithShuffleIntoLoad - Check if a vector extract from a target
15655 /// specific shuffle of a load can be folded into a single element load.
15656 /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
15657 /// shuffles have been customed lowered so we need to handle those here.
15658 static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
15659 TargetLowering::DAGCombinerInfo &DCI) {
15660 if (DCI.isBeforeLegalizeOps())
15663 SDValue InVec = N->getOperand(0);
15664 SDValue EltNo = N->getOperand(1);
15666 if (!isa<ConstantSDNode>(EltNo))
15669 EVT VT = InVec.getValueType();
15671 bool HasShuffleIntoBitcast = false;
15672 if (InVec.getOpcode() == ISD::BITCAST) {
15673 // Don't duplicate a load with other uses.
15674 if (!InVec.hasOneUse())
15676 EVT BCVT = InVec.getOperand(0).getValueType();
15677 if (BCVT.getVectorNumElements() != VT.getVectorNumElements())
15679 InVec = InVec.getOperand(0);
15680 HasShuffleIntoBitcast = true;
15683 if (!isTargetShuffle(InVec.getOpcode()))
15686 // Don't duplicate a load with other uses.
15687 if (!InVec.hasOneUse())
15690 SmallVector<int, 16> ShuffleMask;
15692 if (!getTargetShuffleMask(InVec.getNode(), VT.getSimpleVT(), ShuffleMask,
15696 // Select the input vector, guarding against out of range extract vector.
15697 unsigned NumElems = VT.getVectorNumElements();
15698 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
15699 int Idx = (Elt > (int)NumElems) ? -1 : ShuffleMask[Elt];
15700 SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0)
15701 : InVec.getOperand(1);
15703 // If inputs to shuffle are the same for both ops, then allow 2 uses
15704 unsigned AllowedUses = InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
15706 if (LdNode.getOpcode() == ISD::BITCAST) {
15707 // Don't duplicate a load with other uses.
15708 if (!LdNode.getNode()->hasNUsesOfValue(AllowedUses, 0))
15711 AllowedUses = 1; // only allow 1 load use if we have a bitcast
15712 LdNode = LdNode.getOperand(0);
15715 if (!ISD::isNormalLoad(LdNode.getNode()))
15718 LoadSDNode *LN0 = cast<LoadSDNode>(LdNode);
15720 if (!LN0 ||!LN0->hasNUsesOfValue(AllowedUses, 0) || LN0->isVolatile())
15723 if (HasShuffleIntoBitcast) {
15724 // If there's a bitcast before the shuffle, check if the load type and
15725 // alignment is valid.
15726 unsigned Align = LN0->getAlignment();
15727 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15728 unsigned NewAlign = TLI.getDataLayout()->
15729 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
15731 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, VT))
15735 // All checks match so transform back to vector_shuffle so that DAG combiner
15736 // can finish the job
15739 // Create shuffle node taking into account the case that its a unary shuffle
15740 SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(VT) : InVec.getOperand(1);
15741 Shuffle = DAG.getVectorShuffle(InVec.getValueType(), dl,
15742 InVec.getOperand(0), Shuffle,
15744 Shuffle = DAG.getNode(ISD::BITCAST, dl, VT, Shuffle);
15745 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle,
15749 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
15750 /// generation and convert it from being a bunch of shuffles and extracts
15751 /// to a simple store and scalar loads to extract the elements.
15752 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
15753 TargetLowering::DAGCombinerInfo &DCI) {
15754 SDValue NewOp = XFormVExtractWithShuffleIntoLoad(N, DAG, DCI);
15755 if (NewOp.getNode())
15758 SDValue InputVector = N->getOperand(0);
15759 // Detect whether we are trying to convert from mmx to i32 and the bitcast
15760 // from mmx to v2i32 has a single usage.
15761 if (InputVector.getNode()->getOpcode() == llvm::ISD::BITCAST &&
15762 InputVector.getNode()->getOperand(0).getValueType() == MVT::x86mmx &&
15763 InputVector.hasOneUse() && N->getValueType(0) == MVT::i32)
15764 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
15765 N->getValueType(0),
15766 InputVector.getNode()->getOperand(0));
15768 // Only operate on vectors of 4 elements, where the alternative shuffling
15769 // gets to be more expensive.
15770 if (InputVector.getValueType() != MVT::v4i32)
15773 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
15774 // single use which is a sign-extend or zero-extend, and all elements are
15776 SmallVector<SDNode *, 4> Uses;
15777 unsigned ExtractedElements = 0;
15778 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
15779 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
15780 if (UI.getUse().getResNo() != InputVector.getResNo())
15783 SDNode *Extract = *UI;
15784 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
15787 if (Extract->getValueType(0) != MVT::i32)
15789 if (!Extract->hasOneUse())
15791 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
15792 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
15794 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
15797 // Record which element was extracted.
15798 ExtractedElements |=
15799 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
15801 Uses.push_back(Extract);
15804 // If not all the elements were used, this may not be worthwhile.
15805 if (ExtractedElements != 15)
15808 // Ok, we've now decided to do the transformation.
15809 SDLoc dl(InputVector);
15811 // Store the value to a temporary stack slot.
15812 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
15813 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
15814 MachinePointerInfo(), false, false, 0);
15816 // Replace each use (extract) with a load of the appropriate element.
15817 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
15818 UE = Uses.end(); UI != UE; ++UI) {
15819 SDNode *Extract = *UI;
15821 // cOMpute the element's address.
15822 SDValue Idx = Extract->getOperand(1);
15824 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
15825 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
15826 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15827 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
15829 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(),
15830 StackPtr, OffsetVal);
15832 // Load the scalar.
15833 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch,
15834 ScalarAddr, MachinePointerInfo(),
15835 false, false, false, 0);
15837 // Replace the exact with the load.
15838 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
15841 // The replacement was made in place; don't return anything.
15845 /// \brief Matches a VSELECT onto min/max or return 0 if the node doesn't match.
15846 static unsigned matchIntegerMINMAX(SDValue Cond, EVT VT, SDValue LHS,
15847 SDValue RHS, SelectionDAG &DAG,
15848 const X86Subtarget *Subtarget) {
15849 if (!VT.isVector())
15852 switch (VT.getSimpleVT().SimpleTy) {
15857 if (!Subtarget->hasAVX2())
15862 if (!Subtarget->hasSSE2())
15866 // SSE2 has only a small subset of the operations.
15867 bool hasUnsigned = Subtarget->hasSSE41() ||
15868 (Subtarget->hasSSE2() && VT == MVT::v16i8);
15869 bool hasSigned = Subtarget->hasSSE41() ||
15870 (Subtarget->hasSSE2() && VT == MVT::v8i16);
15872 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
15874 // Check for x CC y ? x : y.
15875 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
15876 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
15881 return hasUnsigned ? X86ISD::UMIN : 0;
15884 return hasUnsigned ? X86ISD::UMAX : 0;
15887 return hasSigned ? X86ISD::SMIN : 0;
15890 return hasSigned ? X86ISD::SMAX : 0;
15892 // Check for x CC y ? y : x -- a min/max with reversed arms.
15893 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
15894 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
15899 return hasUnsigned ? X86ISD::UMAX : 0;
15902 return hasUnsigned ? X86ISD::UMIN : 0;
15905 return hasSigned ? X86ISD::SMAX : 0;
15908 return hasSigned ? X86ISD::SMIN : 0;
15915 /// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT
15917 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
15918 TargetLowering::DAGCombinerInfo &DCI,
15919 const X86Subtarget *Subtarget) {
15921 SDValue Cond = N->getOperand(0);
15922 // Get the LHS/RHS of the select.
15923 SDValue LHS = N->getOperand(1);
15924 SDValue RHS = N->getOperand(2);
15925 EVT VT = LHS.getValueType();
15927 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
15928 // instructions match the semantics of the common C idiom x<y?x:y but not
15929 // x<=y?x:y, because of how they handle negative zero (which can be
15930 // ignored in unsafe-math mode).
15931 if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() &&
15932 VT != MVT::f80 && DAG.getTargetLoweringInfo().isTypeLegal(VT) &&
15933 (Subtarget->hasSSE2() ||
15934 (Subtarget->hasSSE1() && VT.getScalarType() == MVT::f32))) {
15935 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
15937 unsigned Opcode = 0;
15938 // Check for x CC y ? x : y.
15939 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
15940 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
15944 // Converting this to a min would handle NaNs incorrectly, and swapping
15945 // the operands would cause it to handle comparisons between positive
15946 // and negative zero incorrectly.
15947 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
15948 if (!DAG.getTarget().Options.UnsafeFPMath &&
15949 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
15951 std::swap(LHS, RHS);
15953 Opcode = X86ISD::FMIN;
15956 // Converting this to a min would handle comparisons between positive
15957 // and negative zero incorrectly.
15958 if (!DAG.getTarget().Options.UnsafeFPMath &&
15959 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
15961 Opcode = X86ISD::FMIN;
15964 // Converting this to a min would handle both negative zeros and NaNs
15965 // incorrectly, but we can swap the operands to fix both.
15966 std::swap(LHS, RHS);
15970 Opcode = X86ISD::FMIN;
15974 // Converting this to a max would handle comparisons between positive
15975 // and negative zero incorrectly.
15976 if (!DAG.getTarget().Options.UnsafeFPMath &&
15977 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
15979 Opcode = X86ISD::FMAX;
15982 // Converting this to a max would handle NaNs incorrectly, and swapping
15983 // the operands would cause it to handle comparisons between positive
15984 // and negative zero incorrectly.
15985 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
15986 if (!DAG.getTarget().Options.UnsafeFPMath &&
15987 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
15989 std::swap(LHS, RHS);
15991 Opcode = X86ISD::FMAX;
15994 // Converting this to a max would handle both negative zeros and NaNs
15995 // incorrectly, but we can swap the operands to fix both.
15996 std::swap(LHS, RHS);
16000 Opcode = X86ISD::FMAX;
16003 // Check for x CC y ? y : x -- a min/max with reversed arms.
16004 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
16005 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
16009 // Converting this to a min would handle comparisons between positive
16010 // and negative zero incorrectly, and swapping the operands would
16011 // cause it to handle NaNs incorrectly.
16012 if (!DAG.getTarget().Options.UnsafeFPMath &&
16013 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
16014 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
16016 std::swap(LHS, RHS);
16018 Opcode = X86ISD::FMIN;
16021 // Converting this to a min would handle NaNs incorrectly.
16022 if (!DAG.getTarget().Options.UnsafeFPMath &&
16023 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
16025 Opcode = X86ISD::FMIN;
16028 // Converting this to a min would handle both negative zeros and NaNs
16029 // incorrectly, but we can swap the operands to fix both.
16030 std::swap(LHS, RHS);
16034 Opcode = X86ISD::FMIN;
16038 // Converting this to a max would handle NaNs incorrectly.
16039 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
16041 Opcode = X86ISD::FMAX;
16044 // Converting this to a max would handle comparisons between positive
16045 // and negative zero incorrectly, and swapping the operands would
16046 // cause it to handle NaNs incorrectly.
16047 if (!DAG.getTarget().Options.UnsafeFPMath &&
16048 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
16049 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
16051 std::swap(LHS, RHS);
16053 Opcode = X86ISD::FMAX;
16056 // Converting this to a max would handle both negative zeros and NaNs
16057 // incorrectly, but we can swap the operands to fix both.
16058 std::swap(LHS, RHS);
16062 Opcode = X86ISD::FMAX;
16068 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
16071 // If this is a select between two integer constants, try to do some
16073 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
16074 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
16075 // Don't do this for crazy integer types.
16076 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
16077 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
16078 // so that TrueC (the true value) is larger than FalseC.
16079 bool NeedsCondInvert = false;
16081 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
16082 // Efficiently invertible.
16083 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
16084 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
16085 isa<ConstantSDNode>(Cond.getOperand(1))))) {
16086 NeedsCondInvert = true;
16087 std::swap(TrueC, FalseC);
16090 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
16091 if (FalseC->getAPIntValue() == 0 &&
16092 TrueC->getAPIntValue().isPowerOf2()) {
16093 if (NeedsCondInvert) // Invert the condition if needed.
16094 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
16095 DAG.getConstant(1, Cond.getValueType()));
16097 // Zero extend the condition if needed.
16098 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
16100 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
16101 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
16102 DAG.getConstant(ShAmt, MVT::i8));
16105 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
16106 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
16107 if (NeedsCondInvert) // Invert the condition if needed.
16108 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
16109 DAG.getConstant(1, Cond.getValueType()));
16111 // Zero extend the condition if needed.
16112 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
16113 FalseC->getValueType(0), Cond);
16114 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
16115 SDValue(FalseC, 0));
16118 // Optimize cases that will turn into an LEA instruction. This requires
16119 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
16120 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
16121 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
16122 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
16124 bool isFastMultiplier = false;
16126 switch ((unsigned char)Diff) {
16128 case 1: // result = add base, cond
16129 case 2: // result = lea base( , cond*2)
16130 case 3: // result = lea base(cond, cond*2)
16131 case 4: // result = lea base( , cond*4)
16132 case 5: // result = lea base(cond, cond*4)
16133 case 8: // result = lea base( , cond*8)
16134 case 9: // result = lea base(cond, cond*8)
16135 isFastMultiplier = true;
16140 if (isFastMultiplier) {
16141 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
16142 if (NeedsCondInvert) // Invert the condition if needed.
16143 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
16144 DAG.getConstant(1, Cond.getValueType()));
16146 // Zero extend the condition if needed.
16147 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
16149 // Scale the condition by the difference.
16151 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
16152 DAG.getConstant(Diff, Cond.getValueType()));
16154 // Add the base if non-zero.
16155 if (FalseC->getAPIntValue() != 0)
16156 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
16157 SDValue(FalseC, 0));
16164 // Canonicalize max and min:
16165 // (x > y) ? x : y -> (x >= y) ? x : y
16166 // (x < y) ? x : y -> (x <= y) ? x : y
16167 // This allows use of COND_S / COND_NS (see TranslateX86CC) which eliminates
16168 // the need for an extra compare
16169 // against zero. e.g.
16170 // (x - y) > 0 : (x - y) ? 0 -> (x - y) >= 0 : (x - y) ? 0
16172 // testl %edi, %edi
16174 // cmovgl %edi, %eax
16178 // cmovsl %eax, %edi
16179 if (N->getOpcode() == ISD::SELECT && Cond.getOpcode() == ISD::SETCC &&
16180 DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
16181 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
16182 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
16187 ISD::CondCode NewCC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGE;
16188 Cond = DAG.getSetCC(SDLoc(Cond), Cond.getValueType(),
16189 Cond.getOperand(0), Cond.getOperand(1), NewCC);
16190 return DAG.getNode(ISD::SELECT, DL, VT, Cond, LHS, RHS);
16195 // Match VSELECTs into subs with unsigned saturation.
16196 if (!DCI.isBeforeLegalize() &&
16197 N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
16198 // psubus is available in SSE2 and AVX2 for i8 and i16 vectors.
16199 ((Subtarget->hasSSE2() && (VT == MVT::v16i8 || VT == MVT::v8i16)) ||
16200 (Subtarget->hasAVX2() && (VT == MVT::v32i8 || VT == MVT::v16i16)))) {
16201 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
16203 // Check if one of the arms of the VSELECT is a zero vector. If it's on the
16204 // left side invert the predicate to simplify logic below.
16206 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
16208 CC = ISD::getSetCCInverse(CC, true);
16209 } else if (ISD::isBuildVectorAllZeros(RHS.getNode())) {
16213 if (Other.getNode() && Other->getNumOperands() == 2 &&
16214 DAG.isEqualTo(Other->getOperand(0), Cond.getOperand(0))) {
16215 SDValue OpLHS = Other->getOperand(0), OpRHS = Other->getOperand(1);
16216 SDValue CondRHS = Cond->getOperand(1);
16218 // Look for a general sub with unsigned saturation first.
16219 // x >= y ? x-y : 0 --> subus x, y
16220 // x > y ? x-y : 0 --> subus x, y
16221 if ((CC == ISD::SETUGE || CC == ISD::SETUGT) &&
16222 Other->getOpcode() == ISD::SUB && DAG.isEqualTo(OpRHS, CondRHS))
16223 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
16225 // If the RHS is a constant we have to reverse the const canonicalization.
16226 // x > C-1 ? x+-C : 0 --> subus x, C
16227 if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD &&
16228 isSplatVector(CondRHS.getNode()) && isSplatVector(OpRHS.getNode())) {
16229 APInt A = cast<ConstantSDNode>(OpRHS.getOperand(0))->getAPIntValue();
16230 if (CondRHS.getConstantOperandVal(0) == -A-1)
16231 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS,
16232 DAG.getConstant(-A, VT));
16235 // Another special case: If C was a sign bit, the sub has been
16236 // canonicalized into a xor.
16237 // FIXME: Would it be better to use ComputeMaskedBits to determine whether
16238 // it's safe to decanonicalize the xor?
16239 // x s< 0 ? x^C : 0 --> subus x, C
16240 if (CC == ISD::SETLT && Other->getOpcode() == ISD::XOR &&
16241 ISD::isBuildVectorAllZeros(CondRHS.getNode()) &&
16242 isSplatVector(OpRHS.getNode())) {
16243 APInt A = cast<ConstantSDNode>(OpRHS.getOperand(0))->getAPIntValue();
16245 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
16250 // Try to match a min/max vector operation.
16251 if (!DCI.isBeforeLegalize() &&
16252 N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC)
16253 if (unsigned Op = matchIntegerMINMAX(Cond, VT, LHS, RHS, DAG, Subtarget))
16254 return DAG.getNode(Op, DL, N->getValueType(0), LHS, RHS);
16256 // Simplify vector selection if the selector will be produced by CMPP*/PCMP*.
16257 if (!DCI.isBeforeLegalize() && N->getOpcode() == ISD::VSELECT &&
16258 Cond.getOpcode() == ISD::SETCC) {
16260 assert(Cond.getValueType().isVector() &&
16261 "vector select expects a vector selector!");
16263 EVT IntVT = Cond.getValueType();
16264 bool TValIsAllOnes = ISD::isBuildVectorAllOnes(LHS.getNode());
16265 bool FValIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
16267 if (!TValIsAllOnes && !FValIsAllZeros) {
16268 // Try invert the condition if true value is not all 1s and false value
16270 bool TValIsAllZeros = ISD::isBuildVectorAllZeros(LHS.getNode());
16271 bool FValIsAllOnes = ISD::isBuildVectorAllOnes(RHS.getNode());
16273 if (TValIsAllZeros || FValIsAllOnes) {
16274 SDValue CC = Cond.getOperand(2);
16275 ISD::CondCode NewCC =
16276 ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
16277 Cond.getOperand(0).getValueType().isInteger());
16278 Cond = DAG.getSetCC(DL, IntVT, Cond.getOperand(0), Cond.getOperand(1), NewCC);
16279 std::swap(LHS, RHS);
16280 TValIsAllOnes = FValIsAllOnes;
16281 FValIsAllZeros = TValIsAllZeros;
16285 if (TValIsAllOnes || FValIsAllZeros) {
16288 if (TValIsAllOnes && FValIsAllZeros)
16290 else if (TValIsAllOnes)
16291 Ret = DAG.getNode(ISD::OR, DL, IntVT, Cond,
16292 DAG.getNode(ISD::BITCAST, DL, IntVT, RHS));
16293 else if (FValIsAllZeros)
16294 Ret = DAG.getNode(ISD::AND, DL, IntVT, Cond,
16295 DAG.getNode(ISD::BITCAST, DL, IntVT, LHS));
16297 return DAG.getNode(ISD::BITCAST, DL, VT, Ret);
16301 // If we know that this node is legal then we know that it is going to be
16302 // matched by one of the SSE/AVX BLEND instructions. These instructions only
16303 // depend on the highest bit in each word. Try to use SimplifyDemandedBits
16304 // to simplify previous instructions.
16305 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16306 if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() &&
16307 !DCI.isBeforeLegalize() && TLI.isOperationLegal(ISD::VSELECT, VT)) {
16308 unsigned BitWidth = Cond.getValueType().getScalarType().getSizeInBits();
16310 // Don't optimize vector selects that map to mask-registers.
16314 assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size");
16315 APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);
16317 APInt KnownZero, KnownOne;
16318 TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(),
16319 DCI.isBeforeLegalizeOps());
16320 if (TLO.ShrinkDemandedConstant(Cond, DemandedMask) ||
16321 TLI.SimplifyDemandedBits(Cond, DemandedMask, KnownZero, KnownOne, TLO))
16322 DCI.CommitTargetLoweringOpt(TLO);
16328 // Check whether a boolean test is testing a boolean value generated by
16329 // X86ISD::SETCC. If so, return the operand of that SETCC and proper condition
16332 // Simplify the following patterns:
16333 // (Op (CMP (SETCC Cond EFLAGS) 1) EQ) or
16334 // (Op (CMP (SETCC Cond EFLAGS) 0) NEQ)
16335 // to (Op EFLAGS Cond)
16337 // (Op (CMP (SETCC Cond EFLAGS) 0) EQ) or
16338 // (Op (CMP (SETCC Cond EFLAGS) 1) NEQ)
16339 // to (Op EFLAGS !Cond)
16341 // where Op could be BRCOND or CMOV.
16343 static SDValue checkBoolTestSetCCCombine(SDValue Cmp, X86::CondCode &CC) {
16344 // Quit if not CMP and SUB with its value result used.
16345 if (Cmp.getOpcode() != X86ISD::CMP &&
16346 (Cmp.getOpcode() != X86ISD::SUB || Cmp.getNode()->hasAnyUseOfValue(0)))
16349 // Quit if not used as a boolean value.
16350 if (CC != X86::COND_E && CC != X86::COND_NE)
16353 // Check CMP operands. One of them should be 0 or 1 and the other should be
16354 // an SetCC or extended from it.
16355 SDValue Op1 = Cmp.getOperand(0);
16356 SDValue Op2 = Cmp.getOperand(1);
16359 const ConstantSDNode* C = 0;
16360 bool needOppositeCond = (CC == X86::COND_E);
16361 bool checkAgainstTrue = false; // Is it a comparison against 1?
16363 if ((C = dyn_cast<ConstantSDNode>(Op1)))
16365 else if ((C = dyn_cast<ConstantSDNode>(Op2)))
16367 else // Quit if all operands are not constants.
16370 if (C->getZExtValue() == 1) {
16371 needOppositeCond = !needOppositeCond;
16372 checkAgainstTrue = true;
16373 } else if (C->getZExtValue() != 0)
16374 // Quit if the constant is neither 0 or 1.
16377 bool truncatedToBoolWithAnd = false;
16378 // Skip (zext $x), (trunc $x), or (and $x, 1) node.
16379 while (SetCC.getOpcode() == ISD::ZERO_EXTEND ||
16380 SetCC.getOpcode() == ISD::TRUNCATE ||
16381 SetCC.getOpcode() == ISD::AND) {
16382 if (SetCC.getOpcode() == ISD::AND) {
16384 ConstantSDNode *CS;
16385 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(0))) &&
16386 CS->getZExtValue() == 1)
16388 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(1))) &&
16389 CS->getZExtValue() == 1)
16393 SetCC = SetCC.getOperand(OpIdx);
16394 truncatedToBoolWithAnd = true;
16396 SetCC = SetCC.getOperand(0);
16399 switch (SetCC.getOpcode()) {
16400 case X86ISD::SETCC_CARRY:
16401 // Since SETCC_CARRY gives output based on R = CF ? ~0 : 0, it's unsafe to
16402 // simplify it if the result of SETCC_CARRY is not canonicalized to 0 or 1,
16403 // i.e. it's a comparison against true but the result of SETCC_CARRY is not
16404 // truncated to i1 using 'and'.
16405 if (checkAgainstTrue && !truncatedToBoolWithAnd)
16407 assert(X86::CondCode(SetCC.getConstantOperandVal(0)) == X86::COND_B &&
16408 "Invalid use of SETCC_CARRY!");
16410 case X86ISD::SETCC:
16411 // Set the condition code or opposite one if necessary.
16412 CC = X86::CondCode(SetCC.getConstantOperandVal(0));
16413 if (needOppositeCond)
16414 CC = X86::GetOppositeBranchCondition(CC);
16415 return SetCC.getOperand(1);
16416 case X86ISD::CMOV: {
16417 // Check whether false/true value has canonical one, i.e. 0 or 1.
16418 ConstantSDNode *FVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(0));
16419 ConstantSDNode *TVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(1));
16420 // Quit if true value is not a constant.
16423 // Quit if false value is not a constant.
16425 SDValue Op = SetCC.getOperand(0);
16426 // Skip 'zext' or 'trunc' node.
16427 if (Op.getOpcode() == ISD::ZERO_EXTEND ||
16428 Op.getOpcode() == ISD::TRUNCATE)
16429 Op = Op.getOperand(0);
16430 // A special case for rdrand/rdseed, where 0 is set if false cond is
16432 if ((Op.getOpcode() != X86ISD::RDRAND &&
16433 Op.getOpcode() != X86ISD::RDSEED) || Op.getResNo() != 0)
16436 // Quit if false value is not the constant 0 or 1.
16437 bool FValIsFalse = true;
16438 if (FVal && FVal->getZExtValue() != 0) {
16439 if (FVal->getZExtValue() != 1)
16441 // If FVal is 1, opposite cond is needed.
16442 needOppositeCond = !needOppositeCond;
16443 FValIsFalse = false;
16445 // Quit if TVal is not the constant opposite of FVal.
16446 if (FValIsFalse && TVal->getZExtValue() != 1)
16448 if (!FValIsFalse && TVal->getZExtValue() != 0)
16450 CC = X86::CondCode(SetCC.getConstantOperandVal(2));
16451 if (needOppositeCond)
16452 CC = X86::GetOppositeBranchCondition(CC);
16453 return SetCC.getOperand(3);
16460 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
16461 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
16462 TargetLowering::DAGCombinerInfo &DCI,
16463 const X86Subtarget *Subtarget) {
16466 // If the flag operand isn't dead, don't touch this CMOV.
16467 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
16470 SDValue FalseOp = N->getOperand(0);
16471 SDValue TrueOp = N->getOperand(1);
16472 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
16473 SDValue Cond = N->getOperand(3);
16475 if (CC == X86::COND_E || CC == X86::COND_NE) {
16476 switch (Cond.getOpcode()) {
16480 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
16481 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
16482 return (CC == X86::COND_E) ? FalseOp : TrueOp;
16488 Flags = checkBoolTestSetCCCombine(Cond, CC);
16489 if (Flags.getNode() &&
16490 // Extra check as FCMOV only supports a subset of X86 cond.
16491 (FalseOp.getValueType() != MVT::f80 || hasFPCMov(CC))) {
16492 SDValue Ops[] = { FalseOp, TrueOp,
16493 DAG.getConstant(CC, MVT::i8), Flags };
16494 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList(),
16495 Ops, array_lengthof(Ops));
16498 // If this is a select between two integer constants, try to do some
16499 // optimizations. Note that the operands are ordered the opposite of SELECT
16501 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
16502 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
16503 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
16504 // larger than FalseC (the false value).
16505 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
16506 CC = X86::GetOppositeBranchCondition(CC);
16507 std::swap(TrueC, FalseC);
16508 std::swap(TrueOp, FalseOp);
16511 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
16512 // This is efficient for any integer data type (including i8/i16) and
16514 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
16515 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
16516 DAG.getConstant(CC, MVT::i8), Cond);
16518 // Zero extend the condition if needed.
16519 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
16521 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
16522 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
16523 DAG.getConstant(ShAmt, MVT::i8));
16524 if (N->getNumValues() == 2) // Dead flag value?
16525 return DCI.CombineTo(N, Cond, SDValue());
16529 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
16530 // for any integer data type, including i8/i16.
16531 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
16532 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
16533 DAG.getConstant(CC, MVT::i8), Cond);
16535 // Zero extend the condition if needed.
16536 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
16537 FalseC->getValueType(0), Cond);
16538 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
16539 SDValue(FalseC, 0));
16541 if (N->getNumValues() == 2) // Dead flag value?
16542 return DCI.CombineTo(N, Cond, SDValue());
16546 // Optimize cases that will turn into an LEA instruction. This requires
16547 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
16548 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
16549 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
16550 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
16552 bool isFastMultiplier = false;
16554 switch ((unsigned char)Diff) {
16556 case 1: // result = add base, cond
16557 case 2: // result = lea base( , cond*2)
16558 case 3: // result = lea base(cond, cond*2)
16559 case 4: // result = lea base( , cond*4)
16560 case 5: // result = lea base(cond, cond*4)
16561 case 8: // result = lea base( , cond*8)
16562 case 9: // result = lea base(cond, cond*8)
16563 isFastMultiplier = true;
16568 if (isFastMultiplier) {
16569 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
16570 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
16571 DAG.getConstant(CC, MVT::i8), Cond);
16572 // Zero extend the condition if needed.
16573 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
16575 // Scale the condition by the difference.
16577 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
16578 DAG.getConstant(Diff, Cond.getValueType()));
16580 // Add the base if non-zero.
16581 if (FalseC->getAPIntValue() != 0)
16582 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
16583 SDValue(FalseC, 0));
16584 if (N->getNumValues() == 2) // Dead flag value?
16585 return DCI.CombineTo(N, Cond, SDValue());
16592 // Handle these cases:
16593 // (select (x != c), e, c) -> select (x != c), e, x),
16594 // (select (x == c), c, e) -> select (x == c), x, e)
16595 // where the c is an integer constant, and the "select" is the combination
16596 // of CMOV and CMP.
16598 // The rationale for this change is that the conditional-move from a constant
16599 // needs two instructions, however, conditional-move from a register needs
16600 // only one instruction.
16602 // CAVEAT: By replacing a constant with a symbolic value, it may obscure
16603 // some instruction-combining opportunities. This opt needs to be
16604 // postponed as late as possible.
16606 if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) {
16607 // the DCI.xxxx conditions are provided to postpone the optimization as
16608 // late as possible.
16610 ConstantSDNode *CmpAgainst = 0;
16611 if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) &&
16612 (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) &&
16613 !isa<ConstantSDNode>(Cond.getOperand(0))) {
16615 if (CC == X86::COND_NE &&
16616 CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) {
16617 CC = X86::GetOppositeBranchCondition(CC);
16618 std::swap(TrueOp, FalseOp);
16621 if (CC == X86::COND_E &&
16622 CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) {
16623 SDValue Ops[] = { FalseOp, Cond.getOperand(0),
16624 DAG.getConstant(CC, MVT::i8), Cond };
16625 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops,
16626 array_lengthof(Ops));
16634 /// PerformMulCombine - Optimize a single multiply with constant into two
16635 /// in order to implement it with two cheaper instructions, e.g.
16636 /// LEA + SHL, LEA + LEA.
16637 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
16638 TargetLowering::DAGCombinerInfo &DCI) {
16639 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
16642 EVT VT = N->getValueType(0);
16643 if (VT != MVT::i64)
16646 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
16649 uint64_t MulAmt = C->getZExtValue();
16650 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
16653 uint64_t MulAmt1 = 0;
16654 uint64_t MulAmt2 = 0;
16655 if ((MulAmt % 9) == 0) {
16657 MulAmt2 = MulAmt / 9;
16658 } else if ((MulAmt % 5) == 0) {
16660 MulAmt2 = MulAmt / 5;
16661 } else if ((MulAmt % 3) == 0) {
16663 MulAmt2 = MulAmt / 3;
16666 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
16669 if (isPowerOf2_64(MulAmt2) &&
16670 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
16671 // If second multiplifer is pow2, issue it first. We want the multiply by
16672 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
16674 std::swap(MulAmt1, MulAmt2);
16677 if (isPowerOf2_64(MulAmt1))
16678 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
16679 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
16681 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
16682 DAG.getConstant(MulAmt1, VT));
16684 if (isPowerOf2_64(MulAmt2))
16685 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
16686 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
16688 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
16689 DAG.getConstant(MulAmt2, VT));
16691 // Do not add new nodes to DAG combiner worklist.
16692 DCI.CombineTo(N, NewMul, false);
16697 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
16698 SDValue N0 = N->getOperand(0);
16699 SDValue N1 = N->getOperand(1);
16700 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
16701 EVT VT = N0.getValueType();
16703 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
16704 // since the result of setcc_c is all zero's or all ones.
16705 if (VT.isInteger() && !VT.isVector() &&
16706 N1C && N0.getOpcode() == ISD::AND &&
16707 N0.getOperand(1).getOpcode() == ISD::Constant) {
16708 SDValue N00 = N0.getOperand(0);
16709 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
16710 ((N00.getOpcode() == ISD::ANY_EXTEND ||
16711 N00.getOpcode() == ISD::ZERO_EXTEND) &&
16712 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
16713 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
16714 APInt ShAmt = N1C->getAPIntValue();
16715 Mask = Mask.shl(ShAmt);
16717 return DAG.getNode(ISD::AND, SDLoc(N), VT,
16718 N00, DAG.getConstant(Mask, VT));
16722 // Hardware support for vector shifts is sparse which makes us scalarize the
16723 // vector operations in many cases. Also, on sandybridge ADD is faster than
16725 // (shl V, 1) -> add V,V
16726 if (isSplatVector(N1.getNode())) {
16727 assert(N0.getValueType().isVector() && "Invalid vector shift type");
16728 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1->getOperand(0));
16729 // We shift all of the values by one. In many cases we do not have
16730 // hardware support for this operation. This is better expressed as an ADD
16732 if (N1C && (1 == N1C->getZExtValue())) {
16733 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N0);
16740 /// \brief Returns a vector of 0s if the node in input is a vector logical
16741 /// shift by a constant amount which is known to be bigger than or equal
16742 /// to the vector element size in bits.
16743 static SDValue performShiftToAllZeros(SDNode *N, SelectionDAG &DAG,
16744 const X86Subtarget *Subtarget) {
16745 EVT VT = N->getValueType(0);
16747 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16 &&
16748 (!Subtarget->hasInt256() ||
16749 (VT != MVT::v4i64 && VT != MVT::v8i32 && VT != MVT::v16i16)))
16752 SDValue Amt = N->getOperand(1);
16754 if (isSplatVector(Amt.getNode())) {
16755 SDValue SclrAmt = Amt->getOperand(0);
16756 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(SclrAmt)) {
16757 APInt ShiftAmt = C->getAPIntValue();
16758 unsigned MaxAmount = VT.getVectorElementType().getSizeInBits();
16760 // SSE2/AVX2 logical shifts always return a vector of 0s
16761 // if the shift amount is bigger than or equal to
16762 // the element size. The constant shift amount will be
16763 // encoded as a 8-bit immediate.
16764 if (ShiftAmt.trunc(8).uge(MaxAmount))
16765 return getZeroVector(VT, Subtarget, DAG, DL);
16772 /// PerformShiftCombine - Combine shifts.
16773 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
16774 TargetLowering::DAGCombinerInfo &DCI,
16775 const X86Subtarget *Subtarget) {
16776 if (N->getOpcode() == ISD::SHL) {
16777 SDValue V = PerformSHLCombine(N, DAG);
16778 if (V.getNode()) return V;
16781 if (N->getOpcode() != ISD::SRA) {
16782 // Try to fold this logical shift into a zero vector.
16783 SDValue V = performShiftToAllZeros(N, DAG, Subtarget);
16784 if (V.getNode()) return V;
16790 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
16791 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
16792 // and friends. Likewise for OR -> CMPNEQSS.
16793 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
16794 TargetLowering::DAGCombinerInfo &DCI,
16795 const X86Subtarget *Subtarget) {
16798 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
16799 // we're requiring SSE2 for both.
16800 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
16801 SDValue N0 = N->getOperand(0);
16802 SDValue N1 = N->getOperand(1);
16803 SDValue CMP0 = N0->getOperand(1);
16804 SDValue CMP1 = N1->getOperand(1);
16807 // The SETCCs should both refer to the same CMP.
16808 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
16811 SDValue CMP00 = CMP0->getOperand(0);
16812 SDValue CMP01 = CMP0->getOperand(1);
16813 EVT VT = CMP00.getValueType();
16815 if (VT == MVT::f32 || VT == MVT::f64) {
16816 bool ExpectingFlags = false;
16817 // Check for any users that want flags:
16818 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
16819 !ExpectingFlags && UI != UE; ++UI)
16820 switch (UI->getOpcode()) {
16825 ExpectingFlags = true;
16827 case ISD::CopyToReg:
16828 case ISD::SIGN_EXTEND:
16829 case ISD::ZERO_EXTEND:
16830 case ISD::ANY_EXTEND:
16834 if (!ExpectingFlags) {
16835 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
16836 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
16838 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
16839 X86::CondCode tmp = cc0;
16844 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
16845 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
16846 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
16847 X86ISD::NodeType NTOperator = is64BitFP ?
16848 X86ISD::FSETCCsd : X86ISD::FSETCCss;
16849 // FIXME: need symbolic constants for these magic numbers.
16850 // See X86ATTInstPrinter.cpp:printSSECC().
16851 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
16852 SDValue OnesOrZeroesF = DAG.getNode(NTOperator, DL, MVT::f32, CMP00, CMP01,
16853 DAG.getConstant(x86cc, MVT::i8));
16854 SDValue OnesOrZeroesI = DAG.getNode(ISD::BITCAST, DL, MVT::i32,
16856 SDValue ANDed = DAG.getNode(ISD::AND, DL, MVT::i32, OnesOrZeroesI,
16857 DAG.getConstant(1, MVT::i32));
16858 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8, ANDed);
16859 return OneBitOfTruth;
16867 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
16868 /// so it can be folded inside ANDNP.
16869 static bool CanFoldXORWithAllOnes(const SDNode *N) {
16870 EVT VT = N->getValueType(0);
16872 // Match direct AllOnes for 128 and 256-bit vectors
16873 if (ISD::isBuildVectorAllOnes(N))
16876 // Look through a bit convert.
16877 if (N->getOpcode() == ISD::BITCAST)
16878 N = N->getOperand(0).getNode();
16880 // Sometimes the operand may come from a insert_subvector building a 256-bit
16882 if (VT.is256BitVector() &&
16883 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
16884 SDValue V1 = N->getOperand(0);
16885 SDValue V2 = N->getOperand(1);
16887 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
16888 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
16889 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
16890 ISD::isBuildVectorAllOnes(V2.getNode()))
16897 // On AVX/AVX2 the type v8i1 is legalized to v8i16, which is an XMM sized
16898 // register. In most cases we actually compare or select YMM-sized registers
16899 // and mixing the two types creates horrible code. This method optimizes
16900 // some of the transition sequences.
16901 static SDValue WidenMaskArithmetic(SDNode *N, SelectionDAG &DAG,
16902 TargetLowering::DAGCombinerInfo &DCI,
16903 const X86Subtarget *Subtarget) {
16904 EVT VT = N->getValueType(0);
16905 if (!VT.is256BitVector())
16908 assert((N->getOpcode() == ISD::ANY_EXTEND ||
16909 N->getOpcode() == ISD::ZERO_EXTEND ||
16910 N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node");
16912 SDValue Narrow = N->getOperand(0);
16913 EVT NarrowVT = Narrow->getValueType(0);
16914 if (!NarrowVT.is128BitVector())
16917 if (Narrow->getOpcode() != ISD::XOR &&
16918 Narrow->getOpcode() != ISD::AND &&
16919 Narrow->getOpcode() != ISD::OR)
16922 SDValue N0 = Narrow->getOperand(0);
16923 SDValue N1 = Narrow->getOperand(1);
16926 // The Left side has to be a trunc.
16927 if (N0.getOpcode() != ISD::TRUNCATE)
16930 // The type of the truncated inputs.
16931 EVT WideVT = N0->getOperand(0)->getValueType(0);
16935 // The right side has to be a 'trunc' or a constant vector.
16936 bool RHSTrunc = N1.getOpcode() == ISD::TRUNCATE;
16937 bool RHSConst = (isSplatVector(N1.getNode()) &&
16938 isa<ConstantSDNode>(N1->getOperand(0)));
16939 if (!RHSTrunc && !RHSConst)
16942 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16944 if (!TLI.isOperationLegalOrPromote(Narrow->getOpcode(), WideVT))
16947 // Set N0 and N1 to hold the inputs to the new wide operation.
16948 N0 = N0->getOperand(0);
16950 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT.getScalarType(),
16951 N1->getOperand(0));
16952 SmallVector<SDValue, 8> C(WideVT.getVectorNumElements(), N1);
16953 N1 = DAG.getNode(ISD::BUILD_VECTOR, DL, WideVT, &C[0], C.size());
16954 } else if (RHSTrunc) {
16955 N1 = N1->getOperand(0);
16958 // Generate the wide operation.
16959 SDValue Op = DAG.getNode(Narrow->getOpcode(), DL, WideVT, N0, N1);
16960 unsigned Opcode = N->getOpcode();
16962 case ISD::ANY_EXTEND:
16964 case ISD::ZERO_EXTEND: {
16965 unsigned InBits = NarrowVT.getScalarType().getSizeInBits();
16966 APInt Mask = APInt::getAllOnesValue(InBits);
16967 Mask = Mask.zext(VT.getScalarType().getSizeInBits());
16968 return DAG.getNode(ISD::AND, DL, VT,
16969 Op, DAG.getConstant(Mask, VT));
16971 case ISD::SIGN_EXTEND:
16972 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT,
16973 Op, DAG.getValueType(NarrowVT));
16975 llvm_unreachable("Unexpected opcode");
16979 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
16980 TargetLowering::DAGCombinerInfo &DCI,
16981 const X86Subtarget *Subtarget) {
16982 EVT VT = N->getValueType(0);
16983 if (DCI.isBeforeLegalizeOps())
16986 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
16990 // Create BLSI, and BLSR instructions
16991 // BLSI is X & (-X)
16992 // BLSR is X & (X-1)
16993 if (Subtarget->hasBMI() && (VT == MVT::i32 || VT == MVT::i64)) {
16994 SDValue N0 = N->getOperand(0);
16995 SDValue N1 = N->getOperand(1);
16998 // Check LHS for neg
16999 if (N0.getOpcode() == ISD::SUB && N0.getOperand(1) == N1 &&
17000 isZero(N0.getOperand(0)))
17001 return DAG.getNode(X86ISD::BLSI, DL, VT, N1);
17003 // Check RHS for neg
17004 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1) == N0 &&
17005 isZero(N1.getOperand(0)))
17006 return DAG.getNode(X86ISD::BLSI, DL, VT, N0);
17008 // Check LHS for X-1
17009 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1 &&
17010 isAllOnes(N0.getOperand(1)))
17011 return DAG.getNode(X86ISD::BLSR, DL, VT, N1);
17013 // Check RHS for X-1
17014 if (N1.getOpcode() == ISD::ADD && N1.getOperand(0) == N0 &&
17015 isAllOnes(N1.getOperand(1)))
17016 return DAG.getNode(X86ISD::BLSR, DL, VT, N0);
17021 // Want to form ANDNP nodes:
17022 // 1) In the hopes of then easily combining them with OR and AND nodes
17023 // to form PBLEND/PSIGN.
17024 // 2) To match ANDN packed intrinsics
17025 if (VT != MVT::v2i64 && VT != MVT::v4i64)
17028 SDValue N0 = N->getOperand(0);
17029 SDValue N1 = N->getOperand(1);
17032 // Check LHS for vnot
17033 if (N0.getOpcode() == ISD::XOR &&
17034 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
17035 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
17036 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
17038 // Check RHS for vnot
17039 if (N1.getOpcode() == ISD::XOR &&
17040 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
17041 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
17042 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
17047 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
17048 TargetLowering::DAGCombinerInfo &DCI,
17049 const X86Subtarget *Subtarget) {
17050 EVT VT = N->getValueType(0);
17051 if (DCI.isBeforeLegalizeOps())
17054 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
17058 SDValue N0 = N->getOperand(0);
17059 SDValue N1 = N->getOperand(1);
17061 // look for psign/blend
17062 if (VT == MVT::v2i64 || VT == MVT::v4i64) {
17063 if (!Subtarget->hasSSSE3() ||
17064 (VT == MVT::v4i64 && !Subtarget->hasInt256()))
17067 // Canonicalize pandn to RHS
17068 if (N0.getOpcode() == X86ISD::ANDNP)
17070 // or (and (m, y), (pandn m, x))
17071 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
17072 SDValue Mask = N1.getOperand(0);
17073 SDValue X = N1.getOperand(1);
17075 if (N0.getOperand(0) == Mask)
17076 Y = N0.getOperand(1);
17077 if (N0.getOperand(1) == Mask)
17078 Y = N0.getOperand(0);
17080 // Check to see if the mask appeared in both the AND and ANDNP and
17084 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
17085 // Look through mask bitcast.
17086 if (Mask.getOpcode() == ISD::BITCAST)
17087 Mask = Mask.getOperand(0);
17088 if (X.getOpcode() == ISD::BITCAST)
17089 X = X.getOperand(0);
17090 if (Y.getOpcode() == ISD::BITCAST)
17091 Y = Y.getOperand(0);
17093 EVT MaskVT = Mask.getValueType();
17095 // Validate that the Mask operand is a vector sra node.
17096 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
17097 // there is no psrai.b
17098 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
17099 unsigned SraAmt = ~0;
17100 if (Mask.getOpcode() == ISD::SRA) {
17101 SDValue Amt = Mask.getOperand(1);
17102 if (isSplatVector(Amt.getNode())) {
17103 SDValue SclrAmt = Amt->getOperand(0);
17104 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(SclrAmt))
17105 SraAmt = C->getZExtValue();
17107 } else if (Mask.getOpcode() == X86ISD::VSRAI) {
17108 SDValue SraC = Mask.getOperand(1);
17109 SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
17111 if ((SraAmt + 1) != EltBits)
17116 // Now we know we at least have a plendvb with the mask val. See if
17117 // we can form a psignb/w/d.
17118 // psign = x.type == y.type == mask.type && y = sub(0, x);
17119 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
17120 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
17121 X.getValueType() == MaskVT && Y.getValueType() == MaskVT) {
17122 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
17123 "Unsupported VT for PSIGN");
17124 Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask.getOperand(0));
17125 return DAG.getNode(ISD::BITCAST, DL, VT, Mask);
17127 // PBLENDVB only available on SSE 4.1
17128 if (!Subtarget->hasSSE41())
17131 EVT BlendVT = (VT == MVT::v4i64) ? MVT::v32i8 : MVT::v16i8;
17133 X = DAG.getNode(ISD::BITCAST, DL, BlendVT, X);
17134 Y = DAG.getNode(ISD::BITCAST, DL, BlendVT, Y);
17135 Mask = DAG.getNode(ISD::BITCAST, DL, BlendVT, Mask);
17136 Mask = DAG.getNode(ISD::VSELECT, DL, BlendVT, Mask, Y, X);
17137 return DAG.getNode(ISD::BITCAST, DL, VT, Mask);
17141 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
17144 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
17145 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
17147 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
17149 if (!N0.hasOneUse() || !N1.hasOneUse())
17152 SDValue ShAmt0 = N0.getOperand(1);
17153 if (ShAmt0.getValueType() != MVT::i8)
17155 SDValue ShAmt1 = N1.getOperand(1);
17156 if (ShAmt1.getValueType() != MVT::i8)
17158 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
17159 ShAmt0 = ShAmt0.getOperand(0);
17160 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
17161 ShAmt1 = ShAmt1.getOperand(0);
17164 unsigned Opc = X86ISD::SHLD;
17165 SDValue Op0 = N0.getOperand(0);
17166 SDValue Op1 = N1.getOperand(0);
17167 if (ShAmt0.getOpcode() == ISD::SUB) {
17168 Opc = X86ISD::SHRD;
17169 std::swap(Op0, Op1);
17170 std::swap(ShAmt0, ShAmt1);
17173 unsigned Bits = VT.getSizeInBits();
17174 if (ShAmt1.getOpcode() == ISD::SUB) {
17175 SDValue Sum = ShAmt1.getOperand(0);
17176 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
17177 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
17178 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
17179 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
17180 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
17181 return DAG.getNode(Opc, DL, VT,
17183 DAG.getNode(ISD::TRUNCATE, DL,
17186 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
17187 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
17189 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
17190 return DAG.getNode(Opc, DL, VT,
17191 N0.getOperand(0), N1.getOperand(0),
17192 DAG.getNode(ISD::TRUNCATE, DL,
17199 // Generate NEG and CMOV for integer abs.
17200 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
17201 EVT VT = N->getValueType(0);
17203 // Since X86 does not have CMOV for 8-bit integer, we don't convert
17204 // 8-bit integer abs to NEG and CMOV.
17205 if (VT.isInteger() && VT.getSizeInBits() == 8)
17208 SDValue N0 = N->getOperand(0);
17209 SDValue N1 = N->getOperand(1);
17212 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
17213 // and change it to SUB and CMOV.
17214 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
17215 N0.getOpcode() == ISD::ADD &&
17216 N0.getOperand(1) == N1 &&
17217 N1.getOpcode() == ISD::SRA &&
17218 N1.getOperand(0) == N0.getOperand(0))
17219 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
17220 if (Y1C->getAPIntValue() == VT.getSizeInBits()-1) {
17221 // Generate SUB & CMOV.
17222 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32),
17223 DAG.getConstant(0, VT), N0.getOperand(0));
17225 SDValue Ops[] = { N0.getOperand(0), Neg,
17226 DAG.getConstant(X86::COND_GE, MVT::i8),
17227 SDValue(Neg.getNode(), 1) };
17228 return DAG.getNode(X86ISD::CMOV, DL, DAG.getVTList(VT, MVT::Glue),
17229 Ops, array_lengthof(Ops));
17234 // PerformXorCombine - Attempts to turn XOR nodes into BLSMSK nodes
17235 static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
17236 TargetLowering::DAGCombinerInfo &DCI,
17237 const X86Subtarget *Subtarget) {
17238 EVT VT = N->getValueType(0);
17239 if (DCI.isBeforeLegalizeOps())
17242 if (Subtarget->hasCMov()) {
17243 SDValue RV = performIntegerAbsCombine(N, DAG);
17248 // Try forming BMI if it is available.
17249 if (!Subtarget->hasBMI())
17252 if (VT != MVT::i32 && VT != MVT::i64)
17255 assert(Subtarget->hasBMI() && "Creating BLSMSK requires BMI instructions");
17257 // Create BLSMSK instructions by finding X ^ (X-1)
17258 SDValue N0 = N->getOperand(0);
17259 SDValue N1 = N->getOperand(1);
17262 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1 &&
17263 isAllOnes(N0.getOperand(1)))
17264 return DAG.getNode(X86ISD::BLSMSK, DL, VT, N1);
17266 if (N1.getOpcode() == ISD::ADD && N1.getOperand(0) == N0 &&
17267 isAllOnes(N1.getOperand(1)))
17268 return DAG.getNode(X86ISD::BLSMSK, DL, VT, N0);
17273 /// PerformLOADCombine - Do target-specific dag combines on LOAD nodes.
17274 static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
17275 TargetLowering::DAGCombinerInfo &DCI,
17276 const X86Subtarget *Subtarget) {
17277 LoadSDNode *Ld = cast<LoadSDNode>(N);
17278 EVT RegVT = Ld->getValueType(0);
17279 EVT MemVT = Ld->getMemoryVT();
17281 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
17282 unsigned RegSz = RegVT.getSizeInBits();
17284 // On Sandybridge unaligned 256bit loads are inefficient.
17285 ISD::LoadExtType Ext = Ld->getExtensionType();
17286 unsigned Alignment = Ld->getAlignment();
17287 bool IsAligned = Alignment == 0 || Alignment >= MemVT.getSizeInBits()/8;
17288 if (RegVT.is256BitVector() && !Subtarget->hasInt256() &&
17289 !DCI.isBeforeLegalizeOps() && !IsAligned && Ext == ISD::NON_EXTLOAD) {
17290 unsigned NumElems = RegVT.getVectorNumElements();
17294 SDValue Ptr = Ld->getBasePtr();
17295 SDValue Increment = DAG.getConstant(16, TLI.getPointerTy());
17297 EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
17299 SDValue Load1 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
17300 Ld->getPointerInfo(), Ld->isVolatile(),
17301 Ld->isNonTemporal(), Ld->isInvariant(),
17303 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
17304 SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
17305 Ld->getPointerInfo(), Ld->isVolatile(),
17306 Ld->isNonTemporal(), Ld->isInvariant(),
17307 std::min(16U, Alignment));
17308 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
17310 Load2.getValue(1));
17312 SDValue NewVec = DAG.getUNDEF(RegVT);
17313 NewVec = Insert128BitVector(NewVec, Load1, 0, DAG, dl);
17314 NewVec = Insert128BitVector(NewVec, Load2, NumElems/2, DAG, dl);
17315 return DCI.CombineTo(N, NewVec, TF, true);
17318 // If this is a vector EXT Load then attempt to optimize it using a
17319 // shuffle. If SSSE3 is not available we may emit an illegal shuffle but the
17320 // expansion is still better than scalar code.
17321 // We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise we'll
17322 // emit a shuffle and a arithmetic shift.
17323 // TODO: It is possible to support ZExt by zeroing the undef values
17324 // during the shuffle phase or after the shuffle.
17325 if (RegVT.isVector() && RegVT.isInteger() && Subtarget->hasSSE2() &&
17326 (Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)) {
17327 assert(MemVT != RegVT && "Cannot extend to the same type");
17328 assert(MemVT.isVector() && "Must load a vector from memory");
17330 unsigned NumElems = RegVT.getVectorNumElements();
17331 unsigned MemSz = MemVT.getSizeInBits();
17332 assert(RegSz > MemSz && "Register size must be greater than the mem size");
17334 if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256())
17337 // All sizes must be a power of two.
17338 if (!isPowerOf2_32(RegSz * MemSz * NumElems))
17341 // Attempt to load the original value using scalar loads.
17342 // Find the largest scalar type that divides the total loaded size.
17343 MVT SclrLoadTy = MVT::i8;
17344 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
17345 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
17346 MVT Tp = (MVT::SimpleValueType)tp;
17347 if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
17352 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
17353 if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
17355 SclrLoadTy = MVT::f64;
17357 // Calculate the number of scalar loads that we need to perform
17358 // in order to load our vector from memory.
17359 unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
17360 if (Ext == ISD::SEXTLOAD && NumLoads > 1)
17363 unsigned loadRegZize = RegSz;
17364 if (Ext == ISD::SEXTLOAD && RegSz == 256)
17367 // Represent our vector as a sequence of elements which are the
17368 // largest scalar that we can load.
17369 EVT LoadUnitVecVT = EVT::getVectorVT(*DAG.getContext(), SclrLoadTy,
17370 loadRegZize/SclrLoadTy.getSizeInBits());
17372 // Represent the data using the same element type that is stored in
17373 // memory. In practice, we ''widen'' MemVT.
17375 EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
17376 loadRegZize/MemVT.getScalarType().getSizeInBits());
17378 assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
17379 "Invalid vector type");
17381 // We can't shuffle using an illegal type.
17382 if (!TLI.isTypeLegal(WideVecVT))
17385 SmallVector<SDValue, 8> Chains;
17386 SDValue Ptr = Ld->getBasePtr();
17387 SDValue Increment = DAG.getConstant(SclrLoadTy.getSizeInBits()/8,
17388 TLI.getPointerTy());
17389 SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
17391 for (unsigned i = 0; i < NumLoads; ++i) {
17392 // Perform a single load.
17393 SDValue ScalarLoad = DAG.getLoad(SclrLoadTy, dl, Ld->getChain(),
17394 Ptr, Ld->getPointerInfo(),
17395 Ld->isVolatile(), Ld->isNonTemporal(),
17396 Ld->isInvariant(), Ld->getAlignment());
17397 Chains.push_back(ScalarLoad.getValue(1));
17398 // Create the first element type using SCALAR_TO_VECTOR in order to avoid
17399 // another round of DAGCombining.
17401 Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
17403 Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
17404 ScalarLoad, DAG.getIntPtrConstant(i));
17406 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
17409 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Chains[0],
17412 // Bitcast the loaded value to a vector of the original element type, in
17413 // the size of the target vector type.
17414 SDValue SlicedVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Res);
17415 unsigned SizeRatio = RegSz/MemSz;
17417 if (Ext == ISD::SEXTLOAD) {
17418 // If we have SSE4.1 we can directly emit a VSEXT node.
17419 if (Subtarget->hasSSE41()) {
17420 SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
17421 return DCI.CombineTo(N, Sext, TF, true);
17424 // Otherwise we'll shuffle the small elements in the high bits of the
17425 // larger type and perform an arithmetic shift. If the shift is not legal
17426 // it's better to scalarize.
17427 if (!TLI.isOperationLegalOrCustom(ISD::SRA, RegVT))
17430 // Redistribute the loaded elements into the different locations.
17431 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
17432 for (unsigned i = 0; i != NumElems; ++i)
17433 ShuffleVec[i*SizeRatio + SizeRatio-1] = i;
17435 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
17436 DAG.getUNDEF(WideVecVT),
17439 Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
17441 // Build the arithmetic shift.
17442 unsigned Amt = RegVT.getVectorElementType().getSizeInBits() -
17443 MemVT.getVectorElementType().getSizeInBits();
17444 Shuff = DAG.getNode(ISD::SRA, dl, RegVT, Shuff,
17445 DAG.getConstant(Amt, RegVT));
17447 return DCI.CombineTo(N, Shuff, TF, true);
17450 // Redistribute the loaded elements into the different locations.
17451 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
17452 for (unsigned i = 0; i != NumElems; ++i)
17453 ShuffleVec[i*SizeRatio] = i;
17455 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
17456 DAG.getUNDEF(WideVecVT),
17459 // Bitcast to the requested type.
17460 Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
17461 // Replace the original load with the new sequence
17462 // and return the new chain.
17463 return DCI.CombineTo(N, Shuff, TF, true);
17469 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
17470 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
17471 const X86Subtarget *Subtarget) {
17472 StoreSDNode *St = cast<StoreSDNode>(N);
17473 EVT VT = St->getValue().getValueType();
17474 EVT StVT = St->getMemoryVT();
17476 SDValue StoredVal = St->getOperand(1);
17477 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
17479 // If we are saving a concatenation of two XMM registers, perform two stores.
17480 // On Sandy Bridge, 256-bit memory operations are executed by two
17481 // 128-bit ports. However, on Haswell it is better to issue a single 256-bit
17482 // memory operation.
17483 unsigned Alignment = St->getAlignment();
17484 bool IsAligned = Alignment == 0 || Alignment >= VT.getSizeInBits()/8;
17485 if (VT.is256BitVector() && !Subtarget->hasInt256() &&
17486 StVT == VT && !IsAligned) {
17487 unsigned NumElems = VT.getVectorNumElements();
17491 SDValue Value0 = Extract128BitVector(StoredVal, 0, DAG, dl);
17492 SDValue Value1 = Extract128BitVector(StoredVal, NumElems/2, DAG, dl);
17494 SDValue Stride = DAG.getConstant(16, TLI.getPointerTy());
17495 SDValue Ptr0 = St->getBasePtr();
17496 SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
17498 SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
17499 St->getPointerInfo(), St->isVolatile(),
17500 St->isNonTemporal(), Alignment);
17501 SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
17502 St->getPointerInfo(), St->isVolatile(),
17503 St->isNonTemporal(),
17504 std::min(16U, Alignment));
17505 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
17508 // Optimize trunc store (of multiple scalars) to shuffle and store.
17509 // First, pack all of the elements in one place. Next, store to memory
17510 // in fewer chunks.
17511 if (St->isTruncatingStore() && VT.isVector()) {
17512 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
17513 unsigned NumElems = VT.getVectorNumElements();
17514 assert(StVT != VT && "Cannot truncate to the same type");
17515 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
17516 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
17518 // From, To sizes and ElemCount must be pow of two
17519 if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
17520 // We are going to use the original vector elt for storing.
17521 // Accumulated smaller vector elements must be a multiple of the store size.
17522 if (0 != (NumElems * FromSz) % ToSz) return SDValue();
17524 unsigned SizeRatio = FromSz / ToSz;
17526 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
17528 // Create a type on which we perform the shuffle
17529 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
17530 StVT.getScalarType(), NumElems*SizeRatio);
17532 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
17534 SDValue WideVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, St->getValue());
17535 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
17536 for (unsigned i = 0; i != NumElems; ++i)
17537 ShuffleVec[i] = i * SizeRatio;
17539 // Can't shuffle using an illegal type.
17540 if (!TLI.isTypeLegal(WideVecVT))
17543 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
17544 DAG.getUNDEF(WideVecVT),
17546 // At this point all of the data is stored at the bottom of the
17547 // register. We now need to save it to mem.
17549 // Find the largest store unit
17550 MVT StoreType = MVT::i8;
17551 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
17552 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
17553 MVT Tp = (MVT::SimpleValueType)tp;
17554 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToSz)
17558 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
17559 if (TLI.isTypeLegal(MVT::f64) && StoreType.getSizeInBits() < 64 &&
17560 (64 <= NumElems * ToSz))
17561 StoreType = MVT::f64;
17563 // Bitcast the original vector into a vector of store-size units
17564 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
17565 StoreType, VT.getSizeInBits()/StoreType.getSizeInBits());
17566 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
17567 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, dl, StoreVecVT, Shuff);
17568 SmallVector<SDValue, 8> Chains;
17569 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
17570 TLI.getPointerTy());
17571 SDValue Ptr = St->getBasePtr();
17573 // Perform one or more big stores into memory.
17574 for (unsigned i=0, e=(ToSz*NumElems)/StoreType.getSizeInBits(); i!=e; ++i) {
17575 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
17576 StoreType, ShuffWide,
17577 DAG.getIntPtrConstant(i));
17578 SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr,
17579 St->getPointerInfo(), St->isVolatile(),
17580 St->isNonTemporal(), St->getAlignment());
17581 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
17582 Chains.push_back(Ch);
17585 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Chains[0],
17589 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
17590 // the FP state in cases where an emms may be missing.
17591 // A preferable solution to the general problem is to figure out the right
17592 // places to insert EMMS. This qualifies as a quick hack.
17594 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
17595 if (VT.getSizeInBits() != 64)
17598 const Function *F = DAG.getMachineFunction().getFunction();
17599 bool NoImplicitFloatOps = F->getAttributes().
17600 hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat);
17601 bool F64IsLegal = !DAG.getTarget().Options.UseSoftFloat && !NoImplicitFloatOps
17602 && Subtarget->hasSSE2();
17603 if ((VT.isVector() ||
17604 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
17605 isa<LoadSDNode>(St->getValue()) &&
17606 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
17607 St->getChain().hasOneUse() && !St->isVolatile()) {
17608 SDNode* LdVal = St->getValue().getNode();
17609 LoadSDNode *Ld = 0;
17610 int TokenFactorIndex = -1;
17611 SmallVector<SDValue, 8> Ops;
17612 SDNode* ChainVal = St->getChain().getNode();
17613 // Must be a store of a load. We currently handle two cases: the load
17614 // is a direct child, and it's under an intervening TokenFactor. It is
17615 // possible to dig deeper under nested TokenFactors.
17616 if (ChainVal == LdVal)
17617 Ld = cast<LoadSDNode>(St->getChain());
17618 else if (St->getValue().hasOneUse() &&
17619 ChainVal->getOpcode() == ISD::TokenFactor) {
17620 for (unsigned i = 0, e = ChainVal->getNumOperands(); i != e; ++i) {
17621 if (ChainVal->getOperand(i).getNode() == LdVal) {
17622 TokenFactorIndex = i;
17623 Ld = cast<LoadSDNode>(St->getValue());
17625 Ops.push_back(ChainVal->getOperand(i));
17629 if (!Ld || !ISD::isNormalLoad(Ld))
17632 // If this is not the MMX case, i.e. we are just turning i64 load/store
17633 // into f64 load/store, avoid the transformation if there are multiple
17634 // uses of the loaded value.
17635 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
17640 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
17641 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
17643 if (Subtarget->is64Bit() || F64IsLegal) {
17644 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
17645 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
17646 Ld->getPointerInfo(), Ld->isVolatile(),
17647 Ld->isNonTemporal(), Ld->isInvariant(),
17648 Ld->getAlignment());
17649 SDValue NewChain = NewLd.getValue(1);
17650 if (TokenFactorIndex != -1) {
17651 Ops.push_back(NewChain);
17652 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
17655 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
17656 St->getPointerInfo(),
17657 St->isVolatile(), St->isNonTemporal(),
17658 St->getAlignment());
17661 // Otherwise, lower to two pairs of 32-bit loads / stores.
17662 SDValue LoAddr = Ld->getBasePtr();
17663 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
17664 DAG.getConstant(4, MVT::i32));
17666 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
17667 Ld->getPointerInfo(),
17668 Ld->isVolatile(), Ld->isNonTemporal(),
17669 Ld->isInvariant(), Ld->getAlignment());
17670 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
17671 Ld->getPointerInfo().getWithOffset(4),
17672 Ld->isVolatile(), Ld->isNonTemporal(),
17674 MinAlign(Ld->getAlignment(), 4));
17676 SDValue NewChain = LoLd.getValue(1);
17677 if (TokenFactorIndex != -1) {
17678 Ops.push_back(LoLd);
17679 Ops.push_back(HiLd);
17680 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
17684 LoAddr = St->getBasePtr();
17685 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
17686 DAG.getConstant(4, MVT::i32));
17688 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
17689 St->getPointerInfo(),
17690 St->isVolatile(), St->isNonTemporal(),
17691 St->getAlignment());
17692 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
17693 St->getPointerInfo().getWithOffset(4),
17695 St->isNonTemporal(),
17696 MinAlign(St->getAlignment(), 4));
17697 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
17702 /// isHorizontalBinOp - Return 'true' if this vector operation is "horizontal"
17703 /// and return the operands for the horizontal operation in LHS and RHS. A
17704 /// horizontal operation performs the binary operation on successive elements
17705 /// of its first operand, then on successive elements of its second operand,
17706 /// returning the resulting values in a vector. For example, if
17707 /// A = < float a0, float a1, float a2, float a3 >
17709 /// B = < float b0, float b1, float b2, float b3 >
17710 /// then the result of doing a horizontal operation on A and B is
17711 /// A horizontal-op B = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >.
17712 /// In short, LHS and RHS are inspected to see if LHS op RHS is of the form
17713 /// A horizontal-op B, for some already available A and B, and if so then LHS is
17714 /// set to A, RHS to B, and the routine returns 'true'.
17715 /// Note that the binary operation should have the property that if one of the
17716 /// operands is UNDEF then the result is UNDEF.
17717 static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool IsCommutative) {
17718 // Look for the following pattern: if
17719 // A = < float a0, float a1, float a2, float a3 >
17720 // B = < float b0, float b1, float b2, float b3 >
17722 // LHS = VECTOR_SHUFFLE A, B, <0, 2, 4, 6>
17723 // RHS = VECTOR_SHUFFLE A, B, <1, 3, 5, 7>
17724 // then LHS op RHS = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >
17725 // which is A horizontal-op B.
17727 // At least one of the operands should be a vector shuffle.
17728 if (LHS.getOpcode() != ISD::VECTOR_SHUFFLE &&
17729 RHS.getOpcode() != ISD::VECTOR_SHUFFLE)
17732 MVT VT = LHS.getValueType().getSimpleVT();
17734 assert((VT.is128BitVector() || VT.is256BitVector()) &&
17735 "Unsupported vector type for horizontal add/sub");
17737 // Handle 128 and 256-bit vector lengths. AVX defines horizontal add/sub to
17738 // operate independently on 128-bit lanes.
17739 unsigned NumElts = VT.getVectorNumElements();
17740 unsigned NumLanes = VT.getSizeInBits()/128;
17741 unsigned NumLaneElts = NumElts / NumLanes;
17742 assert((NumLaneElts % 2 == 0) &&
17743 "Vector type should have an even number of elements in each lane");
17744 unsigned HalfLaneElts = NumLaneElts/2;
17746 // View LHS in the form
17747 // LHS = VECTOR_SHUFFLE A, B, LMask
17748 // If LHS is not a shuffle then pretend it is the shuffle
17749 // LHS = VECTOR_SHUFFLE LHS, undef, <0, 1, ..., N-1>
17750 // NOTE: in what follows a default initialized SDValue represents an UNDEF of
17753 SmallVector<int, 16> LMask(NumElts);
17754 if (LHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
17755 if (LHS.getOperand(0).getOpcode() != ISD::UNDEF)
17756 A = LHS.getOperand(0);
17757 if (LHS.getOperand(1).getOpcode() != ISD::UNDEF)
17758 B = LHS.getOperand(1);
17759 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(LHS.getNode())->getMask();
17760 std::copy(Mask.begin(), Mask.end(), LMask.begin());
17762 if (LHS.getOpcode() != ISD::UNDEF)
17764 for (unsigned i = 0; i != NumElts; ++i)
17768 // Likewise, view RHS in the form
17769 // RHS = VECTOR_SHUFFLE C, D, RMask
17771 SmallVector<int, 16> RMask(NumElts);
17772 if (RHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
17773 if (RHS.getOperand(0).getOpcode() != ISD::UNDEF)
17774 C = RHS.getOperand(0);
17775 if (RHS.getOperand(1).getOpcode() != ISD::UNDEF)
17776 D = RHS.getOperand(1);
17777 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(RHS.getNode())->getMask();
17778 std::copy(Mask.begin(), Mask.end(), RMask.begin());
17780 if (RHS.getOpcode() != ISD::UNDEF)
17782 for (unsigned i = 0; i != NumElts; ++i)
17786 // Check that the shuffles are both shuffling the same vectors.
17787 if (!(A == C && B == D) && !(A == D && B == C))
17790 // If everything is UNDEF then bail out: it would be better to fold to UNDEF.
17791 if (!A.getNode() && !B.getNode())
17794 // If A and B occur in reverse order in RHS, then "swap" them (which means
17795 // rewriting the mask).
17797 CommuteVectorShuffleMask(RMask, NumElts);
17799 // At this point LHS and RHS are equivalent to
17800 // LHS = VECTOR_SHUFFLE A, B, LMask
17801 // RHS = VECTOR_SHUFFLE A, B, RMask
17802 // Check that the masks correspond to performing a horizontal operation.
17803 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
17804 for (unsigned i = 0; i != NumLaneElts; ++i) {
17805 int LIdx = LMask[i+l], RIdx = RMask[i+l];
17807 // Ignore any UNDEF components.
17808 if (LIdx < 0 || RIdx < 0 ||
17809 (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) ||
17810 (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts)))
17813 // Check that successive elements are being operated on. If not, this is
17814 // not a horizontal operation.
17815 unsigned Src = (i/HalfLaneElts); // each lane is split between srcs
17816 int Index = 2*(i%HalfLaneElts) + NumElts*Src + l;
17817 if (!(LIdx == Index && RIdx == Index + 1) &&
17818 !(IsCommutative && LIdx == Index + 1 && RIdx == Index))
17823 LHS = A.getNode() ? A : B; // If A is 'UNDEF', use B for it.
17824 RHS = B.getNode() ? B : A; // If B is 'UNDEF', use A for it.
17828 /// PerformFADDCombine - Do target-specific dag combines on floating point adds.
17829 static SDValue PerformFADDCombine(SDNode *N, SelectionDAG &DAG,
17830 const X86Subtarget *Subtarget) {
17831 EVT VT = N->getValueType(0);
17832 SDValue LHS = N->getOperand(0);
17833 SDValue RHS = N->getOperand(1);
17835 // Try to synthesize horizontal adds from adds of shuffles.
17836 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
17837 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
17838 isHorizontalBinOp(LHS, RHS, true))
17839 return DAG.getNode(X86ISD::FHADD, SDLoc(N), VT, LHS, RHS);
17843 /// PerformFSUBCombine - Do target-specific dag combines on floating point subs.
17844 static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG,
17845 const X86Subtarget *Subtarget) {
17846 EVT VT = N->getValueType(0);
17847 SDValue LHS = N->getOperand(0);
17848 SDValue RHS = N->getOperand(1);
17850 // Try to synthesize horizontal subs from subs of shuffles.
17851 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
17852 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
17853 isHorizontalBinOp(LHS, RHS, false))
17854 return DAG.getNode(X86ISD::FHSUB, SDLoc(N), VT, LHS, RHS);
17858 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
17859 /// X86ISD::FXOR nodes.
17860 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
17861 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
17862 // F[X]OR(0.0, x) -> x
17863 // F[X]OR(x, 0.0) -> x
17864 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
17865 if (C->getValueAPF().isPosZero())
17866 return N->getOperand(1);
17867 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
17868 if (C->getValueAPF().isPosZero())
17869 return N->getOperand(0);
17873 /// PerformFMinFMaxCombine - Do target-specific dag combines on X86ISD::FMIN and
17874 /// X86ISD::FMAX nodes.
17875 static SDValue PerformFMinFMaxCombine(SDNode *N, SelectionDAG &DAG) {
17876 assert(N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD::FMAX);
17878 // Only perform optimizations if UnsafeMath is used.
17879 if (!DAG.getTarget().Options.UnsafeFPMath)
17882 // If we run in unsafe-math mode, then convert the FMAX and FMIN nodes
17883 // into FMINC and FMAXC, which are Commutative operations.
17884 unsigned NewOp = 0;
17885 switch (N->getOpcode()) {
17886 default: llvm_unreachable("unknown opcode");
17887 case X86ISD::FMIN: NewOp = X86ISD::FMINC; break;
17888 case X86ISD::FMAX: NewOp = X86ISD::FMAXC; break;
17891 return DAG.getNode(NewOp, SDLoc(N), N->getValueType(0),
17892 N->getOperand(0), N->getOperand(1));
17895 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
17896 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
17897 // FAND(0.0, x) -> 0.0
17898 // FAND(x, 0.0) -> 0.0
17899 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
17900 if (C->getValueAPF().isPosZero())
17901 return N->getOperand(0);
17902 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
17903 if (C->getValueAPF().isPosZero())
17904 return N->getOperand(1);
17908 /// PerformFANDNCombine - Do target-specific dag combines on X86ISD::FANDN nodes
17909 static SDValue PerformFANDNCombine(SDNode *N, SelectionDAG &DAG) {
17910 // FANDN(x, 0.0) -> 0.0
17911 // FANDN(0.0, x) -> x
17912 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
17913 if (C->getValueAPF().isPosZero())
17914 return N->getOperand(1);
17915 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
17916 if (C->getValueAPF().isPosZero())
17917 return N->getOperand(1);
17921 static SDValue PerformBTCombine(SDNode *N,
17923 TargetLowering::DAGCombinerInfo &DCI) {
17924 // BT ignores high bits in the bit index operand.
17925 SDValue Op1 = N->getOperand(1);
17926 if (Op1.hasOneUse()) {
17927 unsigned BitWidth = Op1.getValueSizeInBits();
17928 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
17929 APInt KnownZero, KnownOne;
17930 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
17931 !DCI.isBeforeLegalizeOps());
17932 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
17933 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
17934 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
17935 DCI.CommitTargetLoweringOpt(TLO);
17940 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
17941 SDValue Op = N->getOperand(0);
17942 if (Op.getOpcode() == ISD::BITCAST)
17943 Op = Op.getOperand(0);
17944 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
17945 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
17946 VT.getVectorElementType().getSizeInBits() ==
17947 OpVT.getVectorElementType().getSizeInBits()) {
17948 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
17953 static SDValue PerformSIGN_EXTEND_INREGCombine(SDNode *N, SelectionDAG &DAG,
17954 const X86Subtarget *Subtarget) {
17955 EVT VT = N->getValueType(0);
17956 if (!VT.isVector())
17959 SDValue N0 = N->getOperand(0);
17960 SDValue N1 = N->getOperand(1);
17961 EVT ExtraVT = cast<VTSDNode>(N1)->getVT();
17964 // The SIGN_EXTEND_INREG to v4i64 is expensive operation on the
17965 // both SSE and AVX2 since there is no sign-extended shift right
17966 // operation on a vector with 64-bit elements.
17967 //(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) ->
17968 // (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT)))
17969 if (VT == MVT::v4i64 && (N0.getOpcode() == ISD::ANY_EXTEND ||
17970 N0.getOpcode() == ISD::SIGN_EXTEND)) {
17971 SDValue N00 = N0.getOperand(0);
17973 // EXTLOAD has a better solution on AVX2,
17974 // it may be replaced with X86ISD::VSEXT node.
17975 if (N00.getOpcode() == ISD::LOAD && Subtarget->hasInt256())
17976 if (!ISD::isNormalLoad(N00.getNode()))
17979 if (N00.getValueType() == MVT::v4i32 && ExtraVT.getSizeInBits() < 128) {
17980 SDValue Tmp = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32,
17982 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i64, Tmp);
17988 static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
17989 TargetLowering::DAGCombinerInfo &DCI,
17990 const X86Subtarget *Subtarget) {
17991 if (!DCI.isBeforeLegalizeOps())
17994 if (!Subtarget->hasFp256())
17997 EVT VT = N->getValueType(0);
17998 if (VT.isVector() && VT.getSizeInBits() == 256) {
17999 SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget);
18007 static SDValue PerformFMACombine(SDNode *N, SelectionDAG &DAG,
18008 const X86Subtarget* Subtarget) {
18010 EVT VT = N->getValueType(0);
18012 // Let legalize expand this if it isn't a legal type yet.
18013 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
18016 EVT ScalarVT = VT.getScalarType();
18017 if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) ||
18018 (!Subtarget->hasFMA() && !Subtarget->hasFMA4()))
18021 SDValue A = N->getOperand(0);
18022 SDValue B = N->getOperand(1);
18023 SDValue C = N->getOperand(2);
18025 bool NegA = (A.getOpcode() == ISD::FNEG);
18026 bool NegB = (B.getOpcode() == ISD::FNEG);
18027 bool NegC = (C.getOpcode() == ISD::FNEG);
18029 // Negative multiplication when NegA xor NegB
18030 bool NegMul = (NegA != NegB);
18032 A = A.getOperand(0);
18034 B = B.getOperand(0);
18036 C = C.getOperand(0);
18040 Opcode = (!NegC) ? X86ISD::FMADD : X86ISD::FMSUB;
18042 Opcode = (!NegC) ? X86ISD::FNMADD : X86ISD::FNMSUB;
18044 return DAG.getNode(Opcode, dl, VT, A, B, C);
18047 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG,
18048 TargetLowering::DAGCombinerInfo &DCI,
18049 const X86Subtarget *Subtarget) {
18050 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
18051 // (and (i32 x86isd::setcc_carry), 1)
18052 // This eliminates the zext. This transformation is necessary because
18053 // ISD::SETCC is always legalized to i8.
18055 SDValue N0 = N->getOperand(0);
18056 EVT VT = N->getValueType(0);
18058 if (N0.getOpcode() == ISD::AND &&
18060 N0.getOperand(0).hasOneUse()) {
18061 SDValue N00 = N0.getOperand(0);
18062 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
18063 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
18064 if (!C || C->getZExtValue() != 1)
18066 return DAG.getNode(ISD::AND, dl, VT,
18067 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
18068 N00.getOperand(0), N00.getOperand(1)),
18069 DAG.getConstant(1, VT));
18073 if (VT.is256BitVector()) {
18074 SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget);
18082 // Optimize x == -y --> x+y == 0
18083 // x != -y --> x+y != 0
18084 static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG) {
18085 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
18086 SDValue LHS = N->getOperand(0);
18087 SDValue RHS = N->getOperand(1);
18089 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB)
18090 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(LHS.getOperand(0)))
18091 if (C->getAPIntValue() == 0 && LHS.hasOneUse()) {
18092 SDValue addV = DAG.getNode(ISD::ADD, SDLoc(N),
18093 LHS.getValueType(), RHS, LHS.getOperand(1));
18094 return DAG.getSetCC(SDLoc(N), N->getValueType(0),
18095 addV, DAG.getConstant(0, addV.getValueType()), CC);
18097 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && RHS.getOpcode() == ISD::SUB)
18098 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS.getOperand(0)))
18099 if (C->getAPIntValue() == 0 && RHS.hasOneUse()) {
18100 SDValue addV = DAG.getNode(ISD::ADD, SDLoc(N),
18101 RHS.getValueType(), LHS, RHS.getOperand(1));
18102 return DAG.getSetCC(SDLoc(N), N->getValueType(0),
18103 addV, DAG.getConstant(0, addV.getValueType()), CC);
18108 // Helper function of PerformSETCCCombine. It is to materialize "setb reg"
18109 // as "sbb reg,reg", since it can be extended without zext and produces
18110 // an all-ones bit which is more useful than 0/1 in some cases.
18111 static SDValue MaterializeSETB(SDLoc DL, SDValue EFLAGS, SelectionDAG &DAG) {
18112 return DAG.getNode(ISD::AND, DL, MVT::i8,
18113 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
18114 DAG.getConstant(X86::COND_B, MVT::i8), EFLAGS),
18115 DAG.getConstant(1, MVT::i8));
18118 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
18119 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG,
18120 TargetLowering::DAGCombinerInfo &DCI,
18121 const X86Subtarget *Subtarget) {
18123 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0));
18124 SDValue EFLAGS = N->getOperand(1);
18126 if (CC == X86::COND_A) {
18127 // Try to convert COND_A into COND_B in an attempt to facilitate
18128 // materializing "setb reg".
18130 // Do not flip "e > c", where "c" is a constant, because Cmp instruction
18131 // cannot take an immediate as its first operand.
18133 if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() &&
18134 EFLAGS.getValueType().isInteger() &&
18135 !isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
18136 SDValue NewSub = DAG.getNode(X86ISD::SUB, SDLoc(EFLAGS),
18137 EFLAGS.getNode()->getVTList(),
18138 EFLAGS.getOperand(1), EFLAGS.getOperand(0));
18139 SDValue NewEFLAGS = SDValue(NewSub.getNode(), EFLAGS.getResNo());
18140 return MaterializeSETB(DL, NewEFLAGS, DAG);
18144 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
18145 // a zext and produces an all-ones bit which is more useful than 0/1 in some
18147 if (CC == X86::COND_B)
18148 return MaterializeSETB(DL, EFLAGS, DAG);
18152 Flags = checkBoolTestSetCCCombine(EFLAGS, CC);
18153 if (Flags.getNode()) {
18154 SDValue Cond = DAG.getConstant(CC, MVT::i8);
18155 return DAG.getNode(X86ISD::SETCC, DL, N->getVTList(), Cond, Flags);
18161 // Optimize branch condition evaluation.
18163 static SDValue PerformBrCondCombine(SDNode *N, SelectionDAG &DAG,
18164 TargetLowering::DAGCombinerInfo &DCI,
18165 const X86Subtarget *Subtarget) {
18167 SDValue Chain = N->getOperand(0);
18168 SDValue Dest = N->getOperand(1);
18169 SDValue EFLAGS = N->getOperand(3);
18170 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(2));
18174 Flags = checkBoolTestSetCCCombine(EFLAGS, CC);
18175 if (Flags.getNode()) {
18176 SDValue Cond = DAG.getConstant(CC, MVT::i8);
18177 return DAG.getNode(X86ISD::BRCOND, DL, N->getVTList(), Chain, Dest, Cond,
18184 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
18185 const X86TargetLowering *XTLI) {
18186 SDValue Op0 = N->getOperand(0);
18187 EVT InVT = Op0->getValueType(0);
18189 // SINT_TO_FP(v4i8) -> SINT_TO_FP(SEXT(v4i8 to v4i32))
18190 if (InVT == MVT::v8i8 || InVT == MVT::v4i8) {
18192 MVT DstVT = InVT == MVT::v4i8 ? MVT::v4i32 : MVT::v8i32;
18193 SDValue P = DAG.getNode(ISD::SIGN_EXTEND, dl, DstVT, Op0);
18194 return DAG.getNode(ISD::SINT_TO_FP, dl, N->getValueType(0), P);
18197 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
18198 // a 32-bit target where SSE doesn't support i64->FP operations.
18199 if (Op0.getOpcode() == ISD::LOAD) {
18200 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
18201 EVT VT = Ld->getValueType(0);
18202 if (!Ld->isVolatile() && !N->getValueType(0).isVector() &&
18203 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
18204 !XTLI->getSubtarget()->is64Bit() &&
18205 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
18206 SDValue FILDChain = XTLI->BuildFILD(SDValue(N, 0), Ld->getValueType(0),
18207 Ld->getChain(), Op0, DAG);
18208 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
18215 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
18216 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
18217 X86TargetLowering::DAGCombinerInfo &DCI) {
18218 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
18219 // the result is either zero or one (depending on the input carry bit).
18220 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
18221 if (X86::isZeroNode(N->getOperand(0)) &&
18222 X86::isZeroNode(N->getOperand(1)) &&
18223 // We don't have a good way to replace an EFLAGS use, so only do this when
18225 SDValue(N, 1).use_empty()) {
18227 EVT VT = N->getValueType(0);
18228 SDValue CarryOut = DAG.getConstant(0, N->getValueType(1));
18229 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
18230 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
18231 DAG.getConstant(X86::COND_B,MVT::i8),
18233 DAG.getConstant(1, VT));
18234 return DCI.CombineTo(N, Res1, CarryOut);
18240 // fold (add Y, (sete X, 0)) -> adc 0, Y
18241 // (add Y, (setne X, 0)) -> sbb -1, Y
18242 // (sub (sete X, 0), Y) -> sbb 0, Y
18243 // (sub (setne X, 0), Y) -> adc -1, Y
18244 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
18247 // Look through ZExts.
18248 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
18249 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
18252 SDValue SetCC = Ext.getOperand(0);
18253 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
18256 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
18257 if (CC != X86::COND_E && CC != X86::COND_NE)
18260 SDValue Cmp = SetCC.getOperand(1);
18261 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
18262 !X86::isZeroNode(Cmp.getOperand(1)) ||
18263 !Cmp.getOperand(0).getValueType().isInteger())
18266 SDValue CmpOp0 = Cmp.getOperand(0);
18267 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
18268 DAG.getConstant(1, CmpOp0.getValueType()));
18270 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
18271 if (CC == X86::COND_NE)
18272 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
18273 DL, OtherVal.getValueType(), OtherVal,
18274 DAG.getConstant(-1ULL, OtherVal.getValueType()), NewCmp);
18275 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
18276 DL, OtherVal.getValueType(), OtherVal,
18277 DAG.getConstant(0, OtherVal.getValueType()), NewCmp);
18280 /// PerformADDCombine - Do target-specific dag combines on integer adds.
18281 static SDValue PerformAddCombine(SDNode *N, SelectionDAG &DAG,
18282 const X86Subtarget *Subtarget) {
18283 EVT VT = N->getValueType(0);
18284 SDValue Op0 = N->getOperand(0);
18285 SDValue Op1 = N->getOperand(1);
18287 // Try to synthesize horizontal adds from adds of shuffles.
18288 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
18289 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
18290 isHorizontalBinOp(Op0, Op1, true))
18291 return DAG.getNode(X86ISD::HADD, SDLoc(N), VT, Op0, Op1);
18293 return OptimizeConditionalInDecrement(N, DAG);
18296 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG,
18297 const X86Subtarget *Subtarget) {
18298 SDValue Op0 = N->getOperand(0);
18299 SDValue Op1 = N->getOperand(1);
18301 // X86 can't encode an immediate LHS of a sub. See if we can push the
18302 // negation into a preceding instruction.
18303 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
18304 // If the RHS of the sub is a XOR with one use and a constant, invert the
18305 // immediate. Then add one to the LHS of the sub so we can turn
18306 // X-Y -> X+~Y+1, saving one register.
18307 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
18308 isa<ConstantSDNode>(Op1.getOperand(1))) {
18309 APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue();
18310 EVT VT = Op0.getValueType();
18311 SDValue NewXor = DAG.getNode(ISD::XOR, SDLoc(Op1), VT,
18313 DAG.getConstant(~XorC, VT));
18314 return DAG.getNode(ISD::ADD, SDLoc(N), VT, NewXor,
18315 DAG.getConstant(C->getAPIntValue()+1, VT));
18319 // Try to synthesize horizontal adds from adds of shuffles.
18320 EVT VT = N->getValueType(0);
18321 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
18322 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
18323 isHorizontalBinOp(Op0, Op1, true))
18324 return DAG.getNode(X86ISD::HSUB, SDLoc(N), VT, Op0, Op1);
18326 return OptimizeConditionalInDecrement(N, DAG);
18329 /// performVZEXTCombine - Performs build vector combines
18330 static SDValue performVZEXTCombine(SDNode *N, SelectionDAG &DAG,
18331 TargetLowering::DAGCombinerInfo &DCI,
18332 const X86Subtarget *Subtarget) {
18333 // (vzext (bitcast (vzext (x)) -> (vzext x)
18334 SDValue In = N->getOperand(0);
18335 while (In.getOpcode() == ISD::BITCAST)
18336 In = In.getOperand(0);
18338 if (In.getOpcode() != X86ISD::VZEXT)
18341 return DAG.getNode(X86ISD::VZEXT, SDLoc(N), N->getValueType(0),
18345 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
18346 DAGCombinerInfo &DCI) const {
18347 SelectionDAG &DAG = DCI.DAG;
18348 switch (N->getOpcode()) {
18350 case ISD::EXTRACT_VECTOR_ELT:
18351 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, DCI);
18353 case ISD::SELECT: return PerformSELECTCombine(N, DAG, DCI, Subtarget);
18354 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI, Subtarget);
18355 case ISD::ADD: return PerformAddCombine(N, DAG, Subtarget);
18356 case ISD::SUB: return PerformSubCombine(N, DAG, Subtarget);
18357 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
18358 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
18361 case ISD::SRL: return PerformShiftCombine(N, DAG, DCI, Subtarget);
18362 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
18363 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
18364 case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget);
18365 case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget);
18366 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
18367 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, this);
18368 case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
18369 case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
18371 case X86ISD::FOR: return PerformFORCombine(N, DAG);
18373 case X86ISD::FMAX: return PerformFMinFMaxCombine(N, DAG);
18374 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
18375 case X86ISD::FANDN: return PerformFANDNCombine(N, DAG);
18376 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
18377 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
18378 case ISD::ANY_EXTEND:
18379 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG, DCI, Subtarget);
18380 case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
18381 case ISD::SIGN_EXTEND_INREG: return PerformSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
18382 case ISD::TRUNCATE: return PerformTruncateCombine(N, DAG,DCI,Subtarget);
18383 case ISD::SETCC: return PerformISDSETCCCombine(N, DAG);
18384 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
18385 case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
18386 case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);
18387 case X86ISD::SHUFP: // Handle all target specific shuffles
18388 case X86ISD::PALIGNR:
18389 case X86ISD::UNPCKH:
18390 case X86ISD::UNPCKL:
18391 case X86ISD::MOVHLPS:
18392 case X86ISD::MOVLHPS:
18393 case X86ISD::PSHUFD:
18394 case X86ISD::PSHUFHW:
18395 case X86ISD::PSHUFLW:
18396 case X86ISD::MOVSS:
18397 case X86ISD::MOVSD:
18398 case X86ISD::VPERMILP:
18399 case X86ISD::VPERM2X128:
18400 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
18401 case ISD::FMA: return PerformFMACombine(N, DAG, Subtarget);
18407 /// isTypeDesirableForOp - Return true if the target has native support for
18408 /// the specified value type and it is 'desirable' to use the type for the
18409 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
18410 /// instruction encodings are longer and some i16 instructions are slow.
18411 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
18412 if (!isTypeLegal(VT))
18414 if (VT != MVT::i16)
18421 case ISD::SIGN_EXTEND:
18422 case ISD::ZERO_EXTEND:
18423 case ISD::ANY_EXTEND:
18436 /// IsDesirableToPromoteOp - This method query the target whether it is
18437 /// beneficial for dag combiner to promote the specified node. If true, it
18438 /// should return the desired promotion type by reference.
18439 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
18440 EVT VT = Op.getValueType();
18441 if (VT != MVT::i16)
18444 bool Promote = false;
18445 bool Commute = false;
18446 switch (Op.getOpcode()) {
18449 LoadSDNode *LD = cast<LoadSDNode>(Op);
18450 // If the non-extending load has a single use and it's not live out, then it
18451 // might be folded.
18452 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
18453 Op.hasOneUse()*/) {
18454 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
18455 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
18456 // The only case where we'd want to promote LOAD (rather then it being
18457 // promoted as an operand is when it's only use is liveout.
18458 if (UI->getOpcode() != ISD::CopyToReg)
18465 case ISD::SIGN_EXTEND:
18466 case ISD::ZERO_EXTEND:
18467 case ISD::ANY_EXTEND:
18472 SDValue N0 = Op.getOperand(0);
18473 // Look out for (store (shl (load), x)).
18474 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
18487 SDValue N0 = Op.getOperand(0);
18488 SDValue N1 = Op.getOperand(1);
18489 if (!Commute && MayFoldLoad(N1))
18491 // Avoid disabling potential load folding opportunities.
18492 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
18494 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
18504 //===----------------------------------------------------------------------===//
18505 // X86 Inline Assembly Support
18506 //===----------------------------------------------------------------------===//
18509 // Helper to match a string separated by whitespace.
18510 bool matchAsmImpl(StringRef s, ArrayRef<const StringRef *> args) {
18511 s = s.substr(s.find_first_not_of(" \t")); // Skip leading whitespace.
18513 for (unsigned i = 0, e = args.size(); i != e; ++i) {
18514 StringRef piece(*args[i]);
18515 if (!s.startswith(piece)) // Check if the piece matches.
18518 s = s.substr(piece.size());
18519 StringRef::size_type pos = s.find_first_not_of(" \t");
18520 if (pos == 0) // We matched a prefix.
18528 const VariadicFunction1<bool, StringRef, StringRef, matchAsmImpl> matchAsm={};
18531 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
18532 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
18534 std::string AsmStr = IA->getAsmString();
18536 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
18537 if (!Ty || Ty->getBitWidth() % 16 != 0)
18540 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
18541 SmallVector<StringRef, 4> AsmPieces;
18542 SplitString(AsmStr, AsmPieces, ";\n");
18544 switch (AsmPieces.size()) {
18545 default: return false;
18547 // FIXME: this should verify that we are targeting a 486 or better. If not,
18548 // we will turn this bswap into something that will be lowered to logical
18549 // ops instead of emitting the bswap asm. For now, we don't support 486 or
18550 // lower so don't worry about this.
18552 if (matchAsm(AsmPieces[0], "bswap", "$0") ||
18553 matchAsm(AsmPieces[0], "bswapl", "$0") ||
18554 matchAsm(AsmPieces[0], "bswapq", "$0") ||
18555 matchAsm(AsmPieces[0], "bswap", "${0:q}") ||
18556 matchAsm(AsmPieces[0], "bswapl", "${0:q}") ||
18557 matchAsm(AsmPieces[0], "bswapq", "${0:q}")) {
18558 // No need to check constraints, nothing other than the equivalent of
18559 // "=r,0" would be valid here.
18560 return IntrinsicLowering::LowerToByteSwap(CI);
18563 // rorw $$8, ${0:w} --> llvm.bswap.i16
18564 if (CI->getType()->isIntegerTy(16) &&
18565 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
18566 (matchAsm(AsmPieces[0], "rorw", "$$8,", "${0:w}") ||
18567 matchAsm(AsmPieces[0], "rolw", "$$8,", "${0:w}"))) {
18569 const std::string &ConstraintsStr = IA->getConstraintString();
18570 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
18571 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
18572 if (AsmPieces.size() == 4 &&
18573 AsmPieces[0] == "~{cc}" &&
18574 AsmPieces[1] == "~{dirflag}" &&
18575 AsmPieces[2] == "~{flags}" &&
18576 AsmPieces[3] == "~{fpsr}")
18577 return IntrinsicLowering::LowerToByteSwap(CI);
18581 if (CI->getType()->isIntegerTy(32) &&
18582 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
18583 matchAsm(AsmPieces[0], "rorw", "$$8,", "${0:w}") &&
18584 matchAsm(AsmPieces[1], "rorl", "$$16,", "$0") &&
18585 matchAsm(AsmPieces[2], "rorw", "$$8,", "${0:w}")) {
18587 const std::string &ConstraintsStr = IA->getConstraintString();
18588 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
18589 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
18590 if (AsmPieces.size() == 4 &&
18591 AsmPieces[0] == "~{cc}" &&
18592 AsmPieces[1] == "~{dirflag}" &&
18593 AsmPieces[2] == "~{flags}" &&
18594 AsmPieces[3] == "~{fpsr}")
18595 return IntrinsicLowering::LowerToByteSwap(CI);
18598 if (CI->getType()->isIntegerTy(64)) {
18599 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
18600 if (Constraints.size() >= 2 &&
18601 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
18602 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
18603 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
18604 if (matchAsm(AsmPieces[0], "bswap", "%eax") &&
18605 matchAsm(AsmPieces[1], "bswap", "%edx") &&
18606 matchAsm(AsmPieces[2], "xchgl", "%eax,", "%edx"))
18607 return IntrinsicLowering::LowerToByteSwap(CI);
18615 /// getConstraintType - Given a constraint letter, return the type of
18616 /// constraint it is for this target.
18617 X86TargetLowering::ConstraintType
18618 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
18619 if (Constraint.size() == 1) {
18620 switch (Constraint[0]) {
18631 return C_RegisterClass;
18655 return TargetLowering::getConstraintType(Constraint);
18658 /// Examine constraint type and operand type and determine a weight value.
18659 /// This object must already have been set up with the operand type
18660 /// and the current alternative constraint selected.
18661 TargetLowering::ConstraintWeight
18662 X86TargetLowering::getSingleConstraintMatchWeight(
18663 AsmOperandInfo &info, const char *constraint) const {
18664 ConstraintWeight weight = CW_Invalid;
18665 Value *CallOperandVal = info.CallOperandVal;
18666 // If we don't have a value, we can't do a match,
18667 // but allow it at the lowest weight.
18668 if (CallOperandVal == NULL)
18670 Type *type = CallOperandVal->getType();
18671 // Look at the constraint type.
18672 switch (*constraint) {
18674 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
18685 if (CallOperandVal->getType()->isIntegerTy())
18686 weight = CW_SpecificReg;
18691 if (type->isFloatingPointTy())
18692 weight = CW_SpecificReg;
18695 if (type->isX86_MMXTy() && Subtarget->hasMMX())
18696 weight = CW_SpecificReg;
18700 if (((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasSSE1()) ||
18701 ((type->getPrimitiveSizeInBits() == 256) && Subtarget->hasFp256()))
18702 weight = CW_Register;
18705 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
18706 if (C->getZExtValue() <= 31)
18707 weight = CW_Constant;
18711 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
18712 if (C->getZExtValue() <= 63)
18713 weight = CW_Constant;
18717 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
18718 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
18719 weight = CW_Constant;
18723 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
18724 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
18725 weight = CW_Constant;
18729 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
18730 if (C->getZExtValue() <= 3)
18731 weight = CW_Constant;
18735 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
18736 if (C->getZExtValue() <= 0xff)
18737 weight = CW_Constant;
18742 if (dyn_cast<ConstantFP>(CallOperandVal)) {
18743 weight = CW_Constant;
18747 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
18748 if ((C->getSExtValue() >= -0x80000000LL) &&
18749 (C->getSExtValue() <= 0x7fffffffLL))
18750 weight = CW_Constant;
18754 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
18755 if (C->getZExtValue() <= 0xffffffff)
18756 weight = CW_Constant;
18763 /// LowerXConstraint - try to replace an X constraint, which matches anything,
18764 /// with another that has more specific requirements based on the type of the
18765 /// corresponding operand.
18766 const char *X86TargetLowering::
18767 LowerXConstraint(EVT ConstraintVT) const {
18768 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
18769 // 'f' like normal targets.
18770 if (ConstraintVT.isFloatingPoint()) {
18771 if (Subtarget->hasSSE2())
18773 if (Subtarget->hasSSE1())
18777 return TargetLowering::LowerXConstraint(ConstraintVT);
18780 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
18781 /// vector. If it is invalid, don't add anything to Ops.
18782 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
18783 std::string &Constraint,
18784 std::vector<SDValue>&Ops,
18785 SelectionDAG &DAG) const {
18786 SDValue Result(0, 0);
18788 // Only support length 1 constraints for now.
18789 if (Constraint.length() > 1) return;
18791 char ConstraintLetter = Constraint[0];
18792 switch (ConstraintLetter) {
18795 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
18796 if (C->getZExtValue() <= 31) {
18797 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
18803 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
18804 if (C->getZExtValue() <= 63) {
18805 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
18811 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
18812 if (isInt<8>(C->getSExtValue())) {
18813 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
18819 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
18820 if (C->getZExtValue() <= 255) {
18821 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
18827 // 32-bit signed value
18828 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
18829 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
18830 C->getSExtValue())) {
18831 // Widen to 64 bits here to get it sign extended.
18832 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
18835 // FIXME gcc accepts some relocatable values here too, but only in certain
18836 // memory models; it's complicated.
18841 // 32-bit unsigned value
18842 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
18843 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
18844 C->getZExtValue())) {
18845 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
18849 // FIXME gcc accepts some relocatable values here too, but only in certain
18850 // memory models; it's complicated.
18854 // Literal immediates are always ok.
18855 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
18856 // Widen to 64 bits here to get it sign extended.
18857 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
18861 // In any sort of PIC mode addresses need to be computed at runtime by
18862 // adding in a register or some sort of table lookup. These can't
18863 // be used as immediates.
18864 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
18867 // If we are in non-pic codegen mode, we allow the address of a global (with
18868 // an optional displacement) to be used with 'i'.
18869 GlobalAddressSDNode *GA = 0;
18870 int64_t Offset = 0;
18872 // Match either (GA), (GA+C), (GA+C1+C2), etc.
18874 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
18875 Offset += GA->getOffset();
18877 } else if (Op.getOpcode() == ISD::ADD) {
18878 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
18879 Offset += C->getZExtValue();
18880 Op = Op.getOperand(0);
18883 } else if (Op.getOpcode() == ISD::SUB) {
18884 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
18885 Offset += -C->getZExtValue();
18886 Op = Op.getOperand(0);
18891 // Otherwise, this isn't something we can handle, reject it.
18895 const GlobalValue *GV = GA->getGlobal();
18896 // If we require an extra load to get this address, as in PIC mode, we
18897 // can't accept it.
18898 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
18899 getTargetMachine())))
18902 Result = DAG.getTargetGlobalAddress(GV, SDLoc(Op),
18903 GA->getValueType(0), Offset);
18908 if (Result.getNode()) {
18909 Ops.push_back(Result);
18912 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
18915 std::pair<unsigned, const TargetRegisterClass*>
18916 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
18918 // First, see if this is a constraint that directly corresponds to an LLVM
18920 if (Constraint.size() == 1) {
18921 // GCC Constraint Letters
18922 switch (Constraint[0]) {
18924 // TODO: Slight differences here in allocation order and leaving
18925 // RIP in the class. Do they matter any more here than they do
18926 // in the normal allocation?
18927 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
18928 if (Subtarget->is64Bit()) {
18929 if (VT == MVT::i32 || VT == MVT::f32)
18930 return std::make_pair(0U, &X86::GR32RegClass);
18931 if (VT == MVT::i16)
18932 return std::make_pair(0U, &X86::GR16RegClass);
18933 if (VT == MVT::i8 || VT == MVT::i1)
18934 return std::make_pair(0U, &X86::GR8RegClass);
18935 if (VT == MVT::i64 || VT == MVT::f64)
18936 return std::make_pair(0U, &X86::GR64RegClass);
18939 // 32-bit fallthrough
18940 case 'Q': // Q_REGS
18941 if (VT == MVT::i32 || VT == MVT::f32)
18942 return std::make_pair(0U, &X86::GR32_ABCDRegClass);
18943 if (VT == MVT::i16)
18944 return std::make_pair(0U, &X86::GR16_ABCDRegClass);
18945 if (VT == MVT::i8 || VT == MVT::i1)
18946 return std::make_pair(0U, &X86::GR8_ABCD_LRegClass);
18947 if (VT == MVT::i64)
18948 return std::make_pair(0U, &X86::GR64_ABCDRegClass);
18950 case 'r': // GENERAL_REGS
18951 case 'l': // INDEX_REGS
18952 if (VT == MVT::i8 || VT == MVT::i1)
18953 return std::make_pair(0U, &X86::GR8RegClass);
18954 if (VT == MVT::i16)
18955 return std::make_pair(0U, &X86::GR16RegClass);
18956 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
18957 return std::make_pair(0U, &X86::GR32RegClass);
18958 return std::make_pair(0U, &X86::GR64RegClass);
18959 case 'R': // LEGACY_REGS
18960 if (VT == MVT::i8 || VT == MVT::i1)
18961 return std::make_pair(0U, &X86::GR8_NOREXRegClass);
18962 if (VT == MVT::i16)
18963 return std::make_pair(0U, &X86::GR16_NOREXRegClass);
18964 if (VT == MVT::i32 || !Subtarget->is64Bit())
18965 return std::make_pair(0U, &X86::GR32_NOREXRegClass);
18966 return std::make_pair(0U, &X86::GR64_NOREXRegClass);
18967 case 'f': // FP Stack registers.
18968 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
18969 // value to the correct fpstack register class.
18970 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
18971 return std::make_pair(0U, &X86::RFP32RegClass);
18972 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
18973 return std::make_pair(0U, &X86::RFP64RegClass);
18974 return std::make_pair(0U, &X86::RFP80RegClass);
18975 case 'y': // MMX_REGS if MMX allowed.
18976 if (!Subtarget->hasMMX()) break;
18977 return std::make_pair(0U, &X86::VR64RegClass);
18978 case 'Y': // SSE_REGS if SSE2 allowed
18979 if (!Subtarget->hasSSE2()) break;
18981 case 'x': // SSE_REGS if SSE1 allowed or AVX_REGS if AVX allowed
18982 if (!Subtarget->hasSSE1()) break;
18984 switch (VT.SimpleTy) {
18986 // Scalar SSE types.
18989 return std::make_pair(0U, &X86::FR32RegClass);
18992 return std::make_pair(0U, &X86::FR64RegClass);
19000 return std::make_pair(0U, &X86::VR128RegClass);
19008 return std::make_pair(0U, &X86::VR256RegClass);
19013 return std::make_pair(0U, &X86::VR512RegClass);
19019 // Use the default implementation in TargetLowering to convert the register
19020 // constraint into a member of a register class.
19021 std::pair<unsigned, const TargetRegisterClass*> Res;
19022 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
19024 // Not found as a standard register?
19025 if (Res.second == 0) {
19026 // Map st(0) -> st(7) -> ST0
19027 if (Constraint.size() == 7 && Constraint[0] == '{' &&
19028 tolower(Constraint[1]) == 's' &&
19029 tolower(Constraint[2]) == 't' &&
19030 Constraint[3] == '(' &&
19031 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
19032 Constraint[5] == ')' &&
19033 Constraint[6] == '}') {
19035 Res.first = X86::ST0+Constraint[4]-'0';
19036 Res.second = &X86::RFP80RegClass;
19040 // GCC allows "st(0)" to be called just plain "st".
19041 if (StringRef("{st}").equals_lower(Constraint)) {
19042 Res.first = X86::ST0;
19043 Res.second = &X86::RFP80RegClass;
19048 if (StringRef("{flags}").equals_lower(Constraint)) {
19049 Res.first = X86::EFLAGS;
19050 Res.second = &X86::CCRRegClass;
19054 // 'A' means EAX + EDX.
19055 if (Constraint == "A") {
19056 Res.first = X86::EAX;
19057 Res.second = &X86::GR32_ADRegClass;
19063 // Otherwise, check to see if this is a register class of the wrong value
19064 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
19065 // turn into {ax},{dx}.
19066 if (Res.second->hasType(VT))
19067 return Res; // Correct type already, nothing to do.
19069 // All of the single-register GCC register classes map their values onto
19070 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
19071 // really want an 8-bit or 32-bit register, map to the appropriate register
19072 // class and return the appropriate register.
19073 if (Res.second == &X86::GR16RegClass) {
19074 if (VT == MVT::i8 || VT == MVT::i1) {
19075 unsigned DestReg = 0;
19076 switch (Res.first) {
19078 case X86::AX: DestReg = X86::AL; break;
19079 case X86::DX: DestReg = X86::DL; break;
19080 case X86::CX: DestReg = X86::CL; break;
19081 case X86::BX: DestReg = X86::BL; break;
19084 Res.first = DestReg;
19085 Res.second = &X86::GR8RegClass;
19087 } else if (VT == MVT::i32 || VT == MVT::f32) {
19088 unsigned DestReg = 0;
19089 switch (Res.first) {
19091 case X86::AX: DestReg = X86::EAX; break;
19092 case X86::DX: DestReg = X86::EDX; break;
19093 case X86::CX: DestReg = X86::ECX; break;
19094 case X86::BX: DestReg = X86::EBX; break;
19095 case X86::SI: DestReg = X86::ESI; break;
19096 case X86::DI: DestReg = X86::EDI; break;
19097 case X86::BP: DestReg = X86::EBP; break;
19098 case X86::SP: DestReg = X86::ESP; break;
19101 Res.first = DestReg;
19102 Res.second = &X86::GR32RegClass;
19104 } else if (VT == MVT::i64 || VT == MVT::f64) {
19105 unsigned DestReg = 0;
19106 switch (Res.first) {
19108 case X86::AX: DestReg = X86::RAX; break;
19109 case X86::DX: DestReg = X86::RDX; break;
19110 case X86::CX: DestReg = X86::RCX; break;
19111 case X86::BX: DestReg = X86::RBX; break;
19112 case X86::SI: DestReg = X86::RSI; break;
19113 case X86::DI: DestReg = X86::RDI; break;
19114 case X86::BP: DestReg = X86::RBP; break;
19115 case X86::SP: DestReg = X86::RSP; break;
19118 Res.first = DestReg;
19119 Res.second = &X86::GR64RegClass;
19122 } else if (Res.second == &X86::FR32RegClass ||
19123 Res.second == &X86::FR64RegClass ||
19124 Res.second == &X86::VR128RegClass ||
19125 Res.second == &X86::VR256RegClass ||
19126 Res.second == &X86::FR32XRegClass ||
19127 Res.second == &X86::FR64XRegClass ||
19128 Res.second == &X86::VR128XRegClass ||
19129 Res.second == &X86::VR256XRegClass ||
19130 Res.second == &X86::VR512RegClass) {
19131 // Handle references to XMM physical registers that got mapped into the
19132 // wrong class. This can happen with constraints like {xmm0} where the
19133 // target independent register mapper will just pick the first match it can
19134 // find, ignoring the required type.
19136 if (VT == MVT::f32 || VT == MVT::i32)
19137 Res.second = &X86::FR32RegClass;
19138 else if (VT == MVT::f64 || VT == MVT::i64)
19139 Res.second = &X86::FR64RegClass;
19140 else if (X86::VR128RegClass.hasType(VT))
19141 Res.second = &X86::VR128RegClass;
19142 else if (X86::VR256RegClass.hasType(VT))
19143 Res.second = &X86::VR256RegClass;
19144 else if (X86::VR512RegClass.hasType(VT))
19145 Res.second = &X86::VR512RegClass;